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; |
|
|
48 | typedef cl_mem OpenCL__Memory; |
|
|
49 | typedef cl_mem OpenCL__Buffer; |
|
|
50 | typedef cl_mem OpenCL__BufferObj; |
|
|
51 | typedef cl_mem OpenCL__Image; |
|
|
52 | typedef cl_mem OpenCL__Memory_ornull; |
|
|
53 | typedef cl_mem OpenCL__Buffer_ornull; |
|
|
54 | typedef cl_mem OpenCL__Image_ornull; |
|
|
55 | typedef cl_sampler OpenCL__Sampler; |
|
|
56 | typedef cl_program OpenCL__Program; |
|
|
57 | typedef cl_kernel OpenCL__Kernel; |
|
|
58 | typedef cl_event OpenCL__Event; |
|
|
59 | typedef cl_event OpenCL__UserEvent; |
11 | |
60 | |
12 | static const struct { |
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; |
|
|
89 | |
|
|
90 | /*****************************************************************************/ |
|
|
91 | |
|
|
92 | // name must include a leading underscore |
|
|
93 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
94 | // for perl. doh. |
|
|
95 | static void * |
|
|
96 | glsym (const char *name) |
|
|
97 | { |
|
|
98 | void *fun = 0; |
|
|
99 | |
|
|
100 | #if defined I_DLFCN && defined RTLD_DEFAULT |
|
|
101 | fun = dlsym (RTLD_DEFAULT, name + 1); |
|
|
102 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
|
|
103 | |
|
|
104 | if (!fun) |
|
|
105 | { |
|
|
106 | static void *libgl; |
|
|
107 | static const char *glso[] = { |
|
|
108 | "libGL.so.1", |
|
|
109 | "libGL.so.3", |
|
|
110 | "libGL.so.4.0", |
|
|
111 | "libGL.so", |
|
|
112 | "/usr/lib/libGL.so", |
|
|
113 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
114 | }; |
|
|
115 | int i; |
|
|
116 | |
|
|
117 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
118 | { |
|
|
119 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
120 | if (libgl) |
|
|
121 | break; |
|
|
122 | } |
|
|
123 | |
|
|
124 | if (libgl) |
|
|
125 | { |
|
|
126 | fun = dlsym (libgl, name + 1); |
|
|
127 | if (!fun) fun = dlsym (libgl, name); |
|
|
128 | } |
|
|
129 | } |
|
|
130 | #endif |
|
|
131 | |
|
|
132 | return fun; |
|
|
133 | } |
|
|
134 | |
|
|
135 | /*****************************************************************************/ |
|
|
136 | |
|
|
137 | /* up to two temporary buffers */ |
|
|
138 | static void * ecb_noinline |
|
|
139 | tmpbuf (size_t size) |
|
|
140 | { |
|
|
141 | enum { buffers = 4 }; |
|
|
142 | static int idx; |
|
|
143 | static void *buf [buffers]; |
|
|
144 | static size_t len [buffers]; |
|
|
145 | |
|
|
146 | idx = (idx + 1) % buffers; |
|
|
147 | |
|
|
148 | if (len [idx] < size) |
|
|
149 | { |
|
|
150 | free (buf [idx]); |
|
|
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
|
|
152 | buf [idx] = malloc (len [idx]); |
|
|
153 | } |
|
|
154 | |
|
|
155 | return buf [idx]; |
|
|
156 | } |
|
|
157 | |
|
|
158 | /*****************************************************************************/ |
|
|
159 | |
|
|
160 | typedef struct |
|
|
161 | { |
13 | IV iv; |
162 | IV iv; |
14 | const char *name; |
163 | const char *name; |
15 | } cl_error[] = { |
|
|
16 | #define def_error(name) { (IV)CL_ ## name, # name }, |
164 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
17 | #include "invalid.h" |
165 | } ivstr; |
18 | }; |
|
|
19 | |
166 | |
20 | static const char * |
167 | static const char * |
21 | clstrerror (cl_int res) |
168 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
22 | { |
169 | { |
23 | int i; |
170 | int i; |
24 | static char numbuf [32]; |
171 | static char strbuf [32]; |
25 | |
172 | |
26 | for (i = sizeof (cl_error) / sizeof (cl_error [0]); i--; ) |
173 | for (i = count; i--; ) |
27 | if (cl_error [i].iv == res) |
174 | if (base [i].iv == value) |
28 | return cl_error [i].name; |
175 | return base [i].name; |
29 | |
176 | |
30 | snprintf (numbuf, sizeof (numbuf), "ERROR(%d)", res); |
177 | snprintf (strbuf, sizeof (strbuf), fallback, (int)value); |
31 | |
178 | |
32 | return numbuf; |
179 | return strbuf; |
33 | } |
180 | } |
34 | |
181 | |
|
|
182 | static const char * |
|
|
183 | enum2str (cl_uint value) |
|
|
184 | { |
|
|
185 | static const ivstr enumstr[] = { |
|
|
186 | #include "enumstr.h" |
|
|
187 | }; |
|
|
188 | |
|
|
189 | return iv2str (value, enumstr, sizeof (enumstr) / sizeof (enumstr [0]), "ENUM(0x%04x)"); |
|
|
190 | } |
|
|
191 | |
|
|
192 | static const char * |
|
|
193 | err2str (cl_int err) |
|
|
194 | { |
|
|
195 | static const ivstr errstr[] = { |
|
|
196 | #include "errstr.h" |
|
|
197 | }; |
|
|
198 | |
|
|
199 | return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
|
|
200 | } |
|
|
201 | |
|
|
202 | /*****************************************************************************/ |
|
|
203 | |
|
|
204 | static cl_int res; |
|
|
205 | |
35 | #define FAIL(name,res) \ |
206 | #define FAIL(name) \ |
36 | croak (# name ": %s", clstrerror (res)); |
207 | croak ("cl" # name ": %s", err2str (res)); |
37 | |
208 | |
38 | #define NEED_SUCCESS(name,args) \ |
209 | #define NEED_SUCCESS(name,args) \ |
39 | do { \ |
210 | do { \ |
40 | cl_int res = name args; \ |
211 | res = cl ## name args; \ |
41 | \ |
212 | \ |
42 | if (res) \ |
213 | if (res) \ |
43 | FAIL (name, res); \ |
214 | FAIL (name); \ |
44 | } while (0) |
215 | } while (0) |
45 | |
216 | |
|
|
217 | #define NEED_SUCCESS_ARG(retdecl, name, args) \ |
|
|
218 | retdecl = cl ## name args; \ |
|
|
219 | if (res) \ |
|
|
220 | FAIL (name); |
|
|
221 | |
|
|
222 | /*****************************************************************************/ |
|
|
223 | |
|
|
224 | static SV * |
|
|
225 | new_clobj (HV *stash, IV id) |
|
|
226 | { |
|
|
227 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
228 | } |
|
|
229 | |
|
|
230 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
231 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
232 | |
|
|
233 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
234 | /* they can be upgraded at runtime to the array form */ |
|
|
235 | static void * ecb_noinline |
|
|
236 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
237 | { |
|
|
238 | // sv_derived_from is quite slow :( |
|
|
239 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
240 | return (void *)SvIV (SvRV (sv)); |
|
|
241 | |
|
|
242 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
243 | } |
|
|
244 | |
|
|
245 | // the "no-inherit" version of the above |
|
|
246 | static void * ecb_noinline |
|
|
247 | SvCLOBJ_ni (const char *func, const char *svname, SV *sv, HV *stash) |
|
|
248 | { |
|
|
249 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
250 | return (void *)SvIV (SvRV (sv)); |
|
|
251 | |
|
|
252 | croak ("%s: %s is not of type %s", func, svname, HvNAME (stash)); |
|
|
253 | } |
|
|
254 | |
|
|
255 | /*****************************************************************************/ |
|
|
256 | |
|
|
257 | static cl_context_properties * ecb_noinline |
|
|
258 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
|
|
259 | { |
|
|
260 | if (!sv || !SvOK (sv)) |
|
|
261 | if (extra) |
|
|
262 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
|
|
263 | else |
|
|
264 | return 0; |
|
|
265 | |
|
|
266 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
|
|
267 | { |
|
|
268 | AV *av = (AV *)SvRV (sv); |
|
|
269 | int i, len = av_len (av) + 1; |
|
|
270 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
|
|
271 | cl_context_properties *l = p; |
|
|
272 | |
|
|
273 | if (len & 1) |
|
|
274 | croak ("%s: %s is not a property list (must contain an even number of elements)", func, svname); |
|
|
275 | |
|
|
276 | while (extracount--) |
|
|
277 | *l++ = *extra++; |
|
|
278 | |
|
|
279 | for (i = 0; i < len; i += 2) |
|
|
280 | { |
|
|
281 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
|
|
282 | SV *p_sv = *av_fetch (av, i + 1, 0); |
|
|
283 | cl_context_properties v = SvIV (p_sv); // code below can override |
|
|
284 | |
|
|
285 | switch (t) |
|
|
286 | { |
|
|
287 | case CL_CONTEXT_PLATFORM: |
|
|
288 | if (SvROK (p_sv)) |
|
|
289 | v = (cl_context_properties)SvCLOBJ (func, svname, p_sv, "OpenCL::Platform"); |
|
|
290 | break; |
|
|
291 | |
|
|
292 | case CL_GLX_DISPLAY_KHR: |
|
|
293 | if (!SvOK (p_sv)) |
|
|
294 | { |
|
|
295 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
296 | if (func) |
|
|
297 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
298 | } |
|
|
299 | break; |
|
|
300 | |
|
|
301 | case CL_GL_CONTEXT_KHR: |
|
|
302 | if (!SvOK (p_sv)) |
|
|
303 | { |
|
|
304 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
305 | if (func) |
|
|
306 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
307 | } |
|
|
308 | break; |
|
|
309 | |
|
|
310 | default: |
|
|
311 | /* unknown property, treat as int */ |
|
|
312 | break; |
|
|
313 | } |
|
|
314 | |
|
|
315 | *l++ = t; |
|
|
316 | *l++ = v; |
|
|
317 | } |
|
|
318 | |
|
|
319 | *l = 0; |
|
|
320 | |
|
|
321 | return p; |
|
|
322 | } |
|
|
323 | |
|
|
324 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
|
|
325 | } |
|
|
326 | |
|
|
327 | /*****************************************************************************/ |
|
|
328 | /* callback stuff */ |
|
|
329 | |
|
|
330 | /* default context callback, log to stderr */ |
|
|
331 | static void CL_CALLBACK |
|
|
332 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
333 | { |
|
|
334 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
335 | } |
|
|
336 | |
|
|
337 | typedef struct |
|
|
338 | { |
|
|
339 | int free_cb; |
|
|
340 | void (*push)(void *data1, void *data2, void *data3); |
|
|
341 | } eq_vtbl; |
|
|
342 | |
|
|
343 | typedef struct eq_item |
|
|
344 | { |
|
|
345 | struct eq_item *next; |
|
|
346 | eq_vtbl *vtbl; |
|
|
347 | SV *cb; |
|
|
348 | void *data1, *data2, *data3; |
|
|
349 | } eq_item; |
|
|
350 | |
|
|
351 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
352 | static void *eq_signal_arg; |
|
|
353 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
354 | static eq_item *eq_head, *eq_tail; |
|
|
355 | |
|
|
356 | static void ecb_noinline |
|
|
357 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
358 | { |
|
|
359 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
360 | |
|
|
361 | item->next = 0; |
|
|
362 | item->vtbl = vtbl; |
|
|
363 | item->cb = cb; |
|
|
364 | item->data1 = data1; |
|
|
365 | item->data2 = data2; |
|
|
366 | item->data3 = data3; |
|
|
367 | |
|
|
368 | X_LOCK (eq_lock); |
|
|
369 | |
|
|
370 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
371 | eq_tail = item; |
|
|
372 | |
|
|
373 | X_UNLOCK (eq_lock); |
|
|
374 | |
|
|
375 | eq_signal_func (eq_signal_arg, 0); |
|
|
376 | } |
|
|
377 | |
|
|
378 | static eq_item * |
|
|
379 | eq_dec (void) |
|
|
380 | { |
|
|
381 | eq_item *res; |
|
|
382 | |
|
|
383 | X_LOCK (eq_lock); |
|
|
384 | |
|
|
385 | res = eq_head; |
|
|
386 | if (res) |
|
|
387 | eq_head = res->next; |
|
|
388 | |
|
|
389 | X_UNLOCK (eq_lock); |
|
|
390 | |
|
|
391 | return res; |
|
|
392 | } |
|
|
393 | |
|
|
394 | static void |
|
|
395 | eq_poll (void) |
|
|
396 | { |
|
|
397 | eq_item *item; |
|
|
398 | |
|
|
399 | while ((item = eq_dec ())) |
|
|
400 | { |
|
|
401 | ENTER; |
|
|
402 | SAVETMPS; |
|
|
403 | |
|
|
404 | dSP; |
|
|
405 | PUSHMARK (SP); |
|
|
406 | EXTEND (SP, 2); |
|
|
407 | |
|
|
408 | if (item->vtbl->free_cb) |
|
|
409 | sv_2mortal (item->cb); |
|
|
410 | |
|
|
411 | PUTBACK; |
|
|
412 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
413 | |
|
|
414 | SV *cb = item->cb; |
|
|
415 | free (item); |
|
|
416 | |
|
|
417 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
418 | |
|
|
419 | FREETMPS; |
|
|
420 | LEAVE; |
|
|
421 | } |
|
|
422 | } |
|
|
423 | |
|
|
424 | static void |
|
|
425 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
426 | { |
|
|
427 | eq_poll (); |
|
|
428 | } |
|
|
429 | |
|
|
430 | /*****************************************************************************/ |
|
|
431 | /* context notify */ |
|
|
432 | |
|
|
433 | static void ecb_noinline |
|
|
434 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
435 | { |
|
|
436 | dSP; |
|
|
437 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
438 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
439 | PUTBACK; |
|
|
440 | |
|
|
441 | free (data1); |
|
|
442 | free (data2); |
|
|
443 | } |
|
|
444 | |
|
|
445 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
446 | |
|
|
447 | static void CL_CALLBACK |
|
|
448 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
449 | { |
|
|
450 | void *pvt_copy = malloc (cb); |
|
|
451 | memcpy (pvt_copy, pvt, cb); |
|
|
452 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
453 | } |
|
|
454 | |
|
|
455 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
456 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
457 | void *user_data = 0; \ |
|
|
458 | \ |
|
|
459 | if (SvOK (notify)) \ |
|
|
460 | { \ |
|
|
461 | pfn_notify = eq_context_notify; \ |
|
|
462 | user_data = s_get_cv (notify); \ |
|
|
463 | } |
|
|
464 | |
|
|
465 | static SV * ecb_noinline |
|
|
466 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
467 | { |
|
|
468 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
469 | |
|
|
470 | if (user_data) |
|
|
471 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
472 | |
|
|
473 | return sv; |
|
|
474 | } |
|
|
475 | |
|
|
476 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
477 | |
|
|
478 | /*****************************************************************************/ |
|
|
479 | /* build/compile/link notify */ |
|
|
480 | |
|
|
481 | static void |
|
|
