1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #include "ecb.h"//D |
|
|
6 | |
|
|
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
8 | #include "xthread.h" |
|
|
9 | #include "schmorp.h" |
|
|
10 | |
|
|
11 | #ifdef I_DLFCN |
|
|
12 | #include <dlfcn.h> |
|
|
13 | #endif |
|
|
14 | |
|
|
15 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
16 | // but then fail to define the api functions unless you ALSO define |
|
|
17 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
18 | // after all, the functions are deprecated, not removed. |
|
|
19 | // in addition, you cannot test for this in any future-proof way. |
|
|
20 | // each time a new opencl version comes out, you need to make a new |
|
|
21 | // release. |
|
|
22 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
23 | |
|
|
24 | #ifndef PREFER_1_1 |
|
|
25 | #define PREFER_1_1 1 |
|
|
26 | #endif |
|
|
27 | |
|
|
28 | #if PREFER_1_1 |
|
|
29 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
30 | #endif |
|
|
31 | |
|
|
32 | #ifdef __APPLE__ |
|
|
33 | #include <OpenCL/opencl.h> |
|
|
34 | #else |
5 | #include <CL/opencl.h> |
35 | #include <CL/opencl.h> |
|
|
36 | #endif |
6 | |
37 | |
|
|
38 | #ifndef CL_VERSION_1_2 |
|
|
39 | #undef PREFER_1_1 |
|
|
40 | #define PREFER_1_1 1 |
|
|
41 | #endif |
|
|
42 | |
7 | typedef cl_platform_id OpenCL__Platform; |
43 | typedef cl_platform_id OpenCL__Platform; |
8 | typedef cl_device_id OpenCL__Device; |
44 | typedef cl_device_id OpenCL__Device; |
|
|
45 | typedef cl_device_id OpenCL__SubDevice; |
9 | typedef cl_context OpenCL__Context; |
46 | typedef cl_context OpenCL__Context; |
10 | typedef cl_command_queue OpenCL__Queue; |
47 | typedef cl_command_queue OpenCL__Queue; |
11 | typedef cl_mem OpenCL__Memory; |
48 | typedef cl_mem OpenCL__Memory; |
12 | typedef cl_mem OpenCL__Buffer; |
49 | typedef cl_mem OpenCL__Buffer; |
|
|
50 | typedef cl_mem OpenCL__BufferObj; |
13 | typedef cl_mem OpenCL__Image; |
51 | typedef cl_mem OpenCL__Image; |
14 | typedef cl_mem OpenCL__Image2D; |
|
|
15 | typedef cl_mem OpenCL__Image3D; |
|
|
16 | typedef cl_mem OpenCL__Memory_ornull; |
52 | typedef cl_mem OpenCL__Memory_ornull; |
17 | typedef cl_mem OpenCL__Buffer_ornull; |
53 | typedef cl_mem OpenCL__Buffer_ornull; |
18 | typedef cl_mem OpenCL__Image_ornull; |
54 | typedef cl_mem OpenCL__Image_ornull; |
19 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
20 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
21 | typedef cl_sampler OpenCL__Sampler; |
55 | typedef cl_sampler OpenCL__Sampler; |
22 | typedef cl_program OpenCL__Program; |
56 | typedef cl_program OpenCL__Program; |
23 | typedef cl_kernel OpenCL__Kernel; |
57 | typedef cl_kernel OpenCL__Kernel; |
24 | typedef cl_event OpenCL__Event; |
58 | typedef cl_event OpenCL__Event; |
25 | typedef cl_event OpenCL__UserEvent; |
59 | typedef cl_event OpenCL__UserEvent; |
|
|
60 | |
|
|
61 | typedef struct mapped * OpenCL__Mapped; |
|
|
62 | |
|
|
63 | typedef SV *FUTURE; |
|
|
64 | |
|
|
65 | static HV |
|
|
66 | *stash_platform, |
|
|
67 | *stash_device, |
|
|
68 | *stash_subdevice, |
|
|
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; |
26 | |
89 | |
27 | /*****************************************************************************/ |
90 | /*****************************************************************************/ |
28 | |
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 | } |
|
|
134 | |
|
|
135 | /*****************************************************************************/ |
|
|
136 | |
29 | /* up to two temporary buffers */ |
137 | /* up to two temporary buffers */ |
30 | static void * |
138 | static void * ecb_noinline |
31 | tmpbuf (size_t size) |
139 | tmpbuf (size_t size) |
32 | { |
140 | { |
|
|
141 | enum { buffers = 4 }; |
33 | static int idx; |
142 | static int idx; |
34 | static void *buf [2]; |
143 | static void *buf [buffers]; |
35 | static size_t len [2]; |
144 | static size_t len [buffers]; |
36 | |
145 | |
37 | idx ^= 1; |
146 | idx = (idx + 1) % buffers; |
38 | |
147 | |
39 | if (len [idx] < size) |
148 | if (len [idx] < size) |
40 | { |
149 | { |
41 | free (buf [idx]); |
150 | free (buf [idx]); |
42 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
44 | } |
153 | } |
45 | |
154 | |
46 | return buf [idx]; |
155 | return buf [idx]; |
47 | } |
156 | } |
48 | |
157 | |
|
|
158 | static const char * ecb_noinline |
|
|
159 | cv_get_name (CV *cv) |
|
|
160 | { |
|
|
161 | static char fullname [256]; |
|
|
162 | |
|
|
163 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
164 | |
|
|
165 | HV *stash = GvSTASH (gv); |
|
|
166 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
167 | const char *gvname = GvNAME (gv); |
|
|
168 | |
|
|
169 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
170 | return fullname; |
|
|
171 | } |
|
|
172 | |
49 | /*****************************************************************************/ |
173 | /*****************************************************************************/ |
50 | |
174 | |
51 | typedef struct |
175 | typedef struct |
52 | { |
176 | { |
53 | IV iv; |
177 | IV iv; |
… | |
… | |
90 | return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
214 | return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
91 | } |
215 | } |
92 | |
216 | |
93 | /*****************************************************************************/ |
217 | /*****************************************************************************/ |
94 | |
218 | |
95 | static cl_int last_error; |
219 | static cl_int res; |
96 | |
220 | |
97 | #define FAIL(name,err) \ |
221 | #define FAIL(name) \ |
98 | croak ("cl" # name ": %s", err2str (last_error = err)); |
222 | croak ("cl" # name ": %s", err2str (res)); |
99 | |
223 | |
100 | #define NEED_SUCCESS(name,args) \ |
224 | #define NEED_SUCCESS(name,args) \ |
101 | do { \ |
225 | do { \ |
102 | cl_int res = cl ## name args; \ |
226 | res = cl ## name args; \ |
103 | \ |
227 | \ |
104 | if (res) \ |
228 | if (res) \ |
105 | FAIL (name, res); \ |
229 | FAIL (name); \ |
106 | } while (0) |
230 | } while (0) |
107 | |
231 | |
|
|
232 | #define NEED_SUCCESS_ARG(retdecl, name, args) \ |
|
|
233 | retdecl = cl ## name args; \ |
|
|
234 | if (res) \ |
|
|
235 | FAIL (name); |
|
|
236 | |
108 | /*****************************************************************************/ |
237 | /*****************************************************************************/ |
109 | |
238 | |
110 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
111 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
112 | |
|
|
113 | static void * |
239 | static SV * |
|
|
240 | new_clobj (HV *stash, IV id) |
|
|
241 | { |
|
|
242 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
243 | } |
|
|
244 | |
|
|
245 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
246 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
247 | |
|
|
248 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
249 | /* they can be upgraded at runtime to the array form */ |
|
|
250 | static void * ecb_noinline |
114 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
251 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
115 | { |
252 | { |
|
|
253 | // sv_derived_from is quite slow :( |
116 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
254 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
117 | return (void *)SvIV (SvRV (sv)); |
255 | return (void *)SvIV (SvRV (sv)); |
118 | |
256 | |
119 | croak ("%s: %s is not of type %s", func, svname, pkg); |
257 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
258 | } |
|
|
259 | |
|
|
260 | // the "no-inherit" version of the above |
|
|
261 | static void * ecb_noinline |
|
|
262 | SvCLOBJ_ni (const char *func, const char *svname, SV *sv, HV *stash) |
|
|
263 | { |
|
|
264 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
265 | return (void *)SvIV (SvRV (sv)); |
|
|
266 | |
|
|
267 | croak ("%s: %s is not of type %s", func, svname, HvNAME (stash)); |
120 | } |
268 | } |
121 | |
269 | |
122 | /*****************************************************************************/ |
270 | /*****************************************************************************/ |
123 | |
271 | |
124 | static cl_event * |
272 | static cl_context_properties * ecb_noinline |
|
|
273 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
|
|
274 | { |
|
|
275 | if (!sv || !SvOK (sv)) |
|
|
276 | if (extra) |
|
|
277 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
|
|
278 | else |
|
|
279 | return 0; |
|
|
280 | |
|
|
281 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
|
|
282 | { |
|
|
283 | AV *av = (AV *)SvRV (sv); |
|
|
284 | int i, len = av_len (av) + 1; |
|
|
285 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
|
|
286 | cl_context_properties *l = p; |
|
|
287 | |
|
|
288 | if (len & 1) |
|
|
289 | croak ("%s: %s is not a property list (must contain an even number of elements)", func, svname); |
|
|
290 | |
|
|
291 | while (extracount--) |
|
|
292 | *l++ = *extra++; |
|
|
293 | |
|
|
294 | for (i = 0; i < len; i += 2) |
|
|
295 | { |
|
|
296 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
|
|
297 | SV *p_sv = *av_fetch (av, i + 1, 0); |
|
|
298 | cl_context_properties v = SvIV (p_sv); // code below can override |
|
|
299 | |
|
|
300 | switch (t) |
|
|
301 | { |
|
|
302 | case CL_CONTEXT_PLATFORM: |
|
|
303 | if (SvROK (p_sv)) |
|
|
304 | v = (cl_context_properties)SvCLOBJ (func, svname, p_sv, "OpenCL::Platform"); |
|
|
305 | break; |
|
|
306 | |
|
|
307 | case CL_GLX_DISPLAY_KHR: |
|
|
308 | if (!SvOK (p_sv)) |
|
|
309 | { |
|
|
310 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
311 | if (func) |
|
|
312 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
313 | } |
|
|
314 | break; |
|
|
315 | |
|
|
316 | case CL_GL_CONTEXT_KHR: |
|
|
317 | if (!SvOK (p_sv)) |
|
|
318 | { |
|
|
319 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
320 | if (func) |
|
|
321 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
322 | } |
|
|
323 | break; |
|
|
324 | |
|
|
325 | default: |
|
|
326 | /* unknown property, treat as int */ |
|
|
327 | break; |
|
|
328 | } |
|
|
329 | |
|
|
330 | *l++ = t; |
|
|
331 | *l++ = v; |
|
|
332 | } |
|
|
333 | |
|
|
334 | *l = 0; |
|
|
335 | |
|
|
336 | return p; |
|
|
337 | } |
|
|
338 | |
|
|
339 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
|
|
340 | } |
|
|
341 | |
|
|
342 | static void * ecb_noinline |
|
|
343 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
344 | { |
|
|
345 | const char *funcname = cv_get_name (cv); |
|
|
346 | |
|
|
347 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
348 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
349 | funcname, argname, or_undef ? "undef or " : "", klass); |
|
|
350 | |
|
|
351 | AV *av = (AV *)SvRV (arg); |
|
|
352 | void **list = 0; |
|
|
353 | cl_uint count = av_len (av) + 1; |
|
|
354 | |
|
|
355 | if (count) |
|
|
356 | { |
|
|
357 | list = tmpbuf (sizeof (*list) * count); |
|
|
358 | int i; |
|
|
359 | for (i = 0; i < count; ++i) |
|
|
360 | list [i] = SvCLOBJ (funcname, argname, *av_fetch (av, i, 1), klass); |
|
|
361 | } |
|
|
362 | |
|
|
363 | if (!count && !or_undef) |
|
|
364 | croak ("%s: '%s' must contain at least one %s object", |
|
|
365 | funcname, argname, klass); |
|
|
366 | |
|
|
367 | *rcount = count; |
|
|
368 | return (void *)list; |
|
|
369 | } |
|
|
370 | |
|
|
371 | /*****************************************************************************/ |
|
|
372 | /* callback stuff */ |
|
|
373 | |
|
|
374 | /* default context callback, log to stderr */ |
|
|
375 | static void CL_CALLBACK |
|
|
376 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
377 | { |
|
|
378 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
379 | } |
|
|
380 | |
|
|
381 | typedef struct |
|
|
382 | { |
|
|
383 | int free_cb; |
|
|
384 | void (*push)(void *data1, void *data2, void *data3); |
|
|
385 | } eq_vtbl; |
|
|
386 | |
|
|
387 | typedef struct eq_item |
|
|
388 | { |
|
|
389 | struct eq_item *next; |
|
|
390 | eq_vtbl *vtbl; |
|
|
391 | SV *cb; |
|
|
392 | void *data1, *data2, *data3; |
|
|
393 | } eq_item; |
|
|
394 | |
|
|
395 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
396 | static void *eq_signal_arg; |
|
|
397 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
398 | static eq_item *eq_head, *eq_tail; |
|
|
399 | |
|
|
400 | static void ecb_noinline |
|
|
401 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
402 | { |
|
|
403 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
404 | |
|
|
405 | item->next = 0; |
|
|
406 | item->vtbl = vtbl; |
|
|
407 | item->cb = cb; |
|
|
408 | item->data1 = data1; |
|
|
409 | item->data2 = data2; |
|
|
410 | item->data3 = data3; |
|
|
411 | |
|
|
412 | X_LOCK (eq_lock); |
|
|
413 | |
|
|
414 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
415 | eq_tail = item; |
|
|
416 | |
|
|
417 | X_UNLOCK (eq_lock); |
|
|
418 | |
|
|
419 | eq_signal_func (eq_signal_arg, 0); |
|
|
420 | } |
|
|
421 | |
|
|
422 | static eq_item * |
|
|
423 | eq_dec (void) |
|
|
424 | { |
|
|
425 | eq_item *res; |
|
|
426 | |
|
|
427 | X_LOCK (eq_lock); |
|
|
428 | |
|
|
429 | res = eq_head; |
|
|
430 | if (res) |
|
|
431 | eq_head = res->next; |
|
|
432 | |
|
|
433 | X_UNLOCK (eq_lock); |
|
|
434 | |
|
|
435 | return res; |
|
|
436 | } |
|
|
437 | |
|
|
438 | static void |
|
|
439 | eq_poll (void) |
|
|
440 | { |
|
|
441 | eq_item *item; |
|
|
442 | |
|
|
443 | while ((item = eq_dec ())) |
|
|
444 | { |
|
|
445 | ENTER; |
|
|
446 | SAVETMPS; |
|
|
447 | |
|
|
448 | dSP; |
|
|
449 | PUSHMARK (SP); |
|
|
450 | EXTEND (SP, 2); |
|
|
451 | |
|
|
452 | if (item->vtbl->free_cb) |
|
|
453 | sv_2mortal (item->cb); |
|
|
454 | |
|
|
455 | PUTBACK; |
|
|
456 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
457 | |
|
|
458 | SV *cb = item->cb; |
|
|
459 | free (item); |
|
|
460 | |
|
|
461 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
462 | |
|
|
463 | FREETMPS; |
|
|
464 | LEAVE; |
|
|
465 | } |
|
|
466 | } |
|
|
467 | |
|
|
468 | static void |
|
|
469 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
470 | { |
|
|
471 | eq_poll (); |
|
|
472 | } |
|
|
473 | |
|
|
474 | /*****************************************************************************/ |
|
|
475 | /* context notify */ |
|
|
476 | |
|
|
477 | static void ecb_noinline |
|
|
478 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
479 | { |
|
|
480 | dSP; |
|
|
481 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
482 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
483 | PUTBACK; |
|
|
484 | |
|
|
485 | free (data1); |
|
|
486 | free (data2); |
|
|
487 | } |
|
|
488 | |
|
|
489 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
490 | |
|
|
491 | static void CL_CALLBACK |
|
|
492 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
493 | { |
|
|
494 | void *pvt_copy = malloc (cb); |
|
|
495 | memcpy (pvt_copy, pvt, cb); |
|
|
496 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
497 | } |
|
|
498 | |
|
|
499 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
500 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
501 | void *user_data = 0; \ |
|
|
502 | \ |
|
|
503 | if (SvOK (notify)) \ |
|
|
504 | { \ |
|
|
505 | pfn_notify = eq_context_notify; \ |
|
|
506 | user_data = s_get_cv (notify); \ |
|
|
507 | } |
|
|
508 | |
|
|
509 | static SV * ecb_noinline |
|
|
510 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
511 | { |
|
|
512 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
513 | |
|
|
514 | if (user_data) |
|
|
515 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
516 | |
|
|
517 | return sv; |
|
|
518 | } |
|
|
519 | |
|
|
520 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
521 | |
|
|
522 | /*****************************************************************************/ |
|
|
523 | /* build/compile/link notify */ |
|
|
524 | |
|
|
525 | static void |
|
|
526 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
527 | { |
|
|
528 | dSP; |
|
|
