1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #include "ecb.h"//D |
|
|
6 | |
|
|
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
8 | #include "xthread.h" |
|
|
9 | #include "schmorp.h" |
|
|
10 | |
5 | #ifdef I_DLFCN |
11 | #ifdef I_DLFCN |
6 | #include <dlfcn.h> |
12 | #include <dlfcn.h> |
|
|
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 |
7 | #endif |
30 | #endif |
8 | |
31 | |
9 | #ifdef __APPLE__ |
32 | #ifdef __APPLE__ |
10 | #include <OpenCL/opencl.h> |
33 | #include <OpenCL/opencl.h> |
11 | #else |
34 | #else |
12 | #include <CL/opencl.h> |
35 | #include <CL/opencl.h> |
13 | #endif |
36 | #endif |
14 | |
37 | |
|
|
38 | #ifndef CL_VERSION_1_2 |
|
|
39 | #undef PREFER_1_1 |
|
|
40 | #define PREFER_1_1 1 |
|
|
41 | #endif |
|
|
42 | |
15 | typedef cl_platform_id OpenCL__Platform; |
43 | typedef cl_platform_id OpenCL__Platform; |
16 | typedef cl_device_id OpenCL__Device; |
44 | typedef cl_device_id OpenCL__Device; |
|
|
45 | typedef cl_device_id OpenCL__SubDevice; |
17 | typedef cl_context OpenCL__Context; |
46 | typedef cl_context OpenCL__Context; |
18 | typedef cl_command_queue OpenCL__Queue; |
47 | typedef cl_command_queue OpenCL__Queue; |
19 | typedef cl_mem OpenCL__Memory; |
48 | typedef cl_mem OpenCL__Memory; |
20 | typedef cl_mem OpenCL__Buffer; |
49 | typedef cl_mem OpenCL__Buffer; |
21 | typedef cl_mem OpenCL__BufferObj; |
50 | typedef cl_mem OpenCL__BufferObj; |
22 | typedef cl_mem OpenCL__Image; |
51 | typedef cl_mem OpenCL__Image; |
23 | typedef cl_mem OpenCL__Image2D; |
|
|
24 | typedef cl_mem OpenCL__Image3D; |
|
|
25 | typedef cl_mem OpenCL__Memory_ornull; |
52 | typedef cl_mem OpenCL__Memory_ornull; |
26 | typedef cl_mem OpenCL__Buffer_ornull; |
53 | typedef cl_mem OpenCL__Buffer_ornull; |
27 | typedef cl_mem OpenCL__Image_ornull; |
54 | typedef cl_mem OpenCL__Image_ornull; |
28 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
29 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
30 | typedef cl_sampler OpenCL__Sampler; |
55 | typedef cl_sampler OpenCL__Sampler; |
31 | typedef cl_program OpenCL__Program; |
56 | typedef cl_program OpenCL__Program; |
32 | typedef cl_kernel OpenCL__Kernel; |
57 | typedef cl_kernel OpenCL__Kernel; |
33 | typedef cl_event OpenCL__Event; |
58 | typedef cl_event OpenCL__Event; |
34 | typedef cl_event OpenCL__UserEvent; |
59 | typedef cl_event OpenCL__UserEvent; |
35 | |
60 | |
36 | typedef SV *FUTURE; |
61 | typedef struct mapped * OpenCL__Mapped; |
|
|
62 | |
|
|
63 | static HV |
|
|
64 | *stash_platform, |
|
|
65 | *stash_device, |
|
|
66 | *stash_subdevice, |
|
|
67 | *stash_context, |
|
|
68 | *stash_queue, |
|
|
69 | *stash_program, |
|
|
70 | *stash_kernel, |
|
|
71 | *stash_sampler, |
|
|
72 | *stash_event, |
|
|
73 | *stash_userevent, |
|
|
74 | *stash_memory, |
|
|
75 | *stash_buffer, |
|
|
76 | *stash_bufferobj, |
|
|
77 | *stash_image, |
|
|
78 | *stash_image1d, |
|
|
79 | *stash_image1darray, |
|
|
80 | *stash_image1dbuffer, |
|
|
81 | *stash_image2d, |
|
|
82 | *stash_image2darray, |
|
|
83 | *stash_image3d, |
|
|
84 | *stash_mapped, |
|
|
85 | *stash_mappedbuffer, |
|
|
86 | *stash_mappedimage; |
37 | |
87 | |
38 | /*****************************************************************************/ |
88 | /*****************************************************************************/ |
39 | |
89 | |
40 | // name must include a leading underscore |
90 | // name must include a leading underscore |
|
|
91 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
92 | // for perl. doh. |
41 | static void * |
93 | static void * |
42 | getsym (const char *name) |
94 | glsym (const char *name) |
43 | { |
95 | { |
|
|
96 | void *fun = 0; |
|
|
97 | |
44 | #if defined I_DLFCN && defined RTLD_DEFAULT |
98 | #if defined I_DLFCN && defined RTLD_DEFAULT |
45 | #if !DLSYM_NEEDS_UNDERSCORE |
99 | fun = dlsym (RTLD_DEFAULT, name + 1); |
46 | ++name; // skip _ |
|
|
47 | #endif |
|
|
48 | return dlsym (RTLD_DEFAULT, name); |
100 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
49 | #else |
101 | |
50 | return 0; |
102 | if (!fun) |
|
|
103 | { |
|
|
104 | static void *libgl; |
|
|
105 | static const char *glso[] = { |
|
|
106 | "libGL.so.1", |
|
|
107 | "libGL.so.3", |
|
|
108 | "libGL.so.4.0", |
|
|
109 | "libGL.so", |
|
|
110 | "/usr/lib/libGL.so", |
|
|
111 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
112 | }; |
|
|
113 | int i; |
|
|
114 | |
|
|
115 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
116 | { |
|
|
117 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
118 | if (libgl) |
|
|
119 | break; |
|
|
120 | } |
|
|
121 | |
|
|
122 | if (libgl) |
|
|
123 | { |
|
|
124 | fun = dlsym (libgl, name + 1); |
|
|
125 | if (!fun) fun = dlsym (libgl, name); |
|
|
126 | } |
|
|
127 | } |
51 | #endif |
128 | #endif |
|
|
129 | |
|
|
130 | return fun; |
52 | } |
131 | } |
53 | |
132 | |
54 | /*****************************************************************************/ |
133 | /*****************************************************************************/ |
55 | |
134 | |
56 | /* up to two temporary buffers */ |
135 | /* up to two temporary buffers */ |
57 | static void * |
136 | static void * ecb_noinline |
58 | tmpbuf (size_t size) |
137 | tmpbuf (size_t size) |
59 | { |
138 | { |
60 | enum { buffers = 3 }; |
139 | enum { buffers = 4 }; |
61 | static int idx; |
140 | static int idx; |
62 | static void *buf [buffers]; |
141 | static void *buf [buffers]; |
63 | static size_t len [buffers]; |
142 | static size_t len [buffers]; |
64 | |
143 | |
65 | idx = ++idx % buffers; |
144 | idx = (idx + 1) % buffers; |
66 | |
145 | |
67 | if (len [idx] < size) |
146 | if (len [idx] < size) |
68 | { |
147 | { |
69 | free (buf [idx]); |
148 | free (buf [idx]); |
70 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
149 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
71 | buf [idx] = malloc (len [idx]); |
150 | buf [idx] = malloc (len [idx]); |
72 | } |
151 | } |
73 | |
152 | |
74 | return buf [idx]; |
153 | return buf [idx]; |
|
|
154 | } |
|
|
155 | |
|
|
156 | static const char * ecb_noinline |
|
|
157 | cv_get_name (CV *cv) |
|
|
158 | { |
|
|
159 | static char fullname [256]; |
|
|
160 | |
|
|
161 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
162 | |
|
|
163 | HV *stash = GvSTASH (gv); |
|
|
164 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
165 | const char *gvname = GvNAME (gv); |
|
|
166 | |
|
|
167 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
168 | return fullname; |
75 | } |
169 | } |
76 | |
170 | |
77 | /*****************************************************************************/ |
171 | /*****************************************************************************/ |
78 | |
172 | |
79 | typedef struct |
173 | typedef struct |
… | |
… | |
138 | if (res) \ |
232 | if (res) \ |
139 | FAIL (name); |
233 | FAIL (name); |
140 | |
234 | |
141 | /*****************************************************************************/ |
235 | /*****************************************************************************/ |
142 | |
236 | |
143 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
144 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
145 | |
|
|
146 | static void * |
237 | static SV * |
|
|
238 | new_clobj (HV *stash, IV id) |
|
|
239 | { |
|
|
240 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
241 | } |
|
|
242 | |
|
|
243 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
244 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
245 | |
|
|
246 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
247 | /* they can be upgraded at runtime to the array form */ |
|
|
248 | static void * ecb_noinline |
147 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
249 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
148 | { |
250 | { |
|
|
251 | // sv_derived_from is quite slow :( |
149 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
252 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
150 | return (void *)SvIV (SvRV (sv)); |
253 | return (void *)SvIV (SvRV (sv)); |
151 | |
254 | |
152 | croak ("%s: %s is not of type %s", func, svname, pkg); |
255 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
256 | } |
|
|
257 | |
|
|
258 | // the "no-inherit" version of the above |
|
|
259 | static void * ecb_noinline |
|
|
260 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
261 | { |
|
|
262 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
263 | return (void *)SvIV (SvRV (sv)); |
|
|
264 | |
|
|
265 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
153 | } |
266 | } |
154 | |
267 | |
155 | /*****************************************************************************/ |
268 | /*****************************************************************************/ |
156 | |
269 | |
157 | static cl_context_properties * |
270 | static cl_context_properties * ecb_noinline |
158 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
271 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
159 | { |
272 | { |
160 | if (!sv || !SvOK (sv)) |
273 | if (!sv || !SvOK (sv)) |
161 | if (extra) |
274 | if (extra) |
162 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
275 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
163 | else |
276 | else |
… | |
… | |
169 | int i, len = av_len (av) + 1; |
282 | int i, len = av_len (av) + 1; |
170 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
283 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
171 | cl_context_properties *l = p; |
284 | cl_context_properties *l = p; |
172 | |
285 | |
173 | if (len & 1) |
286 | if (len & 1) |
174 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
287 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
175 | |
288 | |
176 | while (extracount--) |
289 | while (extracount--) |
177 | *l++ = *extra++; |
290 | *l++ = *extra++; |
178 | |
291 | |
179 | for (i = 0; i < len; i += 2) |
292 | for (i = 0; i < len; i += 2) |
180 | { |
293 | { |
181 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
294 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
182 | SV *p_sv = *av_fetch (av, i + 1, 0); |
295 | SV *p_sv = *av_fetch (av, i + 1, 0); |
183 | cl_context_properties v; |
296 | cl_context_properties v = SvIV (p_sv); // code below can override |
184 | |
297 | |
185 | switch (t) |
298 | switch (t) |
186 | { |
299 | { |
|
|
300 | case CL_CONTEXT_PLATFORM: |
|
|
301 | if (SvROK (p_sv)) |
|
|
302 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
303 | break; |
|
|
304 | |
|
|
305 | case CL_GLX_DISPLAY_KHR: |
|
|
306 | if (!SvOK (p_sv)) |
|
|
307 | { |
|
|
308 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
309 | if (func) |
|
|
310 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
311 | } |
|
|
312 | break; |
|
|
313 | |
|
|
314 | case CL_GL_CONTEXT_KHR: |
|
|
315 | if (!SvOK (p_sv)) |
|
|
316 | { |
|
|
317 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
318 | if (func) |
|
|
319 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
320 | } |
|
|
321 | break; |
|
|
322 | |
187 | default: |
323 | default: |
188 | /* unknown property, treat as int */ |
324 | /* unknown property, treat as int */ |
189 | v = SvIV (p_sv); |
|
|
190 | break; |
325 | break; |
191 | } |
326 | } |
192 | |
327 | |
193 | *l++ = t; |
328 | *l++ = t; |
194 | *l++ = v; |
329 | *l++ = v; |
… | |
… | |
197 | *l = 0; |
332 | *l = 0; |
198 | |
333 | |
199 | return p; |
334 | return p; |
200 | } |
335 | } |
201 | |
336 | |
202 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
337 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
|
|
338 | } |
|
|
339 | |
|
|
340 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
|
|
341 | // is a pointer (and only on well-behaved systems). |
|
|
342 | static void * ecb_noinline |
|
|
343 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
344 | { |
|
|
345 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
346 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
347 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
|
|
348 | |
|
|
349 | AV *av = (AV *)SvRV (arg); |
|
|
350 | void **list = 0; |
|
|
351 | cl_uint count = av_len (av) + 1; |
|
|
352 | |
|
|
353 | if (count) |
|
|
354 | { |
|
|
355 | list = tmpbuf (sizeof (*list) * count); |
|
|
356 | int i; |
|
|
357 | for (i = 0; i < count; ++i) |
|
|
358 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