482 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
483 | { |
|
|
484 | dSP; |
|
|
485 | PUSH_CLOBJ (stash_program, data1); |
|
|
486 | PUTBACK; |
|
|
487 | } |
|
|
488 | |
|
|
489 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
490 | |
|
|
491 | static void CL_CALLBACK |
|
|
492 | eq_program_notify (cl_program program, void *user_data) |
|
|
493 | { |
|
|
494 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
495 | } |
|
|
496 | |
|
|
497 | struct build_args |
|
|
498 | { |
|
|
499 | cl_program program; |
|
|
500 | char *options; |
|
|
501 | void *user_data; |
|
|
502 | cl_uint num_devices; |
|
|
503 | }; |
|
|
504 | |
|
|
505 | X_THREAD_PROC (build_program_thread) |
|
|
506 | { |
|
|
507 | struct build_args *arg = thr_arg; |
|
|
508 | |
|
|
509 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
510 | |
|
|
511 | if (arg->user_data) |
|
|
512 | eq_program_notify (arg->program, arg->user_data); |
|
|
513 | else |
|
|
514 | clReleaseProgram (arg->program); |
|
|
515 | |
|
|
516 | free (arg->options); |
|
|
517 | free (arg); |
|
|
518 | |
|
|
519 | return 0; |
|
|
520 | } |
|
|
521 | |
|
|
522 | static void |
|
|
523 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
524 | { |
|
|
525 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
526 | |
|
|
527 | arg->program = program; |
|
|
528 | arg->options = strdup (options); |
|
|
529 | arg->user_data = user_data; |
|
|
530 | arg->num_devices = num_devices; |
|
|
531 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
532 | |
|
|
533 | xthread_t id; |
|
|
534 | thread_create (&id, build_program_thread, arg); |
|
|
535 | } |
|
|
536 | |
|
|
537 | /*****************************************************************************/ |
|
|
538 | /* event objects */ |
|
|
539 | |
|
|
540 | static void |
|
|
541 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
542 | { |
|
|
543 | dSP; |
|
|
544 | PUSH_CLOBJ (stash_event, data1); |
|
|
545 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
546 | PUTBACK; |
|
|
547 | } |
|
|
548 | |
|
|
549 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
550 | |
|
|
551 | static void CL_CALLBACK |
|
|
552 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
553 | { |
|
|
554 | clRetainEvent (event); |
|
|
555 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
556 | } |
|
|
557 | |
|
|
558 | /*****************************************************************************/ |
|
|
559 | /* utilities for XS code */ |
|
|
560 | |
|
|
561 | static size_t |
|
|
562 | img_row_pitch (cl_mem img) |
|
|
563 | { |
|
|
564 | size_t res; |
|
|
565 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
|
|
566 | return res; |
|
|
567 | } |
|
|
568 | |
|
|
569 | static cl_event * ecb_noinline |
|
|
570 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
|
|
571 | { |
|
|
572 | cl_uint count = *rcount; |
|
|
573 | |
|
|
574 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
575 | *rcount = count = 0; |
|
|
576 | |
|
|
577 | if (!count && !extra) |
|
|
578 | return 0; |
|
|
579 | |
|
|
580 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
581 | int i = 0; |
|
|
582 | |
|
|
583 | while (count--) |
|
|
584 | if (SvOK (items [count])) |
|
|
585 | list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
|
|
586 | |
|
|
587 | if (extra) |
|
|
588 | list [i++] = extra; |
|
|
589 | |
|
|
590 | *rcount = i; |
|
|
591 | |
|
|
592 | return i ? list : 0; |
|
|
593 | } |
|
|
594 | |
|
|
595 | #define EVENT_LIST(skip) \ |
|
|
596 | cl_uint event_list_count = items - (skip); \ |
|
|
597 | cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) |
|
|
598 | |
|
|
599 | #define INFO(class) \ |
|
|
600 | { \ |
|
|
601 | size_t size; \ |
|
|
602 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
|
|
603 | SV *sv = sv_2mortal (newSV (size)); \ |
|
|
604 | SvUPGRADE (sv, SVt_PV); \ |
|
|
605 | SvPOK_only (sv); \ |
|
|
606 | SvCUR_set (sv, size); \ |
|
|
607 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
|
|
608 | XPUSHs (sv); \ |
|
|
609 | } |
|
|
610 | |
|
|
611 | /*****************************************************************************/ |
|
|
612 | /* mapped_xxx */ |
|
|
613 | |
|
|
614 | static OpenCL__Mapped |
|
|
615 | SvMAPPED (SV *sv) |
|
|
616 | { |
|
|
617 | // no typechecking atm., keep your fingers crossed |
|
|
618 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
619 | } |
|
|
620 | |
|
|
621 | struct mapped |
|
|
622 | { |
|
|
623 | cl_command_queue queue; |
|
|
624 | cl_mem memobj; |
|
|
625 | void *ptr; |
|
|
626 | size_t cb; |
|
|
627 | cl_event event; |
|
|
628 | size_t row_pitch; |
|
|
629 | size_t slice_pitch; |
|
|
630 | }; |
|
|
631 | |
|
|
632 | static SV * |
|
|
633 | 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) |
|
|
634 | { |
|
|
635 | SV *data = newSV (0); |
|
|
636 | SvUPGRADE (data, SVt_PVMG); |
|
|
637 | |
|
|
638 | OpenCL__Mapped mapped; |
|
|
639 | New (0, mapped, 1, struct mapped); |
|
|
640 | |
|
|
641 | clRetainCommandQueue (queue); |
|
|
642 | |
|
|
643 | mapped->queue = queue; |
|
|
644 | mapped->memobj = memobj; |
|
|
645 | mapped->ptr = ptr; |
|
|
646 | mapped->cb = cb; |
|
|
647 | mapped->event = ev; |
|
|
648 | mapped->row_pitch = row_pitch; |
|
|
649 | mapped->slice_pitch = slice_pitch; |
|
|
650 | |
|
|
651 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
652 | |
|
|
653 | if (SvLEN (data)) |
|
|
654 | Safefree (data); |
|
|
655 | |
|
|
656 | SvPVX (data) = (char *)ptr; |
|
|
657 | SvCUR_set (data, cb); |
|
|
658 | SvLEN_set (data, 0); |
|
|
659 | SvPOK_only (data); |
|
|
660 | |
|
|
661 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
662 | |
|
|
663 | if (!(flags & CL_MAP_WRITE)) |
|
|
664 | SvREADONLY_on (data); |
|
|
665 | |
|
|
666 | return obj; |
|
|
667 | } |
|
|
668 | |
|
|
669 | static void |
|
|
670 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
671 | { |
|
|
672 | SV *data = SvRV (sv); |
|
|
673 | |
|
|
674 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
675 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
676 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
677 | else |
|
|
678 | { |
|
|
679 | SvREADONLY_off (data); |
|
|
680 | SvCUR_set (data, 0); |
|
|
681 | SvPVX (data) = 0; |
|
|
682 | SvOK_off (data); |
|
|
683 | } |
|
|
684 | |
|
|
685 | mapped->ptr = 0; |
|
|
686 | } |
|
|
687 | |
|
|
688 | static void |
|
|
689 | mapped_unmap (SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
690 | { |
|
|
691 | cl_event *event_list_ptr = event_list (wait_list, &event_list_count, mapped->event); |
|
|
692 | cl_event ev; |
|
|
693 | |
|
|
694 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
695 | |
|
|
696 | clReleaseEvent (mapped->event); |
|
|
697 | mapped->event = ev; |
|
|
698 | |
|
|
699 | mapped_detach (self, mapped); |
|
|
700 | } |
|
|
701 | |
|
|
702 | /*****************************************************************************/ |
|
|
703 | |
46 | MODULE = OpenCL PACKAGE = OpenCL |
704 | MODULE = OpenCL PACKAGE = OpenCL |
47 | |
705 | |
|
|
706 | PROTOTYPES: ENABLE |
|
|
707 | |
|
|
708 | void |
|
|
709 | poll () |
|
|
710 | CODE: |
|
|
711 | eq_poll (); |
|
|
712 | |
|
|
713 | void |
|
|
714 | _eq_initialise (IV func, IV arg) |
|
|
715 | CODE: |
|
|
716 | eq_signal_func = (void (*)(void *, int))func; |
|
|
717 | eq_signal_arg = (void*)arg; |
|
|
718 | |
48 | BOOT: |
719 | BOOT: |
49 | { |
720 | { |
50 | HV *stash = gv_stashpv ("OpenCL", 1); |
721 | HV *stash = gv_stashpv ("OpenCL", 1); |
51 | static const struct { |
722 | static const ivstr *civ, const_iv[] = { |
52 | const char *name; |
723 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
53 | IV iv; |
724 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
54 | } *civ, const_iv[] = { |
725 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
55 | #define const_iv(name) { # name, (IV)CL_ ## name }, |
726 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
|
|
727 | { sizeof (cl_int ), "SIZEOF_INT" }, |
|
|
728 | { sizeof (cl_uint ), "SIZEOF_UINT" }, |
|
|
729 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
|
|
730 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
|
|
731 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
|
|
732 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
|
|
733 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
56 | #include "constiv.h" |
734 | #include "constiv.h" |
57 | }; |
735 | }; |
|
|
736 | |
58 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
737 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
59 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
738 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
739 | |
|
|
740 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
741 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
742 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
743 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
744 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
745 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
746 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
747 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
748 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
749 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
750 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
751 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
752 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
753 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
754 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
755 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
756 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
757 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
758 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
759 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
760 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
761 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
762 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
763 | |
|
|
764 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
60 | } |
765 | } |
|
|
766 | |
|
|
767 | cl_int |
|
|
768 | errno () |
|
|
769 | CODE: |
|
|
770 | RETVAL = res; |
|
|
771 | OUTPUT: |
|
|
772 | RETVAL |
|
|
773 | |
|
|
774 | const char * |
|
|
775 | err2str (cl_int err = res) |
|
|
776 | |
|
|
777 | const char * |
|
|
778 | enum2str (cl_uint value) |
61 | |
779 | |
62 | void |
780 | void |
63 | platforms () |
781 | platforms () |
64 | PPCODE: |
782 | PPCODE: |
65 | { |
|
|
66 | cl_platform_id *list; |
783 | cl_platform_id *list; |
67 | cl_uint count; |
784 | cl_uint count; |
68 | int i; |
785 | int i; |
69 | |
786 | |
70 | NEED_SUCCESS (clGetPlatformIDs, (0, 0, &count)); |
787 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
71 | Newx (list, count, cl_platform_id); |
788 | list = tmpbuf (sizeof (*list) * count); |
72 | NEED_SUCCESS (clGetPlatformIDs, (count, list, 0)); |
789 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
73 | |
790 | |
74 | EXTEND (SP, count); |
791 | EXTEND (SP, count); |
75 | for (i = 0; i < count; ++i) |
792 | for (i = 0; i < count; ++i) |
76 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Platform", list [i])); |
793 | PUSH_CLOBJ (stash_platform, list [i]); |
77 | |
794 | |
78 | Safefree (list); |
|
|
79 | } |
|
|
80 | |
|
|
81 | void |
795 | void |
82 | context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
796 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
797 | PPCODE: |
|
|
798 | CONTEXT_NOTIFY_CALLBACK; |
|
|
799 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
|
|
800 | XPUSH_CLOBJ_CONTEXT; |
|
|
801 | |
|
|
802 | void |
|
|
803 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
|
|
804 | PPCODE: |
|
|
805 | /* der Gipfel der Kunst */ |
|
|
806 | |
|
|
807 | void |
|
|
808 | wait_for_events (...) |
83 | PPCODE: |
809 | CODE: |
84 | { |
810 | EVENT_LIST (0); |
85 | cl_int res; |
811 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
86 | cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
|
|
87 | |
812 | |
88 | if (res) |
813 | PROTOTYPES: DISABLE |
89 | FAIL (clCreateContextFromType, res); |
|
|
90 | |
|
|
91 | XPUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Context", ctx)); |
|
|
92 | } |
|
|
93 | |
814 | |
94 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
815 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
95 | |
816 | |
96 | void |
817 | void |
97 | info (OpenCL::Platform this, cl_platform_info name) |
818 | info (OpenCL::Platform self, cl_platform_info name) |
|
|
819 | PPCODE: |
|
|
820 | INFO (Platform) |
|
|
821 | |
|
|
822 | void |
|
|
823 | unload_compiler (OpenCL::Platform self) |
|
|
824 | CODE: |
|
|
825 | #if CL_VERSION_1_2 |
|
|
826 | clUnloadPlatformCompiler (self); |
|
|
827 | #endif |
|
|
828 | |
|
|
829 | #BEGIN:platform |
|
|
830 | |
|
|
831 | void |
|
|
832 | profile (OpenCL::Platform self) |
|
|
833 | ALIAS: |
|
|
834 | profile = CL_PLATFORM_PROFILE |
|
|
835 | version = CL_PLATFORM_VERSION |
|
|
836 | name = CL_PLATFORM_NAME |
|
|
837 | vendor = CL_PLATFORM_VENDOR |
|
|
838 | extensions = CL_PLATFORM_EXTENSIONS |
98 | PPCODE: |
839 | PPCODE: |
99 | { |
|
|
100 | size_t size; |
840 | size_t size; |
101 | SV *sv; |
|
|
102 | |
|
|
103 | NEED_SUCCESS (clGetPlatformInfo, (this, name, 0, 0, &size)); |
841 | NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size)); |
|
|
842 | char *value = tmpbuf (size); |
|
|
843 | NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0)); |
|
|
844 | EXTEND (SP, 1); |