529 | PUSH_CLOBJ (stash_program, data1); |
|
|
530 | PUTBACK; |
|
|
531 | } |
|
|
532 | |
|
|
533 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
534 | |
|
|
535 | static void CL_CALLBACK |
|
|
536 | eq_program_notify (cl_program program, void *user_data) |
|
|
537 | { |
|
|
538 | clRetainProgram (program); |
|
|
539 | |
|
|
540 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
541 | } |
|
|
542 | |
|
|
543 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
544 | |
|
|
545 | static program_callback ecb_noinline |
|
|
546 | make_program_callback (SV *notify, void **ruser_data) |
|
|
547 | { |
|
|
548 | if (SvOK (notify)) |
|
|
549 | { |
|
|
550 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
551 | return eq_program_notify; |
|
|
552 | } |
|
|
553 | else |
|
|
554 | { |
|
|
555 | *ruser_data = 0; |
|
|
556 | return 0; |
|
|
557 | } |
|
|
558 | } |
|
|
559 | |
|
|
560 | struct build_args |
|
|
561 | { |
|
|
562 | cl_program program; |
|
|
563 | char *options; |
|
|
564 | void *user_data; |
|
|
565 | cl_uint num_devices; |
|
|
566 | }; |
|
|
567 | |
|
|
568 | X_THREAD_PROC (build_program_thread) |
|
|
569 | { |
|
|
570 | struct build_args *arg = thr_arg; |
|
|
571 | |
|
|
572 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
573 | |
|
|
574 | if (arg->user_data) |
|
|
575 | eq_program_notify (arg->program, arg->user_data); |
|
|
576 | else |
|
|
577 | clReleaseProgram (arg->program); |
|
|
578 | |
|
|
579 | free (arg->options); |
|
|
580 | free (arg); |
|
|
581 | |
|
|
582 | return 0; |
|
|
583 | } |
|
|
584 | |
|
|
585 | static void |
|
|
586 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
587 | { |
|
|
588 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
589 | |
|
|
590 | arg->program = program; |
|
|
591 | arg->options = strdup (options); |
|
|
592 | arg->user_data = user_data; |
|
|
593 | arg->num_devices = num_devices; |
|
|
594 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
595 | |
|
|
596 | xthread_t id; |
|
|
597 | thread_create (&id, build_program_thread, arg); |
|
|
598 | } |
|
|
599 | |
|
|
600 | /*****************************************************************************/ |
|
|
601 | /* event objects */ |
|
|
602 | |
|
|
603 | static void |
|
|
604 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
605 | { |
|
|
606 | dSP; |
|
|
607 | PUSH_CLOBJ (stash_event, data1); |
|
|
608 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
609 | PUTBACK; |
|
|
610 | } |
|
|
611 | |
|
|
612 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
613 | |
|
|
614 | static void CL_CALLBACK |
|
|
615 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
616 | { |
|
|
617 | clRetainEvent (event); |
|
|
618 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
619 | } |
|
|
620 | |
|
|
621 | /*****************************************************************************/ |
|
|
622 | /* utilities for XS code */ |
|
|
623 | |
|
|
624 | static size_t |
|
|
625 | img_row_pitch (cl_mem img) |
|
|
626 | { |
|
|
627 | size_t res; |
|
|
628 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
|
|
629 | return res; |
|
|
630 | } |
|
|
631 | |
|
|
632 | static cl_event * ecb_noinline |
125 | event_list (SV **items, int count) |
633 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
126 | { |
634 | { |
|
|
635 | cl_uint count = *rcount; |
|
|
636 | |
|
|
637 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
638 | *rcount = count = 0; |
|
|
639 | |
|
|
640 | if (!count && !extra) |
|
|
641 | return 0; |
|
|
642 | |
127 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
643 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
644 | int i = 0; |
128 | |
645 | |
129 | while (count--) |
646 | while (count--) |
|
|
647 | if (SvOK (items [count])) |
130 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
648 | list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
131 | |
649 | |
|
|
650 | if (extra) |
|
|
651 | list [i++] = extra; |
|
|
652 | |
|
|
653 | *rcount = i; |
|
|
654 | |
132 | return list; |
655 | return i ? list : 0; |
133 | } |
656 | } |
134 | |
657 | |
135 | #define EVENT_LIST(items,count) \ |
658 | #define EVENT_LIST(skip) \ |
136 | cl_uint event_list_count = (count); \ |
659 | cl_uint event_list_count = items - (skip); \ |
137 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
660 | cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) |
138 | |
661 | |
139 | #define INFO(class) \ |
662 | #define INFO(class) \ |
140 | { \ |
663 | { \ |
141 | size_t size; \ |
664 | size_t size; \ |
142 | SV *sv; \ |
|
|
143 | \ |
|
|
144 | NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ |
665 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
145 | sv = sv_2mortal (newSV (size)); \ |
666 | SV *sv = sv_2mortal (newSV (size)); \ |
146 | SvUPGRADE (sv, SVt_PV); \ |
667 | SvUPGRADE (sv, SVt_PV); \ |
147 | SvPOK_only (sv); \ |
668 | SvPOK_only (sv); \ |
148 | SvCUR_set (sv, size); \ |
669 | SvCUR_set (sv, size); \ |
149 | NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ |
670 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
150 | XPUSHs (sv); \ |
671 | XPUSHs (sv); \ |
151 | } |
672 | } |
152 | |
673 | |
|
|
674 | /*****************************************************************************/ |
|
|
675 | /* mapped_xxx */ |
|
|
676 | |
|
|
677 | static OpenCL__Mapped |
|
|
678 | SvMAPPED (SV *sv) |
|
|
679 | { |
|
|
680 | // no typechecking atm., keep your fingers crossed |
|
|
681 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
682 | } |
|
|
683 | |
|
|
684 | struct mapped |
|
|
685 | { |
|
|
686 | cl_command_queue queue; |
|
|
687 | cl_mem memobj; |
|
|
688 | void *ptr; |
|
|
689 | size_t cb; |
|
|
690 | cl_event event; |
|
|
691 | size_t row_pitch; |
|
|
692 | size_t slice_pitch; |
|
|
693 | }; |
|
|
694 | |
|
|
695 | static SV * |
|
|
696 | 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) |
|
|
697 | { |
|
|
698 | SV *data = newSV (0); |
|
|
699 | SvUPGRADE (data, SVt_PVMG); |
|
|
700 | |
|
|
701 | OpenCL__Mapped mapped; |
|
|
702 | New (0, mapped, 1, struct mapped); |
|
|
703 | |
|
|
704 | clRetainCommandQueue (queue); |
|
|
705 | |
|
|
706 | mapped->queue = queue; |
|
|
707 | mapped->memobj = memobj; |
|
|
708 | mapped->ptr = ptr; |
|
|
709 | mapped->cb = cb; |
|
|
710 | mapped->event = ev; |
|
|
711 | mapped->row_pitch = row_pitch; |
|
|
712 | mapped->slice_pitch = slice_pitch; |
|
|
713 | |
|
|
714 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
715 | |
|
|
716 | if (SvLEN (data)) |
|
|
717 | Safefree (data); |
|
|
718 | |
|
|
719 | SvPVX (data) = (char *)ptr; |
|
|
720 | SvCUR_set (data, cb); |
|
|
721 | SvLEN_set (data, 0); |
|
|
722 | SvPOK_only (data); |
|
|
723 | |
|
|
724 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
725 | |
|
|
726 | if (!(flags & CL_MAP_WRITE)) |
|
|
727 | SvREADONLY_on (data); |
|
|
728 | |
|
|
729 | return obj; |
|
|
730 | } |
|
|
731 | |
|
|
732 | static void |
|
|
733 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
734 | { |
|
|
735 | SV *data = SvRV (sv); |
|
|
736 | |
|
|
737 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
738 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
739 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
740 | else |
|
|
741 | { |
|
|
742 | SvREADONLY_off (data); |
|
|
743 | SvCUR_set (data, 0); |
|
|
744 | SvPVX (data) = 0; |
|
|
745 | SvOK_off (data); |
|
|
746 | } |
|
|
747 | |
|
|
748 | mapped->ptr = 0; |
|
|
749 | } |
|
|
750 | |
|
|
751 | static void |
|
|
752 | mapped_unmap (SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
753 | { |
|
|
754 | cl_event *event_list_ptr = event_list (wait_list, &event_list_count, mapped->event); |
|
|
755 | cl_event ev; |
|
|
756 | |
|
|
757 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
758 | |
|
|
759 | clReleaseEvent (mapped->event); |
|
|
760 | mapped->event = ev; |
|
|
761 | |
|
|
762 | mapped_detach (self, mapped); |
|
|
763 | } |
|
|
764 | |
|
|
765 | /*****************************************************************************/ |
|
|
766 | |
153 | MODULE = OpenCL PACKAGE = OpenCL |
767 | MODULE = OpenCL PACKAGE = OpenCL |
154 | |
768 | |
155 | PROTOTYPES: ENABLE |
769 | PROTOTYPES: ENABLE |
156 | |
770 | |
|
|
771 | void |
|
|
772 | poll () |
|
|
773 | CODE: |
|
|
774 | eq_poll (); |
|
|
775 | |
|
|
776 | void |
|
|
777 | _eq_initialise (IV func, IV arg) |
|
|
778 | CODE: |
|
|
779 | eq_signal_func = (void (*)(void *, int))func; |
|
|
780 | eq_signal_arg = (void*)arg; |
|
|
781 | |
157 | BOOT: |
782 | BOOT: |
158 | { |
783 | { |
159 | HV *stash = gv_stashpv ("OpenCL", 1); |
784 | HV *stash = gv_stashpv ("OpenCL", 1); |
160 | static const ivstr *civ, const_iv[] = { |
785 | static const ivstr *civ, const_iv[] = { |
161 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
786 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
162 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
787 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
163 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
788 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
164 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
789 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
165 | { sizeof (cl_int ), "SIZEOF_INT" }, |
790 | { sizeof (cl_int ), "SIZEOF_INT" }, |
166 | { sizeof (cl_uint ), "SIZEOF_UINT" }, |
791 | { sizeof (cl_uint ), "SIZEOF_UINT" }, |
167 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
792 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
168 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
793 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
169 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
794 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
170 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
795 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
171 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
796 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
172 | #include "constiv.h" |
797 | #include "constiv.h" |
173 | }; |
798 | }; |
|
|
799 | |
174 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
800 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
175 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
801 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
802 | |
|
|
803 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
804 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
805 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
806 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
807 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
808 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
809 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
810 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
811 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
812 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
813 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
814 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
815 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
816 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
817 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
818 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
819 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
820 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
821 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
822 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
823 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
824 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
825 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
826 | |
|
|
827 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
176 | } |
828 | } |
177 | |
829 | |
178 | cl_int |
830 | cl_int |
179 | errno () |
831 | errno () |
180 | CODE: |
832 | CODE: |
181 | errno = last_error; |
833 | RETVAL = res; |
|
|
834 | OUTPUT: |
|
|
835 | RETVAL |
182 | |
836 | |
183 | const char * |
837 | const char * |
184 | err2str (cl_int err) |
838 | err2str (cl_int err = res) |
185 | |
839 | |
186 | const char * |
840 | const char * |
187 | enum2str (cl_uint value) |
841 | enum2str (cl_uint value) |
188 | |
842 | |
189 | void |
843 | void |
190 | platforms () |
844 | platforms () |
191 | PPCODE: |
845 | PPCODE: |
192 | { |
|
|
193 | cl_platform_id *list; |
846 | cl_platform_id *list; |
194 | cl_uint count; |
847 | cl_uint count; |
195 | int i; |
848 | int i; |
196 | |
849 | |
197 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
850 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
198 | list = tmpbuf (sizeof (*list) * count); |
851 | list = tmpbuf (sizeof (*list) * count); |
199 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
852 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
200 | |
853 | |
201 | EXTEND (SP, count); |
854 | EXTEND (SP, count); |
202 | for (i = 0; i < count; ++i) |
855 | for (i = 0; i < count; ++i) |
203 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
856 | PUSH_CLOBJ (stash_platform, list [i]); |
204 | } |
|
|
205 | |
857 | |
206 | void |
858 | void |
207 | context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
859 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
208 | PPCODE: |
860 | PPCODE: |
209 | { |
861 | CONTEXT_NOTIFY_CALLBACK; |
210 | cl_int res; |
862 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
211 | cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
863 | XPUSH_CLOBJ_CONTEXT; |
212 | |
864 | |
213 | if (res) |
865 | void |
214 | FAIL (CreateContextFromType, res); |
866 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
215 | |
867 | PPCODE: |
216 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
868 | /* der Gipfel der Kunst */ |
217 | } |
|
|
218 | |
869 | |
219 | void |
870 | void |
220 | wait_for_events (...) |
871 | wait_for_events (...) |
221 | CODE: |
872 | CODE: |
222 | { |
|
|
223 | EVENT_LIST (0, items); |
873 | EVENT_LIST (0); |
224 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
874 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
225 | } |
|
|
226 | |
875 | |
227 | PROTOTYPES: DISABLE |
876 | PROTOTYPES: DISABLE |
228 | |
877 | |
229 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
878 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
230 | |
879 | |
231 | void |
880 | void |
232 | info (OpenCL::Platform this, cl_platform_info name) |
881 | info (OpenCL::Platform self, cl_platform_info name) |
233 | PPCODE: |
882 | PPCODE: |
234 | INFO (Platform) |
883 | INFO (Platform) |
235 | |
884 | |
236 | void |
885 | void |
|
|
886 | unload_compiler (OpenCL::Platform self) |
|
|
887 | CODE: |
|
|
888 | #if CL_VERSION_1_2 |
|
|
889 | clUnloadPlatformCompiler (self); |
|
|
890 | #endif |
|
|
891 | |
|
|
892 | #BEGIN:platform |
|
|
893 | |
|
|
894 | void |
|
|
895 | profile (OpenCL::Platform self) |
|
|
896 | ALIAS: |
|
|
897 | profile = CL_PLATFORM_PROFILE |
|
|
898 | version = CL_PLATFORM_VERSION |
|
|
899 | name = CL_PLATFORM_NAME |
|
|
900 | vendor = CL_PLATFORM_VENDOR |
|
|
901 | extensions = CL_PLATFORM_EXTENSIONS |
|
|
902 | PPCODE: |
|
|
903 | size_t size; |
|
|
904 | NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size)); |
|
|
905 | char *value = tmpbuf (size); |
|
|
906 | NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0)); |
|
|
907 | EXTEND (SP, 1); |
|
|
908 | const int i = 0; |
|
|
909 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
910 | |
|
|
911 | #END:platform |
|
|
912 | |
|
|
913 | void |
237 | devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) |
914 | devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL) |
238 | PPCODE: |
915 | PPCODE: |
239 | { |
|
|
240 | cl_device_id *list; |
916 | cl_device_id *list; |
241 | cl_uint count; |
917 | cl_uint count; |
242 | int i; |
918 | int i; |
243 | |
919 | |
244 | NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); |
920 | NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count)); |
245 | list = tmpbuf (sizeof (*list) * count); |
921 | list = tmpbuf (sizeof (*list) * count); |
246 | NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); |