|
|
359 | } |
|
|
360 | |
|
|
361 | if (!count && !or_undef) |
|
|
362 | croak ("%s: '%s' must contain at least one %s object", |
|
|
363 | cv_get_name (cv), argname, klass); |
|
|
364 | |
|
|
365 | *rcount = count; |
|
|
366 | return (void *)list; |
203 | } |
367 | } |
204 | |
368 | |
205 | /*****************************************************************************/ |
369 | /*****************************************************************************/ |
|
|
370 | /* callback stuff */ |
|
|
371 | |
|
|
372 | /* default context callback, log to stderr */ |
|
|
373 | static void CL_CALLBACK |
|
|
374 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
375 | { |
|
|
376 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
377 | } |
|
|
378 | |
|
|
379 | typedef struct |
|
|
380 | { |
|
|
381 | int free_cb; |
|
|
382 | void (*push)(void *data1, void *data2, void *data3); |
|
|
383 | } eq_vtbl; |
|
|
384 | |
|
|
385 | typedef struct eq_item |
|
|
386 | { |
|
|
387 | struct eq_item *next; |
|
|
388 | eq_vtbl *vtbl; |
|
|
389 | SV *cb; |
|
|
390 | void *data1, *data2, *data3; |
|
|
391 | } eq_item; |
|
|
392 | |
|
|
393 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
394 | static void *eq_signal_arg; |
|
|
395 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
396 | static eq_item *eq_head, *eq_tail; |
|
|
397 | |
|
|
398 | static void ecb_noinline |
|
|
399 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
400 | { |
|
|
401 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
402 | |
|
|
403 | item->next = 0; |
|
|
404 | item->vtbl = vtbl; |
|
|
405 | item->cb = cb; |
|
|
406 | item->data1 = data1; |
|
|
407 | item->data2 = data2; |
|
|
408 | item->data3 = data3; |
|
|
409 | |
|
|
410 | X_LOCK (eq_lock); |
|
|
411 | |
|
|
412 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
413 | eq_tail = item; |
|
|
414 | |
|
|
415 | X_UNLOCK (eq_lock); |
|
|
416 | |
|
|
417 | eq_signal_func (eq_signal_arg, 0); |
|
|
418 | } |
|
|
419 | |
|
|
420 | static eq_item * |
|
|
421 | eq_dec (void) |
|
|
422 | { |
|
|
423 | eq_item *res; |
|
|
424 | |
|
|
425 | X_LOCK (eq_lock); |
|
|
426 | |
|
|
427 | res = eq_head; |
|
|
428 | if (res) |
|
|
429 | eq_head = res->next; |
|
|
430 | |
|
|
431 | X_UNLOCK (eq_lock); |
|
|
432 | |
|
|
433 | return res; |
|
|
434 | } |
|
|
435 | |
|
|
436 | static void |
|
|
437 | eq_poll (void) |
|
|
438 | { |
|
|
439 | eq_item *item; |
|
|
440 | |
|
|
441 | while ((item = eq_dec ())) |
|
|
442 | { |
|
|
443 | ENTER; |
|
|
444 | SAVETMPS; |
|
|
445 | |
|
|
446 | dSP; |
|
|
447 | PUSHMARK (SP); |
|
|
448 | EXTEND (SP, 2); |
|
|
449 | |
|
|
450 | if (item->vtbl->free_cb) |
|
|
451 | sv_2mortal (item->cb); |
|
|
452 | |
|
|
453 | PUTBACK; |
|
|
454 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
455 | |
|
|
456 | SV *cb = item->cb; |
|
|
457 | free (item); |
|
|
458 | |
|
|
459 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
460 | |
|
|
461 | FREETMPS; |
|
|
462 | LEAVE; |
|
|
463 | } |
|
|
464 | } |
|
|
465 | |
|
|
466 | static void |
|
|
467 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
468 | { |
|
|
469 | eq_poll (); |
|
|
470 | } |
|
|
471 | |
|
|
472 | /*****************************************************************************/ |
|
|
473 | /* context notify */ |
|
|
474 | |
|
|
475 | static void ecb_noinline |
|
|
476 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
477 | { |
|
|
478 | dSP; |
|
|
479 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
480 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
481 | PUTBACK; |
|
|
482 | |
|
|
483 | free (data1); |
|
|
484 | free (data2); |
|
|
485 | } |
|
|
486 | |
|
|
487 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
488 | |
|
|
489 | static void CL_CALLBACK |
|
|
490 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
491 | { |
|
|
492 | void *pvt_copy = malloc (cb); |
|
|
493 | memcpy (pvt_copy, pvt, cb); |
|
|
494 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
495 | } |
|
|
496 | |
|
|
497 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
498 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
499 | void *user_data = 0; \ |
|
|
500 | \ |
|
|
501 | if (SvOK (notify)) \ |
|
|
502 | { \ |
|
|
503 | pfn_notify = eq_context_notify; \ |
|
|
504 | user_data = s_get_cv (notify); \ |
|
|
505 | } |
|
|
506 | |
|
|
507 | static SV * ecb_noinline |
|
|
508 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
509 | { |
|
|
510 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
511 | |
|
|
512 | if (user_data) |
|
|
513 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
514 | |
|
|
515 | return sv; |
|
|
516 | } |
|
|
517 | |
|
|
518 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
519 | |
|
|
520 | /*****************************************************************************/ |
|
|
521 | /* build/compile/link notify */ |
|
|
522 | |
|
|
523 | static void |
|
|
524 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
525 | { |
|
|
526 | dSP; |
|
|
527 | PUSH_CLOBJ (stash_program, data1); |
|
|
528 | PUTBACK; |
|
|
529 | } |
|
|
530 | |
|
|
531 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
532 | |
|
|
533 | static void CL_CALLBACK |
|
|
534 | eq_program_notify (cl_program program, void *user_data) |
|
|
535 | { |
|
|
536 | clRetainProgram (program); |
|
|
537 | |
|
|
538 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
539 | } |
|
|
540 | |
|
|
541 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
542 | |
|
|
543 | static program_callback ecb_noinline |
|
|
544 | make_program_callback (SV *notify, void **ruser_data) |
|
|
545 | { |
|
|
546 | if (SvOK (notify)) |
|
|
547 | { |
|
|
548 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
549 | return eq_program_notify; |
|
|
550 | } |
|
|
551 | else |
|
|
552 | { |
|
|
553 | *ruser_data = 0; |
|
|
554 | return 0; |
|
|
555 | } |
|
|
556 | } |
|
|
557 | |
|
|
558 | struct build_args |
|
|
559 | { |
|
|
560 | cl_program program; |
|
|
561 | char *options; |
|
|
562 | void *user_data; |
|
|
563 | cl_uint num_devices; |
|
|
564 | }; |
|
|
565 | |
|
|
566 | X_THREAD_PROC (build_program_thread) |
|
|
567 | { |
|
|
568 | struct build_args *arg = thr_arg; |
|
|
569 | |
|
|
570 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
571 | |
|
|
572 | if (arg->user_data) |
|
|
573 | eq_program_notify (arg->program, arg->user_data); |
|
|
574 | else |
|
|
575 | clReleaseProgram (arg->program); |
|
|
576 | |
|
|
577 | free (arg->options); |
|
|
578 | free (arg); |
|
|
579 | |
|
|
580 | return 0; |
|
|
581 | } |
|
|
582 | |
|
|
583 | static void |
|
|
584 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
585 | { |
|
|
586 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
587 | |
|
|
588 | arg->program = program; |
|
|
589 | arg->options = strdup (options); |
|
|
590 | arg->user_data = user_data; |
|
|
591 | arg->num_devices = num_devices; |
|
|
592 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
593 | |
|
|
594 | xthread_t id; |
|
|
595 | thread_create (&id, build_program_thread, arg); |
|
|
596 | } |
|
|
597 | |
|
|
598 | /*****************************************************************************/ |
|
|
599 | /* event objects */ |
|
|
600 | |
|
|
601 | static void |
|
|
602 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
603 | { |
|
|
604 | dSP; |
|
|
605 | PUSH_CLOBJ (stash_event, data1); |
|
|
606 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
607 | PUTBACK; |
|
|
608 | } |
|
|
609 | |
|
|
610 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
611 | |
|
|
612 | static void CL_CALLBACK |
|
|
613 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
614 | { |
|
|
615 | clRetainEvent (event); |
|
|
616 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
617 | } |
|
|
618 | |
|
|
619 | /*****************************************************************************/ |
|
|
620 | /* utilities for XS code */ |
206 | |
621 | |
207 | static size_t |
622 | static size_t |
208 | img_row_pitch (cl_mem img) |
623 | img_row_pitch (cl_mem img) |
209 | { |
624 | { |
210 | size_t res; |
625 | size_t res; |
211 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
626 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
212 | return res; |
627 | return res; |
213 | } |
628 | } |
214 | |
629 | |
215 | static cl_event * |
630 | static cl_event * ecb_noinline |
216 | event_list (SV **items, int count) |
631 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
217 | { |
632 | { |
218 | if (!count) |
633 | cl_uint count = *rcount; |
|
|
634 | |
|
|
635 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
636 | *rcount = count = 0; |
|
|
637 | |
|
|
638 | if (!count && !extra) |
219 | return 0; |
639 | return 0; |
220 | |
640 | |
221 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
641 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
642 | int i = 0; |
222 | |
643 | |
223 | while (count--) |
644 | while (count--) |
|
|
645 | if (SvOK (items [count])) |
224 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
646 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
225 | |
647 | |
|
|
648 | if (extra) |
|
|
649 | list [i++] = extra; |
|
|
650 | |
|
|
651 | *rcount = i; |
|
|
652 | |
226 | return list; |
653 | return i ? list : 0; |
227 | } |
654 | } |
228 | |
655 | |
229 | #define EVENT_LIST(items,count) \ |
656 | #define EVENT_LIST(skip) \ |
230 | cl_uint event_list_count = (count); \ |
657 | cl_uint event_list_count = items - (skip); \ |
231 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
658 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
232 | |
659 | |
233 | #define INFO(class) \ |
660 | #define INFO(class) \ |
234 | { \ |
661 | { \ |
235 | size_t size; \ |
662 | size_t size; \ |
236 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
663 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
240 | SvCUR_set (sv, size); \ |
667 | SvCUR_set (sv, size); \ |
241 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
668 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
242 | XPUSHs (sv); \ |
669 | XPUSHs (sv); \ |
243 | } |
670 | } |
244 | |
671 | |
|
|
672 | /*****************************************************************************/ |
|
|
673 | /* mapped_xxx */ |
|
|
674 | |
|
|
675 | static OpenCL__Mapped |
|
|
676 | SvMAPPED (SV *sv) |
|
|
677 | { |
|
|
678 | // no typechecking atm., keep your fingers crossed |
|
|
679 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
680 | } |
|
|
681 | |
|
|
682 | struct mapped |
|
|
683 | { |
|
|
684 | cl_command_queue queue; |
|
|
685 | cl_mem memobj; |
|
|
686 | void *ptr; |
|
|
687 | size_t cb; |
|
|
688 | cl_event event; |
|
|
689 | size_t row_pitch; |
|
|
690 | size_t slice_pitch; |
|
|
691 | }; |
|
|
692 | |
|
|
693 | static SV * |
|
|
694 | 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) |
|
|
695 | { |
|
|
696 | SV *data = newSV (0); |
|
|
697 | SvUPGRADE (data, SVt_PVMG); |
|
|
698 | |
|
|
699 | OpenCL__Mapped mapped; |
|
|
700 | New (0, mapped, 1, struct mapped); |
|
|
701 | |
|
|
702 | clRetainCommandQueue (queue); |
|
|
703 | |
|
|
704 | mapped->queue = queue; |
|
|
705 | mapped->memobj = memobj; |
|
|
706 | mapped->ptr = ptr; |
|
|
707 | mapped->cb = cb; |
|
|
708 | mapped->event = ev; |
|
|
709 | mapped->row_pitch = row_pitch; |
|
|
710 | mapped->slice_pitch = slice_pitch; |
|
|
711 | |
|
|
712 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
713 | |
|
|
714 | if (SvLEN (data)) |
|
|
715 | Safefree (data); |
|
|
716 | |
|
|
717 | SvPVX (data) = (char *)ptr; |
|
|
718 | SvCUR_set (data, cb); |
|
|
719 | SvLEN_set (data, 0); |
|
|
720 | SvPOK_only (data); |
|
|
721 | |
|
|
722 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
723 | |
|
|
724 | if (!(flags & CL_MAP_WRITE)) |
|
|
725 | SvREADONLY_on (data); |
|
|
726 | |
|
|
727 | return obj; |
|
|
728 | } |
|
|
729 | |
|
|
730 | static void |
|
|
731 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
732 | { |
|
|
733 | SV *data = SvRV (sv); |
|
|
734 | |
|
|
735 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
736 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
737 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
738 | else |
|
|
739 | { |
|
|
740 | SvREADONLY_off (data); |
|
|
741 | SvCUR_set (data, 0); |
|
|
742 | SvPVX (data) = 0; |
|
|
743 | SvOK_off (data); |
|
|
744 | } |
|
|
745 | |
|
|