|
|
845 | const int i = 0; |
|
|
846 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
847 | |
|
|
848 | #END:platform |
|
|
849 | |
|
|
850 | void |
|
|
851 | devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL) |
|
|
852 | PPCODE: |
|
|
853 | cl_device_id *list; |
|
|
854 | cl_uint count; |
|
|
855 | int i; |
|
|
856 | |
|
|
857 | NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count)); |
|
|
858 | list = tmpbuf (sizeof (*list) * count); |
|
|
859 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
|
|
860 | |
|
|
861 | EXTEND (SP, count); |
|
|
862 | for (i = 0; i < count; ++i) |
|
|
863 | PUSH_CLOBJ (stash_device, list [i]); |
|
|
864 | |
|
|
865 | void |
|
|
866 | context (OpenCL::Platform self, SV *properties = 0, SV *devices, SV *notify = &PL_sv_undef) |
|
|
867 | PPCODE: |
|
|
868 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
869 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
|
|
870 | |
|
|
871 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
872 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context", "properties", properties, extra, 2); |
|
|
873 | |
|
|
874 | AV *av = (AV *)SvRV (devices); |
|
|
875 | cl_uint num_devices = av_len (av) + 1; |
|
|
876 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
|
|
877 | |
|
|
878 | int i; |
|
|
879 | for (i = num_devices; i--; ) |
|
|
880 | device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
881 | |
|
|
882 | CONTEXT_NOTIFY_CALLBACK; |
|
|
883 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, num_devices, device_list, pfn_notify, user_data, &res)); |
|
|
884 | XPUSH_CLOBJ_CONTEXT; |
|
|
885 | |
|
|
886 | void |
|
|
887 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
888 | PPCODE: |
|
|
889 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
890 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
891 | |
|
|
892 | CONTEXT_NOTIFY_CALLBACK; |
|
|
893 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
894 | XPUSH_CLOBJ_CONTEXT; |
|
|
895 | |
|
|
896 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
897 | |
|
|
898 | void |
|
|
899 | info (OpenCL::Device self, cl_device_info name) |
|
|
900 | PPCODE: |
|
|
901 | INFO (Device) |
|
|
902 | |
|
|
903 | #if CL_VERSION_1_2 |
|
|
904 | |
|
|
905 | void |
|
|
906 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
907 | PPCODE: |
|
|
908 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
|
|
909 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
|
|
910 | |
|
|
911 | properties = SvRV (properties); |
|
|
912 | |
|
|
913 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
914 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
915 | |
|
|
916 | int i; |
|
|
917 | for (i = 0; i < count; ++i) |
|
|
918 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
|
|
919 | |
|
|
920 | props [count] = 0; |
|
|
921 | |
|
|
922 | cl_uint num_devices; |
|
|
923 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
924 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
925 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
|
|
926 | |
|
|
927 | EXTEND (SP, num_devices); |
|
|
928 | for (i = 0; i < count; ++i) |
|
|
929 | PUSH_CLOBJ (stash_subdevice, list [i]); |
|
|
930 | |
|
|
931 | #endif |
|
|
932 | |
|
|
933 | #BEGIN:device |
|
|
934 | |
|
|
935 | void |
|
|
936 | type (OpenCL::Device self) |
|
|
937 | PPCODE: |
|
|
938 | cl_device_type value [1]; |
|
|
939 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
|
|
940 | EXTEND (SP, 1); |
|
|
941 | const int i = 0; |
|
|
942 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
943 | |
|
|
944 | void |
|
|
945 | vendor_id (OpenCL::Device self) |
|
|
946 | ALIAS: |
|
|
947 | vendor_id = CL_DEVICE_VENDOR_ID |
|
|
948 | max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS |
|
|
949 | max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS |
|
|
950 | preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR |
|
|
951 | preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT |
|
|
952 | preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT |
|
|
953 | preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG |
|
|
954 | preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT |
|
|
955 | preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE |
|
|
956 | max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY |
|
|
957 | max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS |
|
|
958 | max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS |
|
|
959 | image_support = CL_DEVICE_IMAGE_SUPPORT |
|
|
960 | max_samplers = CL_DEVICE_MAX_SAMPLERS |
|
|
961 | mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN |
|
|
962 | min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE |
|
|
963 | global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
|
|
964 | max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS |
|
|
965 | preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF |
|
|
966 | native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR |
|
|
967 | native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT |
|
|
968 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
|
|
969 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
|
|
970 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
|
|
971 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
|
|
972 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
|
|
973 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
|
|
974 | PPCODE: |
|
|
975 | cl_uint value [1]; |
|
|
976 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
977 | EXTEND (SP, 1); |
|
|
978 | const int i = 0; |
|
|
979 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
980 | |
|
|
981 | void |
|
|
982 | max_work_group_size (OpenCL::Device self) |
|
|
983 | ALIAS: |
|
|
984 | max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE |
|
|
985 | image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH |
|
|
986 | image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT |
|
|
987 | image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH |
|
|
988 | image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT |
|
|
989 | image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH |
|
|
990 | max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE |
|
|
991 | profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION |
|
|
992 | PPCODE: |
|
|
993 | size_t value [1]; |
|
|
994 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
995 | EXTEND (SP, 1); |
|
|
996 | const int i = 0; |
|
|
997 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
998 | |
|
|
999 | void |
|
|
1000 | max_work_item_sizes (OpenCL::Device self) |
|
|
1001 | PPCODE: |
|
|
1002 | size_t size; |
|
|
1003 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size)); |
|
|
1004 | size_t *value = tmpbuf (size); |
|
|
1005 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0)); |
|
|
1006 | int i, n = size / sizeof (*value); |
|
|
1007 | EXTEND (SP, n); |
|
|
1008 | for (i = 0; i < n; ++i) |
|
|
1009 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1010 | |
|
|
1011 | void |
|
|
1012 | address_bits (OpenCL::Device self) |
|
|
1013 | PPCODE: |
|
|
1014 | cl_bitfield value [1]; |
|
|
1015 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_ADDRESS_BITS, sizeof (value), value, 0)); |
|
|
1016 | EXTEND (SP, 1); |
|
|
1017 | const int i = 0; |
|
|
1018 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1019 | |
|
|
1020 | void |
|
|
1021 | max_mem_alloc_size (OpenCL::Device self) |
|
|
1022 | ALIAS: |
|
|
1023 | max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE |
|
|
1024 | global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE |
|
|
1025 | global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE |
|
|
1026 | max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE |
|
|
1027 | local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE |
|
|
1028 | PPCODE: |
|
|
1029 | cl_ulong value [1]; |
|
|
1030 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1031 | EXTEND (SP, 1); |
|
|
1032 | const int i = 0; |
|
|
1033 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1034 | |
|
|
1035 | void |
|
|
1036 | single_fp_config (OpenCL::Device self) |
|
|
1037 | ALIAS: |
|
|
1038 | single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG |
|
|
1039 | double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG |
|
|
1040 | half_fp_config = CL_DEVICE_HALF_FP_CONFIG |
|
|
1041 | PPCODE: |
|
|
1042 | cl_device_fp_config value [1]; |
|
|
1043 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1044 | EXTEND (SP, 1); |
|
|
1045 | const int i = 0; |
|
|
1046 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1047 | |
|
|
1048 | void |
|
|
1049 | global_mem_cache_type (OpenCL::Device self) |
|
|
1050 | PPCODE: |
|
|
1051 | cl_device_mem_cache_type value [1]; |
|
|
1052 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, sizeof (value), value, 0)); |
|
|
1053 | EXTEND (SP, 1); |
|
|
1054 | const int i = 0; |
|
|
1055 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1056 | |
|
|
1057 | void |
|
|
1058 | local_mem_type (OpenCL::Device self) |
|
|
1059 | PPCODE: |
|
|
1060 | cl_device_local_mem_type value [1]; |
|
|
1061 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_LOCAL_MEM_TYPE, sizeof (value), value, 0)); |
|
|
1062 | EXTEND (SP, 1); |
|
|
1063 | const int i = 0; |
|
|
1064 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1065 | |
|
|
1066 | void |
|
|
1067 | error_correction_support (OpenCL::Device self) |
|
|
1068 | ALIAS: |
|
|
1069 | error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT |
|
|
1070 | endian_little = CL_DEVICE_ENDIAN_LITTLE |
|
|
1071 | available = CL_DEVICE_AVAILABLE |
|
|
1072 | compiler_available = CL_DEVICE_COMPILER_AVAILABLE |
|
|
1073 | host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY |
|
|
1074 | PPCODE: |
|
|
1075 | cl_bool value [1]; |
|
|
1076 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1077 | EXTEND (SP, 1); |
|
|
1078 | const int i = 0; |
|
|
1079 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
1080 | |
|
|
1081 | void |
|
|
1082 | execution_capabilities (OpenCL::Device self) |
|
|
1083 | PPCODE: |
|
|
1084 | cl_device_exec_capabilities value [1]; |
|
|
1085 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_EXECUTION_CAPABILITIES, sizeof (value), value, 0)); |
|
|
1086 | EXTEND (SP, 1); |
|
|
1087 | const int i = 0; |
|
|
1088 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1089 | |
|
|
1090 | void |
|
|
1091 | properties (OpenCL::Device self) |
|
|
1092 | PPCODE: |
|
|
1093 | cl_command_queue_properties value [1]; |
|
|
1094 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
1095 | EXTEND (SP, 1); |
|
|
1096 | const int i = 0; |
|
|
1097 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1098 | |
|
|
1099 | void |
|
|
1100 | platform (OpenCL::Device self) |
|
|
1101 | PPCODE: |
|
|
1102 | cl_platform_id value [1]; |
|
|
1103 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
|
|
1104 | EXTEND (SP, 1); |
|
|
1105 | const int i = 0; |
|
|
1106 | { |
|
|
1107 | PUSH_CLOBJ (stash_platform, value [i]); |
|
|
1108 | } |
|
|
1109 | |
|
|
1110 | void |
|
|
1111 | name (OpenCL::Device self) |
|
|
1112 | ALIAS: |
|
|
1113 | name = CL_DEVICE_NAME |
|
|
1114 | vendor = CL_DEVICE_VENDOR |
|
|
1115 | driver_version = CL_DRIVER_VERSION |
|
|
1116 | profile = CL_DEVICE_PROFILE |
|
|
1117 | version = CL_DEVICE_VERSION |
|
|
1118 | extensions = CL_DEVICE_EXTENSIONS |
|
|
1119 | PPCODE: |
|
|
1120 | size_t size; |
|
|
1121 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1122 | char *value = tmpbuf (size); |
|
|
1123 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1124 | EXTEND (SP, 1); |
|
|
1125 | const int i = 0; |
|
|
1126 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
1127 | |
|
|
1128 | void |
|
|
1129 | parent_device_ext (OpenCL::Device self) |
|
|
1130 | PPCODE: |
|
|
1131 | cl_device_id value [1]; |
|
|
1132 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
|
|
1133 | EXTEND (SP, 1); |
|
|
1134 | const int i = 0; |
|
|
1135 | { |
|
|
1136 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1137 | } |
|
|
1138 | |
|
|
1139 | void |
|
|
1140 | partition_types_ext (OpenCL::Device self) |
|
|
1141 | ALIAS: |
|
|
1142 | partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT |
|
|
1143 | affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT |
|
|
1144 | partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT |
|
|
1145 | PPCODE: |
|
|
1146 | size_t size; |
|
|
1147 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1148 | cl_device_partition_property_ext *value = tmpbuf (size); |
|
|
1149 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1150 | int i, n = size / sizeof (*value); |
|
|
1151 | EXTEND (SP, n); |
|
|
1152 | for (i = 0; i < n; ++i) |
|
|
1153 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1154 | |
|
|
1155 | #END:device |
|
|
1156 | |
|
|
1157 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1158 | |
|
|
1159 | #if CL_VERSION_1_2 |
|
|
1160 | |
|
|
1161 | void |
|
|
1162 | DESTROY (OpenCL::SubDevice self) |
|
|
1163 | CODE: |
|
|
1164 | clReleaseDevice (self); |
|
|
1165 | |
|
|
1166 | #endif |
|
|
1167 | |
|
|
1168 | MODULE = OpenCL PACKAGE = OpenCL::Context |
|
|
1169 | |
|
|
1170 | void |
|
|
1171 | DESTROY (OpenCL::Context self) |
|
|
1172 | CODE: |
|
|
1173 | clReleaseContext (self); |
|
|
1174 | |
|
|
1175 | void |
|
|