922 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
247 | |
923 | |
248 | EXTEND (SP, count); |
924 | EXTEND (SP, count); |
249 | for (i = 0; i < count; ++i) |
925 | for (i = 0; i < count; ++i) |
250 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
926 | PUSH_CLOBJ (stash_device, list [i]); |
251 | } |
|
|
252 | |
927 | |
253 | void |
928 | void |
254 | context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
929 | context (OpenCL::Platform self, SV *properties = 0, SV *devices, SV *notify = &PL_sv_undef) |
255 | PPCODE: |
930 | PPCODE: |
256 | { |
|
|
257 | cl_int res; |
|
|
258 | cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
931 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
259 | cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
932 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context", "properties", properties, extra, 2); |
260 | |
933 | |
261 | if (res) |
934 | cl_uint device_count = 0; |
262 | FAIL (CreateContextFromType, res); |
935 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
263 | |
936 | |
264 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
937 | CONTEXT_NOTIFY_CALLBACK; |
265 | } |
938 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
939 | XPUSH_CLOBJ_CONTEXT; |
|
|
940 | |
|
|
941 | void |
|
|
942 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
943 | PPCODE: |
|
|
944 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
945 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
946 | |
|
|
947 | CONTEXT_NOTIFY_CALLBACK; |
|
|
948 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
949 | XPUSH_CLOBJ_CONTEXT; |
266 | |
950 | |
267 | MODULE = OpenCL PACKAGE = OpenCL::Device |
951 | MODULE = OpenCL PACKAGE = OpenCL::Device |
268 | |
952 | |
269 | void |
953 | void |
270 | info (OpenCL::Device this, cl_device_info name) |
954 | info (OpenCL::Device self, cl_device_info name) |
271 | PPCODE: |
955 | PPCODE: |
272 | INFO (Device) |
956 | INFO (Device) |
273 | |
957 | |
|
|
958 | #if CL_VERSION_1_2 |
|
|
959 | |
274 | void |
960 | void |
275 | context_simple (OpenCL::Device this) |
961 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
962 | PPCODE: |
|
|
963 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
|
|
964 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
|
|
965 | |
|
|
966 | properties = SvRV (properties); |
|
|
967 | |
|
|
968 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
969 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
970 | |
|
|
971 | int i; |
|
|
972 | for (i = 0; i < count; ++i) |
|
|
973 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
|
|
974 | |
|
|
975 | props [count] = 0; |
|
|
976 | |
|
|
977 | cl_uint num_devices; |
|
|
978 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
979 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
980 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
|
|
981 | |
|
|
982 | EXTEND (SP, num_devices); |
|
|
983 | for (i = 0; i < count; ++i) |
|
|
984 | PUSH_CLOBJ (stash_subdevice, list [i]); |
|
|
985 | |
|
|
986 | #endif |
|
|
987 | |
|
|
988 | #BEGIN:device |
|
|
989 | |
|
|
990 | void |
|
|
991 | type (OpenCL::Device self) |
276 | PPCODE: |
992 | PPCODE: |
277 | { |
993 | cl_device_type value [1]; |
278 | cl_int res; |
994 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
279 | cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
995 | EXTEND (SP, 1); |
|
|
996 | const int i = 0; |
|
|
997 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
280 | |
998 | |
281 | if (res) |
999 | void |
282 | FAIL (CreateContext, res); |
1000 | vendor_id (OpenCL::Device self) |
|
|
1001 | ALIAS: |
|
|
1002 | vendor_id = CL_DEVICE_VENDOR_ID |
|
|
1003 | max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS |
|
|
1004 | max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS |
|
|
1005 | preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR |
|
|
1006 | preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT |
|
|
1007 | preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT |
|
|
1008 | preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG |
|
|
1009 | preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT |
|
|
1010 | preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE |
|
|
1011 | max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY |
|
|
1012 | max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS |
|
|
1013 | max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS |
|
|
1014 | image_support = CL_DEVICE_IMAGE_SUPPORT |
|
|
1015 | max_samplers = CL_DEVICE_MAX_SAMPLERS |
|
|
1016 | mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN |
|
|
1017 | min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE |
|
|
1018 | global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
|
|
1019 | max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS |
|
|
1020 | preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF |
|
|
1021 | native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR |
|
|
1022 | native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT |
|
|
1023 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
|
|
1024 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
|
|
1025 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
|
|
1026 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
|
|
1027 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
|
|
1028 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
|
|
1029 | PPCODE: |
|
|
1030 | cl_uint value [1]; |
|
|
1031 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1032 | EXTEND (SP, 1); |
|
|
1033 | const int i = 0; |
|
|
1034 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
283 | |
1035 | |
284 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
1036 | void |
285 | } |
1037 | max_work_group_size (OpenCL::Device self) |
|
|
1038 | ALIAS: |
|
|
1039 | max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE |
|
|
1040 | image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH |
|
|
1041 | image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT |
|
|
1042 | image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH |
|
|
1043 | image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT |
|
|
1044 | image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH |
|
|
1045 | max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE |
|
|
1046 | profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION |
|
|
1047 | PPCODE: |
|
|
1048 | size_t value [1]; |
|
|
1049 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1050 | EXTEND (SP, 1); |
|
|
1051 | const int i = 0; |
|
|
1052 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1053 | |
|
|
1054 | void |
|
|
1055 | max_work_item_sizes (OpenCL::Device self) |
|
|
1056 | PPCODE: |
|
|
1057 | size_t size; |
|
|
1058 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size)); |
|
|
1059 | size_t *value = tmpbuf (size); |
|
|
1060 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0)); |
|
|
1061 | int i, n = size / sizeof (*value); |
|
|
1062 | EXTEND (SP, n); |
|
|
1063 | for (i = 0; i < n; ++i) |
|
|
1064 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1065 | |
|
|
1066 | void |
|
|
1067 | address_bits (OpenCL::Device self) |
|
|
1068 | PPCODE: |
|
|
1069 | cl_bitfield value [1]; |
|
|
1070 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_ADDRESS_BITS, sizeof (value), value, 0)); |
|
|
1071 | EXTEND (SP, 1); |
|
|
1072 | const int i = 0; |
|
|
1073 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1074 | |
|
|
1075 | void |
|
|
1076 | max_mem_alloc_size (OpenCL::Device self) |
|
|
1077 | ALIAS: |
|
|
1078 | max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE |
|
|
1079 | global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE |
|
|
1080 | global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE |
|
|
1081 | max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE |
|
|
1082 | local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE |
|
|
1083 | PPCODE: |
|
|
1084 | cl_ulong value [1]; |
|
|
1085 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1086 | EXTEND (SP, 1); |
|
|
1087 | const int i = 0; |
|
|
1088 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1089 | |
|
|
1090 | void |
|
|
1091 | single_fp_config (OpenCL::Device self) |
|
|
1092 | ALIAS: |
|
|
1093 | single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG |
|
|
1094 | double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG |
|
|
1095 | half_fp_config = CL_DEVICE_HALF_FP_CONFIG |
|
|
1096 | PPCODE: |
|
|
1097 | cl_device_fp_config value [1]; |
|
|
1098 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1099 | EXTEND (SP, 1); |
|
|
1100 | const int i = 0; |
|
|
1101 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1102 | |
|
|
1103 | void |
|
|
1104 | global_mem_cache_type (OpenCL::Device self) |
|
|
1105 | PPCODE: |
|
|
1106 | cl_device_mem_cache_type value [1]; |
|
|
1107 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, sizeof (value), value, 0)); |
|
|
1108 | EXTEND (SP, 1); |
|
|
1109 | const int i = 0; |
|
|
1110 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1111 | |
|
|
1112 | void |
|
|
1113 | local_mem_type (OpenCL::Device self) |
|
|
1114 | PPCODE: |
|
|
1115 | cl_device_local_mem_type value [1]; |
|
|
1116 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_LOCAL_MEM_TYPE, sizeof (value), value, 0)); |
|
|
1117 | EXTEND (SP, 1); |
|
|
1118 | const int i = 0; |
|
|
1119 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1120 | |
|
|
1121 | void |
|
|
1122 | error_correction_support (OpenCL::Device self) |
|
|
1123 | ALIAS: |
|
|
1124 | error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT |
|
|
1125 | endian_little = CL_DEVICE_ENDIAN_LITTLE |
|
|
1126 | available = CL_DEVICE_AVAILABLE |
|
|
1127 | compiler_available = CL_DEVICE_COMPILER_AVAILABLE |
|
|
1128 | host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY |
|
|
1129 | PPCODE: |
|
|
1130 | cl_bool value [1]; |
|
|
1131 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1132 | EXTEND (SP, 1); |
|
|
1133 | const int i = 0; |
|
|
1134 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
1135 | |
|
|
1136 | void |
|
|
1137 | execution_capabilities (OpenCL::Device self) |
|
|
1138 | PPCODE: |
|
|
1139 | cl_device_exec_capabilities value [1]; |
|
|
1140 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_EXECUTION_CAPABILITIES, sizeof (value), value, 0)); |
|
|
1141 | EXTEND (SP, 1); |
|
|
1142 | const int i = 0; |
|
|
1143 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1144 | |
|
|
1145 | void |
|
|
1146 | properties (OpenCL::Device self) |
|
|
1147 | PPCODE: |
|
|
1148 | cl_command_queue_properties value [1]; |
|
|
1149 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
1150 | EXTEND (SP, 1); |
|
|
1151 | const int i = 0; |
|
|
1152 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1153 | |
|
|
1154 | void |
|
|
1155 | platform (OpenCL::Device self) |
|
|
1156 | PPCODE: |
|
|
1157 | cl_platform_id value [1]; |
|
|
1158 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
|
|
1159 | EXTEND (SP, 1); |
|
|
1160 | const int i = 0; |
|
|
1161 | { |
|
|
1162 | PUSH_CLOBJ (stash_platform, value [i]); |
|
|
1163 | } |
|
|
1164 | |
|
|
1165 | void |
|
|
1166 | name (OpenCL::Device self) |
|
|
1167 | ALIAS: |
|
|
1168 | name = CL_DEVICE_NAME |
|
|
1169 | vendor = CL_DEVICE_VENDOR |
|
|
1170 | driver_version = CL_DRIVER_VERSION |
|
|
1171 | profile = CL_DEVICE_PROFILE |
|
|
1172 | version = CL_DEVICE_VERSION |
|
|
1173 | extensions = CL_DEVICE_EXTENSIONS |
|
|
1174 | PPCODE: |
|
|
1175 | size_t size; |
|
|
1176 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1177 | char *value = tmpbuf (size); |
|
|
1178 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1179 | EXTEND (SP, 1); |
|
|
1180 | const int i = 0; |
|
|
1181 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
1182 | |
|
|
1183 | void |
|
|
1184 | parent_device_ext (OpenCL::Device self) |
|
|
1185 | PPCODE: |
|
|
1186 | cl_device_id value [1]; |
|
|
1187 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
|
|
1188 | EXTEND (SP, 1); |
|
|
1189 | const int i = 0; |
|
|
1190 | { |
|
|
1191 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1192 | } |
|
|
1193 | |
|
|
1194 | void |
|
|
1195 | partition_types_ext (OpenCL::Device self) |
|
|
1196 | ALIAS: |
|
|
1197 | partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT |
|
|
1198 | affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT |
|
|
1199 | partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT |
|
|
1200 | PPCODE: |
|
|
1201 | size_t size; |
|
|
1202 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1203 | cl_device_partition_property_ext *value = tmpbuf (size); |
|
|
1204 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1205 | int i, n = size / sizeof (*value); |
|
|
1206 | EXTEND (SP, n); |
|
|
1207 | for (i = 0; i < n; ++i) |
|
|
1208 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1209 | |
|
|
1210 | #END:device |
|
|
1211 | |
|
|
1212 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1213 | |
|
|
1214 | #if CL_VERSION_1_2 |
|
|
1215 | |
|
|
1216 | void |
|
|
1217 | DESTROY (OpenCL::SubDevice self) |
|
|
1218 | CODE: |
|
|
1219 | clReleaseDevice (self); |
|
|
1220 | |
|
|
1221 | #endif |
286 | |
1222 | |
287 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1223 | MODULE = OpenCL PACKAGE = OpenCL::Context |
288 | |
1224 | |
289 | void |
1225 | void |
290 | DESTROY (OpenCL::Context context) |
1226 | DESTROY (OpenCL::Context self) |
291 | CODE: |
1227 | CODE: |
292 | clReleaseContext (context); |
1228 | clReleaseContext (self); |
293 | |
1229 | |
294 | void |
1230 | void |
295 | info (OpenCL::Context this, cl_context_info name) |
1231 | info (OpenCL::Context self, cl_context_info name) |
296 | PPCODE: |
1232 | PPCODE: |
297 | INFO (Context) |
1233 | INFO (Context) |
298 | |
1234 | |
299 | void |
1235 | void |
300 | queue (OpenCL::Context this, OpenCL::Device device, cl_command_queue_properties properties) |
1236 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
301 | PPCODE: |
1237 | PPCODE: |
302 | { |
|
|
303 | cl_int res; |
|
|
304 | cl_command_queue queue = clCreateCommandQueue (this, device, properties, &res); |
1238 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
|
|
1239 | XPUSH_CLOBJ (stash_queue, queue); |
305 | |
1240 | |
306 | if (res) |
|
|
307 | FAIL (CreateCommandQueue, res); |
|
|
308 | |
|
|
309 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
|
|
310 | } |
|
|
311 | |
|
|
312 | void |
1241 | void |
313 | user_event (OpenCL::Context this) |
1242 | user_event (OpenCL::Context self) |
314 | PPCODE: |
1243 | PPCODE: |
315 | { |
1244 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
316 | cl_int res; |
1245 | XPUSH_CLOBJ (stash_userevent, ev); |
317 | cl_event ev = clCreateUserEvent (this, &res); |
|
|
318 | |
1246 | |
319 | if (res) |
|
|
320 | FAIL (CreateUserevent, res); |
|
|
321 | |
|
|
322 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
|
|
323 | } |
|
|
324 | |
|
|
325 | void |
1247 | void |
326 | buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) |
1248 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
327 | PPCODE: |
1249 | PPCODE: |
328 | { |
|
|
329 | cl_int res; |
|
|
330 | cl_mem mem; |
|
|
331 | |
|
|
332 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1250 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
333 | croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1251 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
334 | |
1252 | |
335 | mem = clCreateBuffer (this, flags, len, 0, &res); |
1253 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
|
|
1254 | XPUSH_CLOBJ (stash_bufferobj, mem); |
336 | |
1255 | |
337 | if (res) |
|
|
338 | FAIL (CreateBuffer, res); |
|
|
339 | |
|
|
340 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
|
|
341 | } |
|
|
342 | |
|
|
343 | void |
1256 | void |
344 | buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) |