746 | mapped->ptr = 0; |
|
|
747 | } |
|
|
748 | |
|
|
749 | static void |
|
|
750 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
751 | { |
|
|
752 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
753 | cl_event ev; |
|
|
754 | |
|
|
755 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
756 | |
|
|
757 | clReleaseEvent (mapped->event); |
|
|
758 | mapped->event = ev; |
|
|
759 | |
|
|
760 | mapped_detach (self, mapped); |
|
|
761 | } |
|
|
762 | |
|
|
763 | /*****************************************************************************/ |
|
|
764 | |
245 | MODULE = OpenCL PACKAGE = OpenCL |
765 | MODULE = OpenCL PACKAGE = OpenCL |
246 | |
766 | |
247 | PROTOTYPES: ENABLE |
767 | PROTOTYPES: ENABLE |
|
|
768 | |
|
|
769 | void |
|
|
770 | poll () |
|
|
771 | CODE: |
|
|
772 | eq_poll (); |
|
|
773 | |
|
|
774 | void |
|
|
775 | _eq_initialise (IV func, IV arg) |
|
|
776 | CODE: |
|
|
777 | eq_signal_func = (void (*)(void *, int))func; |
|
|
778 | eq_signal_arg = (void*)arg; |
248 | |
779 | |
249 | BOOT: |
780 | BOOT: |
250 | { |
781 | { |
251 | HV *stash = gv_stashpv ("OpenCL", 1); |
782 | HV *stash = gv_stashpv ("OpenCL", 1); |
252 | static const ivstr *civ, const_iv[] = { |
783 | static const ivstr *civ, const_iv[] = { |
… | |
… | |
259 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
790 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
260 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
791 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
261 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
792 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
262 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
793 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
263 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
794 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
795 | { PREFER_1_1 , "PREFER_1_1" }, |
264 | #include "constiv.h" |
796 | #include "constiv.h" |
265 | }; |
797 | }; |
|
|
798 | |
266 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
799 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
267 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
800 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
801 | |
|
|
802 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
803 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
804 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
805 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
806 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
807 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
808 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
809 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
810 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
811 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
812 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
813 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
814 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
815 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
816 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
817 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
818 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
819 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
820 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
821 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
822 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
823 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
824 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
825 | |
|
|
826 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
268 | } |
827 | } |
269 | |
828 | |
270 | cl_int |
829 | cl_int |
271 | errno () |
830 | errno () |
272 | CODE: |
831 | CODE: |
273 | errno = res; |
832 | RETVAL = res; |
|
|
833 | OUTPUT: |
|
|
834 | RETVAL |
274 | |
835 | |
275 | const char * |
836 | const char * |
276 | err2str (cl_int err) |
837 | err2str (cl_int err = res) |
277 | |
838 | |
278 | const char * |
839 | const char * |
279 | enum2str (cl_uint value) |
840 | enum2str (cl_uint value) |
280 | |
841 | |
281 | void |
842 | void |
… | |
… | |
289 | list = tmpbuf (sizeof (*list) * count); |
850 | list = tmpbuf (sizeof (*list) * count); |
290 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
851 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
291 | |
852 | |
292 | EXTEND (SP, count); |
853 | EXTEND (SP, count); |
293 | for (i = 0; i < count; ++i) |
854 | for (i = 0; i < count; ++i) |
294 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
855 | PUSH_CLOBJ (stash_platform, list [i]); |
295 | |
856 | |
296 | void |
857 | void |
297 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
858 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
298 | PPCODE: |
859 | PPCODE: |
|
|
860 | CONTEXT_NOTIFY_CALLBACK; |
299 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
861 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
300 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
862 | XPUSH_CLOBJ_CONTEXT; |
301 | |
863 | |
302 | void |
864 | void |
303 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
865 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
304 | PPCODE: |
866 | PPCODE: |
305 | /* der Gipfel der Kunst */ |
867 | cl_uint device_count; |
|
|
868 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
869 | |
|
|
870 | CONTEXT_NOTIFY_CALLBACK; |
|
|
871 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
872 | XPUSH_CLOBJ_CONTEXT; |
306 | |
873 | |
307 | void |
874 | void |
308 | wait_for_events (...) |
875 | wait_for_events (...) |
309 | CODE: |
876 | CODE: |
310 | EVENT_LIST (0, items); |
877 | EVENT_LIST (0); |
311 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
878 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
312 | |
879 | |
313 | PROTOTYPES: DISABLE |
880 | PROTOTYPES: DISABLE |
314 | |
881 | |
315 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
882 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
316 | |
883 | |
317 | void |
884 | void |
318 | info (OpenCL::Platform self, cl_platform_info name) |
885 | info (OpenCL::Platform self, cl_platform_info name) |
319 | PPCODE: |
886 | PPCODE: |
320 | INFO (Platform) |
887 | INFO (Platform) |
|
|
888 | |
|
|
889 | void |
|
|
890 | unload_compiler (OpenCL::Platform self) |
|
|
891 | CODE: |
|
|
892 | #if CL_VERSION_1_2 |
|
|
893 | clUnloadPlatformCompiler (self); |
|
|
894 | #endif |
321 | |
895 | |
322 | #BEGIN:platform |
896 | #BEGIN:platform |
323 | |
897 | |
324 | void |
898 | void |
325 | profile (OpenCL::Platform self) |
899 | profile (OpenCL::Platform self) |
… | |
… | |
351 | list = tmpbuf (sizeof (*list) * count); |
925 | list = tmpbuf (sizeof (*list) * count); |
352 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
926 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
353 | |
927 | |
354 | EXTEND (SP, count); |
928 | EXTEND (SP, count); |
355 | for (i = 0; i < count; ++i) |
929 | for (i = 0; i < count; ++i) |
356 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
930 | PUSH_CLOBJ (stash_device, list [i]); |
357 | |
931 | |
358 | void |
932 | void |
359 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
933 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
360 | PPCODE: |
934 | PPCODE: |
|
|
935 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
936 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
937 | |
|
|
938 | cl_uint device_count; |
|
|
939 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
940 | |
|
|
941 | CONTEXT_NOTIFY_CALLBACK; |
|
|
942 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
943 | XPUSH_CLOBJ_CONTEXT; |
|
|
944 | |
|
|
945 | void |
|
|
946 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
947 | PPCODE: |
|
|
948 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
949 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
950 | |
|
|
951 | CONTEXT_NOTIFY_CALLBACK; |
|
|
952 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
953 | XPUSH_CLOBJ_CONTEXT; |
|
|
954 | |
|
|
955 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
956 | |
|
|
957 | void |
|
|
958 | info (OpenCL::Device self, cl_device_info name) |
|
|
959 | PPCODE: |
|
|
960 | INFO (Device) |
|
|
961 | |
|
|
962 | #if CL_VERSION_1_2 |
|
|
963 | |
|
|
964 | void |
|
|
965 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
966 | PPCODE: |
361 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
967 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
362 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
968 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
363 | |
969 | |
364 | AV *av = (AV *)SvRV (devices); |
970 | properties = SvRV (properties); |
365 | cl_uint num_devices = av_len (av) + 1; |
971 | |
366 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
972 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
973 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
974 | |
367 | int i; |
975 | int i; |
|
|
976 | for (i = 0; i < count; ++i) |
|
|
977 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
368 | |
978 | |
369 | for (i = num_devices; i--; ) |
979 | props [count] = 0; |
370 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
371 | |
980 | |
372 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
981 | cl_uint num_devices; |
373 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
982 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
983 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
984 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
374 | |
985 | |
375 | void |
986 | EXTEND (SP, num_devices); |
376 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
987 | for (i = 0; i < count; ++i) |
377 | PPCODE: |
988 | PUSH_CLOBJ (stash_subdevice, list [i]); |
378 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
379 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
380 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
|
|
381 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
|
|
382 | |
989 | |
383 | MODULE = OpenCL PACKAGE = OpenCL::Device |
990 | #endif |
384 | |
|
|
385 | void |
|
|
386 | info (OpenCL::Device self, cl_device_info name) |
|
|
387 | PPCODE: |
|
|
388 | INFO (Device) |
|
|
389 | |
991 | |
390 | #BEGIN:device |
992 | #BEGIN:device |
391 | |
993 | |
392 | void |
994 | void |
393 | type (OpenCL::Device self) |
995 | type (OpenCL::Device self) |
394 | PPCODE: |
996 | PPCODE: |
395 | cl_device_type value [1]; |
997 | cl_device_type value [1]; |
396 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
998 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
397 | EXTEND (SP, 1); |
999 | EXTEND (SP, 1); |
398 | const int i = 0; |
1000 | const int i = 0; |
399 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1001 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
400 | |
1002 | |
401 | void |
1003 | void |
402 | vendor_id (OpenCL::Device self) |
1004 | vendor_id (OpenCL::Device self) |
403 | ALIAS: |
1005 | ALIAS: |
404 | vendor_id = CL_DEVICE_VENDOR_ID |
1006 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
425 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
1027 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
426 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
1028 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
427 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
1029 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
428 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
1030 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
429 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
1031 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
430 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
1032 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
431 | PPCODE: |
1033 | PPCODE: |
432 | cl_uint value [1]; |
1034 | cl_uint value [1]; |
433 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1035 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
434 | EXTEND (SP, 1); |
1036 | EXTEND (SP, 1); |
435 | const int i = 0; |
1037 | const int i = 0; |
… | |
… | |
549 | PPCODE: |
1151 | PPCODE: |
550 | cl_command_queue_properties value [1]; |
1152 | cl_command_queue_properties value [1]; |
551 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1153 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
552 | EXTEND (SP, 1); |