1176 | info (OpenCL::Context self, cl_context_info name) |
|
|
1177 | PPCODE: |
|
|
1178 | INFO (Context) |
|
|
1179 | |
|
|
1180 | void |
|
|
1181 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
|
|
1182 | PPCODE: |
|
|
1183 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
|
|
1184 | XPUSH_CLOBJ (stash_queue, queue); |
|
|
1185 | |
|
|
1186 | void |
|
|
1187 | user_event (OpenCL::Context self) |
|
|
1188 | PPCODE: |
|
|
1189 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
|
|
1190 | XPUSH_CLOBJ (stash_userevent, ev); |
|
|
1191 | |
|
|
1192 | void |
|
|
1193 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
|
|
1194 | PPCODE: |
|
|
1195 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
|
|
1196 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
|
|
1197 | |
|
|
1198 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
|
|
1199 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1200 | |
|
|
1201 | void |
|
|
1202 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
|
|
1203 | PPCODE: |
|
|
1204 | STRLEN len; |
|
|
1205 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1206 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
|
|
1207 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
|
|
1208 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
|
|
1209 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1210 | |
|
|
1211 | #if CL_VERSION_1_2 |
|
|
1212 | |
|
|
1213 | void |
|
|
1214 | 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) |
|
|
1215 | PPCODE: |
|
|
1216 | STRLEN len; |
|
|
1217 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1218 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1219 | const cl_image_desc desc = { |
|
|
1220 | type, |
|
|
1221 | width, height, depth, |
|
|
1222 | array_size, row_pitch, slice_pitch, |
|
|
1223 | num_mip_level, num_samples, |
|
|
1224 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
1225 | }; |
|
|
1226 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1227 | HV *stash = stash_image; |
|
|
1228 | switch (type) |
|
|
1229 | { |
|
|
1230 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1231 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1232 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1233 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1234 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1235 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1236 | } |
|
|
1237 | XPUSH_CLOBJ (stash, mem); |
|
|
1238 | |
|
|
1239 | #endif |
|
|
1240 | |
|
|
1241 | void |
|
|
1242 | 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) |
|
|
1243 | PPCODE: |
|
|
1244 | STRLEN len; |
|
|
1245 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1246 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1247 | #if PREFER_1_1 |
|
|
1248 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1249 | #else |
|
|
1250 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1251 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1252 | #endif |
|
|
1253 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1254 | |
|
|
1255 | void |
|
|
1256 | 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) |
|
|
1257 | PPCODE: |
|
|
1258 | STRLEN len; |
|
|
1259 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1260 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1261 | #if PREFER_1_1 |
|
|
1262 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
|
|
1263 | #else |
|
|
1264 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1265 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1266 | #endif |
|
|
1267 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1268 | |
|
|
1269 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1270 | |
|
|
1271 | void |
|
|
1272 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
|
|
1273 | PPCODE: |
|
|
1274 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
|
|
1275 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1276 | |
|
|
1277 | void |
|
|
1278 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1279 | PPCODE: |
|
|
1280 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1281 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1282 | |
|
|
1283 | #if CL_VERSION_1_2 |
|
|
1284 | |
|
|
1285 | void |
|
|
1286 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1287 | ALIAS: |
|
|
1288 | PPCODE: |
|
|
1289 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1290 | cl_gl_object_type type; |
|
|
1291 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1292 | HV *stash = stash_memory; |
|
|
1293 | switch (type) |
|
|
1294 | { |
|
|
1295 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1296 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1297 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1298 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1299 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1300 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1301 | } |
|
|
1302 | XPUSH_CLOBJ (stash, mem); |
|
|
1303 | |
|
|
1304 | #endif |
|
|
1305 | |
|
|
1306 | void |
|
|
1307 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1308 | PPCODE: |
|
|
1309 | #if PREFER_1_1 |
|
|
1310 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
1311 | #else |
|
|
1312 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1313 | #endif |
|
|
1314 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1315 | |
|
|
1316 | void |
|
|
1317 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1318 | PPCODE: |
|
|
1319 | #if PREFER_1_1 |
|
|
1320 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
1321 | #else |
|
|
1322 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1323 | #endif |
|
|
1324 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1325 | |
|
|
1326 | #endif |
|
|
1327 | |
|
|
1328 | void |
|
|
1329 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
|
|
1330 | PPCODE: |
|
|
1331 | { |
|
|
1332 | cl_uint count; |
|
|
1333 | cl_image_format *list; |
|
|
1334 | int i; |
|
|
1335 | |
|
|
1336 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count)); |
|
|
1337 | Newx (list, count, cl_image_format); |
|
|
1338 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0)); |
|
|
1339 | |
|
|
1340 | EXTEND (SP, count); |
|
|
1341 | for (i = 0; i < count; ++i) |
|
|
1342 | { |
|
|
1343 | AV *av = newAV (); |
|
|
1344 | av_store (av, 1, newSVuv (list [i].image_channel_data_type)); |
|
|
1345 | av_store (av, 0, newSVuv (list [i].image_channel_order)); |
|
|
1346 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1347 | } |
|
|
1348 | } |
|
|
1349 | |
|
|
1350 | void |
|
|
1351 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
|
|
1352 | PPCODE: |
|
|
1353 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
|
|
1354 | XPUSH_CLOBJ (stash_sampler, sampler); |
|
|
1355 | |
|
|
1356 | void |
|
|
1357 | program_with_source (OpenCL::Context self, SV *program) |
|
|
1358 | PPCODE: |
|
|
1359 | STRLEN len; |
|
|
1360 | size_t len2; |
|
|
1361 | const char *ptr = SvPVbyte (program, len); |
|
|
1362 | |
|
|
1363 | len2 = len; |
|
|
1364 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
|
|
1365 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1366 | |
|
|
1367 | void |
|
|
1368 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1369 | PPCODE: |
|
|
1370 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
1371 | croak ("OpenCL::Context::program_with_binary: devices must be specified as reference to an array of device objects"); |
|
|
1372 | |
|
|
1373 | devices = SvRV (devices); |
|
|
1374 | |
|
|
1375 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1376 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1377 | |
|
|
1378 | binaries = SvRV (binaries); |
|
|
1379 | |
|
|
1380 | if (av_len ((AV *)devices) != av_len ((AV *)binaries)) |
|
|
1381 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1382 | |
|
|
1383 | int count = av_len ((AV *)devices) + 1; |
|
|
1384 | cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count); |
|
|
1385 | size_t *length_list = tmpbuf (sizeof (*length_list) * count); |
|
|
1386 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * count); |
|
|
1387 | cl_int *status_list = tmpbuf (sizeof (*status_list) * count); |
|
|
1388 | |
|
|
1389 | int i; |
|
|
1390 | for (i = 0; i < count; ++i) |
|
|
1391 | { |
|
|
1392 | device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_binary", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device"); |
|
|
1393 | STRLEN len; |
|
|
1394 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1395 | length_list [i] = len; |
|
|
1396 | } |
|
|
1397 | |
|
|
1398 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, count, device_list, length_list, binary_list, GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1399 | |
|
|
1400 | EXTEND (SP, 2); |
|
|
1401 | PUSH_CLOBJ (stash_program, prog); |
|
|
1402 | |
|
|
1403 | if (GIMME_V == G_ARRAY) |
|
|
1404 | { |
|
|
1405 | AV *av = newAV (); |
|
|
1406 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1407 | |
|
|
1408 | for (i = count; i--; ) |
|
|
1409 | av_store (av, i, newSViv (status_list [i])); |
|
|
1410 | } |
|
|
1411 | |
|
|
1412 | #if CL_VERSION_1_2 |
|
|
1413 | |
|
|
1414 | void |
|
|
1415 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1416 | PPCODE: |
|
|
1417 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
1418 | croak ("OpenCL::Context::program_with_built_in_kernels: devices must be specified as reference to an array of device objects"); |
|
|
1419 | |
|
|
1420 | devices = SvRV (devices); |
|
|
1421 | |
|
|
1422 | int count = av_len ((AV *)devices) + 1; |
|
|
1423 | cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count); |
|
|
1424 | |
|
|
1425 | int i; |
|
|
1426 | for (i = 0; i < count; ++i) |
|
|
1427 | device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_built_in_kernels", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device"); |
|
|
1428 | |
|
|
1429 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1430 | |
|
|
1431 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1432 | |
|
|
1433 | #endif |
|
|
1434 | |
|
|
1435 | #BEGIN:context |
|
|
1436 | |
|
|
1437 | void |
|
|
1438 | reference_count (OpenCL::Context self) |
|
|
1439 | ALIAS: |
|
|
1440 | reference_count = CL_CONTEXT_REFERENCE_COUNT |
|
|
1441 | num_devices = CL_CONTEXT_NUM_DEVICES |
|
|
1442 | PPCODE: |
|
|
1443 | cl_uint value [1]; |
|
|
1444 | NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1445 | EXTEND (SP, 1); |
|
|
1446 | const int i = 0; |
|
|
1447 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1448 | |
|
|
1449 | void |
|
|
1450 | devices (OpenCL::Context self) |
|
|
1451 | PPCODE: |
|
|
1452 | size_t size; |
|
|
1453 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size)); |
|
|
1454 | cl_device_id *value = tmpbuf (size); |
|
|
1455 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
|
|
1456 | int i, n = size / sizeof (*value); |
|
|
1457 | EXTEND (SP, n); |
|
|
1458 | for (i = 0; i < n; ++i) |
|
|
1459 | { |
|
|
1460 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1461 | } |
|
|
1462 | |
|
|
1463 | void |
|
|
1464 | properties (OpenCL::Context self) |
|
|
1465 | PPCODE: |
|
|
1466 | size_t size; |
|
|
1467 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size)); |
|
|
1468 | cl_context_properties *value = tmpbuf (size); |
|
|
1469 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0)); |
|
|
1470 | int i, n = size / sizeof (*value); |
|
|
1471 | EXTEND (SP, n); |
|
|
1472 | for (i = 0; i < n; ++i) |
|
|
1473 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
1474 | |
|
|
1475 | #END:context |
|
|
1476 | |
|
|
1477 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
|
|
1478 | |
|
|
1479 | void |
|
|
1480 | DESTROY (OpenCL::Queue self) |
|
|
1481 | CODE: |
|
|
1482 | clReleaseCommandQueue (self); |
|
|
1483 | |
|
|
1484 | void |
|
|
1485 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1486 | ALIAS: |
|
|
1487 | enqueue_read_buffer = 0 |
|
|
1488 | PPCODE: |
|
|
1489 | cl_event ev = 0; |
|
|
1490 | EVENT_LIST (6); |
|
|
1491 | |
|
|
1492 | SvUPGRADE (data, SVt_PV); |
|
|
1493 | SvGROW (data, len); |
|
|
1494 | SvPOK_only (data); |
|
|
1495 | SvCUR_set (data, len); |
|
|
1496 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1497 | |
|
|
1498 | if (ev) |
|
|
1499 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1500 | |
|
|
1501 | void |
|
|
1502 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1503 | ALIAS: |
|
|
1504 | enqueue_write_buffer = 0 |
|
|
1505 | PPCODE: |
|
|
1506 | cl_event ev = 0; |
|
|
1507 | STRLEN len; |
|
|
1508 | char *ptr = SvPVbyte (data, len); |
|
|
1509 | EVENT_LIST (5); |
|
|
1510 | |
|
|
1511 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1512 | |
|
|
1513 | if (ev) |
|
|
1514 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1515 | |
|
|
1516 | #if CL_VERSION_1_2 |
|
|
1517 | |
|
|
1518 | void |
|
|
1519 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1520 | ALIAS: |
|
|
1521 | enqueue_fill_buffer = 0 |
|
|
1522 | PPCODE: |
|
|
1523 | cl_event ev = 0; |
|
|
1524 | STRLEN len; |
|
|
1525 | char *ptr = SvPVbyte (data, len); |
|
|
1526 | EVENT_LIST (5); |
|
|
1527 | |
|
|
1528 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1529 | |
|
|
1530 | if (ev) |
|
|
1531 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1532 | |
|
|
1533 | void |
|
|
1534 | 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, ...) |
|
|
1535 | ALIAS: |
|
|