1257 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
345 | PPCODE: |
1258 | PPCODE: |
346 | { |
|
|
347 | STRLEN len; |
1259 | STRLEN len; |
348 | char *ptr = SvPVbyte (data, len); |
1260 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
349 | cl_int res; |
|
|
350 | cl_mem mem; |
|
|
351 | |
|
|
352 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1261 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
353 | croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1262 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
354 | |
1263 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
355 | mem = clCreateBuffer (this, flags, len, ptr, &res); |
1264 | XPUSH_CLOBJ (stash_bufferobj, mem); |
356 | |
1265 | |
357 | if (res) |
1266 | #if CL_VERSION_1_2 |
358 | FAIL (CreateBuffer, res); |
|
|
359 | |
1267 | |
360 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
|
|
361 | } |
|
|
362 | |
|
|
363 | void |
1268 | void |
364 | image2d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, SV *data) |
1269 | 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) |
365 | PPCODE: |
1270 | PPCODE: |
366 | { |
|
|
367 | STRLEN len; |
1271 | STRLEN len; |
368 | char *ptr = SvPVbyte (data, len); |
1272 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
369 | const cl_image_format format = { channel_order, channel_type }; |
1273 | const cl_image_format format = { channel_order, channel_type }; |
370 | cl_int res; |
1274 | const cl_image_desc desc = { |
371 | cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); |
1275 | type, |
|
|
1276 | width, height, depth, |
|
|
1277 | array_size, row_pitch, slice_pitch, |
|
|
1278 | num_mip_level, num_samples, |
|
|
1279 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
1280 | }; |
|
|
1281 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1282 | HV *stash = stash_image; |
|
|
1283 | switch (type) |
|
|
1284 | { |
|
|
1285 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1286 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1287 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1288 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1289 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1290 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1291 | } |
|
|
1292 | XPUSH_CLOBJ (stash, mem); |
372 | |
1293 | |
373 | if (res) |
1294 | #endif |
374 | FAIL (CreateImage2D, res); |
|
|
375 | |
1295 | |
376 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
377 | } |
|
|
378 | |
|
|
379 | void |
1296 | void |
380 | image3d (OpenCL::Context this, 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 slice_pitch, SV *data) |
1297 | 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) |
381 | PPCODE: |
1298 | PPCODE: |
382 | { |
|
|
383 | STRLEN len; |
1299 | STRLEN len; |
384 | char *ptr = SvPVbyte (data, len); |
1300 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
385 | const cl_image_format format = { channel_order, channel_type }; |
1301 | const cl_image_format format = { channel_order, channel_type }; |
386 | cl_int res; |
1302 | #if PREFER_1_1 |
387 | cl_mem mem = clCreateImage3D (this, flags, &format, width, height, |
1303 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
388 | depth, len / (height * slice_pitch), slice_pitch, ptr, &res); |
1304 | #else |
|
|
1305 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1306 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1307 | #endif |
|
|
1308 | XPUSH_CLOBJ (stash_image2d, mem); |
389 | |
1309 | |
390 | if (res) |
|
|
391 | FAIL (CreateImage3D, res); |
|
|
392 | |
|
|
393 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
|
|
394 | } |
|
|
395 | |
|
|
396 | void |
1310 | void |
|
|
1311 | 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) |
|
|
1312 | PPCODE: |
|
|
1313 | STRLEN len; |
|
|
1314 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1315 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1316 | #if PREFER_1_1 |
|
|
1317 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
|
|
1318 | #else |
|
|
1319 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1320 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1321 | #endif |
|
|
1322 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1323 | |
|
|
1324 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1325 | |
|
|
1326 | void |
|
|
1327 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
|
|
1328 | PPCODE: |
|
|
1329 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
|
|
1330 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1331 | |
|
|
1332 | void |
|
|
1333 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1334 | PPCODE: |
|
|
1335 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1336 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1337 | |
|
|
1338 | #if CL_VERSION_1_2 |
|
|
1339 | |
|
|
1340 | void |
|
|
1341 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1342 | ALIAS: |
|
|
1343 | PPCODE: |
|
|
1344 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1345 | cl_gl_object_type type; |
|
|
1346 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1347 | HV *stash = stash_memory; |
|
|
1348 | switch (type) |
|
|
1349 | { |
|
|
1350 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1351 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1352 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1353 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1354 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1355 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1356 | } |
|
|
1357 | XPUSH_CLOBJ (stash, mem); |
|
|
1358 | |
|
|
1359 | #endif |
|
|
1360 | |
|
|
1361 | void |
|
|
1362 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1363 | PPCODE: |
|
|
1364 | #if PREFER_1_1 |
|
|
1365 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
1366 | #else |
|
|
1367 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1368 | #endif |
|
|
1369 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1370 | |
|
|
1371 | void |
|
|
1372 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1373 | PPCODE: |
|
|
1374 | #if PREFER_1_1 |
|
|
1375 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
1376 | #else |
|
|
1377 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1378 | #endif |
|
|
1379 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1380 | |
|
|
1381 | #endif |
|
|
1382 | |
|
|
1383 | void |
397 | supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) |
1384 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
398 | PPCODE: |
1385 | PPCODE: |
399 | { |
1386 | { |
400 | cl_uint count; |
1387 | cl_uint count; |
401 | cl_image_format *list; |
1388 | cl_image_format *list; |
402 | int i; |
1389 | int i; |
403 | |
1390 | |
404 | NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); |
1391 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count)); |
405 | Newx (list, count, cl_image_format); |
1392 | Newx (list, count, cl_image_format); |
406 | NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); |
1393 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0)); |
407 | |
1394 | |
408 | EXTEND (SP, count); |
1395 | EXTEND (SP, count); |
409 | for (i = 0; i < count; ++i) |
1396 | for (i = 0; i < count; ++i) |
410 | { |
1397 | { |
411 | AV *av = newAV (); |
1398 | AV *av = newAV (); |
… | |
… | |
414 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
1401 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
415 | } |
1402 | } |
416 | } |
1403 | } |
417 | |
1404 | |
418 | void |
1405 | void |
419 | sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1406 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
420 | PPCODE: |
1407 | PPCODE: |
421 | { |
|
|
422 | cl_int res; |
|
|
423 | cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); |
1408 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
424 | |
|
|
425 | if (res) |
|
|
426 | FAIL (CreateSampler, res); |
|
|
427 | |
|
|
428 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1409 | XPUSH_CLOBJ (stash_sampler, sampler); |
429 | } |
|
|
430 | |
1410 | |
431 | void |
1411 | void |
432 | program_with_source (OpenCL::Context this, SV *program) |
1412 | program_with_source (OpenCL::Context self, SV *program) |
433 | PPCODE: |
1413 | PPCODE: |
434 | { |
|
|
435 | STRLEN len; |
1414 | STRLEN len; |
436 | size_t len2; |
1415 | size_t len2; |
437 | const char *ptr = SvPVbyte (program, len); |
1416 | const char *ptr = SvPVbyte (program, len); |
438 | cl_int res; |
|
|
439 | cl_program prog; |
|
|
440 | |
1417 | |
441 | len2 = len; |
1418 | len2 = len; |
442 | prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); |
1419 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
443 | |
|
|
444 | if (res) |
|
|
445 | FAIL (CreateProgramWithSource, res); |
|
|
446 | |
|
|
447 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1420 | XPUSH_CLOBJ (stash_program, prog); |
448 | } |
1421 | |
|
|
1422 | void |
|
|
1423 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1424 | PPCODE: |
|
|
1425 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
1426 | croak ("OpenCL::Context::program_with_binary: devices must be specified as reference to an array of device objects"); |
|
|
1427 | |
|
|
1428 | devices = SvRV (devices); |
|
|
1429 | |
|
|
1430 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1431 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1432 | |
|
|
1433 | binaries = SvRV (binaries); |
|
|
1434 | |
|
|
1435 | if (av_len ((AV *)devices) != av_len ((AV *)binaries)) |
|
|
1436 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1437 | |
|
|
1438 | int count = av_len ((AV *)devices) + 1; |
|
|
1439 | cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count); |
|
|
1440 | size_t *length_list = tmpbuf (sizeof (*length_list) * count); |
|
|
1441 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * count); |
|
|
1442 | cl_int *status_list = tmpbuf (sizeof (*status_list) * count); |
|
|
1443 | |
|
|
1444 | int i; |
|
|
1445 | for (i = 0; i < count; ++i) |
|
|
1446 | { |
|
|
1447 | device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_binary", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device"); |
|
|
1448 | STRLEN len; |
|
|
1449 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1450 | length_list [i] = len; |
|
|
1451 | } |
|
|
1452 | |
|
|
1453 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, count, device_list, length_list, binary_list, GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1454 | |
|
|
1455 | EXTEND (SP, 2); |
|
|
1456 | PUSH_CLOBJ (stash_program, prog); |
|
|
1457 | |
|
|
1458 | if (GIMME_V == G_ARRAY) |
|
|
1459 | { |
|
|
1460 | AV *av = newAV (); |
|
|
1461 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1462 | |
|
|
1463 | for (i = count; i--; ) |
|
|
1464 | av_store (av, i, newSViv (status_list [i])); |
|
|
1465 | } |
|
|
1466 | |
|
|
1467 | #if CL_VERSION_1_2 |
|
|
1468 | |
|
|
1469 | void |
|
|
1470 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1471 | PPCODE: |
|
|
1472 | cl_uint device_count = 0; |
|
|
1473 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1474 | |
|
|
1475 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1476 | |
|
|
1477 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1478 | |
|
|
1479 | void |
|
|
1480 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1481 | CODE: |
|
|
1482 | cl_uint device_count = 0; |
|
|
1483 | cl_device_id *device_list = 0; |
|
|
1484 | |
|
|
1485 | if (SvOK (devices)) |
|
|
1486 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1487 | |
|
|
1488 | cl_uint program_count; |
|
|
1489 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1490 | |
|
|
1491 | void *user_data; |
|
|
1492 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1493 | |
|
|
1494 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1495 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1496 | |
|
|
1497 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1498 | |
|
|
1499 | #endif |
|
|
1500 | |
|
|
1501 | #BEGIN:context |
|
|
1502 | |
|
|
1503 | void |
|
|
1504 | reference_count (OpenCL::Context self) |
|
|
1505 | ALIAS: |
|
|
1506 | reference_count = CL_CONTEXT_REFERENCE_COUNT |
|
|
1507 | num_devices = CL_CONTEXT_NUM_DEVICES |
|
|
1508 | PPCODE: |
|
|
1509 | cl_uint value [1]; |
|
|
1510 | NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1511 | EXTEND (SP, 1); |
|
|
1512 | const int i = 0; |
|
|
1513 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1514 | |
|
|
1515 | void |
|
|
1516 | devices (OpenCL::Context self) |
|
|
1517 | PPCODE: |
|
|
1518 | size_t size; |
|
|
1519 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size)); |
|
|
1520 | cl_device_id *value = tmpbuf (size); |
|
|
1521 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
|
|
1522 | int i, n = size / sizeof (*value); |
|
|
1523 | EXTEND (SP, n); |
|
|
1524 | for (i = 0; i < n; ++i) |
|
|
1525 | { |
|
|
1526 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1527 | } |
|
|
1528 | |
|
|
1529 | void |
|
|
1530 | properties (OpenCL::Context self) |
|
|
1531 | PPCODE: |
|
|
1532 | size_t size; |
|
|
1533 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size)); |
|
|
1534 | cl_context_properties *value = tmpbuf (size); |
|
|
1535 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0)); |
|
|
1536 | int i, n = size / sizeof (*value); |
|
|
1537 | EXTEND (SP, n); |
|
|
1538 | for (i = 0; i < n; ++i) |
|
|
1539 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
1540 | |
|
|
1541 | #END:context |
449 | |
1542 | |
450 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
1543 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
451 | |
1544 | |
452 | void |
1545 | void |
453 | DESTROY (OpenCL::Queue this) |
1546 | DESTROY (OpenCL::Queue self) |
454 | CODE: |
1547 | CODE: |
455 | clReleaseCommandQueue (this); |
1548 | clReleaseCommandQueue (self); |
456 | |
1549 | |
457 | void |
1550 | void |
458 | info (OpenCL::Queue this, cl_command_queue_info name) |
|
|
459 | PPCODE: |
|
|
460 | INFO (CommandQueue) |
|
|
461 | |
|
|
462 | void |
|
|
463 | enqueue_read_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1551 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1552 | ALIAS: |
|
|
1553 | enqueue_read_buffer = 0 |
464 | PPCODE: |
1554 | PPCODE: |
465 | { |
|
|
466 | cl_event ev = 0; |
|
|
467 | EVENT_LIST (6, items - 6); |
1555 | EVENT_LIST (6); |
468 | |
1556 | |
469 | SvUPGRADE (data, SVt_PV); |
1557 | SvUPGRADE (data, SVt_PV); |
470 | SvGROW (data, len); |
1558 | SvGROW (data, len); |
471 | SvPOK_only (data); |
1559 | SvPOK_only (data); |
472 | SvCUR_set (data, len); |
1560 | SvCUR_set (data, len); |
|
|
1561 | |
|
|
1562 | cl_event ev = 0; |
473 | NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1563 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
474 | |
1564 | |
475 | if (ev) |
1565 | if (ev) |
476 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1566 | XPUSH_CLOBJ (stash_event, ev); |
477 | } |
|
|
478 | |
1567 | |
479 | void |
1568 | void |
480 | enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1569 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1570 | ALIAS: |
|
|
1571 | enqueue_write_buffer = 0 |
481 | PPCODE: |
1572 | PPCODE: |
482 | { |
1573 | EVENT_LIST (5); |
483 | cl_event ev = 0; |
1574 | |
484 | STRLEN len; |
1575 | STRLEN len; |
485 | char *ptr = SvPVbyte (data, len); |
1576 | char *ptr = SvPVbyte (data, len); |
486 | EVENT_LIST (5, items - 5); |
|
|
487 | |
1577 | |
|
|
1578 | cl_event ev = 0; |
488 | NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1579 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
489 | |
1580 | |
490 | if (ev) |
1581 | if (ev) |
491 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1582 | XPUSH_CLOBJ (stash_event, ev); |
492 | } |
|
|
493 | |
1583 | |
|
|
1584 | #if CL_VERSION_1_2 |
|
|
1585 | |
494 | void |
1586 | void |
495 | enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1587 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1588 | ALIAS: |
|
|
1589 | enqueue_fill_buffer = 0 |