1154 | EXTEND (SP, 1); |
553 | const int i = 0; |
1155 | const int i = 0; |
554 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1156 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
555 | |
1157 | |
556 | void |
1158 | void |
557 | platform (OpenCL::Device self) |
1159 | platform (OpenCL::Device self) |
558 | PPCODE: |
1160 | PPCODE: |
559 | cl_platform_id value [1]; |
1161 | cl_platform_id value [1]; |
560 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1162 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
561 | EXTEND (SP, 1); |
1163 | EXTEND (SP, 1); |
562 | const int i = 0; |
1164 | const int i = 0; |
563 | { |
1165 | { |
564 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1166 | PUSH_CLOBJ (stash_platform, value [i]); |
565 | } |
1167 | } |
566 | |
1168 | |
567 | void |
1169 | void |
568 | name (OpenCL::Device self) |
1170 | name (OpenCL::Device self) |
569 | ALIAS: |
1171 | ALIAS: |
… | |
… | |
588 | cl_device_id value [1]; |
1190 | cl_device_id value [1]; |
589 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1191 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
590 | EXTEND (SP, 1); |
1192 | EXTEND (SP, 1); |
591 | const int i = 0; |
1193 | const int i = 0; |
592 | { |
1194 | { |
593 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1195 | PUSH_CLOBJ (stash_device, value [i]); |
594 | } |
1196 | } |
595 | |
1197 | |
596 | void |
1198 | void |
597 | partition_types_ext (OpenCL::Device self) |
1199 | partition_types_ext (OpenCL::Device self) |
598 | ALIAS: |
1200 | ALIAS: |
… | |
… | |
609 | for (i = 0; i < n; ++i) |
1211 | for (i = 0; i < n; ++i) |
610 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1212 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
611 | |
1213 | |
612 | #END:device |
1214 | #END:device |
613 | |
1215 | |
|
|
1216 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1217 | |
|
|
1218 | #if CL_VERSION_1_2 |
|
|
1219 | |
|
|
1220 | void |
|
|
1221 | DESTROY (OpenCL::SubDevice self) |
|
|
1222 | CODE: |
|
|
1223 | clReleaseDevice (self); |
|
|
1224 | |
|
|
1225 | #endif |
|
|
1226 | |
614 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1227 | MODULE = OpenCL PACKAGE = OpenCL::Context |
615 | |
1228 | |
616 | void |
1229 | void |
617 | DESTROY (OpenCL::Context context) |
1230 | DESTROY (OpenCL::Context self) |
618 | CODE: |
1231 | CODE: |
619 | clReleaseContext (context); |
1232 | clReleaseContext (self); |
620 | |
1233 | |
621 | void |
1234 | void |
622 | info (OpenCL::Context self, cl_context_info name) |
1235 | info (OpenCL::Context self, cl_context_info name) |
623 | PPCODE: |
1236 | PPCODE: |
624 | INFO (Context) |
1237 | INFO (Context) |
625 | |
1238 | |
626 | void |
1239 | void |
627 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1240 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
628 | PPCODE: |
1241 | PPCODE: |
629 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1242 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
630 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1243 | XPUSH_CLOBJ (stash_queue, queue); |
631 | |
1244 | |
632 | void |
1245 | void |
633 | user_event (OpenCL::Context self) |
1246 | user_event (OpenCL::Context self) |
634 | PPCODE: |
1247 | PPCODE: |
635 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1248 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
636 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1249 | XPUSH_CLOBJ (stash_userevent, ev); |
637 | |
1250 | |
638 | void |
1251 | void |
639 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1252 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
640 | PPCODE: |
1253 | PPCODE: |
641 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1254 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
642 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1255 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
643 | |
1256 | |
644 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1257 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
645 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1258 | XPUSH_CLOBJ (stash_bufferobj, mem); |
646 | |
1259 | |
647 | void |
1260 | void |
648 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1261 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
649 | PPCODE: |
1262 | PPCODE: |
650 | STRLEN len; |
1263 | STRLEN len; |
651 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1264 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
652 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1265 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
653 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1266 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
654 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1267 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
655 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1268 | XPUSH_CLOBJ (stash_bufferobj, mem); |
656 | |
1269 | |
|
|
1270 | #if CL_VERSION_1_2 |
|
|
1271 | |
657 | void |
1272 | void |
658 | 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) |
1273 | 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) |
659 | PPCODE: |
1274 | PPCODE: |
660 | STRLEN len; |
1275 | STRLEN len; |
661 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1276 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
662 | const cl_image_format format = { channel_order, channel_type }; |
1277 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1278 | const cl_image_desc desc = { |
|
|
1279 | type, |
|
|
1280 | width, height, depth, |
|
|
1281 | array_size, row_pitch, slice_pitch, |
|
|
1282 | num_mip_level, num_samples, |
|
|
1283 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
|
|
1284 | }; |
663 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1285 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
664 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1286 | HV *stash = stash_image; |
|
|
1287 | switch (type) |
|
|
1288 | { |
|
|
1289 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1290 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1291 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1292 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1293 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1294 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1295 | } |
|
|
1296 | XPUSH_CLOBJ (stash, mem); |
665 | |
1297 | |
|
|
1298 | #endif |
|
|
1299 | |
666 | void |
1300 | void |
667 | 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) |
1301 | 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) |
668 | PPCODE: |
1302 | PPCODE: |
669 | STRLEN len; |
1303 | STRLEN len; |
670 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1304 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
671 | const cl_image_format format = { channel_order, channel_type }; |
1305 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1306 | #if PREFER_1_1 |
|
|
1307 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1308 | #else |
|
|
1309 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1310 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1311 | #endif |
|
|
1312 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1313 | |
|
|
1314 | void |
|
|
1315 | 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) |
|
|
1316 | PPCODE: |
|
|
1317 | STRLEN len; |
|
|
1318 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1319 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1320 | #if PREFER_1_1 |
672 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1321 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
673 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1322 | #else |
|
|
1323 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1324 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1325 | #endif |
|
|
1326 | XPUSH_CLOBJ (stash_image3d, mem); |
674 | |
1327 | |
675 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1328 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
676 | |
1329 | |
677 | void |
1330 | void |
678 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
1331 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
679 | PPCODE: |
1332 | PPCODE: |
680 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
1333 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
681 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1334 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1335 | |
|
|
1336 | void |
|
|
1337 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1338 | PPCODE: |
|
|
1339 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1340 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1341 | |
|
|
1342 | #if CL_VERSION_1_2 |
|
|
1343 | |
|
|
1344 | void |
|
|
1345 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1346 | ALIAS: |
|
|
1347 | PPCODE: |
|
|
1348 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1349 | cl_gl_object_type type; |
|
|
1350 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1351 | HV *stash = stash_memory; |
|
|
1352 | switch (type) |
|
|
1353 | { |
|
|
1354 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1355 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1356 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1357 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1358 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1359 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1360 | } |
|
|
1361 | XPUSH_CLOBJ (stash, mem); |
|
|
1362 | |
|
|
1363 | #endif |
682 | |
1364 | |
683 | void |
1365 | void |
684 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1366 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
685 | PPCODE: |
1367 | PPCODE: |
|
|
1368 | #if PREFER_1_1 |
686 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
1369 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
687 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1370 | #else |
|
|
1371 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1372 | #endif |
|
|
1373 | XPUSH_CLOBJ (stash_image2d, mem); |
688 | |
1374 | |
689 | void |
1375 | void |
690 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1376 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
691 | PPCODE: |
1377 | PPCODE: |
|
|
1378 | #if PREFER_1_1 |
692 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
1379 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
693 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1380 | #else |
694 | |
1381 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
695 | void |
1382 | #endif |
696 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
1383 | XPUSH_CLOBJ (stash_image3d, mem); |
697 | PPCODE: |
|
|
698 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
699 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
700 | |
1384 | |
701 | #endif |
1385 | #endif |
702 | |
1386 | |
703 | void |
1387 | void |
704 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1388 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
724 | |
1408 | |
725 | void |
1409 | void |
726 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1410 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
727 | PPCODE: |
1411 | PPCODE: |
728 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1412 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
729 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1413 | XPUSH_CLOBJ (stash_sampler, sampler); |
730 | |
1414 | |
731 | void |
1415 | void |
732 | program_with_source (OpenCL::Context self, SV *program) |
1416 | program_with_source (OpenCL::Context self, SV *program) |
733 | PPCODE: |
1417 | PPCODE: |
734 | STRLEN len; |
1418 | STRLEN len; |
735 | size_t len2; |
1419 | size_t len2; |
736 | const char *ptr = SvPVbyte (program, len); |
1420 | const char *ptr = SvPVbyte (program, len); |
737 | |
1421 | |
738 | len2 = len; |
1422 | len2 = len; |
739 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1423 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
740 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1424 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1425 | |
|
|
1426 | void |
|
|
1427 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1428 | PPCODE: |
|
|
1429 | cl_uint device_count; |
|
|
1430 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1431 | |
|
|