1536 | enqueue_fill_image = 0 |
|
|
1537 | PPCODE: |
|
|
1538 | cl_event ev = 0; |
|
|
1539 | const size_t origin [3] = { x, y, z }; |
|
|
1540 | const size_t region [3] = { width, height, depth }; |
|
|
1541 | EVENT_LIST (12); |
|
|
1542 | |
|
|
1543 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1544 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1545 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1546 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1547 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1548 | cl_image_format format; |
|
|
1549 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1550 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1551 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1552 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1553 | |
|
|
1554 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1555 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1556 | |
|
|
1557 | if (ev) |
|
|
1558 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1559 | |
|
|
1560 | #endif |
|
|
1561 | |
|
|
1562 | void |
|
|
1563 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1564 | ALIAS: |
|
|
1565 | enqueue_copy_buffer = 0 |
|
|
1566 | PPCODE: |
|
|
1567 | cl_event ev = 0; |
|
|
1568 | EVENT_LIST (6); |
|
|
1569 | |
|
|
1570 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1571 | |
|
|
1572 | if (ev) |
|
|
1573 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1574 | |
|
|
1575 | void |
|
|
1576 | 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, ...) |
|
|
1577 | ALIAS: |
|
|
1578 | enqueue_read_buffer_rect = 0 |
|
|
1579 | PPCODE: |
|
|
1580 | cl_event ev = 0; |
|
|
1581 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
|
|
1582 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
|
|
1583 | const size_t region[3] = { width, height, depth }; |
|
|
1584 | EVENT_LIST (17); |
|
|
1585 | |
|
|
1586 | if (!buf_row_pitch) |
|
|
1587 | buf_row_pitch = region [0]; |
|
|
1588 | |
|
|
1589 | if (!buf_slice_pitch) |
|
|
1590 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1591 | |
|
|
1592 | if (!host_row_pitch) |
|
|
1593 | host_row_pitch = region [0]; |
|
|
1594 | |
|
|
1595 | if (!host_slice_pitch) |
|
|
1596 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1597 | |
|
|
1598 | size_t len = host_row_pitch * host_slice_pitch * region [2]; |
|
|
1599 | |
|
|
1600 | SvUPGRADE (data, SVt_PV); |
|
|
1601 | SvGROW (data, len); |
|
|
1602 | SvPOK_only (data); |
|
|
1603 | SvCUR_set (data, len); |
|
|
1604 | 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)); |
|
|
1605 | |
|
|
1606 | if (ev) |
|
|
1607 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1608 | |
|
|
1609 | void |
|
|
1610 | 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, ...) |
|
|
1611 | ALIAS: |
|
|
1612 | enqueue_write_buffer_rect = 0 |
|
|
1613 | PPCODE: |
|
|
1614 | cl_event ev = 0; |
|
|
1615 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
|
|
1616 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
|
|
1617 | const size_t region[3] = { width, height, depth }; |
|
|
1618 | STRLEN len; |
|
|
1619 | char *ptr = SvPVbyte (data, len); |
|
|
1620 | EVENT_LIST (17); |
|
|
1621 | |
|
|
1622 | if (!buf_row_pitch) |
|
|
1623 | buf_row_pitch = region [0]; |
|
|
1624 | |
|
|
1625 | if (!buf_slice_pitch) |
|
|
1626 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1627 | |
|
|
1628 | if (!host_row_pitch) |
|
|
1629 | host_row_pitch = region [0]; |
|
|
1630 | |
|
|
1631 | if (!host_slice_pitch) |
|
|
1632 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1633 | |
|
|
1634 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
|
|
1635 | |
|
|
1636 | if (len < min_len) |
|
|
1637 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1638 | |
|
|
1639 | 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)); |
|
|
1640 | |
|
|
1641 | if (ev) |
|
|
1642 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1643 | |
|
|
1644 | void |
|
|
1645 | 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, ...) |
|
|
1646 | ALIAS: |
|
|
1647 | enqueue_copy_buffer_rect = 0 |
|
|
1648 | PPCODE: |
|
|
1649 | cl_event ev = 0; |
|
|
1650 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1651 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1652 | const size_t region[3] = { width, height, depth }; |
|
|
1653 | EVENT_LIST (16); |
|
|
1654 | |
|
|
1655 | 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)); |
|
|
1656 | |
|
|
1657 | if (ev) |
|
|
1658 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1659 | |
|
|
1660 | void |
|
|
1661 | 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, ...) |
|
|
1662 | ALIAS: |
|
|
1663 | enqueue_read_image = 0 |
|
|
1664 | PPCODE: |
|
|
1665 | cl_event ev = 0; |
|
|
1666 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1667 | const size_t region[3] = { width, height, depth }; |
|
|
1668 | EVENT_LIST (12); |
|
|
1669 | |
|
|
1670 | if (!row_pitch) |
|
|
1671 | row_pitch = img_row_pitch (src); |
|
|
1672 | |
|
|
1673 | if (depth > 1 && !slice_pitch) |
|
|
1674 | slice_pitch = row_pitch * height; |
|
|
1675 | |
|
|
1676 | size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1677 | |
|
|
1678 | SvUPGRADE (data, SVt_PV); |
|
|
1679 | SvGROW (data, len); |
|
|
1680 | SvPOK_only (data); |
|
|
1681 | SvCUR_set (data, len); |
|
|
1682 | 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)); |
|
|
1683 | |
|
|
1684 | if (ev) |
|
|
1685 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1686 | |
|
|
1687 | void |
|
|
1688 | 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, ...) |
|
|
1689 | ALIAS: |
|
|
1690 | enqueue_write_image = 0 |
|
|
1691 | PPCODE: |
|
|
1692 | cl_event ev = 0; |
|
|
1693 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1694 | const size_t region[3] = { width, height, depth }; |
|
|
1695 | STRLEN len; |
|
|
1696 | char *ptr = SvPVbyte (data, len); |
|
|
1697 | EVENT_LIST (12); |
|
|
1698 | |
|
|
1699 | if (!row_pitch) |
|
|
1700 | row_pitch = img_row_pitch (dst); |
|
|
1701 | |
|
|
1702 | if (depth > 1 && !slice_pitch) |
|
|
1703 | slice_pitch = row_pitch * height; |
|
|
1704 | |
|
|
1705 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1706 | |
|
|
1707 | if (len < min_len) |
|
|
1708 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1709 | |
|
|
1710 | 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)); |
|
|
1711 | |
|
|
1712 | if (ev) |
|
|
1713 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1714 | |
|
|
1715 | void |
|
|
1716 | 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, ...) |
|
|
1717 | ALIAS: |
|
|
1718 | enqueue_copy_image = 0 |
|
|
1719 | PPCODE: |
|
|
1720 | cl_event ev = 0; |
|
|
1721 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1722 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1723 | const size_t region[3] = { width, height, depth }; |
|
|
1724 | EVENT_LIST (12); |
|
|
1725 | |
|
|
1726 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1727 | |
|
|
1728 | if (ev) |
|
|
1729 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1730 | |
|
|
1731 | void |
|
|
1732 | 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, ...) |
|
|
1733 | ALIAS: |
|
|
1734 | enqueue_copy_image_to_buffer = 0 |
|
|
1735 | PPCODE: |
|
|
1736 | cl_event ev = 0; |
|
|
1737 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1738 | const size_t region [3] = { width, height, depth }; |
|
|
1739 | EVENT_LIST (10); |
|
|
1740 | |
|
|
1741 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1742 | |
|
|
1743 | if (ev) |
|
|
1744 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1745 | |
|
|
1746 | void |
|
|
1747 | 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, ...) |
|
|
1748 | ALIAS: |
|
|
1749 | enqueue_copy_buffer_to_image = 0 |
|
|
1750 | PPCODE: |
|
|
1751 | cl_event ev = 0; |
|
|
1752 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1753 | const size_t region [3] = { width, height, depth }; |
|
|
1754 | EVENT_LIST (10); |
|
|
1755 | |
|
|
1756 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1757 | |
|
|
1758 | if (ev) |
|
|
1759 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1760 | |
|
|
1761 | void |
|
|
1762 | 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, ...) |
|
|
1763 | ALIAS: |
|
|
1764 | enqueue_map_buffer = 0 |
|
|
1765 | PPCODE: |
|
|
1766 | cl_event ev; |
|
|
1767 | EVENT_LIST (6); |
|
|
1768 | size_t cb = SvIV (cb_); |
|
|
1769 | |
|
|
1770 | if (!SvOK (cb_)) |
|
|
1771 | { |
|
|
1772 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1773 | cb -= offset; |
|
|
1774 | } |
|
|
1775 | |
|
|
1776 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1777 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1778 | |
|
|
1779 | void |
|
|
1780 | 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, ...) |
|
|
1781 | ALIAS: |
|
|
1782 | enqueue_map_image = 0 |
|
|
1783 | PPCODE: |
|
|
1784 | size_t width = SvIV (width_); |
|
|
1785 | if (!SvOK (width_)) |
|
|
1786 | { |
|
|
1787 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1788 | width -= x; |
|
|
1789 | } |
|
|
1790 | |
|
|
1791 | size_t height = SvIV (width_); |
|
|
1792 | if (!SvOK (height_)) |
|
|
1793 | { |
|
|
1794 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1795 | height -= y; |
|
|
1796 | } |
|
|
1797 | |
|
|
1798 | size_t depth = SvIV (width_); |
|
|
1799 | if (!SvOK (depth_)) |
|
|
1800 | { |
|
|
1801 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1802 | depth -= z; |
|
|
1803 | |
|
|
1804 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1805 | if (!depth) |
|
|
1806 | depth = 1; |
|
|
1807 | } |
|
|
1808 | |
|
|
1809 | const size_t origin[3] = { x, y, z }; |
|
|
1810 | const size_t region[3] = { width, height, depth }; |
|
|
1811 | size_t row_pitch, slice_pitch; |
|
|
1812 | EVENT_LIST (10); |
|
|
1813 | |
|
|
1814 | cl_event ev; |
|
|
1815 | 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)); |
|
|
1816 | |
|
|
1817 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1818 | : row_pitch ? row_pitch * region [1] |
|
|
1819 | : region [0]; |
|
|
1820 | |
|
|
1821 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1822 | |
|
|
1823 | void |
|
|
1824 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1825 | PPCODE: |
|
|
1826 | mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); |
|
|
1827 | if (GIMME_V != G_VOID) |
|
|
1828 | { |
|
|
1829 | clRetainEvent (mapped->event); |
|
|
1830 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1831 | } |
|
|
1832 | |
|
|
1833 | void |
|
|
1834 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1835 | ALIAS: |
|
|
1836 | enqueue_task = 0 |
|
|
1837 | PPCODE: |
|
|
1838 | cl_event ev = 0; |
|
|
1839 | EVENT_LIST (2); |
|
|
1840 | |
|
|
1841 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1842 | |
|
|
1843 | if (ev) |
|
|
1844 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1845 | |
|
|
1846 | void |
|
|
1847 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1848 | ALIAS: |
|
|
1849 | enqueue_nd_range_kernel = 0 |
|
|
1850 | PPCODE: |
|
|
1851 | cl_event ev = 0; |
|
|
1852 | size_t *gwo = 0, *gws, *lws = 0; |
|
|
1853 | int gws_len; |
|
|
1854 | size_t *lists; |
|
|
1855 | int i; |
|
|
1856 | EVENT_LIST (5); |
|
|
1857 | |
|
|
1858 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
|
|
1859 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
|
|
1860 | |
|
|
1861 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
|
|
1862 | |
|
|
1863 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
|
|
1864 | |
|
|
1865 | gws = lists + gws_len * 0; |
|
|
1866 | for (i = 0; i < gws_len; ++i) |
|
|
1867 | { |
|
|
1868 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1869 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1870 | if (!gws [i]) |
|
|
1871 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1872 | } |
|
|
1873 | |
|
|
1874 | if (SvOK (global_work_offset)) |
|
|
1875 | { |
|
|
1876 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
|
|
1877 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
|
|
1878 | |
|
|
1879 | if (AvFILLp (SvRV (global_work_size)) + 1 != gws_len) |
|
|
1880 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array of same size as global_work_size"); |
|
|
1881 | |
|
|
1882 | gwo = lists + gws_len * 1; |
|
|
1883 | for (i = 0; i < gws_len; ++i) |
|
|
1884 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
|
|
1885 | } |
|
|
1886 | |
|
|
1887 | if (SvOK (local_work_size)) |
|
|
1888 | { |
|
|
1889 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
|
|
1890 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
|
|
1891 | |
|
|
1892 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
|
|
1893 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
|
|
1894 | |
|
|
1895 | lws = lists + gws_len * 2; |
|
|
1896 | for (i = 0; i < gws_len; ++i) |
|
|
1897 | { |
|
|
1898 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1899 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1900 | if (!lws [i]) |
|
|
1901 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1902 | } |
|
|
1903 | } |
|
|
1904 | |
|
|
1905 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1906 | |
|
|
1907 | if (ev) |
|
|
1908 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1909 | |
|
|
1910 | #if CL_VERSION_1_2 |
|
|
1911 | |
|
|
1912 | void |
|
|
1913 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
1914 | ALIAS: |
|
|
1915 | enqueue_migrate_mem_objects = 0 |
|
|
1916 | PPCODE: |
|