496 | PPCODE: |
1590 | PPCODE: |
497 | { |
1591 | EVENT_LIST (5); |
|
|
1592 | |
|
|
1593 | STRLEN len; |
|
|
1594 | char *ptr = SvPVbyte (data, len); |
|
|
1595 | |
498 | cl_event ev = 0; |
1596 | cl_event ev = 0; |
499 | EVENT_LIST (6, items - 6); |
|
|
500 | |
|
|
501 | NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1597 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
502 | |
1598 | |
503 | if (ev) |
1599 | if (ev) |
504 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1600 | XPUSH_CLOBJ (stash_event, ev); |
505 | } |
|
|
506 | |
1601 | |
507 | /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueReadBufferRect.html */ |
|
|
508 | /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueWriteBufferRect.html */ |
|
|
509 | |
|
|
510 | void |
1602 | void |
511 | enqueue_read_image (OpenCL::Queue this, 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, ...) |
1603 | 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, ...) |
|
|
1604 | ALIAS: |
|
|
1605 | enqueue_fill_image = 0 |
512 | PPCODE: |
1606 | PPCODE: |
513 | { |
1607 | EVENT_LIST (12); |
|
|
1608 | |
|
|
1609 | const size_t origin [3] = { x, y, z }; |
|
|
1610 | const size_t region [3] = { width, height, depth }; |
|
|
1611 | |
|
|
1612 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1613 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1614 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1615 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1616 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1617 | cl_image_format format; |
|
|
1618 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1619 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1620 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1621 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1622 | |
514 | cl_event ev = 0; |
1623 | cl_event ev = 0; |
|
|
1624 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1625 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1626 | |
|
|
1627 | if (ev) |
|
|
1628 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1629 | |
|
|
1630 | #endif |
|
|
1631 | |
|
|
1632 | void |
|
|
1633 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1634 | ALIAS: |
|
|
1635 | enqueue_copy_buffer = 0 |
|
|
1636 | PPCODE: |
|
|
1637 | EVENT_LIST (6); |
|
|
1638 | |
|
|
1639 | cl_event ev = 0; |
|
|
1640 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1641 | |
|
|
1642 | if (ev) |
|
|
1643 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1644 | |
|
|
1645 | void |
|
|
1646 | 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, ...) |
|
|
1647 | ALIAS: |
|
|
1648 | enqueue_read_buffer_rect = 0 |
|
|
1649 | PPCODE: |
|
|
1650 | EVENT_LIST (17); |
|
|
1651 | |
|
|
1652 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
515 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1653 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
516 | const size_t region[3] = { width, height, depth }; |
1654 | const size_t region[3] = { width, height, depth }; |
|
|
1655 | |
|
|
1656 | if (!buf_row_pitch) |
|
|
1657 | buf_row_pitch = region [0]; |
|
|
1658 | |
|
|
1659 | if (!buf_slice_pitch) |
|
|
1660 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1661 | |
|
|
1662 | if (!host_row_pitch) |
|
|
1663 | host_row_pitch = region [0]; |
|
|
1664 | |
|
|
1665 | if (!host_slice_pitch) |
|
|
1666 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1667 | |
517 | size_t len = row_pitch * slice_pitch * depth; |
1668 | size_t len = host_row_pitch * host_slice_pitch * region [2]; |
518 | EVENT_LIST (11, items - 11); |
|
|
519 | |
1669 | |
520 | SvUPGRADE (data, SVt_PV); |
1670 | SvUPGRADE (data, SVt_PV); |
521 | SvGROW (data, len); |
1671 | SvGROW (data, len); |
522 | SvPOK_only (data); |
1672 | SvPOK_only (data); |
523 | SvCUR_set (data, len); |
1673 | SvCUR_set (data, len); |
524 | NEED_SUCCESS (EnqueueReadImage, (this, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1674 | |
|
|
1675 | cl_event ev = 0; |
|
|
1676 | 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)); |
525 | |
1677 | |
526 | if (ev) |
1678 | if (ev) |
527 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1679 | XPUSH_CLOBJ (stash_event, ev); |
528 | } |
|
|
529 | |
1680 | |
530 | void |
1681 | void |
531 | enqueue_write_image (OpenCL::Queue this, 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, SV *data, ...) |
1682 | 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, ...) |
|
|
1683 | ALIAS: |
|
|
1684 | enqueue_write_buffer_rect = 0 |
532 | PPCODE: |
1685 | PPCODE: |
533 | { |
1686 | EVENT_LIST (17); |
|
|
1687 | |
|
|
1688 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
|
|
1689 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
|
|
1690 | const size_t region[3] = { width, height, depth }; |
|
|
1691 | STRLEN len; |
|
|
1692 | char *ptr = SvPVbyte (data, len); |
|
|
1693 | |
|
|
1694 | if (!buf_row_pitch) |
|
|
1695 | buf_row_pitch = region [0]; |
|
|
1696 | |
|
|
1697 | if (!buf_slice_pitch) |
|
|
1698 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1699 | |
|
|
1700 | if (!host_row_pitch) |
|
|
1701 | host_row_pitch = region [0]; |
|
|
1702 | |
|
|
1703 | if (!host_slice_pitch) |
|
|
1704 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1705 | |
|
|
1706 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
|
|
1707 | |
|
|
1708 | if (len < min_len) |
|
|
1709 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1710 | |
534 | cl_event ev = 0; |
1711 | cl_event ev = 0; |
|
|
1712 | 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)); |
|
|
1713 | |
|
|
1714 | if (ev) |
|
|
1715 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1716 | |
|
|
1717 | void |
|
|
1718 | 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, ...) |
|
|
1719 | ALIAS: |
|
|
1720 | enqueue_copy_buffer_rect = 0 |
|
|
1721 | PPCODE: |
|
|
1722 | EVENT_LIST (16); |
|
|
1723 | |
|
|
1724 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1725 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1726 | const size_t region[3] = { width, height, depth }; |
|
|
1727 | |
|
|
1728 | cl_event ev = 0; |
|
|
1729 | 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)); |
|
|
1730 | |
|
|
1731 | if (ev) |
|
|
1732 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1733 | |
|
|
1734 | void |
|
|
1735 | 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, ...) |
|
|
1736 | ALIAS: |
|
|
1737 | enqueue_read_image = 0 |
|
|
1738 | PPCODE: |
|
|
1739 | EVENT_LIST (12); |
|
|
1740 | |
|
|
1741 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1742 | const size_t region[3] = { width, height, depth }; |
|
|
1743 | |
|
|
1744 | if (!row_pitch) |
|
|
1745 | row_pitch = img_row_pitch (src); |
|
|
1746 | |
|
|
1747 | if (depth > 1 && !slice_pitch) |
|
|
1748 | slice_pitch = row_pitch * height; |
|
|
1749 | |
|
|
1750 | size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1751 | |
|
|
1752 | SvUPGRADE (data, SVt_PV); |
|
|
1753 | SvGROW (data, len); |
|
|
1754 | SvPOK_only (data); |
|
|
1755 | SvCUR_set (data, len); |
|
|
1756 | |
|
|
1757 | cl_event ev = 0; |
|
|
1758 | 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)); |
|
|
1759 | |
|
|
1760 | if (ev) |
|
|
1761 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1762 | |
|
|
1763 | void |
|
|
1764 | 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, ...) |
|
|
1765 | ALIAS: |
|
|
1766 | enqueue_write_image = 0 |
|
|
1767 | PPCODE: |
|
|
1768 | EVENT_LIST (12); |
|
|
1769 | |
535 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1770 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
536 | const size_t region[3] = { width, height, depth }; |
1771 | const size_t region[3] = { width, height, depth }; |
537 | STRLEN len; |
1772 | STRLEN len; |
538 | char *ptr = SvPVbyte (data, len); |
1773 | char *ptr = SvPVbyte (data, len); |
|
|
1774 | |
|
|
1775 | if (!row_pitch) |
|
|
1776 | row_pitch = img_row_pitch (dst); |
|
|
1777 | |
|
|
1778 | if (depth > 1 && !slice_pitch) |
539 | size_t slice_pitch = len / (row_pitch * height); |
1779 | slice_pitch = row_pitch * height; |
540 | EVENT_LIST (11, items - 11); |
|
|
541 | |
1780 | |
|
|
1781 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1782 | |
|
|
1783 | if (len < min_len) |
|
|
1784 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1785 | |
|
|
1786 | cl_event ev = 0; |
542 | NEED_SUCCESS (EnqueueWriteImage, (this, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1787 | 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)); |
543 | |
1788 | |
544 | if (ev) |
1789 | if (ev) |
545 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1790 | XPUSH_CLOBJ (stash_event, ev); |
546 | } |
|
|
547 | |
1791 | |
548 | void |
1792 | void |
549 | enqueue_copy_buffer_rect (OpenCL::Queue this, 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, ...) |
1793 | 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, ...) |
|
|
1794 | ALIAS: |
|
|
1795 | enqueue_copy_image = 0 |
550 | PPCODE: |
1796 | PPCODE: |
551 | { |
1797 | EVENT_LIST (12); |
552 | cl_event ev = 0; |
1798 | |
553 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1799 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
554 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1800 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
555 | const size_t region[3] = { width, height, depth }; |
1801 | const size_t region[3] = { width, height, depth }; |
556 | EVENT_LIST (16, items - 16); |
|
|
557 | |
1802 | |
|
|
1803 | cl_event ev = 0; |
558 | NEED_SUCCESS (EnqueueCopyBufferRect, (this, 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)); |
1804 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
559 | |
1805 | |
560 | if (ev) |
1806 | if (ev) |
561 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1807 | XPUSH_CLOBJ (stash_event, ev); |
562 | } |
|
|
563 | |
1808 | |
564 | void |
1809 | void |
565 | enqueue_copy_buffer_to_image (OpenCL::Queue this, 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, ...) |
1810 | 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, ...) |
|
|
1811 | ALIAS: |
|
|
1812 | enqueue_copy_image_to_buffer = 0 |
566 | PPCODE: |
1813 | PPCODE: |
567 | { |
1814 | EVENT_LIST (10); |
|
|
1815 | |
|
|
1816 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1817 | const size_t region [3] = { width, height, depth }; |
|
|
1818 | |
568 | cl_event ev = 0; |
1819 | cl_event ev = 0; |
|
|
1820 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1821 | |
|
|
1822 | if (ev) |
|
|
1823 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1824 | |
|
|
1825 | void |
|
|
1826 | 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, ...) |
|
|
1827 | ALIAS: |
|
|
1828 | enqueue_copy_buffer_to_image = 0 |
|
|
1829 | PPCODE: |
|
|
1830 | EVENT_LIST (10); |
|
|
1831 | |
569 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1832 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1833 | const size_t region [3] = { width, height, depth }; |
|
|
1834 | |
|
|
1835 | cl_event ev = 0; |
|
|
1836 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1837 | |
|
|
1838 | if (ev) |
|
|
1839 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1840 | |
|
|
1841 | void |
|
|
1842 | map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...) |
|
|
1843 | ALIAS: |
|
|
1844 | enqueue_map_buffer = 0 |
|
|
1845 | PPCODE: |
|
|
1846 | EVENT_LIST (6); |
|
|
1847 | |
|
|
1848 | size_t cb = SvIV (cb_); |
|
|
1849 | |
|
|
1850 | if (!SvOK (cb_)) |
|
|
1851 | { |
|
|
1852 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1853 | cb -= offset; |
|
|
1854 | } |
|
|
1855 | |
|
|
1856 | cl_event ev; |
|
|
1857 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1858 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1859 | |
|
|
1860 | void |
|
|
1861 | map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...) |
|
|
1862 | ALIAS: |
|
|
1863 | enqueue_map_image = 0 |
|
|
1864 | PPCODE: |
|
|
1865 | size_t width = SvIV (width_); |
|
|
1866 | if (!SvOK (width_)) |
|
|
1867 | { |
|
|
1868 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1869 | width -= x; |
|
|
1870 | } |
|
|
1871 | |
|
|
1872 | size_t height = SvIV (width_); |
|
|
1873 | if (!SvOK (height_)) |
|
|
1874 | { |
|
|
1875 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1876 | height -= y; |
|
|
1877 | } |
|
|
1878 | |
|
|
1879 | size_t depth = SvIV (width_); |
|
|
1880 | if (!SvOK (depth_)) |
|
|
1881 | { |
|
|
1882 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1883 | depth -= z; |
|
|
1884 | |
|
|
1885 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1886 | if (!depth) |
|
|
1887 | depth = 1; |
|
|
1888 | } |
|
|
1889 | |
|
|
1890 | const size_t origin[3] = { x, y, z }; |
570 | const size_t region[3] = { width, height, depth }; |
1891 | const size_t region[3] = { width, height, depth }; |
|
|
1892 | size_t row_pitch, slice_pitch; |
571 | EVENT_LIST (10, items - 10); |
1893 | EVENT_LIST (10); |
572 | |
1894 | |
573 | NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1895 | cl_event ev; |
|
|
1896 | 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)); |
|
|
1897 | |
|
|
1898 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1899 | : row_pitch ? row_pitch * region [1] |
|
|
1900 | : region [0]; |
|
|
1901 | |
|
|
1902 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1903 | |
|
|
1904 | void |
|
|
1905 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1906 | PPCODE: |
|
|
1907 | mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); |
|
|
1908 | if (GIMME_V != G_VOID) |
|
|
1909 | { |
|
|
1910 | clRetainEvent (mapped->event); |
|
|
1911 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1912 | } |
|
|
1913 | |
|
|
1914 | void |
|
|
1915 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1916 | ALIAS: |
|
|
1917 | enqueue_task = 0 |
|
|
1918 | PPCODE: |
|
|
1919 | EVENT_LIST (2); |
|
|
1920 | |
|
|
1921 | cl_event ev = 0; |
|
|
1922 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
574 | |
1923 | |
575 | if (ev) |
1924 | if (ev) |
576 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1925 | XPUSH_CLOBJ (stash_event, ev); |
577 | } |
|
|
578 | |
1926 | |
579 | void |
1927 | void |
580 | enqueue_copy_image (OpenCL::Queue this, OpenCL::Image 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, ...) |
|
|
581 | PPCODE: |
|
|
582 | { |
|
|
583 | cl_event ev = 0; |
|
|
584 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
585 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
586 | const size_t region[3] = { width, height, depth }; |
|
|
587 | EVENT_LIST (12, items - 12); |
|
|
588 | |
|
|
589 | NEED_SUCCESS (EnqueueCopyImage, (this, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
590 | |
|
|
591 | if (ev) |
|
|
592 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
593 | } |
|
|
594 | |
|
|
595 | void |
|
|
596 | enqueue_copy_image_to_buffer (OpenCL::Queue this, 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, ...) |
|
|
597 | PPCODE: |
|
|
598 | { |
|
|
599 | cl_event ev = 0; |
|
|
600 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
601 | const size_t region[3] = { width, height, depth }; |
|
|
602 | EVENT_LIST (10, items - 10); |
|
|
603 | |
|
|
604 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (this, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
605 | |
|
|
606 | if (ev) |
|
|
607 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
608 | } |
|
|
609 | |
|
|
610 | void |
|
|
611 | enqueue_task (OpenCL::Queue this, OpenCL::Kernel kernel, ...) |
|
|
612 | PPCODE: |
|
|
613 | { |
|
|
614 | cl_event ev = 0; |
|
|
615 | EVENT_LIST (2, items - 2); |
|
|
616 | |
|
|
617 | NEED_SUCCESS (EnqueueTask, (this, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
618 | |
|
|
619 | if (ev) |
|
|
620 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
621 | } |
|
|
622 | |
|
|
623 | void |
|
|
624 | enqueue_nd_range_kernel (OpenCL::Queue this, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1928 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1929 | ALIAS: |
|
|
1930 | enqueue_nd_range_kernel = 0 |