1432 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1433 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1434 | |
|
|
1435 | binaries = SvRV (binaries); |
|
|
1436 | |
|
|
1437 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1438 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1439 | |
|
|
1440 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1441 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1442 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1443 | |
|
|
1444 | int i; |
|
|
1445 | for (i = 0; i < device_count; ++i) |
|
|
1446 | { |
|
|
1447 | STRLEN len; |
|
|
1448 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1449 | length_list [i] = len; |
|
|
1450 | } |
|
|
1451 | |
|
|
1452 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1453 | length_list, binary_list, |
|
|
1454 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1455 | |
|
|
1456 | EXTEND (SP, 2); |
|
|
1457 | PUSH_CLOBJ (stash_program, prog); |
|
|
1458 | |
|
|
1459 | if (GIMME_V == G_ARRAY) |
|
|
1460 | { |
|
|
1461 | AV *av = newAV (); |
|
|
1462 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1463 | |
|
|
1464 | for (i = device_count; i--; ) |
|
|
1465 | av_store (av, i, newSViv (status_list [i])); |
|
|
1466 | } |
|
|
1467 | |
|
|
1468 | #if CL_VERSION_1_2 |
|
|
1469 | |
|
|
1470 | void |
|
|
1471 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1472 | PPCODE: |
|
|
1473 | cl_uint device_count; |
|
|
1474 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1475 | |
|
|
1476 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1477 | |
|
|
1478 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1479 | |
|
|
1480 | void |
|
|
1481 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1482 | CODE: |
|
|
1483 | cl_uint device_count = 0; |
|
|
1484 | cl_device_id *device_list = 0; |
|
|
1485 | |
|
|
1486 | if (SvOK (devices)) |
|
|
1487 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1488 | |
|
|
1489 | cl_uint program_count; |
|
|
1490 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1491 | |
|
|
1492 | void *user_data; |
|
|
1493 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1494 | |
|
|
1495 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1496 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1497 | |
|
|
1498 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1499 | |
|
|
1500 | #endif |
741 | |
1501 | |
742 | #BEGIN:context |
1502 | #BEGIN:context |
743 | |
1503 | |
744 | void |
1504 | void |
745 | reference_count (OpenCL::Context self) |
1505 | reference_count (OpenCL::Context self) |
… | |
… | |
762 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1522 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
763 | int i, n = size / sizeof (*value); |
1523 | int i, n = size / sizeof (*value); |
764 | EXTEND (SP, n); |
1524 | EXTEND (SP, n); |
765 | for (i = 0; i < n; ++i) |
1525 | for (i = 0; i < n; ++i) |
766 | { |
1526 | { |
767 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1527 | PUSH_CLOBJ (stash_device, value [i]); |
768 | } |
1528 | } |
769 | |
1529 | |
770 | void |
1530 | void |
771 | properties (OpenCL::Context self) |
1531 | properties (OpenCL::Context self) |
772 | PPCODE: |
1532 | PPCODE: |
… | |
… | |
787 | DESTROY (OpenCL::Queue self) |
1547 | DESTROY (OpenCL::Queue self) |
788 | CODE: |
1548 | CODE: |
789 | clReleaseCommandQueue (self); |
1549 | clReleaseCommandQueue (self); |
790 | |
1550 | |
791 | void |
1551 | void |
792 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1552 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1553 | ALIAS: |
|
|
1554 | enqueue_read_buffer = 0 |
793 | PPCODE: |
1555 | PPCODE: |
794 | cl_event ev = 0; |
|
|
795 | EVENT_LIST (6, items - 6); |
1556 | EVENT_LIST (6); |
796 | |
1557 | |
797 | SvUPGRADE (data, SVt_PV); |
1558 | SvUPGRADE (data, SVt_PV); |
798 | SvGROW (data, len); |
1559 | SvGROW (data, len); |
799 | SvPOK_only (data); |
1560 | SvPOK_only (data); |
800 | SvCUR_set (data, len); |
1561 | SvCUR_set (data, len); |
|
|
1562 | |
|
|
1563 | cl_event ev = 0; |
801 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1564 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
802 | |
1565 | |
803 | if (ev) |
1566 | if (ev) |
804 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1567 | XPUSH_CLOBJ (stash_event, ev); |
805 | |
1568 | |
806 | void |
1569 | void |
807 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1570 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1571 | ALIAS: |
|
|
1572 | enqueue_write_buffer = 0 |
808 | PPCODE: |
1573 | PPCODE: |
809 | cl_event ev = 0; |
1574 | EVENT_LIST (5); |
|
|
1575 | |
810 | STRLEN len; |
1576 | STRLEN len; |
811 | char *ptr = SvPVbyte (data, len); |
1577 | char *ptr = SvPVbyte (data, len); |
812 | EVENT_LIST (5, items - 5); |
|
|
813 | |
1578 | |
|
|
1579 | cl_event ev = 0; |
814 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1580 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
815 | |
1581 | |
816 | if (ev) |
1582 | if (ev) |
817 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1583 | XPUSH_CLOBJ (stash_event, ev); |
818 | |
1584 | |
|
|
1585 | #if CL_VERSION_1_2 |
|
|
1586 | |
819 | void |
1587 | void |
820 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1588 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1589 | ALIAS: |
|
|
1590 | enqueue_fill_buffer = 0 |
821 | PPCODE: |
1591 | PPCODE: |
|
|
1592 | EVENT_LIST (5); |
|
|
1593 | |
|
|
1594 | STRLEN len; |
|
|
1595 | char *ptr = SvPVbyte (data, len); |
|
|
1596 | |
822 | cl_event ev = 0; |
1597 | cl_event ev = 0; |
|
|
1598 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1599 | |
|
|
1600 | if (ev) |
|
|
1601 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1602 | |
|
|
1603 | void |
|
|
1604 | 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, ...) |
|
|
1605 | ALIAS: |
|
|
1606 | enqueue_fill_image = 0 |
|
|
1607 | PPCODE: |
|
|
1608 | EVENT_LIST (12); |
|
|
1609 | |
|
|
1610 | const size_t origin [3] = { x, y, z }; |
|
|
1611 | const size_t region [3] = { width, height, depth }; |
|
|
1612 | |
|
|
1613 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1614 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1615 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1616 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1617 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1618 | cl_image_format format; |
|
|
1619 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1620 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1621 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1622 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1623 | |
|
|
1624 | cl_event ev = 0; |
|
|
1625 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1626 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1627 | |
|
|
1628 | if (ev) |
|
|
1629 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1630 | |
|
|
1631 | #endif |
|
|
1632 | |
|
|
1633 | void |
|
|
1634 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1635 | ALIAS: |
|
|
1636 | enqueue_copy_buffer = 0 |
|
|
1637 | PPCODE: |
823 | EVENT_LIST (6, items - 6); |
1638 | EVENT_LIST (6); |
824 | |
1639 | |
|
|
1640 | cl_event ev = 0; |
825 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1641 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
826 | |
1642 | |
827 | if (ev) |
1643 | if (ev) |
828 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1644 | XPUSH_CLOBJ (stash_event, ev); |
829 | |
1645 | |
830 | void |
1646 | void |
831 | enqueue_read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1647 | 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, ...) |
|
|
1648 | ALIAS: |
|
|
1649 | enqueue_read_buffer_rect = 0 |
832 | PPCODE: |
1650 | PPCODE: |
833 | cl_event ev = 0; |
1651 | EVENT_LIST (17); |
|
|
1652 | |
834 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1653 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
835 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1654 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
836 | const size_t region[3] = { width, height, depth }; |
1655 | const size_t region[3] = { width, height, depth }; |
837 | EVENT_LIST (17, items - 17); |
|
|
838 | |
1656 | |
839 | if (!buf_row_pitch) |
1657 | if (!buf_row_pitch) |
840 | buf_row_pitch = region [0]; |
1658 | buf_row_pitch = region [0]; |
841 | |
1659 | |
842 | if (!buf_slice_pitch) |
1660 | if (!buf_slice_pitch) |
… | |
… | |
852 | |
1670 | |
853 | SvUPGRADE (data, SVt_PV); |
1671 | SvUPGRADE (data, SVt_PV); |
854 | SvGROW (data, len); |
1672 | SvGROW (data, len); |
855 | SvPOK_only (data); |
1673 | SvPOK_only (data); |
856 | SvCUR_set (data, len); |
1674 | SvCUR_set (data, len); |
|
|
1675 | |
|
|
1676 | cl_event ev = 0; |
857 | 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)); |
1677 | 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)); |
858 | |
1678 | |
859 | if (ev) |
1679 | if (ev) |
860 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1680 | XPUSH_CLOBJ (stash_event, ev); |
861 | |
1681 | |
862 | void |
1682 | void |
863 | enqueue_write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1683 | 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, ...) |
|
|
1684 | ALIAS: |
|
|
1685 | enqueue_write_buffer_rect = 0 |
864 | PPCODE: |
1686 | PPCODE: |
865 | cl_event ev = 0; |
1687 | EVENT_LIST (17); |
|
|
1688 | |
866 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1689 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
867 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1690 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
868 | const size_t region[3] = { width, height, depth }; |
1691 | const size_t region[3] = { width, height, depth }; |
869 | STRLEN len; |
1692 | STRLEN len; |
870 | char *ptr = SvPVbyte (data, len); |
1693 | char *ptr = SvPVbyte (data, len); |
871 | EVENT_LIST (17, items - 17); |
|
|
872 | |
1694 | |
873 | if (!buf_row_pitch) |
1695 | if (!buf_row_pitch) |
874 | buf_row_pitch = region [0]; |
1696 | buf_row_pitch = region [0]; |
875 | |
1697 | |
876 | if (!buf_slice_pitch) |
1698 | if (!buf_slice_pitch) |
… | |
… | |
885 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1707 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
886 | |
1708 | |
887 | if (len < min_len) |
1709 | if (len < min_len) |
888 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1710 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
889 | |
1711 | |
|
|
1712 | cl_event ev = 0; |
890 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1713 | 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)); |
891 | |
1714 | |
892 | if (ev) |
1715 | if (ev) |
893 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1716 | XPUSH_CLOBJ (stash_event, ev); |
894 | |
1717 | |
895 | void |
1718 | void |
896 | enqueue_copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
1719 | 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, ...) |
|
|
1720 | ALIAS: |
|
|
1721 | enqueue_copy_buffer_rect = 0 |
897 | PPCODE: |
1722 | PPCODE: |
898 | cl_event ev = 0; |
1723 | EVENT_LIST (16); |
|
|
1724 | |
899 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1725 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
900 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1726 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
901 | const size_t region[3] = { width, height, depth }; |
1727 | const size_t region[3] = { width, height, depth }; |
902 | EVENT_LIST (16, items - 16); |
|
|
903 | |
1728 | |
|
|
1729 | cl_event ev = 0; |
904 | NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1730 | 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)); |
905 | |
1731 | |
906 | if (ev) |
1732 | if (ev) |
907 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1733 | XPUSH_CLOBJ (stash_event, ev); |
908 | |
1734 | |
909 | void |
1735 | void |
910 | enqueue_read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1736 | 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, ...) |
|
|
1737 | ALIAS: |
|
|
1738 | enqueue_read_image = 0 |
911 | PPCODE: |
1739 | PPCODE: |
912 | cl_event ev = 0; |
1740 | EVENT_LIST (12); |
|
|
1741 | |
913 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1742 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
914 | const size_t region[3] = { width, height, depth }; |
1743 | const size_t region[3] = { width, height, depth }; |
915 | EVENT_LIST (12, items - 12); |