|
1917 | EVENT_LIST (3); |
|
|
1918 | |
|
|
1919 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
|
|
1920 | croak ("OpenCL::Queue::migrate_mem_objects: objects must be an array reference with OpenCL::Memory objects"); |
|
|
1921 | |
|
|
1922 | objects = SvRV (objects); |
|
|
1923 | |
|
|
1924 | cl_uint object_count = av_len ((AV *)objects) + 1; |
|
|
1925 | cl_mem *object_list = tmpbuf (sizeof (*object_list) * object_count); |
|
|
1926 | |
|
|
1927 | int i; |
|
|
1928 | for (i = object_count; i--; ) |
|
|
1929 | object_list [i] = SvCLOBJ ("OpenCL::Queue::migrate_mem_objects", "objects", *av_fetch ((AV *)objects, i, 0), "OpenCL::Memory"); |
|
|
1930 | |
|
|
1931 | cl_event ev = 0; |
|
|
1932 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, object_count, object_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1933 | |
|
|
1934 | if (ev) |
|
|
1935 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1936 | |
|
|
1937 | #endif |
|
|
1938 | |
|
|
1939 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1940 | |
|
|
1941 | void |
|
|
1942 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
1943 | ALIAS: |
|
|
1944 | release_gl_objects = 1 |
|
|
1945 | enqueue_acquire_gl_objects = 0 |
|
|
1946 | enqueue_release_gl_objects = 1 |
|
|
1947 | PPCODE: |
|
|
1948 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
|
|
1949 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
|
|
1950 | |
|
|
1951 | cl_event ev = 0; |
|
|
1952 | EVENT_LIST (2); |
|
|
1953 | AV *av = (AV *)SvRV (objects); |
|
|
1954 | cl_uint num_objects = av_len (av) + 1; |
|
|
1955 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1956 | |
|
|
1957 | int i; |
|
|
1958 | for (i = num_objects; i--; ) |
|
|
1959 | object_list [i] = SvCLOBJ ("OpenCL::Queue::acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1960 | |
|
|
1961 | if (ix) |
|
|
1962 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1963 | else |
|
|
1964 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1965 | |
|
|
1966 | if (ev) |
|
|
1967 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1968 | |
|
|
1969 | #endif |
|
|
1970 | |
|
|
1971 | void |
|
|
1972 | wait_for_events (OpenCL::Queue self, ...) |
|
|
1973 | ALIAS: |
|
|
1974 | enqueue_wait_for_events = 0 |
|
|
1975 | CODE: |
|
|
1976 | EVENT_LIST (1); |
|
|
1977 | #if PREFER_1_1 |
|
|
1978 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1979 | #else |
|
|
1980 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
1981 | #endif |
|
|
1982 | |
|
|
1983 | void |
|
|
1984 | marker (OpenCL::Queue self, ...) |
|
|
1985 | ALIAS: |
|
|
1986 | enqueue_marker = 0 |
|
|
1987 | PPCODE: |
|
|
1988 | cl_event ev = 0; |
|
|
1989 | EVENT_LIST (1); |
|
|
1990 | #if PREFER_1_1 |
|
|
1991 | if (!event_list_count) |
|
|
1992 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1993 | else |
|
|
1994 | #if CL_VERSION_1_2 |
|
|
1995 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1996 | #else |
|
|
1997 | { |
|
|
1998 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
1999 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2000 | } |
|
|
2001 | #endif |
|
|
2002 | #else |
|
|
2003 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2004 | #endif |
|
|
2005 | if (ev) |
|
|
2006 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2007 | |
|
|
2008 | void |
|
|
2009 | barrier (OpenCL::Queue self, ...) |
|
|
2010 | ALIAS: |
|
|
2011 | enqueue_barrier = 0 |
|
|
2012 | PPCODE: |
|
|
2013 | cl_event ev = 0; |
|
|
2014 | EVENT_LIST (1); |
|
|
2015 | #if PREFER_1_1 |
|
|
2016 | if (!event_list_count && GIMME_V == G_VOID) |
|
|
2017 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2018 | else |
|
|
2019 | #if CL_VERSION_1_2 |
|
|
2020 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2021 | #else |
|
|
2022 | { |
|
|
2023 | if (event_list_count) |
|
|
2024 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2025 | |
|
|
2026 | if (GIMME_V != G_VOID) |
|
|
2027 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2028 | } |
|
|
2029 | #endif |
|
|
2030 | #else |
|
|
2031 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2032 | #endif |
|
|
2033 | if (ev) |
|
|
2034 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2035 | |
|
|
2036 | void |
|
|
2037 | flush (OpenCL::Queue self) |
|
|
2038 | CODE: |
|
|
2039 | NEED_SUCCESS (Flush, (self)); |
|
|
2040 | |
|
|
2041 | void |
|
|
2042 | finish (OpenCL::Queue self) |
|
|
2043 | CODE: |
|
|
2044 | NEED_SUCCESS (Finish, (self)); |
|
|
2045 | |
|
|
2046 | void |
|
|
2047 | info (OpenCL::Queue self, cl_command_queue_info name) |
|
|
2048 | PPCODE: |
|
|
2049 | INFO (CommandQueue) |
|
|
2050 | |
|
|
2051 | #BEGIN:command_queue |
|
|
2052 | |
|
|
2053 | void |
|
|
2054 | context (OpenCL::Queue self) |
|
|
2055 | PPCODE: |
|
|
2056 | cl_context value [1]; |
|
|
2057 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
|
|
2058 | EXTEND (SP, 1); |
|
|
2059 | const int i = 0; |
|
|
2060 | { |
|
|
2061 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2062 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2063 | } |
|
|
2064 | |
|
|
2065 | void |
|
|
2066 | device (OpenCL::Queue self) |
|
|
2067 | PPCODE: |
|
|
2068 | cl_device_id value [1]; |
|
|
2069 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
|
|
2070 | EXTEND (SP, 1); |
|
|
2071 | const int i = 0; |
|
|
2072 | { |
|
|
2073 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2074 | } |
|
|
2075 | |
|
|
2076 | void |
|
|
2077 | reference_count (OpenCL::Queue self) |
|
|
2078 | PPCODE: |
|
|
2079 | cl_uint value [1]; |
|
|
2080 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0)); |
|
|
2081 | EXTEND (SP, 1); |
|
|
2082 | const int i = 0; |
|
|
2083 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2084 | |
|
|
2085 | void |
|
|
2086 | properties (OpenCL::Queue self) |
|
|
2087 | PPCODE: |
|
|
2088 | cl_command_queue_properties value [1]; |
|
|
2089 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
2090 | EXTEND (SP, 1); |
|
|
2091 | const int i = 0; |
|
|
2092 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2093 | |
|
|
2094 | #END:command_queue |
|
|
2095 | |
|
|
2096 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
|
|
2097 | |
|
|
2098 | void |
|
|
2099 | DESTROY (OpenCL::Memory self) |
|
|
2100 | CODE: |
|
|
2101 | clReleaseMemObject (self); |
|
|
2102 | |
|
|
2103 | void |
|
|
2104 | info (OpenCL::Memory self, cl_mem_info name) |
|
|
2105 | PPCODE: |
|
|
2106 | INFO (MemObject) |
|
|
2107 | |
|
|
2108 | #BEGIN:mem |
|
|
2109 | |
|
|
2110 | void |
|
|
2111 | type (OpenCL::Memory self) |
|
|
2112 | PPCODE: |
|
|
2113 | cl_mem_object_type value [1]; |
|
|
2114 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
|
|
2115 | EXTEND (SP, 1); |
|
|
2116 | const int i = 0; |
|
|
2117 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2118 | |
|
|
2119 | void |
|
|
2120 | flags (OpenCL::Memory self) |
|
|
2121 | PPCODE: |
|
|
2122 | cl_mem_flags value [1]; |
|
|
2123 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
|
|
2124 | EXTEND (SP, 1); |
|
|
2125 | const int i = 0; |
|
|
2126 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2127 | |
|
|
2128 | void |
|
|
2129 | size (OpenCL::Memory self) |
|
|
2130 | ALIAS: |
|
|
2131 | size = CL_MEM_SIZE |
|
|
2132 | offset = CL_MEM_OFFSET |
|
|
2133 | PPCODE: |
|
|
2134 | size_t value [1]; |
|
|
2135 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2136 | EXTEND (SP, 1); |
|
|
2137 | const int i = 0; |
|
|
2138 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2139 | |
|
|
2140 | void |
|
|
2141 | host_ptr (OpenCL::Memory self) |
|
|
2142 | PPCODE: |
|
|
2143 | void * value [1]; |
|
|
2144 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0)); |
|
|
2145 | EXTEND (SP, 1); |
|
|
2146 | const int i = 0; |
|
|
2147 | PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i]))); |
|
|
2148 | |
|
|
2149 | void |
|
|
2150 | map_count (OpenCL::Memory self) |
|
|
2151 | ALIAS: |
|
|
2152 | map_count = CL_MEM_MAP_COUNT |
|
|
2153 | reference_count = CL_MEM_REFERENCE_COUNT |
|
|
2154 | PPCODE: |
|
|
2155 | cl_uint value [1]; |
|
|
2156 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2157 | EXTEND (SP, 1); |
|
|
2158 | const int i = 0; |
|
|
2159 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2160 | |
|
|
2161 | void |
|
|
2162 | context (OpenCL::Memory self) |
|
|
2163 | PPCODE: |
|
|
2164 | cl_context value [1]; |
|
|
2165 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
|
|
2166 | EXTEND (SP, 1); |
|
|
2167 | const int i = 0; |
|
|
2168 | { |
|
|
2169 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2170 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2171 | } |
|
|
2172 | |
|
|
2173 | void |
|
|
2174 | associated_memobject (OpenCL::Memory self) |
|
|
2175 | PPCODE: |
|
|
2176 | cl_mem value [1]; |
|
|
2177 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
|
|
2178 | EXTEND (SP, 1); |
|
|
2179 | const int i = 0; |
|
|
2180 | { |
|
|
2181 | NEED_SUCCESS (RetainMemObject, (value [i])); |
|
|
2182 | PUSH_CLOBJ (stash_memory, value [i]); |
|
|
2183 | } |
|
|
2184 | |
|
|
2185 | #END:mem |
|
|
2186 | |
|
|
2187 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2188 | |
|
|
2189 | void |
|
|
2190 | gl_object_info (OpenCL::Memory self) |
|
|
2191 | PPCODE: |
|
|
2192 | cl_gl_object_type type; |
|
|
2193 | cl_GLuint name; |
|
|
2194 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
|
|
2195 | EXTEND (SP, 2); |
|
|
2196 | PUSHs (sv_2mortal (newSVuv (type))); |
|
|
2197 | PUSHs (sv_2mortal (newSVuv (name))); |
|
|
2198 | |
|
|
2199 | #endif |
|
|
2200 | |
|
|
2201 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
|
|
2202 | |
|
|
2203 | void |
|
|
2204 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
|
|
2205 | PPCODE: |
|
|
2206 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR)) |
|
|
2207 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
|
|
2208 | |
|
|
2209 | cl_buffer_region crdata = { origin, size }; |
|
|
2210 | |
|
|
2211 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
|
|
2212 | XPUSH_CLOBJ (stash_buffer, mem); |
|
|
2213 | |
|
|
2214 | MODULE = OpenCL PACKAGE = OpenCL::Image |
|
|
2215 | |
|
|
2216 | void |
|
|
2217 | image_info (OpenCL::Image self, cl_image_info name) |
|
|
2218 | PPCODE: |
|
|
2219 | INFO (Image) |
|
|
2220 | |
|
|
2221 | void |
|
|
2222 | format (OpenCL::Image self) |
|
|
2223 | PPCODE: |
|
|
2224 | cl_image_format format; |
|
|
2225 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2226 | EXTEND (SP, 2); |
|
|
2227 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2228 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
|
|
2229 | |
|
|
2230 | #BEGIN:image |
|
|
2231 | |
|
|
2232 | void |
|
|
2233 | element_size (OpenCL::Image self) |
|
|
2234 | ALIAS: |
|
|
2235 | element_size = CL_IMAGE_ELEMENT_SIZE |
|
|
2236 | row_pitch = CL_IMAGE_ROW_PITCH |
|
|
2237 | slice_pitch = CL_IMAGE_SLICE_PITCH |
|
|
2238 | width = CL_IMAGE_WIDTH |
|
|
2239 | height = CL_IMAGE_HEIGHT |
|
|
2240 | depth = CL_IMAGE_DEPTH |
|
|
2241 | PPCODE: |
|
|
2242 | size_t value [1]; |
|
|
2243 | NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2244 | EXTEND (SP, 1); |
|
|
2245 | const int i = 0; |
|
|
2246 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2247 | |
|
|
2248 | #END:image |
|
|
2249 | |
|
|
2250 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2251 | |
|
|
2252 | #BEGIN:gl_texture |
|
|
2253 | |
|
|
2254 | void |
|
|
2255 | target (OpenCL::Image self) |
|
|
2256 | PPCODE: |
|
|
2257 | cl_GLenum value [1]; |
|
|
2258 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
|
|
2259 | EXTEND (SP, 1); |
|
|
2260 | const int i = 0; |
|
|
2261 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2262 | |
|
|
2263 | void |
|
|
2264 | gl_mipmap_level (OpenCL::Image self) |
|
|
2265 | PPCODE: |
|
|
2266 | cl_GLint value [1]; |
|
|
2267 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
|
|
2268 | EXTEND (SP, 1); |
|
|
2269 | const int i = 0; |
|
|
2270 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2271 | |
|
|
2272 | #END:gl_texture |
|
|
2273 | |
|
|
2274 | #endif |
|
|
2275 | |
|
|
2276 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
|
|
2277 | |
|
|
2278 | void |
|
|
2279 | DESTROY (OpenCL::Sampler self) |
|
|
2280 | CODE: |
|
|
2281 | clReleaseSampler (self); |
|
|
2282 | |
|
|
2283 | void |
|
|
2284 | info (OpenCL::Sampler self, cl_sampler_info name) |
|
|
2285 | PPCODE: |
|
|
2286 | INFO (Sampler) |
|
|
2287 | |
|
|
2288 | #BEGIN:sampler |
|
|
2289 | |
|
|
2290 | void |
|
|
2291 | reference_count (OpenCL::Sampler self) |
|
|
2292 | PPCODE: |
|
|
2293 | cl_uint value [1]; |
|
|
2294 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_REFERENCE_COUNT, sizeof (value), value, 0)); |
|
|
2295 | EXTEND (SP, 1); |
|
|
2296 | const int i = 0; |
|
|
2297 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2298 | |
|
|
2299 | void |
|
|
2300 | context (OpenCL::Sampler self) |
|
|
2301 | PPCODE: |
|
|
2302 | cl_context value [1]; |
|
|
2303 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
|
|
2304 | EXTEND (SP, 1); |
|
|
2305 | const int i = 0; |
|
|
2306 | { |
|
|
2307 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2308 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2309 | } |
|
|
2310 | |
|
|
2311 | void |
|
|
2312 | normalized_coords (OpenCL::Sampler self) |
|
|
2313 | PPCODE: |
|
|
2314 | cl_addressing_mode value [1]; |
|
|
2315 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