625 | PPCODE: |
1931 | PPCODE: |
626 | { |
1932 | EVENT_LIST (5); |
627 | cl_event ev = 0; |
1933 | |
628 | size_t *gwo = 0, *gws, *lws = 0; |
1934 | size_t *gwo = 0, *gws, *lws = 0; |
629 | int gws_len; |
1935 | int gws_len; |
630 | size_t *lists; |
1936 | size_t *lists; |
631 | int i; |
1937 | int i; |
632 | EVENT_LIST (5, items - 5); |
|
|
633 | |
1938 | |
634 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1939 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
635 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1940 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
636 | |
1941 | |
637 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1942 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
638 | |
1943 | |
639 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1944 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
640 | |
1945 | |
641 | gws = lists + gws_len * 0; |
1946 | gws = lists + gws_len * 0; |
642 | for (i = 0; i < gws_len; ++i) |
1947 | for (i = 0; i < gws_len; ++i) |
|
|
1948 | { |
643 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1949 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1950 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1951 | if (!gws [i]) |
|
|
1952 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1953 | } |
644 | |
1954 | |
645 | if (SvOK (global_work_offset)) |
1955 | if (SvOK (global_work_offset)) |
646 | { |
1956 | { |
647 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1957 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
648 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1958 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
655 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1965 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
656 | } |
1966 | } |
657 | |
1967 | |
658 | if (SvOK (local_work_size)) |
1968 | if (SvOK (local_work_size)) |
659 | { |
1969 | { |
660 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1970 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
661 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1971 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
662 | |
1972 | |
663 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1973 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
664 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1974 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
665 | |
1975 | |
666 | lws = lists + gws_len * 2; |
1976 | lws = lists + gws_len * 2; |
667 | for (i = 0; i < gws_len; ++i) |
1977 | for (i = 0; i < gws_len; ++i) |
|
|
1978 | { |
668 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
1979 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1980 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1981 | if (!lws [i]) |
|
|
1982 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1983 | } |
669 | } |
1984 | } |
670 | |
1985 | |
|
|
1986 | cl_event ev = 0; |
671 | NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1987 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
672 | |
1988 | |
673 | if (ev) |
1989 | if (ev) |
674 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1990 | XPUSH_CLOBJ (stash_event, ev); |
675 | } |
|
|
676 | |
1991 | |
|
|
1992 | #if CL_VERSION_1_2 |
|
|
1993 | |
677 | void |
1994 | void |
678 | enqueue_marker (OpenCL::Queue this) |
1995 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
1996 | ALIAS: |
|
|
1997 | enqueue_migrate_mem_objects = 0 |
679 | PPCODE: |
1998 | PPCODE: |
680 | { |
1999 | EVENT_LIST (3); |
|
|
2000 | |
|
|
2001 | cl_uint obj_count; |
|
|
2002 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2003 | |
|
|
2004 | cl_event ev = 0; |
|
|
2005 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2006 | |
|
|
2007 | if (ev) |
|
|
2008 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2009 | |
|
|
2010 | #endif |
|
|
2011 | |
|
|
2012 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2013 | |
|
|
2014 | void |
|
|
2015 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
2016 | ALIAS: |
|
|
2017 | release_gl_objects = 1 |
|
|
2018 | enqueue_acquire_gl_objects = 0 |
|
|
2019 | enqueue_release_gl_objects = 1 |
|
|
2020 | PPCODE: |
|
|
2021 | EVENT_LIST (2); |
|
|
2022 | |
|
|
2023 | cl_uint obj_count; |
|
|
2024 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2025 | |
681 | cl_event ev; |
2026 | cl_event ev = 0; |
682 | NEED_SUCCESS (EnqueueMarker, (this, &ev)); |
|
|
683 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
684 | } |
|
|
685 | |
2027 | |
|
|
2028 | if (ix) |
|
|
2029 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2030 | else |
|
|
2031 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2032 | |
|
|
2033 | if (ev) |
|
|
2034 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2035 | |
|
|
2036 | #endif |
|
|
2037 | |
686 | void |
2038 | void |
687 | enqueue_wait_for_events (OpenCL::Queue this, ...) |
2039 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2040 | ALIAS: |
|
|
2041 | enqueue_wait_for_events = 0 |
688 | CODE: |
2042 | CODE: |
689 | { |
|
|
690 | EVENT_LIST (1, items - 1); |
2043 | EVENT_LIST (1); |
|
|
2044 | #if PREFER_1_1 |
691 | NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); |
2045 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
692 | } |
2046 | #else |
|
|
2047 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2048 | #endif |
693 | |
2049 | |
694 | void |
2050 | void |
695 | enqueue_barrier (OpenCL::Queue this) |
2051 | marker (OpenCL::Queue self, ...) |
|
|
2052 | ALIAS: |
|
|
2053 | enqueue_marker = 0 |
696 | CODE: |
2054 | PPCODE: |
|
|
2055 | EVENT_LIST (1); |
|
|
2056 | cl_event ev = 0; |
|
|
2057 | #if PREFER_1_1 |
|
|
2058 | if (!event_list_count) |
|
|
2059 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2060 | else |
|
|
2061 | #if CL_VERSION_1_2 |
|
|
2062 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2063 | #else |
|
|
2064 | { |
|
|
2065 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2066 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2067 | } |
|
|
2068 | #endif |
|
|
2069 | #else |
|
|
2070 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2071 | #endif |
|
|
2072 | if (ev) |
|
|
2073 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2074 | |
|
|
2075 | void |
|
|
2076 | barrier (OpenCL::Queue self, ...) |
|
|
2077 | ALIAS: |
|
|
2078 | enqueue_barrier = 0 |
|
|
2079 | PPCODE: |
|
|
2080 | EVENT_LIST (1); |
|
|
2081 | cl_event ev = 0; |
|
|
2082 | #if PREFER_1_1 |
|
|
2083 | if (!event_list_count && GIMME_V == G_VOID) |
697 | NEED_SUCCESS (EnqueueBarrier, (this)); |
2084 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2085 | else |
|
|
2086 | #if CL_VERSION_1_2 |
|
|
2087 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2088 | #else |
|
|
2089 | { |
|
|
2090 | if (event_list_count) |
|
|
2091 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
698 | |
2092 | |
|
|
2093 | if (GIMME_V != G_VOID) |
|
|
2094 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2095 | } |
|
|
2096 | #endif |
|
|
2097 | #else |
|
|
2098 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2099 | #endif |
|
|
2100 | if (ev) |
|
|
2101 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2102 | |
699 | void |
2103 | void |
700 | flush (OpenCL::Queue this) |
2104 | flush (OpenCL::Queue self) |
701 | CODE: |
2105 | CODE: |
702 | NEED_SUCCESS (Flush, (this)); |
2106 | NEED_SUCCESS (Flush, (self)); |
703 | |
2107 | |
704 | void |
2108 | void |
705 | finish (OpenCL::Queue this) |
2109 | finish (OpenCL::Queue self) |
706 | CODE: |
2110 | CODE: |
707 | NEED_SUCCESS (Finish, (this)); |
2111 | NEED_SUCCESS (Finish, (self)); |
|
|
2112 | |
|
|
2113 | void |
|
|
2114 | info (OpenCL::Queue self, cl_command_queue_info name) |
|
|
2115 | PPCODE: |
|
|
2116 | INFO (CommandQueue) |
|
|
2117 | |
|
|
2118 | #BEGIN:command_queue |
|
|
2119 | |
|
|
2120 | void |
|
|
2121 | context (OpenCL::Queue self) |
|
|
2122 | PPCODE: |
|
|
2123 | cl_context value [1]; |
|
|
2124 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
|
|
2125 | EXTEND (SP, 1); |
|
|
2126 | const int i = 0; |
|
|
2127 | { |
|
|
2128 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2129 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2130 | } |
|
|
2131 | |
|
|
2132 | void |
|
|
2133 | device (OpenCL::Queue self) |
|
|
2134 | PPCODE: |
|
|
2135 | cl_device_id value [1]; |
|
|
2136 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
|
|
2137 | EXTEND (SP, 1); |
|
|
2138 | const int i = 0; |
|
|
2139 | { |
|
|
2140 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2141 | } |
|
|
2142 | |
|
|
2143 | void |
|
|
2144 | reference_count (OpenCL::Queue self) |
|
|
2145 | PPCODE: |
|
|
2146 | cl_uint value [1]; |
|
|
2147 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0)); |
|
|
2148 | EXTEND (SP, 1); |
|
|
2149 | const int i = 0; |
|
|
2150 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2151 | |
|
|
2152 | void |
|
|
2153 | properties (OpenCL::Queue self) |
|
|
2154 | PPCODE: |
|
|
2155 | cl_command_queue_properties value [1]; |
|
|
2156 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
2157 | EXTEND (SP, 1); |
|
|
2158 | const int i = 0; |
|
|
2159 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2160 | |
|
|
2161 | #END:command_queue |
708 | |
2162 | |
709 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2163 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
710 | |
2164 | |
711 | void |
2165 | void |
712 | DESTROY (OpenCL::Memory this) |
2166 | DESTROY (OpenCL::Memory self) |
713 | CODE: |
2167 | CODE: |
714 | clReleaseMemObject (this); |
2168 | clReleaseMemObject (self); |
715 | |
2169 | |
716 | void |
2170 | void |
717 | info (OpenCL::Memory this, cl_mem_info name) |
2171 | info (OpenCL::Memory self, cl_mem_info name) |
718 | PPCODE: |
2172 | PPCODE: |
719 | INFO (MemObject) |
2173 | INFO (MemObject) |
720 | |
2174 | |
|
|
2175 | #BEGIN:mem |
|
|
2176 | |
|
|
2177 | void |
|
|
2178 | type (OpenCL::Memory self) |
|
|
2179 | PPCODE: |
|
|
2180 | cl_mem_object_type value [1]; |
|
|
2181 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
|
|
2182 | EXTEND (SP, 1); |
|
|
2183 | const int i = 0; |
|
|
2184 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2185 | |
|
|
2186 | void |
|
|
2187 | flags (OpenCL::Memory self) |
|
|
2188 | PPCODE: |
|
|
2189 | cl_mem_flags value [1]; |
|
|
2190 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
|
|
2191 | EXTEND (SP, 1); |
|
|
2192 | const int i = 0; |
|
|
2193 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2194 | |
|
|
2195 | void |
|
|
2196 | size (OpenCL::Memory self) |
|
|
2197 | ALIAS: |
|
|
2198 | size = CL_MEM_SIZE |
|
|
2199 | offset = CL_MEM_OFFSET |
|
|
2200 | PPCODE: |
|
|
2201 | size_t value [1]; |
|
|
2202 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2203 | EXTEND (SP, 1); |
|
|
2204 | const int i = 0; |
|
|
2205 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2206 | |
|
|
2207 | void |
|
|
2208 | host_ptr (OpenCL::Memory self) |
|
|
2209 | PPCODE: |
|
|
2210 | void * value [1]; |
|
|
2211 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0)); |
|
|
2212 | EXTEND (SP, 1); |
|
|
2213 | const int i = 0; |
|
|
2214 | PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i]))); |
|
|
2215 | |
|
|
2216 | void |
|
|
2217 | map_count (OpenCL::Memory self) |
|
|
2218 | ALIAS: |
|
|
2219 | map_count = CL_MEM_MAP_COUNT |
|
|
2220 | reference_count = CL_MEM_REFERENCE_COUNT |
|
|
2221 | PPCODE: |
|
|
2222 | cl_uint value [1]; |
|
|
2223 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2224 | EXTEND (SP, 1); |
|
|
2225 | const int i = 0; |
|
|
2226 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2227 | |
|
|
2228 | void |
|
|
2229 | context (OpenCL::Memory self) |
|
|
2230 | PPCODE: |
|
|
2231 | cl_context value [1]; |
|
|
2232 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
|
|
2233 | EXTEND (SP, 1); |
|
|
2234 | const int i = 0; |
|
|
2235 | { |
|
|
2236 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2237 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2238 | } |
|
|
2239 | |
|
|
2240 | void |
|
|
2241 | associated_memobject (OpenCL::Memory self) |
|
|
2242 | PPCODE: |
|
|
2243 | cl_mem value [1]; |
|
|
2244 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
|
|
2245 | EXTEND (SP, 1); |
|
|
2246 | const int i = 0; |
|
|
2247 | { |
|
|
2248 | NEED_SUCCESS (RetainMemObject, (value [i])); |
|
|
2249 | PUSH_CLOBJ (stash_memory, value [i]); |
|
|
2250 | } |
|
|
2251 | |
|
|
2252 | #END:mem |
|
|
2253 | |
|
|
2254 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2255 | |
|
|
2256 | void |
|
|
2257 | gl_object_info (OpenCL::Memory self) |
|
|
2258 | PPCODE: |
|
|
2259 | cl_gl_object_type type; |
|
|
2260 | cl_GLuint name; |
|
|
2261 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
|
|
2262 | EXTEND (SP, 2); |
|
|
2263 | PUSHs (sv_2mortal (newSVuv (type))); |
|
|
2264 | PUSHs (sv_2mortal (newSVuv (name))); |
|
|
2265 | |
|
|
2266 | #endif |
|
|
2267 | |
|
|
2268 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
|
|
2269 | |
|
|
2270 | void |
|
|
2271 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
|
|
2272 | PPCODE: |
|
|
2273 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR)) |
|
|
2274 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
|
|
2275 | |
|
|
2276 | cl_buffer_region crdata = { origin, size }; |
|
|
2277 | |
|
|
2278 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
|
|
2279 | XPUSH_CLOBJ (stash_buffer, mem); |
|
|
2280 | |
|
|
2281 | MODULE = OpenCL PACKAGE = OpenCL::Image |
|
|
2282 | |
|
|
2283 | void |
|
|
2284 | image_info (OpenCL::Image self, cl_image_info name) |
|
|
2285 | PPCODE: |
|
|
2286 | INFO (Image) |
|
|
2287 | |
|
|
2288 | void |
|
|
2289 | format (OpenCL::Image self) |
|
|
2290 | PPCODE: |
|
|
2291 | cl_image_format format; |
|
|
2292 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2293 | EXTEND (SP, 2); |
|
|
2294 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2295 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
|
|
2296 | |
|
|
2297 | #BEGIN:image |
|
|
2298 | |
|
|
2299 | void |
|
|
2300 | element_size (OpenCL::Image self) |
|
|
2301 | ALIAS: |
|
|
2302 | element_size = CL_IMAGE_ELEMENT_SIZE |
|
|
2303 | row_pitch = CL_IMAGE_ROW_PITCH |
|
|
2304 | slice_pitch = CL_IMAGE_SLICE_PITCH |
|
|
2305 | width = CL_IMAGE_WIDTH |
|
|
2306 | height = CL_IMAGE_HEIGHT |
|
|
2307 | depth = CL_IMAGE_DEPTH |
|
|
2308 | PPCODE: |
|
|
2309 | size_t value [1]; |
|
|
2310 | NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2311 | EXTEND (SP, 1); |
|
|
2312 | const int i = 0; |
|
|
2313 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2314 | |
|
|
2315 | #END:image |
|
|
2316 | |
|
|
2317 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2318 | |
|
|
2319 | #BEGIN:gl_texture |
|
|
2320 | |
|
|
2321 | void |
|
|
2322 | target (OpenCL::Image self) |
|
|
2323 | PPCODE: |
|
|
2324 | cl_GLenum value [1]; |
|
|
2325 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
|
|
2326 | EXTEND (SP, 1); |
|
|
2327 | const int i = 0; |
|
|
2328 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2329 | |
|
|
2330 | void |
|
|
2331 | gl_mipmap_level (OpenCL::Image self) |
|
|
2332 | PPCODE: |
|
|
2333 | cl_GLint value [1]; |
|
|
2334 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
|
|
2335 | EXTEND (SP, 1); |
|
|
2336 | const int i = 0; |
|
|
2337 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2338 | |
|
|
2339 | #END:gl_texture |
|
|
2340 | |
|
|
2341 | #endif |
|
|
2342 | |
721 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
2343 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
722 | |
2344 | |
723 | void |
2345 | void |
724 | DESTROY (OpenCL::Sampler this) |
2346 | DESTROY (OpenCL::Sampler self) |
725 | CODE: |
2347 | CODE: |
726 | clReleaseSampler (this); |
2348 | clReleaseSampler (self); |
727 | |
2349 | |
728 | void |
2350 | void |
729 | info (OpenCL::Sampler this, cl_sampler_info name) |
2351 | info (OpenCL::Sampler self, cl_sampler_info name) |
730 | PPCODE: |