|
|
916 | |
1744 | |
917 | if (!row_pitch) |
1745 | if (!row_pitch) |
918 | row_pitch = img_row_pitch (src); |
1746 | row_pitch = img_row_pitch (src); |
919 | |
1747 | |
920 | if (depth > 1 && !slice_pitch) |
1748 | if (depth > 1 && !slice_pitch) |
… | |
… | |
924 | |
1752 | |
925 | SvUPGRADE (data, SVt_PV); |
1753 | SvUPGRADE (data, SVt_PV); |
926 | SvGROW (data, len); |
1754 | SvGROW (data, len); |
927 | SvPOK_only (data); |
1755 | SvPOK_only (data); |
928 | SvCUR_set (data, len); |
1756 | SvCUR_set (data, len); |
|
|
1757 | |
|
|
1758 | cl_event ev = 0; |
929 | NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1759 | 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)); |
930 | |
1760 | |
931 | if (ev) |
1761 | if (ev) |
932 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1762 | XPUSH_CLOBJ (stash_event, ev); |
933 | |
1763 | |
934 | void |
1764 | void |
935 | enqueue_write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1765 | 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, ...) |
|
|
1766 | ALIAS: |
|
|
1767 | enqueue_write_image = 0 |
936 | PPCODE: |
1768 | PPCODE: |
937 | cl_event ev = 0; |
1769 | EVENT_LIST (12); |
|
|
1770 | |
938 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1771 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
939 | const size_t region[3] = { width, height, depth }; |
1772 | const size_t region[3] = { width, height, depth }; |
940 | STRLEN len; |
1773 | STRLEN len; |
941 | char *ptr = SvPVbyte (data, len); |
1774 | char *ptr = SvPVbyte (data, len); |
942 | EVENT_LIST (12, items - 12); |
|
|
943 | |
1775 | |
944 | if (!row_pitch) |
1776 | if (!row_pitch) |
945 | row_pitch = img_row_pitch (dst); |
1777 | row_pitch = img_row_pitch (dst); |
946 | |
1778 | |
947 | if (depth > 1 && !slice_pitch) |
1779 | if (depth > 1 && !slice_pitch) |
… | |
… | |
950 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1782 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
951 | |
1783 | |
952 | if (len < min_len) |
1784 | if (len < min_len) |
953 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1785 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
954 | |
1786 | |
|
|
1787 | cl_event ev = 0; |
955 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1788 | 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)); |
956 | |
1789 | |
957 | if (ev) |
1790 | if (ev) |
958 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1791 | XPUSH_CLOBJ (stash_event, ev); |
959 | |
1792 | |
960 | void |
1793 | void |
961 | enqueue_copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
1794 | 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, ...) |
|
|
1795 | ALIAS: |
|
|
1796 | enqueue_copy_image = 0 |
962 | PPCODE: |
1797 | PPCODE: |
963 | cl_event ev = 0; |
1798 | EVENT_LIST (12); |
|
|
1799 | |
964 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1800 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
965 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1801 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
966 | const size_t region[3] = { width, height, depth }; |
1802 | const size_t region[3] = { width, height, depth }; |
967 | EVENT_LIST (12, items - 12); |
|
|
968 | |
1803 | |
|
|
1804 | cl_event ev = 0; |
969 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1805 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
970 | |
1806 | |
971 | if (ev) |
1807 | if (ev) |
972 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1808 | XPUSH_CLOBJ (stash_event, ev); |
973 | |
1809 | |
974 | void |
1810 | void |
975 | enqueue_copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
1811 | 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, ...) |
|
|
1812 | ALIAS: |
|
|
1813 | enqueue_copy_image_to_buffer = 0 |
976 | PPCODE: |
1814 | PPCODE: |
|
|
1815 | EVENT_LIST (10); |
|
|
1816 | |
|
|
1817 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1818 | const size_t region [3] = { width, height, depth }; |
|
|
1819 | |
977 | cl_event ev = 0; |
1820 | cl_event ev = 0; |
|
|
1821 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1822 | |
|
|
1823 | if (ev) |
|
|
1824 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1825 | |
|
|
1826 | void |
|
|
1827 | 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, ...) |
|
|
1828 | ALIAS: |
|
|
1829 | enqueue_copy_buffer_to_image = 0 |
|
|
1830 | PPCODE: |
|
|
1831 | EVENT_LIST (10); |
|
|
1832 | |
978 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1833 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1834 | const size_t region [3] = { width, height, depth }; |
|
|
1835 | |
|
|
1836 | cl_event ev = 0; |
|
|
1837 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1838 | |
|
|
1839 | if (ev) |
|
|
1840 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1841 | |
|
|
1842 | void |
|
|
1843 | 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, ...) |
|
|
1844 | ALIAS: |
|
|
1845 | enqueue_map_buffer = 0 |
|
|
1846 | PPCODE: |
|
|
1847 | EVENT_LIST (6); |
|
|
1848 | |
|
|
1849 | size_t cb = SvIV (cb_); |
|
|
1850 | |
|
|
1851 | if (!SvOK (cb_)) |
|
|
1852 | { |
|
|
1853 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1854 | cb -= offset; |
|
|
1855 | } |
|
|
1856 | |
|
|
1857 | cl_event ev; |
|
|
1858 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1859 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1860 | |
|
|
1861 | void |
|
|
1862 | 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, ...) |
|
|
1863 | ALIAS: |
|
|
1864 | enqueue_map_image = 0 |
|
|
1865 | PPCODE: |
|
|
1866 | size_t width = SvIV (width_); |
|
|
1867 | if (!SvOK (width_)) |
|
|
1868 | { |
|
|
1869 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1870 | width -= x; |
|
|
1871 | } |
|
|
1872 | |
|
|
1873 | size_t height = SvIV (width_); |
|
|
1874 | if (!SvOK (height_)) |
|
|
1875 | { |
|
|
1876 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1877 | height -= y; |
|
|
1878 | } |
|
|
1879 | |
|
|
1880 | size_t depth = SvIV (width_); |
|
|
1881 | if (!SvOK (depth_)) |
|
|
1882 | { |
|
|
1883 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1884 | depth -= z; |
|
|
1885 | |
|
|
1886 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1887 | if (!depth) |
|
|
1888 | depth = 1; |
|
|
1889 | } |
|
|
1890 | |
|
|
1891 | const size_t origin[3] = { x, y, z }; |
979 | const size_t region[3] = { width, height, depth }; |
1892 | const size_t region[3] = { width, height, depth }; |
|
|
1893 | size_t row_pitch, slice_pitch; |
980 | EVENT_LIST (10, items - 10); |
1894 | EVENT_LIST (10); |
981 | |
1895 | |
|
|
1896 | cl_event ev; |
|
|
1897 | 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)); |
|
|
1898 | |
|
|
1899 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1900 | : row_pitch ? row_pitch * region [1] |
|
|
1901 | : region [0]; |
|
|
1902 | |
|
|
1903 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1904 | |
|
|
1905 | void |
|
|
1906 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1907 | PPCODE: |
|
|
1908 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2); |
|
|
1909 | if (GIMME_V != G_VOID) |
|
|
1910 | { |
|
|
1911 | clRetainEvent (mapped->event); |
|
|
1912 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1913 | } |
|
|
1914 | |
|
|
1915 | void |
|
|
1916 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1917 | ALIAS: |
|
|
1918 | enqueue_task = 0 |
|
|
1919 | PPCODE: |
|
|
1920 | EVENT_LIST (2); |
|
|
1921 | |
|
|
1922 | cl_event ev = 0; |
982 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1923 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
983 | |
1924 | |
984 | if (ev) |
1925 | if (ev) |
985 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1926 | XPUSH_CLOBJ (stash_event, ev); |
986 | |
1927 | |
987 | void |
1928 | void |
988 | enqueue_copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
989 | PPCODE: |
|
|
990 | cl_event ev = 0; |
|
|
991 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
992 | const size_t region[3] = { width, height, depth }; |
|
|
993 | EVENT_LIST (10, items - 10); |
|
|
994 | |
|
|
995 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
996 | |
|
|
997 | if (ev) |
|
|
998 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
999 | |
|
|
1000 | void |
|
|
1001 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1002 | PPCODE: |
|
|
1003 | cl_event ev = 0; |
|
|
1004 | EVENT_LIST (2, items - 2); |
|
|
1005 | |
|
|
1006 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1007 | |
|
|
1008 | if (ev) |
|
|
1009 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1010 | |
|
|
1011 | void |
|
|
1012 | enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1929 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1930 | ALIAS: |
|
|
1931 | enqueue_nd_range_kernel = 0 |
1013 | PPCODE: |
1932 | PPCODE: |
1014 | cl_event ev = 0; |
1933 | EVENT_LIST (5); |
|
|
1934 | |
1015 | size_t *gwo = 0, *gws, *lws = 0; |
1935 | size_t *gwo = 0, *gws, *lws = 0; |
1016 | int gws_len; |
1936 | int gws_len; |
1017 | size_t *lists; |
1937 | size_t *lists; |
1018 | int i; |
1938 | int i; |
1019 | EVENT_LIST (5, items - 5); |
|
|
1020 | |
1939 | |
1021 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1940 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1022 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1941 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1023 | |
1942 | |
1024 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1943 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1025 | |
1944 | |
1026 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1945 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1027 | |
1946 | |
1028 | gws = lists + gws_len * 0; |
1947 | gws = lists + gws_len * 0; |
1029 | for (i = 0; i < gws_len; ++i) |
1948 | for (i = 0; i < gws_len; ++i) |
|
|
1949 | { |
1030 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1950 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1951 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1952 | if (!gws [i]) |
|
|
1953 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1954 | } |
1031 | |
1955 | |
1032 | if (SvOK (global_work_offset)) |
1956 | if (SvOK (global_work_offset)) |
1033 | { |
1957 | { |
1034 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1958 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1035 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1959 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1042 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1966 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1043 | } |
1967 | } |
1044 | |
1968 | |
1045 | if (SvOK (local_work_size)) |
1969 | if (SvOK (local_work_size)) |
1046 | { |
1970 | { |
1047 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1971 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1048 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1972 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1049 | |
1973 | |
1050 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1974 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1051 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1975 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1052 | |
1976 | |
1053 | lws = lists + gws_len * 2; |
1977 | lws = lists + gws_len * 2; |
1054 | for (i = 0; i < gws_len; ++i) |
1978 | for (i = 0; i < gws_len; ++i) |
|
|
1979 | { |
1055 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
1980 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1981 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1982 | if (!lws [i]) |
|
|
1983 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1984 | } |
1056 | } |
1985 | } |
1057 | |
1986 | |
|
|
1987 | cl_event ev = 0; |
1058 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1988 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1059 | |
1989 | |
1060 | if (ev) |
1990 | if (ev) |
1061 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1991 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1992 | |
|
|
1993 | #if CL_VERSION_1_2 |
|
|
1994 | |
|
|
1995 | void |
|
|
1996 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
1997 | ALIAS: |
|
|
1998 | enqueue_migrate_mem_objects = 0 |
|
|
1999 | PPCODE: |
|
|
2000 | EVENT_LIST (3); |
|
|
2001 | |
|
|
2002 | cl_uint obj_count; |
|
|
2003 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2004 | |
|
|