|
|
2316 | EXTEND (SP, 1); |
|
|
2317 | const int i = 0; |
|
|
2318 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2319 | |
|
|
2320 | void |
|
|
2321 | addressing_mode (OpenCL::Sampler self) |
|
|
2322 | PPCODE: |
|
|
2323 | cl_filter_mode value [1]; |
|
|
2324 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
|
|
2325 | EXTEND (SP, 1); |
|
|
2326 | const int i = 0; |
|
|
2327 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2328 | |
|
|
2329 | void |
|
|
2330 | filter_mode (OpenCL::Sampler self) |
|
|
2331 | PPCODE: |
|
|
2332 | cl_bool value [1]; |
|
|
2333 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0)); |
|
|
2334 | EXTEND (SP, 1); |
|
|
2335 | const int i = 0; |
|
|
2336 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
2337 | |
|
|
2338 | #END:sampler |
|
|
2339 | |
|
|
2340 | MODULE = OpenCL PACKAGE = OpenCL::Program |
|
|
2341 | |
|
|
2342 | void |
|
|
2343 | DESTROY (OpenCL::Program self) |
|
|
2344 | CODE: |
|
|
2345 | clReleaseProgram (self); |
|
|
2346 | |
|
|
2347 | void |
|
|
2348 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2349 | ALIAS: |
|
|
2350 | build_async = 1 |
|
|
2351 | CODE: |
|
|
2352 | void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0; |
|
|
2353 | void *user_data = 0; |
|
|
2354 | cl_uint num_devices = 0; |
|
|
2355 | cl_device_id *device_list = 0; |
|
|
2356 | |
|
|
2357 | if (SvOK (devices)) |
|
|
2358 | { |
|
|
2359 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
2360 | croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); |
|
|
2361 | |
|
|
2362 | AV *av = (AV *)SvRV (devices); |
|
|
2363 | num_devices = av_len (av) + 1; |
|
|
2364 | |
|
|
2365 | if (num_devices) |
|
|
2366 | { |
|
|
2367 | device_list = tmpbuf (sizeof (*device_list) * num_devices); |
|
|
2368 | int count; |
|
|
2369 | for (count = 0; count < num_devices; ++count) |
|
|
2370 | device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device"); |
|
|
2371 | } |
|
|
2372 | } |
|
|
2373 | |
|
|
2374 | if (SvOK (notify)) |
|
|
2375 | { |
|
|
2376 | NEED_SUCCESS (RetainProgram, (self)); |
|
|
2377 | pfn_notify = eq_program_notify; |
|
|
2378 | user_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
2379 | } |
|
|
2380 | |
|
|
2381 | if (ix) |
|
|
2382 | build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2383 | else |
|
|
2384 | NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2385 | |
|
|
2386 | void |
|
|
2387 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
|
|
2388 | PPCODE: |
|
|
2389 | size_t size; |
|
|
2390 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size)); |
104 | sv = sv_2mortal (newSV (size)); |
2391 | SV *sv = sv_2mortal (newSV (size)); |
105 | SvUPGRADE (sv, SVt_PV); |
2392 | SvUPGRADE (sv, SVt_PV); |
106 | SvPOK_only (sv); |
2393 | SvPOK_only (sv); |
107 | SvCUR_set (sv, size); |
2394 | SvCUR_set (sv, size); |
108 | NEED_SUCCESS (clGetPlatformInfo, (this, name, size, SvPVX (sv), 0)); |
2395 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0)); |
109 | XPUSHs (sv); |
2396 | XPUSHs (sv); |
110 | } |
|
|
111 | |
2397 | |
|
|
2398 | #BEGIN:program_build |
|
|
2399 | |
112 | void |
2400 | void |
113 | devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) |
2401 | build_status (OpenCL::Program self, OpenCL::Device device) |
114 | PPCODE: |
2402 | PPCODE: |
115 | { |
2403 | cl_build_status value [1]; |
116 | cl_device_id *list; |
2404 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0)); |
|
|
2405 | EXTEND (SP, 1); |
|
|
2406 | const int i = 0; |
|
|
2407 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2408 | |
|
|
2409 | void |
|
|
2410 | build_options (OpenCL::Program self, OpenCL::Device device) |
|
|
2411 | ALIAS: |
|
|
2412 | build_options = CL_PROGRAM_BUILD_OPTIONS |
|
|
2413 | build_log = CL_PROGRAM_BUILD_LOG |
|
|
2414 | PPCODE: |
|
|
2415 | size_t size; |
|
|
2416 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size)); |
|
|
2417 | char *value = tmpbuf (size); |
|
|
2418 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
|
|
2419 | EXTEND (SP, 1); |
|
|
2420 | const int i = 0; |
|
|
2421 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2422 | |
|
|
2423 | #END:program_build |
|
|
2424 | |
|
|
2425 | void |
|
|
2426 | kernel (OpenCL::Program program, SV *function) |
|
|
2427 | PPCODE: |
|
|
2428 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
|
|
2429 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2430 | |
|
|
2431 | void |
|
|
2432 | kernels_in_program (OpenCL::Program program) |
|
|
2433 | PPCODE: |
117 | cl_uint count; |
2434 | cl_uint num_kernels; |
|
|
2435 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2436 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2437 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2438 | |
118 | int i; |
2439 | int i; |
|
|
2440 | EXTEND (SP, num_kernels); |
|
|
2441 | for (i = 0; i < num_kernels; ++i) |
|
|
2442 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
119 | |
2443 | |
120 | NEED_SUCCESS (clGetDeviceIDs, (this, type, 0, 0, &count)); |
2444 | void |
121 | Newx (list, count, cl_device_id); |
2445 | info (OpenCL::Program self, cl_program_info name) |
122 | NEED_SUCCESS (clGetDeviceIDs, (this, type, count, list, 0)); |
2446 | PPCODE: |
|
|
2447 | INFO (Program) |
123 | |
2448 | |
|
|
2449 | void |
|
|
2450 | binaries (OpenCL::Program self) |
|
|
2451 | PPCODE: |
|
|
2452 | cl_uint n, i; |
|
|
2453 | size_t size; |
|
|
2454 | |
|
|
2455 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0)); |
|
|
2456 | if (!n) XSRETURN_EMPTY; |
|
|
2457 | |
|
|
2458 | size_t *sizes = tmpbuf (sizeof (*sizes) * n); |
|
|
2459 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size)); |
|
|
2460 | if (size != sizeof (*sizes) * n) XSRETURN_EMPTY; |
|
|
2461 | unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n); |
|
|
2462 | |
124 | EXTEND (SP, count); |
2463 | EXTEND (SP, n); |
125 | for (i = 0; i < count; ++i) |
2464 | for (i = 0; i < n; ++i) |
126 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
2465 | { |
|
|
2466 | SV *sv = sv_2mortal (newSV (sizes [i])); |
|
|
2467 | SvUPGRADE (sv, SVt_PV); |
|
|
2468 | SvPOK_only (sv); |
|
|
2469 | SvCUR_set (sv, sizes [i]); |
|
|
2470 | ptrs [i] = (void *)SvPVX (sv); |
|
|
2471 | PUSHs (sv); |
|
|
2472 | } |
127 | |
2473 | |
128 | Safefree (list); |
2474 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
129 | } |
2475 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
130 | |
2476 | |
|
|
2477 | #BEGIN:program |
|
|
2478 | |
131 | void |
2479 | void |
132 | context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
2480 | reference_count (OpenCL::Program self) |
|
|
2481 | ALIAS: |
|
|
2482 | reference_count = CL_PROGRAM_REFERENCE_COUNT |
|
|
2483 | num_devices = CL_PROGRAM_NUM_DEVICES |
133 | PPCODE: |
2484 | PPCODE: |
134 | { |
2485 | cl_uint value [1]; |
135 | cl_int res; |
2486 | NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0)); |
136 | cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
2487 | EXTEND (SP, 1); |
137 | cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
2488 | const int i = 0; |
|
|
2489 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
138 | |
2490 | |
139 | if (res) |
|
|
140 | FAIL (clCreateContextFromType, res); |
|
|
141 | |
|
|
142 | XPUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Context", ctx)); |
|
|
143 | } |
|
|
144 | |
|
|
145 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
146 | |
|
|
147 | void |
2491 | void |
148 | info (OpenCL::Device this, cl_device_info name) |
2492 | context (OpenCL::Program self) |
149 | PPCODE: |
2493 | PPCODE: |
150 | { |
2494 | cl_context value [1]; |
|
|
2495 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
|
|
2496 | EXTEND (SP, 1); |
|
|
2497 | const int i = 0; |
|
|
2498 | { |
|
|
2499 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2500 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2501 | } |
|
|
2502 | |
|
|
2503 | void |
|
|
2504 | devices (OpenCL::Program self) |
|
|
2505 | PPCODE: |
151 | size_t size; |
2506 | size_t size; |
152 | SV *sv; |
2507 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size)); |
|
|
2508 | cl_device_id *value = tmpbuf (size); |
|
|
2509 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
|
|
2510 | int i, n = size / sizeof (*value); |
|
|
2511 | EXTEND (SP, n); |
|
|
2512 | for (i = 0; i < n; ++i) |
|
|
2513 | { |
|
|
2514 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2515 | } |
153 | |
2516 | |
154 | NEED_SUCCESS (clGetDeviceInfo, (this, name, 0, 0, &size)); |
2517 | void |
|
|
2518 | source (OpenCL::Program self) |
|
|
2519 | PPCODE: |
|
|
2520 | size_t size; |
|
|
2521 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size)); |
|
|
2522 | char *value = tmpbuf (size); |
|
|
2523 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0)); |
|
|
2524 | EXTEND (SP, 1); |
|
|
2525 | const int i = 0; |
|
|
2526 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2527 | |
|
|
2528 | void |
|
|
2529 | binary_sizes (OpenCL::Program self) |
|
|
2530 | PPCODE: |
|
|
2531 | size_t size; |
|
|
2532 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size)); |
|
|
2533 | size_t *value = tmpbuf (size); |
|
|
2534 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0)); |
|
|
2535 | int i, n = size / sizeof (*value); |
|
|
2536 | EXTEND (SP, n); |
|
|
2537 | for (i = 0; i < n; ++i) |
|
|
2538 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2539 | |
|
|
2540 | #END:program |
|
|
2541 | |
|
|
2542 | MODULE = OpenCL PACKAGE = OpenCL::Kernel |
|
|
2543 | |
|
|
2544 | void |
|
|
2545 | DESTROY (OpenCL::Kernel self) |
|
|
2546 | CODE: |
|
|
2547 | clReleaseKernel (self); |
|
|
2548 | |
|
|
2549 | void |
|
|
2550 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2551 | CODE: |
|
|
2552 | int i; |
|
|
2553 | for (i = 2; ; ++i) |
|
|
2554 | { |
|
|
2555 | while (*format == ' ') |
|
|
2556 | ++format; |
|
|
2557 | |
|
|
2558 | char type = *format++; |
|
|
2559 | |
|
|
2560 | if (!type) |
|
|
2561 | break; |
|
|
2562 | |
|
|
2563 | if (i >= items) |
|
|
2564 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2565 | |
|
|
2566 | SV *sv = ST (i); |
|
|
2567 | |
|
|
2568 | union |
|
|
2569 | { |
|
|
2570 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2571 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2572 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2573 | cl_mem cm; |
|
|
2574 | cl_sampler ca; |
|
|
2575 | size_t cz; |
|
|
2576 | cl_event ce; |
|
|
2577 | } arg; |
|
|
2578 | size_t size; |
|
|
2579 | int nullarg = 0; |
|
|
2580 | |
|
|
2581 | switch (type) |
|
|
2582 | { |
|
|
2583 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2584 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2585 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2586 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2587 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2588 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2589 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2590 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2591 | |
|
|
2592 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2593 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2594 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2595 | |
|
|
2596 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2597 | |
|
|
2598 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2599 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2600 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2601 | |
|
|
2602 | default: |
|
|
2603 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2604 | } |
|
|
2605 | |
|
|
2606 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2607 | if (res) |
|
|
2608 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2609 | } |
|
|
2610 | |
|
|
2611 | if (i != items) |
|
|
2612 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2613 | |
|
|
2614 | void |
|
|
2615 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
|
|
2616 | CODE: |
|
|
2617 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2618 | |
|
|
2619 | void |
|
|
2620 | set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value) |
|
|
2621 | CODE: |
|
|
2622 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2623 | |
|
|
2624 | void |
|
|
2625 | set_short (OpenCL::Kernel self, cl_uint idx, cl_short value) |
|
|
2626 | CODE: |
|
|
2627 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2628 | |
|
|
2629 | void |
|
|
2630 | set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value) |
|
|
2631 | CODE: |
|
|
2632 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2633 | |
|
|
2634 | void |
|
|
2635 | set_int (OpenCL::Kernel self, cl_uint idx, cl_int value) |
|
|
2636 | CODE: |
|
|
2637 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2638 | |
|
|
2639 | void |
|
|
2640 | set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value) |
|
|
2641 | CODE: |
|
|
2642 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2643 | |
|
|
2644 | void |
|
|
2645 | set_long (OpenCL::Kernel self, cl_uint idx, cl_long value) |
|
|
2646 | CODE: |
|
|
2647 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2648 | |
|
|
2649 | void |
|
|
2650 | set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value) |
|
|
2651 | CODE: |
|
|
2652 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2653 | |
|
|
2654 | void |
|
|
2655 | set_half (OpenCL::Kernel self, cl_uint idx, cl_half value) |