2352 | PPCODE: |
731 | INFO (Sampler) |
2353 | INFO (Sampler) |
732 | |
2354 | |
|
|
2355 | #BEGIN:sampler |
|
|
2356 | |
|
|
2357 | void |
|
|
2358 | reference_count (OpenCL::Sampler self) |
|
|
2359 | PPCODE: |
|
|
2360 | cl_uint value [1]; |
|
|
2361 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_REFERENCE_COUNT, sizeof (value), value, 0)); |
|
|
2362 | EXTEND (SP, 1); |
|
|
2363 | const int i = 0; |
|
|
2364 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2365 | |
|
|
2366 | void |
|
|
2367 | context (OpenCL::Sampler self) |
|
|
2368 | PPCODE: |
|
|
2369 | cl_context value [1]; |
|
|
2370 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
|
|
2371 | EXTEND (SP, 1); |
|
|
2372 | const int i = 0; |
|
|
2373 | { |
|
|
2374 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2375 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2376 | } |
|
|
2377 | |
|
|
2378 | void |
|
|
2379 | normalized_coords (OpenCL::Sampler self) |
|
|
2380 | PPCODE: |
|
|
2381 | cl_addressing_mode value [1]; |
|
|
2382 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
|
|
2383 | EXTEND (SP, 1); |
|
|
2384 | const int i = 0; |
|
|
2385 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2386 | |
|
|
2387 | void |
|
|
2388 | addressing_mode (OpenCL::Sampler self) |
|
|
2389 | PPCODE: |
|
|
2390 | cl_filter_mode value [1]; |
|
|
2391 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
|
|
2392 | EXTEND (SP, 1); |
|
|
2393 | const int i = 0; |
|
|
2394 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2395 | |
|
|
2396 | void |
|
|
2397 | filter_mode (OpenCL::Sampler self) |
|
|
2398 | PPCODE: |
|
|
2399 | cl_bool value [1]; |
|
|
2400 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0)); |
|
|
2401 | EXTEND (SP, 1); |
|
|
2402 | const int i = 0; |
|
|
2403 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
2404 | |
|
|
2405 | #END:sampler |
|
|
2406 | |
733 | MODULE = OpenCL PACKAGE = OpenCL::Program |
2407 | MODULE = OpenCL PACKAGE = OpenCL::Program |
734 | |
2408 | |
735 | void |
2409 | void |
736 | DESTROY (OpenCL::Program this) |
2410 | DESTROY (OpenCL::Program self) |
737 | CODE: |
2411 | CODE: |
738 | clReleaseProgram (this); |
2412 | clReleaseProgram (self); |
739 | |
2413 | |
740 | void |
2414 | void |
741 | info (OpenCL::Program this, cl_program_info name) |
2415 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2416 | ALIAS: |
|
|
2417 | build_async = 1 |
742 | PPCODE: |
2418 | CODE: |
743 | INFO (Program) |
2419 | cl_uint device_count = 0; |
|
|
2420 | cl_device_id *device_list = 0; |
744 | |
2421 | |
|
|
2422 | if (SvOK (devices)) |
|
|
2423 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2424 | |
|
|
2425 | void *user_data; |
|
|
2426 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2427 | |
|
|
2428 | if (ix) |
|
|
2429 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2430 | else |
|
|
2431 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2432 | |
|
|
2433 | #if CL_VERSION_1_2 |
|
|
2434 | |
745 | void |
2435 | void |
746 | build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) |
2436 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
747 | CODE: |
2437 | CODE: |
|
|
2438 | cl_uint device_count = 0; |
|
|
2439 | cl_device_id *device_list = 0; |
|
|
2440 | |
|
|
2441 | if (SvOK (devices)) |
|
|
2442 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2443 | |
|
|
2444 | cl_uint header_count = 0; |
|
|
2445 | cl_program *header_list = 0; |
|
|
2446 | const char **header_name = 0; |
|
|
2447 | |
|
|
2448 | if (SvOK (headers)) |
|
|
2449 | { |
|
|
2450 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2451 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2452 | |
|
|
2453 | HV *hv = (HV *)SvRV (devices); |
|
|
2454 | |
|
|
2455 | header_count = hv_iterinit (hv); |
|
|
2456 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2457 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2458 | |
|
|
2459 | HE *he; |
|
|
2460 | int i = 0; |
|
|
2461 | while (he = hv_iternext (hv)) |
|
|
2462 | { |
|
|
2463 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2464 | header_list [i] = SvCLOBJ ("clCompileProgram", "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2465 | ++i; |
|
|
2466 | } |
|
|
2467 | } |
|
|
2468 | |
|
|
2469 | void *user_data; |
|
|
2470 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2471 | |
748 | NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2472 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2473 | header_count, header_list, header_name, pfn_notify, user_data)); |
749 | |
2474 | |
|
|
2475 | #endif |
|
|
2476 | |
750 | void |
2477 | void |
751 | build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) |
2478 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
752 | PPCODE: |
2479 | PPCODE: |
753 | { |
|
|
754 | size_t size; |
2480 | size_t size; |
755 | SV *sv; |
|
|
756 | |
|
|
757 | NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); |
2481 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size)); |
758 | sv = sv_2mortal (newSV (size)); |
2482 | SV *sv = sv_2mortal (newSV (size)); |
759 | SvUPGRADE (sv, SVt_PV); |
2483 | SvUPGRADE (sv, SVt_PV); |
760 | SvPOK_only (sv); |
2484 | SvPOK_only (sv); |
761 | SvCUR_set (sv, size); |
2485 | SvCUR_set (sv, size); |
762 | NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); |
2486 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0)); |
763 | XPUSHs (sv); |
2487 | XPUSHs (sv); |
764 | } |
2488 | |
|
|
2489 | #BEGIN:program_build |
|
|
2490 | |
|
|
2491 | void |
|
|
2492 | build_status (OpenCL::Program self, OpenCL::Device device) |
|
|
2493 | PPCODE: |
|
|
2494 | cl_build_status value [1]; |
|
|
2495 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0)); |
|
|
2496 | EXTEND (SP, 1); |
|
|
2497 | const int i = 0; |
|
|
2498 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2499 | |
|
|
2500 | void |
|
|
2501 | build_options (OpenCL::Program self, OpenCL::Device device) |
|
|
2502 | ALIAS: |
|
|
2503 | build_options = CL_PROGRAM_BUILD_OPTIONS |
|
|
2504 | build_log = CL_PROGRAM_BUILD_LOG |
|
|
2505 | PPCODE: |
|
|
2506 | size_t size; |
|
|
2507 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size)); |
|
|
2508 | char *value = tmpbuf (size); |
|
|
2509 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
|
|
2510 | EXTEND (SP, 1); |
|
|
2511 | const int i = 0; |
|
|
2512 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2513 | |
|
|
2514 | #END:program_build |
765 | |
2515 | |
766 | void |
2516 | void |
767 | kernel (OpenCL::Program program, SV *function) |
2517 | kernel (OpenCL::Program program, SV *function) |
768 | PPCODE: |
2518 | PPCODE: |
769 | { |
|
|
770 | cl_int res; |
|
|
771 | cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); |
2519 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
|
|
2520 | XPUSH_CLOBJ (stash_kernel, kernel); |
772 | |
2521 | |
773 | if (res) |
2522 | void |
774 | FAIL (CreateKernel, res); |
2523 | kernels_in_program (OpenCL::Program program) |
|
|
2524 | PPCODE: |
|
|
2525 | cl_uint num_kernels; |
|
|
2526 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2527 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2528 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
775 | |
2529 | |
776 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2530 | int i; |
777 | } |
2531 | EXTEND (SP, num_kernels); |
|
|
2532 | for (i = 0; i < num_kernels; ++i) |
|
|
2533 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
|
|
2534 | |
|
|
2535 | void |
|
|
2536 | info (OpenCL::Program self, cl_program_info name) |
|
|
2537 | PPCODE: |
|
|
2538 | INFO (Program) |
|
|
2539 | |
|
|
2540 | void |
|
|
2541 | binaries (OpenCL::Program self) |
|
|
2542 | PPCODE: |
|
|
2543 | cl_uint n, i; |
|
|
2544 | size_t size; |
|
|
2545 | |
|
|
2546 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0)); |
|
|
2547 | if (!n) XSRETURN_EMPTY; |
|
|
2548 | |
|
|
2549 | size_t *sizes = tmpbuf (sizeof (*sizes) * n); |
|
|
2550 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size)); |
|
|
2551 | if (size != sizeof (*sizes) * n) XSRETURN_EMPTY; |
|
|
2552 | unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n); |
|
|
2553 | |
|
|
2554 | EXTEND (SP, n); |
|
|
2555 | for (i = 0; i < n; ++i) |
|
|
2556 | { |
|
|
2557 | SV *sv = sv_2mortal (newSV (sizes [i])); |
|
|
2558 | SvUPGRADE (sv, SVt_PV); |
|
|
2559 | SvPOK_only (sv); |
|
|
2560 | SvCUR_set (sv, sizes [i]); |
|
|
2561 | ptrs [i] = (void *)SvPVX (sv); |
|
|
2562 | PUSHs (sv); |
|
|
2563 | } |
|
|
2564 | |
|
|
2565 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
|
|
2566 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
|
|
2567 | |
|
|
2568 | #BEGIN:program |
|
|
2569 | |
|
|
2570 | void |
|
|
2571 | reference_count (OpenCL::Program self) |
|
|
2572 | ALIAS: |
|
|
2573 | reference_count = CL_PROGRAM_REFERENCE_COUNT |
|
|
2574 | num_devices = CL_PROGRAM_NUM_DEVICES |
|
|
2575 | PPCODE: |
|
|
2576 | cl_uint value [1]; |
|
|
2577 | NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2578 | EXTEND (SP, 1); |
|
|
2579 | const int i = 0; |
|
|
2580 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2581 | |
|
|
2582 | void |
|
|
2583 | context (OpenCL::Program self) |
|
|
2584 | PPCODE: |
|
|
2585 | cl_context value [1]; |
|
|
2586 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
|
|
2587 | EXTEND (SP, 1); |
|
|
2588 | const int i = 0; |
|
|
2589 | { |
|
|
2590 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2591 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2592 | } |
|
|
2593 | |
|
|
2594 | void |
|
|
2595 | devices (OpenCL::Program self) |
|
|
2596 | PPCODE: |
|
|
2597 | size_t size; |
|
|
2598 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size)); |
|
|
2599 | cl_device_id *value = tmpbuf (size); |
|
|
2600 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
|
|
2601 | int i, n = size / sizeof (*value); |
|
|
2602 | EXTEND (SP, n); |
|
|
2603 | for (i = 0; i < n; ++i) |
|
|
2604 | { |
|
|
2605 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2606 | } |
|
|
2607 | |
|
|
2608 | void |
|
|
2609 | source (OpenCL::Program self) |
|
|
2610 | PPCODE: |
|
|
2611 | size_t size; |
|
|
2612 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size)); |
|
|
2613 | char *value = tmpbuf (size); |
|
|
2614 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0)); |
|
|
2615 | EXTEND (SP, 1); |
|
|
2616 | const int i = 0; |
|
|
2617 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2618 | |
|
|
2619 | void |
|
|
2620 | binary_sizes (OpenCL::Program self) |
|
|
2621 | PPCODE: |
|
|
2622 | size_t size; |
|
|
2623 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size)); |
|
|
2624 | size_t *value = tmpbuf (size); |
|
|
2625 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0)); |
|
|
2626 | int i, n = size / sizeof (*value); |
|
|
2627 | EXTEND (SP, n); |
|
|
2628 | for (i = 0; i < n; ++i) |
|
|
2629 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2630 | |
|
|
2631 | #END:program |
778 | |
2632 | |
779 | MODULE = OpenCL PACKAGE = OpenCL::Kernel |
2633 | MODULE = OpenCL PACKAGE = OpenCL::Kernel |
780 | |
2634 | |
781 | void |
2635 | void |
782 | DESTROY (OpenCL::Kernel this) |
2636 | DESTROY (OpenCL::Kernel self) |
783 | CODE: |
2637 | CODE: |
784 | clReleaseKernel (this); |
2638 | clReleaseKernel (self); |
785 | |
2639 | |
786 | void |
2640 | void |
|
|
2641 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2642 | CODE: |
|
|
2643 | int i; |
|
|
2644 | for (i = 2; ; ++i) |
|
|
2645 | { |
|
|
2646 | while (*format == ' ') |
|
|
2647 | ++format; |
|
|
2648 | |
|
|
2649 | char type = *format++; |
|
|
2650 | |
|
|
2651 | if (!type) |
|
|
2652 | break; |
|
|
2653 | |
|
|
2654 | if (i >= items) |
|
|
2655 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2656 | |
|
|
2657 | SV *sv = ST (i); |
|
|
2658 | |
|
|
2659 | union |
|
|
2660 | { |
|
|
2661 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2662 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2663 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2664 | cl_mem cm; |
|
|
2665 | cl_sampler ca; |
|
|
2666 | size_t cz; |
|
|
2667 | cl_event ce; |
|
|
2668 | } arg; |
|
|
2669 | size_t size; |
|
|
2670 | int nullarg = 0; |
|
|
2671 | |
|
|
2672 | switch (type) |
|
|
2673 | { |
|
|
2674 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2675 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2676 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2677 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2678 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2679 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2680 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2681 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2682 | |
|
|
2683 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2684 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2685 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2686 | |
|
|
2687 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2688 | |
|
|
2689 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2690 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2691 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2692 | |
|
|
2693 | default: |
|
|
2694 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2695 | } |
|
|
2696 | |
|
|
2697 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2698 | if (res) |
|
|
2699 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2700 | } |
|
|
2701 | |
|
|
2702 | if (i != items) |
|
|
2703 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2704 | |
|
|
2705 | void |
|
|
2706 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
|
|
2707 | CODE: |
|
|
2708 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2709 | |
|
|
2710 | void |
|
|
2711 | set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value) |
|
|
2712 | CODE: |
|
|
2713 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2714 | |
|
|
2715 | void |
|
|
2716 | set_short (OpenCL::Kernel self, cl_uint idx, cl_short value) |
|
|
2717 | CODE: |
|
|
2718 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2719 | |
|
|
2720 | void |
|
|
2721 | set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value) |
|
|
2722 | CODE: |
|
|
2723 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2724 | |
|
|
2725 | void |
|
|
2726 | set_int (OpenCL::Kernel self, cl_uint idx, cl_int value) |
|
|
2727 | CODE: |
|
|
2728 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2729 | |
|
|
2730 | void |
|
|
2731 | set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value) |
|
|
2732 | CODE: |
|
|
2733 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2734 | |
|
|
2735 | void |
|
|
2736 | set_long (OpenCL::Kernel self, cl_uint idx, cl_long value) |
|
|
2737 | CODE: |
|
|
2738 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2739 | |
|
|
2740 | void |
|
|
2741 | set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value) |
|
|
2742 | CODE: |
|
|
2743 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2744 | |
|
|
2745 | void |
|
|
2746 | set_half (OpenCL::Kernel self, cl_uint idx, cl_half value) |
|
|
2747 | CODE: |
|
|
2748 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2749 | |
|
|
2750 | void |
|
|
2751 | set_float (OpenCL::Kernel self, cl_uint idx, cl_float value) |
|
|
2752 | CODE: |
|
|
2753 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2754 | |
|
|
2755 | void |
|
|
2756 | set_double (OpenCL::Kernel self, cl_uint idx, cl_double value) |
|
|
2757 | CODE: |
|
|
2758 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2759 | |
|
|
2760 | void |
|
|
2761 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
|
|
2762 | CODE: |
|
|
2763 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2764 | |
|
|
2765 | void |
|
|
2766 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2767 | CODE: |
|
|
2768 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2769 | |
|
|
2770 | void |
|
|
2771 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2772 | ALIAS: |
|
|
2773 | set_image2d = 0 |