2005 | cl_event ev = 0; |
|
|
2006 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2007 | |
|
|
2008 | if (ev) |
|
|
2009 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2010 | |
|
|
2011 | #endif |
1062 | |
2012 | |
1063 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2013 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1064 | |
2014 | |
1065 | void |
2015 | void |
1066 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2016 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1067 | ALIAS: |
2017 | ALIAS: |
|
|
2018 | release_gl_objects = 1 |
|
|
2019 | enqueue_acquire_gl_objects = 0 |
1068 | enqueue_release_gl_objects = 1 |
2020 | enqueue_release_gl_objects = 1 |
1069 | CODE: |
2021 | PPCODE: |
1070 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
2022 | EVENT_LIST (2); |
1071 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
2023 | |
|
|
2024 | cl_uint obj_count; |
|
|
2025 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
1072 | |
2026 | |
1073 | cl_event ev = 0; |
2027 | cl_event ev = 0; |
1074 | EVENT_LIST (2, items - 2); |
|
|
1075 | AV *av = (AV *)SvRV (objects); |
|
|
1076 | cl_uint num_objects = av_len (av) + 1; |
|
|
1077 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1078 | int i; |
|
|
1079 | |
|
|
1080 | for (i = num_objects; i--; ) |
|
|
1081 | object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1082 | |
2028 | |
1083 | if (ix) |
2029 | if (ix) |
1084 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2030 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1085 | else |
2031 | else |
1086 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2032 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1087 | |
2033 | |
1088 | if (ev) |
2034 | if (ev) |
1089 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2035 | XPUSH_CLOBJ (stash_event, ev); |
1090 | |
2036 | |
1091 | #endif |
2037 | #endif |
1092 | |
2038 | |
1093 | void |
2039 | void |
1094 | enqueue_marker (OpenCL::Queue self) |
|
|
1095 | PPCODE: |
|
|
1096 | cl_event ev; |
|
|
1097 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1098 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1099 | |
|
|
1100 | void |
|
|
1101 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
2040 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2041 | ALIAS: |
|
|
2042 | enqueue_wait_for_events = 0 |
1102 | CODE: |
2043 | CODE: |
1103 | EVENT_LIST (1, items - 1); |
2044 | EVENT_LIST (1); |
|
|
2045 | #if PREFER_1_1 |
1104 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
2046 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2047 | #else |
|
|
2048 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2049 | #endif |
1105 | |
2050 | |
1106 | void |
2051 | void |
1107 | enqueue_barrier (OpenCL::Queue self) |
2052 | marker (OpenCL::Queue self, ...) |
|
|
2053 | ALIAS: |
|
|
2054 | enqueue_marker = 0 |
1108 | CODE: |
2055 | PPCODE: |
|
|
2056 | EVENT_LIST (1); |
|
|
2057 | cl_event ev = 0; |
|
|
2058 | #if PREFER_1_1 |
|
|
2059 | if (!event_list_count) |
|
|
2060 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2061 | else |
|
|
2062 | #if CL_VERSION_1_2 |
|
|
2063 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2064 | #else |
|
|
2065 | { |
|
|
2066 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2067 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2068 | } |
|
|
2069 | #endif |
|
|
2070 | #else |
|
|
2071 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2072 | #endif |
|
|
2073 | if (ev) |
|
|
2074 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2075 | |
|
|
2076 | void |
|
|
2077 | barrier (OpenCL::Queue self, ...) |
|
|
2078 | ALIAS: |
|
|
2079 | enqueue_barrier = 0 |
|
|
2080 | PPCODE: |
|
|
2081 | EVENT_LIST (1); |
|
|
2082 | cl_event ev = 0; |
|
|
2083 | #if PREFER_1_1 |
|
|
2084 | if (!event_list_count && GIMME_V == G_VOID) |
1109 | NEED_SUCCESS (EnqueueBarrier, (self)); |
2085 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2086 | else |
|
|
2087 | #if CL_VERSION_1_2 |
|
|
2088 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2089 | #else |
|
|
2090 | { |
|
|
2091 | if (event_list_count) |
|
|
2092 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2093 | |
|
|
2094 | if (GIMME_V != G_VOID) |
|
|
2095 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2096 | } |
|
|
2097 | #endif |
|
|
2098 | #else |
|
|
2099 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2100 | #endif |
|
|
2101 | if (ev) |
|
|
2102 | XPUSH_CLOBJ (stash_event, ev); |
1110 | |
2103 | |
1111 | void |
2104 | void |
1112 | flush (OpenCL::Queue self) |
2105 | flush (OpenCL::Queue self) |
1113 | CODE: |
2106 | CODE: |
1114 | NEED_SUCCESS (Flush, (self)); |
2107 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1132 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
2125 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1133 | EXTEND (SP, 1); |
2126 | EXTEND (SP, 1); |
1134 | const int i = 0; |
2127 | const int i = 0; |
1135 | { |
2128 | { |
1136 | NEED_SUCCESS (RetainContext, (value [i])); |
2129 | NEED_SUCCESS (RetainContext, (value [i])); |
1137 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2130 | PUSH_CLOBJ (stash_context, value [i]); |
1138 | } |
2131 | } |
1139 | |
2132 | |
1140 | void |
2133 | void |
1141 | device (OpenCL::Queue self) |
2134 | device (OpenCL::Queue self) |
1142 | PPCODE: |
2135 | PPCODE: |
1143 | cl_device_id value [1]; |
2136 | cl_device_id value [1]; |
1144 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
2137 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1145 | EXTEND (SP, 1); |
2138 | EXTEND (SP, 1); |
1146 | const int i = 0; |
2139 | const int i = 0; |
1147 | { |
2140 | { |
1148 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2141 | PUSH_CLOBJ (stash_device, value [i]); |
1149 | } |
2142 | } |
1150 | |
2143 | |
1151 | void |
2144 | void |
1152 | reference_count (OpenCL::Queue self) |
2145 | reference_count (OpenCL::Queue self) |
1153 | PPCODE: |
2146 | PPCODE: |
… | |
… | |
1162 | PPCODE: |
2155 | PPCODE: |
1163 | cl_command_queue_properties value [1]; |
2156 | cl_command_queue_properties value [1]; |
1164 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
2157 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1165 | EXTEND (SP, 1); |
2158 | EXTEND (SP, 1); |
1166 | const int i = 0; |
2159 | const int i = 0; |
1167 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2160 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1168 | |
2161 | |
1169 | #END:command_queue |
2162 | #END:command_queue |
1170 | |
2163 | |
1171 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2164 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1172 | |
2165 | |
… | |
… | |
1187 | PPCODE: |
2180 | PPCODE: |
1188 | cl_mem_object_type value [1]; |
2181 | cl_mem_object_type value [1]; |
1189 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
2182 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1190 | EXTEND (SP, 1); |
2183 | EXTEND (SP, 1); |
1191 | const int i = 0; |
2184 | const int i = 0; |
1192 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2185 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1193 | |
2186 | |
1194 | void |
2187 | void |
1195 | flags (OpenCL::Memory self) |
2188 | flags (OpenCL::Memory self) |
1196 | PPCODE: |
2189 | PPCODE: |
1197 | cl_mem_flags value [1]; |
2190 | cl_mem_flags value [1]; |
1198 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
2191 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1199 | EXTEND (SP, 1); |
2192 | EXTEND (SP, 1); |
1200 | const int i = 0; |
2193 | const int i = 0; |
1201 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2194 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1202 | |
2195 | |
1203 | void |
2196 | void |
1204 | size (OpenCL::Memory self) |
2197 | size (OpenCL::Memory self) |
1205 | ALIAS: |
2198 | ALIAS: |
1206 | size = CL_MEM_SIZE |
2199 | size = CL_MEM_SIZE |
… | |
… | |
1240 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
2233 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1241 | EXTEND (SP, 1); |
2234 | EXTEND (SP, 1); |
1242 | const int i = 0; |
2235 | const int i = 0; |
1243 | { |
2236 | { |
1244 | NEED_SUCCESS (RetainContext, (value [i])); |
2237 | NEED_SUCCESS (RetainContext, (value [i])); |
1245 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2238 | PUSH_CLOBJ (stash_context, value [i]); |
1246 | } |
2239 | } |
1247 | |
2240 | |
1248 | void |
2241 | void |
1249 | associated_memobject (OpenCL::Memory self) |
2242 | associated_memobject (OpenCL::Memory self) |
1250 | PPCODE: |
2243 | PPCODE: |
… | |
… | |
1252 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
2245 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1253 | EXTEND (SP, 1); |
2246 | EXTEND (SP, 1); |
1254 | const int i = 0; |
2247 | const int i = 0; |
1255 | { |
2248 | { |
1256 | NEED_SUCCESS (RetainMemObject, (value [i])); |
2249 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1257 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
2250 | PUSH_CLOBJ (stash_memory, value [i]); |
1258 | } |
2251 | } |
1259 | |
2252 | |
1260 | #END:mem |
2253 | #END:mem |
1261 | |
2254 | |
1262 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2255 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
… | |
… | |
1264 | void |
2257 | void |
1265 | gl_object_info (OpenCL::Memory self) |
2258 | gl_object_info (OpenCL::Memory self) |
1266 | PPCODE: |
2259 | PPCODE: |
1267 | cl_gl_object_type type; |
2260 | cl_gl_object_type type; |
1268 | cl_GLuint name; |
2261 | cl_GLuint name; |
1269 | NEED_SUCCESS (clGetGLObjectInfo, (self, &type, &name)); |
2262 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
1270 | EXTEND (SP, 2); |
2263 | EXTEND (SP, 2); |
1271 | PUSHs (sv_2mortal (newSVuv (type))); |
2264 | PUSHs (sv_2mortal (newSVuv (type))); |
1272 | PUSHs (sv_2mortal (newSVuv (name))); |
2265 | PUSHs (sv_2mortal (newSVuv (name))); |
1273 | |
2266 | |
1274 | #endif |
2267 | #endif |
… | |
… | |
1282 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
2275 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1283 | |
2276 | |
1284 | cl_buffer_region crdata = { origin, size }; |
2277 | cl_buffer_region crdata = { origin, size }; |
1285 | |
2278 | |
1286 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2279 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1287 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2280 | XPUSH_CLOBJ (stash_buffer, mem); |
1288 | |
2281 | |
1289 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2282 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1290 | |
2283 | |
1291 | void |
2284 | void |
1292 | image_info (OpenCL::Image self, cl_image_info name) |
2285 | image_info (OpenCL::Image self, cl_image_info name) |
1293 | PPCODE: |
2286 | PPCODE: |
1294 | INFO (Image) |
2287 | INFO (Image) |
|
|
2288 | |
|
|
2289 | void |
|
|
2290 | format (OpenCL::Image self) |
|
|
2291 | PPCODE: |
|
|
2292 | cl_image_format format; |
|
|
2293 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2294 | EXTEND (SP, 2); |
|
|
2295 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2296 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1295 | |
2297 | |
1296 | #BEGIN:image |
2298 | #BEGIN:image |
1297 | |
2299 | |
1298 | void |
2300 | void |
1299 | element_size (OpenCL::Image self) |
2301 | element_size (OpenCL::Image self) |
… | |
… | |
1319 | |
2321 | |
1320 | void |
2322 | void |
1321 | target (OpenCL::Image self) |
2323 | target (OpenCL::Image self) |
1322 | PPCODE: |
2324 | PPCODE: |
1323 | cl_GLenum value [1]; |
2325 | cl_GLenum value [1]; |
1324 | NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
2326 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
1325 | EXTEND (SP, 1); |
2327 | EXTEND (SP, 1); |
1326 | const int i = 0; |
2328 | const int i = 0; |
1327 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2329 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1328 | |
2330 | |
1329 | void |
2331 | void |
1330 | gl_mipmap_level (OpenCL::Image self) |
2332 | gl_mipmap_level (OpenCL::Image self) |
1331 | PPCODE: |
2333 | PPCODE: |
1332 | cl_GLint value [1]; |
2334 | cl_GLint value [1]; |
1333 | NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
2335 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
1334 | EXTEND (SP, 1); |
2336 | EXTEND (SP, 1); |
1335 | const int i = 0; |
2337 | const int i = 0; |