|
|
2656 | CODE: |
|
|
2657 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2658 | |
|
|
2659 | void |
|
|
2660 | set_float (OpenCL::Kernel self, cl_uint idx, cl_float value) |
|
|
2661 | CODE: |
|
|
2662 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2663 | |
|
|
2664 | void |
|
|
2665 | set_double (OpenCL::Kernel self, cl_uint idx, cl_double value) |
|
|
2666 | CODE: |
|
|
2667 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2668 | |
|
|
2669 | void |
|
|
2670 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
|
|
2671 | CODE: |
|
|
2672 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2673 | |
|
|
2674 | void |
|
|
2675 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2676 | CODE: |
|
|
2677 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2678 | |
|
|
2679 | void |
|
|
2680 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2681 | ALIAS: |
|
|
2682 | set_image2d = 0 |
|
|
2683 | set_image3d = 0 |
|
|
2684 | CODE: |
|
|
2685 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2686 | |
|
|
2687 | void |
|
|
2688 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2689 | CODE: |
|
|
2690 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2691 | |
|
|
2692 | void |
|
|
2693 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
|
|
2694 | CODE: |
|
|
2695 | clSetKernelArg (self, idx, size, 0); |
|
|
2696 | |
|
|
2697 | void |
|
|
2698 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
|
|
2699 | CODE: |
|
|
2700 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2701 | |
|
|
2702 | void |
|
|
2703 | info (OpenCL::Kernel self, cl_kernel_info name) |
|
|
2704 | PPCODE: |
|
|
2705 | INFO (Kernel) |
|
|
2706 | |
|
|
2707 | #BEGIN:kernel |
|
|
2708 | |
|
|
2709 | void |
|
|
2710 | function_name (OpenCL::Kernel self) |
|
|
2711 | PPCODE: |
|
|
2712 | size_t size; |
|
|
2713 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size)); |
|
|
2714 | char *value = tmpbuf (size); |
|
|
2715 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0)); |
|
|
2716 | EXTEND (SP, 1); |
|
|
2717 | const int i = 0; |
|
|
2718 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2719 | |
|
|
2720 | void |
|
|
2721 | num_args (OpenCL::Kernel self) |
|
|
2722 | ALIAS: |
|
|
2723 | num_args = CL_KERNEL_NUM_ARGS |
|
|
2724 | reference_count = CL_KERNEL_REFERENCE_COUNT |
|
|
2725 | PPCODE: |
|
|
2726 | cl_uint value [1]; |
|
|
2727 | NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2728 | EXTEND (SP, 1); |
|
|
2729 | const int i = 0; |
|
|
2730 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2731 | |
|
|
2732 | void |
|
|
2733 | context (OpenCL::Kernel self) |
|
|
2734 | PPCODE: |
|
|
2735 | cl_context value [1]; |
|
|
2736 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
|
|
2737 | EXTEND (SP, 1); |
|
|
2738 | const int i = 0; |
|
|
2739 | { |
|
|
2740 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2741 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2742 | } |
|
|
2743 | |
|
|
2744 | void |
|
|
2745 | program (OpenCL::Kernel self) |
|
|
2746 | PPCODE: |
|
|
2747 | cl_program value [1]; |
|
|
2748 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
|
|
2749 | EXTEND (SP, 1); |
|
|
2750 | const int i = 0; |
|
|
2751 | { |
|
|
2752 | NEED_SUCCESS (RetainProgram, (value [i])); |
|
|
2753 | PUSH_CLOBJ (stash_program, value [i]); |
|
|
2754 | } |
|
|
2755 | |
|
|
2756 | #END:kernel |
|
|
2757 | |
|
|
2758 | void |
|
|
2759 | work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name) |
|
|
2760 | PPCODE: |
|
|
2761 | size_t size; |
|
|
2762 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size)); |
155 | sv = sv_2mortal (newSV (size)); |
2763 | SV *sv = sv_2mortal (newSV (size)); |
156 | SvUPGRADE (sv, SVt_PV); |
2764 | SvUPGRADE (sv, SVt_PV); |
157 | SvPOK_only (sv); |
2765 | SvPOK_only (sv); |
158 | SvCUR_set (sv, size); |
2766 | SvCUR_set (sv, size); |
159 | NEED_SUCCESS (clGetDeviceInfo, (this, name, size, SvPVX (sv), 0)); |
2767 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0)); |
160 | XPUSHs (sv); |
2768 | XPUSHs (sv); |
161 | } |
|
|
162 | |
2769 | |
|
|
2770 | #BEGIN:kernel_work_group |
|
|
2771 | |
163 | void |
2772 | void |
164 | context_simple (OpenCL::Device this) |
2773 | work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2774 | ALIAS: |
|
|
2775 | work_group_size = CL_KERNEL_WORK_GROUP_SIZE |
|
|
2776 | preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE |
165 | PPCODE: |
2777 | PPCODE: |
166 | { |
2778 | size_t value [1]; |
167 | cl_int res; |
2779 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
168 | cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
2780 | EXTEND (SP, 1); |
|
|
2781 | const int i = 0; |
|
|
2782 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
169 | |
2783 | |
170 | if (res) |
|
|
171 | FAIL (clCreateContext, res); |
|
|
172 | |
|
|
173 | XPUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Context", ctx)); |
|
|
174 | } |
|
|
175 | |
|
|
176 | MODULE = OpenCL PACKAGE = OpenCL::Context |
|
|
177 | |
|
|
178 | void |
2784 | void |
179 | DESTROY (OpenCL::Context context) |
2785 | compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
180 | CODE: |
|
|
181 | clReleaseContext (context); |
|
|
182 | |
|
|
183 | void |
|
|
184 | info (OpenCL::Context this, cl_context_info name) |
|
|
185 | PPCODE: |
2786 | PPCODE: |
186 | { |
|
|
187 | size_t size; |
2787 | size_t size; |
188 | SV *sv; |
2788 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size)); |
|
|
2789 | size_t *value = tmpbuf (size); |
|
|
2790 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0)); |
|
|
2791 | int i, n = size / sizeof (*value); |
|
|
2792 | EXTEND (SP, n); |
|
|
2793 | for (i = 0; i < n; ++i) |
|
|
2794 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
189 | |
2795 | |
190 | NEED_SUCCESS (clGetContextInfo, (this, name, 0, 0, &size)); |
|
|
191 | sv = sv_2mortal (newSV (size)); |
|
|
192 | SvUPGRADE (sv, SVt_PV); |
|
|
193 | SvPOK_only (sv); |
|
|
194 | SvCUR_set (sv, size); |
|
|
195 | NEED_SUCCESS (clGetContextInfo, (this, name, size, SvPVX (sv), 0)); |
|
|
196 | XPUSHs (sv); |
|
|
197 | } |
|
|
198 | |
|
|
199 | void |
2796 | void |
200 | command_queue_simple (OpenCL::Context this, OpenCL::Device device) |
2797 | local_mem_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2798 | ALIAS: |
|
|
2799 | local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE |
|
|
2800 | private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE |
201 | PPCODE: |
2801 | PPCODE: |
202 | { |
2802 | cl_ulong value [1]; |
203 | cl_int res; |
2803 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
204 | cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res); |
2804 | EXTEND (SP, 1); |
|
|
2805 | const int i = 0; |
|
|
2806 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
205 | |
2807 | |
206 | if (res) |
2808 | #END:kernel_work_group |
207 | FAIL (clCreateCommandQueue, res); |
|
|
208 | |
2809 | |
209 | XPUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Queue", queue)); |
|
|
210 | } |
|
|
211 | |
|
|
212 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
2810 | MODULE = OpenCL PACKAGE = OpenCL::Event |
213 | |
2811 | |
214 | void |
2812 | void |
215 | DESTROY (OpenCL::Queue this) |
2813 | DESTROY (OpenCL::Event self) |
|
|
2814 | CODE: |
|
|
2815 | clReleaseEvent (self); |
|
|
2816 | |
|
|
2817 | void |
|
|
2818 | wait (OpenCL::Event self) |
|
|
2819 | CODE: |
|
|
2820 | clWaitForEvents (1, &self); |
|
|
2821 | |
|
|
2822 | void |
|
|
2823 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2824 | CODE: |
|
|
2825 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
|
|
2826 | |
|
|
2827 | void |
|
|
2828 | info (OpenCL::Event self, cl_event_info name) |
216 | CODE: |
2829 | PPCODE: |
217 | clReleaseCommandQueue (this); |
2830 | INFO (Event) |
218 | |
2831 | |
|
|
2832 | #BEGIN:event |
|
|
2833 | |
219 | void |
2834 | void |
220 | info (OpenCL::Queue this, cl_command_queue_info name) |
2835 | command_queue (OpenCL::Event self) |
221 | PPCODE: |
2836 | PPCODE: |
222 | { |
2837 | cl_command_queue value [1]; |
223 | size_t size; |
2838 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
224 | SV *sv; |
2839 | EXTEND (SP, 1); |
|
|
2840 | const int i = 0; |
|
|
2841 | { |
|
|
2842 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
|
|
2843 | PUSH_CLOBJ (stash_queue, value [i]); |
|
|
2844 | } |
225 | |
2845 | |
226 | NEED_SUCCESS (clGetCommandQueueInfo, (this, name, 0, 0, &size)); |
2846 | void |
227 | sv = sv_2mortal (newSV (size)); |
2847 | command_type (OpenCL::Event self) |
228 | SvUPGRADE (sv, SVt_PV); |
2848 | PPCODE: |
229 | SvPOK_only (sv); |
2849 | cl_command_type value [1]; |
230 | SvCUR_set (sv, size); |
2850 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_TYPE, sizeof (value), value, 0)); |
231 | NEED_SUCCESS (clGetCommandQueueInfo, (this, name, size, SvPVX (sv), 0)); |
2851 | EXTEND (SP, 1); |
232 | XPUSHs (sv); |
2852 | const int i = 0; |
233 | } |
2853 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
234 | |
2854 | |
|
|
2855 | void |
|
|
2856 | reference_count (OpenCL::Event self) |
|
|
2857 | ALIAS: |
|
|
2858 | reference_count = CL_EVENT_REFERENCE_COUNT |
|
|
2859 | command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS |
|
|
2860 | PPCODE: |
|
|
2861 | cl_uint value [1]; |
|
|
2862 | NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2863 | EXTEND (SP, 1); |
|
|
2864 | const int i = 0; |
|
|
2865 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2866 | |
|
|
2867 | void |
|
|
2868 | context (OpenCL::Event self) |
|
|
2869 | PPCODE: |
|
|
2870 | cl_context value [1]; |
|
|
2871 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
|
|
2872 | EXTEND (SP, 1); |
|
|
2873 | const int i = 0; |
|
|
2874 | { |
|
|
2875 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2876 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2877 | } |
|
|
2878 | |
|
|
2879 | #END:event |
|
|
2880 | |
|
|
2881 | void |
|
|
2882 | profiling_info (OpenCL::Event self, cl_profiling_info name) |
|
|
2883 | PPCODE: |
|
|
2884 | INFO (EventProfiling) |
|
|
2885 | |
|
|
2886 | #BEGIN:profiling |
|
|
2887 | |
|
|
2888 | void |
|
|
2889 | profiling_command_queued (OpenCL::Event self) |
|
|
2890 | ALIAS: |
|
|
2891 | profiling_command_queued = CL_PROFILING_COMMAND_QUEUED |
|
|
2892 | profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT |
|
|
2893 | profiling_command_start = CL_PROFILING_COMMAND_START |
|
|
2894 | profiling_command_end = CL_PROFILING_COMMAND_END |
|
|
2895 | PPCODE: |
|
|
2896 | cl_ulong value [1]; |
|
|
2897 | NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2898 | EXTEND (SP, 1); |
|
|
2899 | const int i = 0; |
|
|
2900 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2901 | |
|
|
2902 | #END:profiling |
|
|
2903 | |
|
|
2904 | MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
|
|
2905 | |
|
|
2906 | void |
|
|
2907 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
|
|
2908 | CODE: |
|
|
2909 | clSetUserEventStatus (self, execution_status); |
|
|
2910 | |
|
|
2911 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
2912 | |
|
|
2913 | void |
|
|
2914 | DESTROY (SV *self) |
|
|
2915 | CODE: |
|
|
2916 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
2917 | |
|
|
2918 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
2919 | mapped_detach (self, mapped); |
|
|
2920 | |
|
|
2921 | clReleaseCommandQueue (mapped->queue); |
|
|
2922 | clReleaseEvent (mapped->event); |
|
|
2923 | Safefree (mapped); |
|
|
2924 | |
|
|
2925 | void |
|
|
2926 | unmap (OpenCL::Mapped self, ...) |
|
|
2927 | CODE: |
|
|
2928 | mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); |
|
|
2929 | |
|
|
2930 | bool |
|
|
2931 | mapped (OpenCL::Mapped self) |
|
|
2932 | CODE: |
|
|
2933 | RETVAL = !!self->ptr; |
|
|
2934 | OUTPUT: |
|
|
2935 | RETVAL |
|
|
2936 | |
|
|
2937 | void |
|
|
2938 | wait (OpenCL::Mapped self) |
|
|
2939 | PPCODE: |
|
|
2940 | if (self->event) |
|
|
2941 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
2942 | |
|
|
2943 | void |
|
|
2944 | event (OpenCL::Mapped self) |
|
|
2945 | PPCODE: |
|
|
2946 | if (!self->event) |
|
|
2947 | XSRETURN_UNDEF; |
|
|
2948 | |
|
|
2949 | clRetainEvent (self->event); |
|
|
2950 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
2951 | |
|
|
2952 | size_t |
|
|
2953 | size (OpenCL::Mapped self) |
|
|
2954 | CODE: |
|
|
2955 | RETVAL = self->cb; |
|
|
2956 | OUTPUT: |
|
|
2957 | RETVAL |
|
|
2958 | |
|
|
2959 | IV |
|
|
2960 | ptr (OpenCL::Mapped self) |
|
|
2961 | CODE: |
|
|
2962 | RETVAL = PTR2IV (self->ptr); |
|
|
2963 | OUTPUT: |
|
|
2964 | RETVAL |
|
|
2965 | |
|
|
2966 | void |
|
|
2967 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
2968 | CODE: |
|
|
2969 | STRLEN len; |
|
|
2970 | const char *ptr = SvPVbyte (data, len); |
|
|
2971 | |
|
|
2972 | if (offset + len > self->cb) |
|
|
2973 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
2974 | |
|
|
2975 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
2976 | |
|
|
2977 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
2978 | |
|
|
2979 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
2980 | |
|
|
2981 | IV |
|
|
2982 | row_pitch (OpenCL::Mapped self) |
|
|
2983 | ALIAS: |
|
|
2984 | slice_pitch = 1 |
|
|
2985 | CODE: |
|
|
2986 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
2987 | OUTPUT: |
|
|
2988 | RETVAL |
|
|
2989 | |
|
|
2990 | |
|
|
2991 | |