|
|
2774 | set_image3d = 0 |
|
|
2775 | CODE: |
|
|
2776 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2777 | |
|
|
2778 | void |
|
|
2779 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2780 | CODE: |
|
|
2781 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2782 | |
|
|
2783 | void |
|
|
2784 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
|
|
2785 | CODE: |
|
|
2786 | clSetKernelArg (self, idx, size, 0); |
|
|
2787 | |
|
|
2788 | void |
|
|
2789 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
|
|
2790 | CODE: |
|
|
2791 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2792 | |
|
|
2793 | void |
787 | info (OpenCL::Kernel this, cl_kernel_info name) |
2794 | info (OpenCL::Kernel self, cl_kernel_info name) |
788 | PPCODE: |
2795 | PPCODE: |
789 | INFO (Kernel) |
2796 | INFO (Kernel) |
790 | |
2797 | |
|
|
2798 | #BEGIN:kernel |
|
|
2799 | |
791 | void |
2800 | void |
792 | set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) |
2801 | function_name (OpenCL::Kernel self) |
|
|
2802 | PPCODE: |
|
|
2803 | size_t size; |
|
|
2804 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size)); |
|
|
2805 | char *value = tmpbuf (size); |
|
|
2806 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0)); |
|
|
2807 | EXTEND (SP, 1); |
|
|
2808 | const int i = 0; |
|
|
2809 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2810 | |
|
|
2811 | void |
|
|
2812 | num_args (OpenCL::Kernel self) |
|
|
2813 | ALIAS: |
|
|
2814 | num_args = CL_KERNEL_NUM_ARGS |
|
|
2815 | reference_count = CL_KERNEL_REFERENCE_COUNT |
|
|
2816 | PPCODE: |
|
|
2817 | cl_uint value [1]; |
|
|
2818 | NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2819 | EXTEND (SP, 1); |
|
|
2820 | const int i = 0; |
|
|
2821 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2822 | |
|
|
2823 | void |
|
|
2824 | context (OpenCL::Kernel self) |
|
|
2825 | PPCODE: |
|
|
2826 | cl_context value [1]; |
|
|
2827 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
|
|
2828 | EXTEND (SP, 1); |
|
|
2829 | const int i = 0; |
|
|
2830 | { |
|
|
2831 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2832 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2833 | } |
|
|
2834 | |
|
|
2835 | void |
|
|
2836 | program (OpenCL::Kernel self) |
|
|
2837 | PPCODE: |
|
|
2838 | cl_program value [1]; |
|
|
2839 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
|
|
2840 | EXTEND (SP, 1); |
|
|
2841 | const int i = 0; |
|
|
2842 | { |
|
|
2843 | NEED_SUCCESS (RetainProgram, (value [i])); |
|
|
2844 | PUSH_CLOBJ (stash_program, value [i]); |
|
|
2845 | } |
|
|
2846 | |
|
|
2847 | #END:kernel |
|
|
2848 | |
|
|
2849 | void |
|
|
2850 | work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name) |
793 | CODE: |
2851 | PPCODE: |
794 | clSetKernelArg (this, idx, sizeof (value), &value); |
2852 | size_t size; |
|
|
2853 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size)); |
|
|
2854 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2855 | SvUPGRADE (sv, SVt_PV); |
|
|
2856 | SvPOK_only (sv); |
|
|
2857 | SvCUR_set (sv, size); |
|
|
2858 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0)); |
|
|
2859 | XPUSHs (sv); |
795 | |
2860 | |
|
|
2861 | #BEGIN:kernel_work_group |
|
|
2862 | |
796 | void |
2863 | void |
797 | set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value) |
2864 | work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2865 | ALIAS: |
|
|
2866 | work_group_size = CL_KERNEL_WORK_GROUP_SIZE |
|
|
2867 | preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE |
|
|
2868 | PPCODE: |
|
|
2869 | size_t value [1]; |
|
|
2870 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
|
|
2871 | EXTEND (SP, 1); |
|
|
2872 | const int i = 0; |
|
|
2873 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2874 | |
|
|
2875 | void |
|
|
2876 | compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2877 | PPCODE: |
|
|
2878 | size_t size; |
|
|
2879 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size)); |
|
|
2880 | size_t *value = tmpbuf (size); |
|
|
2881 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0)); |
|
|
2882 | int i, n = size / sizeof (*value); |
|
|
2883 | EXTEND (SP, n); |
|
|
2884 | for (i = 0; i < n; ++i) |
|
|
2885 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2886 | |
|
|
2887 | void |
|
|
2888 | local_mem_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2889 | ALIAS: |
|
|
2890 | local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE |
|
|
2891 | private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE |
|
|
2892 | PPCODE: |
|
|
2893 | cl_ulong value [1]; |
|
|
2894 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
|
|
2895 | EXTEND (SP, 1); |
|
|
2896 | const int i = 0; |
|
|
2897 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2898 | |
|
|
2899 | #END:kernel_work_group |
|
|
2900 | |
|
|
2901 | #if CL_VERSION_1_2 |
|
|
2902 | |
|
|
2903 | void |
|
|
2904 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
798 | CODE: |
2905 | PPCODE: |
799 | clSetKernelArg (this, idx, sizeof (value), &value); |
2906 | size_t size; |
|
|
2907 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2908 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2909 | SvUPGRADE (sv, SVt_PV); |
|
|
2910 | SvPOK_only (sv); |
|
|
2911 | SvCUR_set (sv, size); |
|
|
2912 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2913 | XPUSHs (sv); |
800 | |
2914 | |
801 | void |
2915 | #BEGIN:kernel_arg |
802 | set_short (OpenCL::Kernel this, cl_uint idx, cl_short value) |
|
|
803 | CODE: |
|
|
804 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
805 | |
2916 | |
806 | void |
2917 | void |
807 | set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) |
2918 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
808 | CODE: |
2919 | PPCODE: |
809 | clSetKernelArg (this, idx, sizeof (value), &value); |
2920 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2921 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2922 | EXTEND (SP, 1); |
|
|
2923 | const int i = 0; |
|
|
2924 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
810 | |
2925 | |
811 | void |
2926 | void |
812 | set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) |
2927 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
813 | CODE: |
2928 | PPCODE: |
814 | clSetKernelArg (this, idx, sizeof (value), &value); |
2929 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2930 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2931 | EXTEND (SP, 1); |
|
|
2932 | const int i = 0; |
|
|
2933 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
815 | |
2934 | |
816 | void |
2935 | void |
817 | set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value) |
2936 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
818 | CODE: |
2937 | ALIAS: |
819 | clSetKernelArg (this, idx, sizeof (value), &value); |
2938 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2939 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2940 | PPCODE: |
|
|
2941 | size_t size; |
|
|
2942 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2943 | char *value = tmpbuf (size); |
|
|
2944 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2945 | EXTEND (SP, 1); |
|
|
2946 | const int i = 0; |
|
|
2947 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
820 | |
2948 | |
821 | void |
2949 | void |
822 | set_long (OpenCL::Kernel this, cl_uint idx, cl_long value) |
2950 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
823 | CODE: |
2951 | PPCODE: |
824 | clSetKernelArg (this, idx, sizeof (value), &value); |
2952 | cl_kernel_arg_type_qualifier value [1]; |
|
|
2953 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2954 | EXTEND (SP, 1); |
|
|
2955 | const int i = 0; |
|
|
2956 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
825 | |
2957 | |
826 | void |
2958 | #END:kernel_arg |
827 | set_ulong (OpenCL::Kernel this, cl_uint idx, cl_ulong value) |
|
|
828 | CODE: |
|
|
829 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
830 | |
2959 | |
831 | void |
2960 | #endif |
832 | set_half (OpenCL::Kernel this, cl_uint idx, cl_half value) |
|
|
833 | CODE: |
|
|
834 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
835 | |
|
|
836 | void |
|
|
837 | set_float (OpenCL::Kernel this, cl_uint idx, cl_float value) |
|
|
838 | CODE: |
|
|
839 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
840 | |
|
|
841 | void |
|
|
842 | set_double (OpenCL::Kernel this, cl_uint idx, cl_double value) |
|
|
843 | CODE: |
|
|
844 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
845 | |
|
|
846 | void |
|
|
847 | set_memory (OpenCL::Kernel this, cl_uint idx, OpenCL::Memory_ornull value) |
|
|
848 | CODE: |
|
|
849 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
850 | |
|
|
851 | void |
|
|
852 | set_buffer (OpenCL::Kernel this, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
853 | CODE: |
|
|
854 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
855 | |
|
|
856 | void |
|
|
857 | set_image2d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
858 | CODE: |
|
|
859 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
860 | |
|
|
861 | void |
|
|
862 | set_image3d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
863 | CODE: |
|
|
864 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
865 | |
|
|
866 | void |
|
|
867 | set_sampler (OpenCL::Kernel this, cl_uint idx, OpenCL::Sampler value) |
|
|
868 | CODE: |
|
|
869 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
870 | |
|
|
871 | void |
|
|
872 | set_event (OpenCL::Kernel this, cl_uint idx, OpenCL::Event value) |
|
|
873 | CODE: |
|
|
874 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
875 | |
2961 | |
876 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2962 | MODULE = OpenCL PACKAGE = OpenCL::Event |
877 | |
2963 | |
878 | void |
2964 | void |
879 | DESTROY (OpenCL::Event this) |
2965 | DESTROY (OpenCL::Event self) |
880 | CODE: |
2966 | CODE: |
881 | clReleaseEvent (this); |
2967 | clReleaseEvent (self); |
882 | |
2968 | |
883 | void |
2969 | void |
|
|
2970 | wait (OpenCL::Event self) |
|
|
2971 | CODE: |
|
|
2972 | clWaitForEvents (1, &self); |
|
|
2973 | |
|
|
2974 | void |
|
|
2975 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2976 | CODE: |
|
|
2977 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
|
|
2978 | |
|
|
2979 | void |
884 | info (OpenCL::Event this, cl_event_info name) |
2980 | info (OpenCL::Event self, cl_event_info name) |
885 | PPCODE: |
2981 | PPCODE: |
886 | INFO (Event) |
2982 | INFO (Event) |
887 | |
2983 | |
|
|
2984 | #BEGIN:event |
|
|
2985 | |
888 | void |
2986 | void |
889 | wait (OpenCL::Event this) |
2987 | command_queue (OpenCL::Event self) |
|
|
2988 | PPCODE: |
|
|
2989 | cl_command_queue value [1]; |
|
|
2990 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
|
|
2991 | EXTEND (SP, 1); |
|
|
2992 | const int i = 0; |
|
|
2993 | { |
|
|
2994 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
|
|
2995 | PUSH_CLOBJ (stash_queue, value [i]); |
|
|
2996 | } |
|
|
2997 | |
|
|
2998 | void |
|
|
2999 | command_type (OpenCL::Event self) |
|
|
3000 | PPCODE: |
|
|
3001 | cl_command_type value [1]; |
|
|
3002 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_TYPE, sizeof (value), value, 0)); |
|
|
3003 | EXTEND (SP, 1); |
|
|
3004 | const int i = 0; |
|
|
3005 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3006 | |
|
|
3007 | void |
|
|
3008 | reference_count (OpenCL::Event self) |
|
|
3009 | ALIAS: |
|
|
3010 | reference_count = CL_EVENT_REFERENCE_COUNT |
|
|
3011 | command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS |
|
|
3012 | PPCODE: |
|
|
3013 | cl_uint value [1]; |
|
|
3014 | NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0)); |
|
|
3015 | EXTEND (SP, 1); |
|
|
3016 | const int i = 0; |
|
|
3017 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3018 | |
|
|
3019 | void |
|
|
3020 | context (OpenCL::Event self) |
|
|
3021 | PPCODE: |
|
|
3022 | cl_context value [1]; |
|
|
3023 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
|
|
3024 | EXTEND (SP, 1); |
|
|
3025 | const int i = 0; |
|
|
3026 | { |
|
|
3027 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
3028 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
3029 | } |
|
|
3030 | |
|
|
3031 | #END:event |
|
|
3032 | |
|
|
3033 | void |
|
|
3034 | profiling_info (OpenCL::Event self, cl_profiling_info name) |
890 | CODE: |
3035 | PPCODE: |
891 | clWaitForEvents (1, &this); |
3036 | INFO (EventProfiling) |
|
|
3037 | |
|
|
3038 | #BEGIN:profiling |
|
|
3039 | |
|
|
3040 | void |
|
|
3041 | profiling_command_queued (OpenCL::Event self) |
|
|
3042 | ALIAS: |
|
|
3043 | profiling_command_queued = CL_PROFILING_COMMAND_QUEUED |
|
|
3044 | profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT |
|
|
3045 | profiling_command_start = CL_PROFILING_COMMAND_START |
|
|
3046 | profiling_command_end = CL_PROFILING_COMMAND_END |
|
|
3047 | PPCODE: |
|
|
3048 | cl_ulong value [1]; |
|
|
3049 | NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0)); |
|
|
3050 | EXTEND (SP, 1); |
|
|
3051 | const int i = 0; |
|
|
3052 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3053 | |
|
|
3054 | #END:profiling |
892 | |
3055 | |
893 | MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
3056 | MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
894 | |
3057 | |
895 | void |
3058 | void |
896 | set_status (OpenCL::UserEvent this, cl_int execution_status) |
3059 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
897 | CODE: |
3060 | CODE: |
898 | clSetUserEventStatus (this, execution_status); |
3061 | clSetUserEventStatus (self, execution_status); |
899 | |
3062 | |
|
|
3063 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3064 | |
|
|
3065 | void |
|
|
3066 | DESTROY (SV *self) |
|
|
3067 | CODE: |
|
|
3068 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3069 | |
|
|
3070 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3071 | mapped_detach (self, mapped); |
|
|
3072 | |
|
|
3073 | clReleaseCommandQueue (mapped->queue); |
|
|
3074 | clReleaseEvent (mapped->event); |
|
|
3075 | Safefree (mapped); |
|
|
3076 | |
|
|
3077 | void |
|
|
3078 | unmap (OpenCL::Mapped self, ...) |
|
|
3079 | CODE: |
|
|
3080 | mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3081 | |
|
|
3082 | bool |
|
|
3083 | mapped (OpenCL::Mapped self) |
|
|
3084 | CODE: |
|
|
3085 | RETVAL = !!self->ptr; |
|
|
3086 | OUTPUT: |
|
|
3087 | RETVAL |
|
|
3088 | |
|
|
3089 | void |
|
|
3090 | wait (OpenCL::Mapped self) |
|
|
3091 | PPCODE: |
|
|
3092 | if (self->event) |
|
|
3093 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3094 | |
|
|
3095 | void |
|
|
3096 | event (OpenCL::Mapped self) |
|
|
3097 | PPCODE: |
|
|
3098 | if (!self->event) |
|
|
3099 | XSRETURN_UNDEF; |
|
|
3100 | |
|
|
3101 | clRetainEvent (self->event); |
|
|
3102 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3103 | |
|
|
3104 | size_t |
|
|
3105 | size (OpenCL::Mapped self) |
|
|
3106 | CODE: |
|
|
3107 | RETVAL = self->cb; |
|
|
3108 | OUTPUT: |
|
|
3109 | RETVAL |
|
|
3110 | |
|
|
3111 | IV |
|
|
3112 | ptr (OpenCL::Mapped self) |
|
|
3113 | CODE: |
|
|
3114 | RETVAL = PTR2IV (self->ptr); |
|
|
3115 | OUTPUT: |
|
|
3116 | RETVAL |
|
|
3117 | |
|
|
3118 | void |
|
|
3119 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3120 | CODE: |
|
|
3121 | STRLEN len; |
|
|
3122 | const char *ptr = SvPVbyte (data, len); |
|
|
3123 | |
|
|
3124 | if (offset + len > self->cb) |
|
|
3125 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3126 | |
|
|
3127 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3128 | |
|
|
3129 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3130 | |
|
|
3131 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3132 | |
|
|
3133 | IV |
|
|
3134 | row_pitch (OpenCL::Mapped self) |
|
|
3135 | ALIAS: |
|
|
3136 | slice_pitch = 1 |
|
|
3137 | CODE: |
|
|
3138 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
3139 | OUTPUT: |
|
|
3140 | RETVAL |
|
|
3141 | |
|
|
3142 | |
|
|
3143 | |