1336 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2338 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1337 | |
2339 | |
1338 | #END:gl_texture |
2340 | #END:gl_texture |
… | |
… | |
1369 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2371 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1370 | EXTEND (SP, 1); |
2372 | EXTEND (SP, 1); |
1371 | const int i = 0; |
2373 | const int i = 0; |
1372 | { |
2374 | { |
1373 | NEED_SUCCESS (RetainContext, (value [i])); |
2375 | NEED_SUCCESS (RetainContext, (value [i])); |
1374 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2376 | PUSH_CLOBJ (stash_context, value [i]); |
1375 | } |
2377 | } |
1376 | |
2378 | |
1377 | void |
2379 | void |
1378 | normalized_coords (OpenCL::Sampler self) |
2380 | normalized_coords (OpenCL::Sampler self) |
1379 | PPCODE: |
2381 | PPCODE: |
1380 | cl_addressing_mode value [1]; |
2382 | cl_addressing_mode value [1]; |
1381 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2383 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1382 | EXTEND (SP, 1); |
2384 | EXTEND (SP, 1); |
1383 | const int i = 0; |
2385 | const int i = 0; |
1384 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2386 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1385 | |
2387 | |
1386 | void |
2388 | void |
1387 | addressing_mode (OpenCL::Sampler self) |
2389 | addressing_mode (OpenCL::Sampler self) |
1388 | PPCODE: |
2390 | PPCODE: |
1389 | cl_filter_mode value [1]; |
2391 | cl_filter_mode value [1]; |
1390 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2392 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1391 | EXTEND (SP, 1); |
2393 | EXTEND (SP, 1); |
1392 | const int i = 0; |
2394 | const int i = 0; |
1393 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2395 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1394 | |
2396 | |
1395 | void |
2397 | void |
1396 | filter_mode (OpenCL::Sampler self) |
2398 | filter_mode (OpenCL::Sampler self) |
1397 | PPCODE: |
2399 | PPCODE: |
1398 | cl_bool value [1]; |
2400 | cl_bool value [1]; |
… | |
… | |
1409 | DESTROY (OpenCL::Program self) |
2411 | DESTROY (OpenCL::Program self) |
1410 | CODE: |
2412 | CODE: |
1411 | clReleaseProgram (self); |
2413 | clReleaseProgram (self); |
1412 | |
2414 | |
1413 | void |
2415 | void |
1414 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2416 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2417 | ALIAS: |
|
|
2418 | build_async = 1 |
1415 | CODE: |
2419 | CODE: |
|
|
2420 | cl_uint device_count = 0; |
|
|
2421 | cl_device_id *device_list = 0; |
|
|
2422 | |
|
|
2423 | if (SvOK (devices)) |
|
|
2424 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2425 | |
|
|
2426 | void *user_data; |
|
|
2427 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2428 | |
|
|
2429 | if (ix) |
|
|
2430 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2431 | else |
|
|
2432 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2433 | |
|
|
2434 | #if CL_VERSION_1_2 |
|
|
2435 | |
|
|
2436 | void |
|
|
2437 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2438 | CODE: |
|
|
2439 | cl_uint device_count = 0; |
|
|
2440 | cl_device_id *device_list = 0; |
|
|
2441 | |
|
|
2442 | if (SvOK (devices)) |
|
|
2443 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2444 | |
|
|
2445 | cl_uint header_count = 0; |
|
|
2446 | cl_program *header_list = 0; |
|
|
2447 | const char **header_name = 0; |
|
|
2448 | |
|
|
2449 | if (SvOK (headers)) |
|
|
2450 | { |
|
|
2451 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2452 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2453 | |
|
|
2454 | HV *hv = (HV *)SvRV (devices); |
|
|
2455 | |
|
|
2456 | header_count = hv_iterinit (hv); |
|
|
2457 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2458 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2459 | |
|
|
2460 | HE *he; |
|
|
2461 | int i = 0; |
|
|
2462 | while (he = hv_iternext (hv)) |
|
|
2463 | { |
|
|
2464 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2465 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2466 | ++i; |
|
|
2467 | } |
|
|
2468 | } |
|
|
2469 | |
|
|
2470 | void *user_data; |
|
|
2471 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2472 | |
1416 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2473 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2474 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2475 | |
|
|
2476 | #endif |
1417 | |
2477 | |
1418 | void |
2478 | void |
1419 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2479 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1420 | PPCODE: |
2480 | PPCODE: |
1421 | size_t size; |
2481 | size_t size; |
… | |
… | |
1450 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2510 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
1451 | EXTEND (SP, 1); |
2511 | EXTEND (SP, 1); |
1452 | const int i = 0; |
2512 | const int i = 0; |
1453 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
2513 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
1454 | |
2514 | |
|
|
2515 | void |
|
|
2516 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2517 | PPCODE: |
|
|
2518 | cl_program_binary_type value [1]; |
|
|
2519 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2520 | EXTEND (SP, 1); |
|
|
2521 | const int i = 0; |
|
|
2522 | PUSHs (sv_2mortal (newSVuv (UV)value [i]))); |
|
|
2523 | |
1455 | #END:program_build |
2524 | #END:program_build |
1456 | |
2525 | |
1457 | void |
2526 | void |
1458 | kernel (OpenCL::Program program, SV *function) |
2527 | kernel (OpenCL::Program program, SV *function) |
1459 | PPCODE: |
2528 | PPCODE: |
1460 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2529 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1461 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2530 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2531 | |
|
|
2532 | void |
|
|
2533 | kernels_in_program (OpenCL::Program program) |
|
|
2534 | PPCODE: |
|
|
2535 | cl_uint num_kernels; |
|
|
2536 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2537 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2538 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2539 | |
|
|
2540 | int i; |
|
|
2541 | EXTEND (SP, num_kernels); |
|
|
2542 | for (i = 0; i < num_kernels; ++i) |
|
|
2543 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1462 | |
2544 | |
1463 | void |
2545 | void |
1464 | info (OpenCL::Program self, cl_program_info name) |
2546 | info (OpenCL::Program self, cl_program_info name) |
1465 | PPCODE: |
2547 | PPCODE: |
1466 | INFO (Program) |
2548 | INFO (Program) |
… | |
… | |
1484 | { |
2566 | { |
1485 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2567 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1486 | SvUPGRADE (sv, SVt_PV); |
2568 | SvUPGRADE (sv, SVt_PV); |
1487 | SvPOK_only (sv); |
2569 | SvPOK_only (sv); |
1488 | SvCUR_set (sv, sizes [i]); |
2570 | SvCUR_set (sv, sizes [i]); |
1489 | ptrs [i] = SvPVX (sv); |
2571 | ptrs [i] = (void *)SvPVX (sv); |
1490 | PUSHs (sv); |
2572 | PUSHs (sv); |
1491 | } |
2573 | } |
1492 | |
2574 | |
1493 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2575 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1494 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2576 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1514 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2596 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1515 | EXTEND (SP, 1); |
2597 | EXTEND (SP, 1); |
1516 | const int i = 0; |
2598 | const int i = 0; |
1517 | { |
2599 | { |
1518 | NEED_SUCCESS (RetainContext, (value [i])); |
2600 | NEED_SUCCESS (RetainContext, (value [i])); |
1519 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2601 | PUSH_CLOBJ (stash_context, value [i]); |
1520 | } |
2602 | } |
1521 | |
2603 | |
1522 | void |
2604 | void |
1523 | devices (OpenCL::Program self) |
2605 | devices (OpenCL::Program self) |
1524 | PPCODE: |
2606 | PPCODE: |
… | |
… | |
1528 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2610 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1529 | int i, n = size / sizeof (*value); |
2611 | int i, n = size / sizeof (*value); |
1530 | EXTEND (SP, n); |
2612 | EXTEND (SP, n); |
1531 | for (i = 0; i < n; ++i) |
2613 | for (i = 0; i < n; ++i) |
1532 | { |
2614 | { |
1533 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2615 | PUSH_CLOBJ (stash_device, value [i]); |
1534 | } |
2616 | } |
1535 | |
2617 | |
1536 | void |
2618 | void |
1537 | source (OpenCL::Program self) |
2619 | source (OpenCL::Program self) |
1538 | PPCODE: |
2620 | PPCODE: |
… | |
… | |
1564 | DESTROY (OpenCL::Kernel self) |
2646 | DESTROY (OpenCL::Kernel self) |
1565 | CODE: |
2647 | CODE: |
1566 | clReleaseKernel (self); |
2648 | clReleaseKernel (self); |
1567 | |
2649 | |
1568 | void |
2650 | void |
|
|
2651 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2652 | CODE: |
|
|
2653 | int i; |
|
|
2654 | for (i = 2; ; ++i) |
|
|
2655 | { |
|
|
2656 | while (*format == ' ') |
|
|
2657 | ++format; |
|
|
2658 | |
|
|
2659 | char type = *format++; |
|
|
2660 | |
|
|
2661 | if (!type) |
|
|
2662 | break; |
|
|
2663 | |
|
|
2664 | if (i >= items) |
|
|
2665 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2666 | |
|
|
2667 | SV *sv = ST (i); |
|
|
2668 | |
|
|
2669 | union |
|
|
2670 | { |
|
|
2671 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2672 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2673 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2674 | cl_mem cm; |
|
|
2675 | cl_sampler ca; |
|
|
2676 | size_t cz; |
|
|
2677 | cl_event ce; |
|
|
2678 | } arg; |
|
|
2679 | size_t size; |
|
|
2680 | int nullarg = 0; |
|
|
2681 | |
|
|
2682 | switch (type) |
|
|
2683 | { |
|
|
2684 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2685 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2686 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2687 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2688 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2689 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2690 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2691 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2692 | |
|
|
2693 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2694 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2695 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2696 | |
|
|
2697 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2698 | |
|
|
2699 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2700 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2701 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2702 | |
|
|
2703 | default: |
|
|
2704 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2705 | } |
|
|
2706 | |
|
|
2707 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2708 | if (res) |
|
|
2709 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2710 | } |
|
|
2711 | |
|
|
2712 | if (i != items) |
|
|
2713 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2714 | |
|
|
2715 | void |
1569 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2716 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1570 | CODE: |
2717 | CODE: |
1571 | clSetKernelArg (self, idx, sizeof (value), &value); |
2718 | clSetKernelArg (self, idx, sizeof (value), &value); |
1572 | |
2719 | |
1573 | void |
2720 | void |
… | |
… | |
1621 | clSetKernelArg (self, idx, sizeof (value), &value); |
2768 | clSetKernelArg (self, idx, sizeof (value), &value); |
1622 | |
2769 | |
1623 | void |
2770 | void |
1624 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2771 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1625 | CODE: |
2772 | CODE: |
|
|
2773 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2774 | |
|
|
2775 | void |
|
|
2776 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2777 | CODE: |
|
|
2778 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2779 | |
|
|
2780 | void |
|
|
2781 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2782 | ALIAS: |
|
|
2783 | set_image2d = 0 |
|
|
2784 | set_image3d = 0 |
|
|
2785 | CODE: |
|
|
2786 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2787 | |
|
|
2788 | void |
|
|
2789 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2790 | CODE: |
1626 | clSetKernelArg (self, idx, sizeof (value), &value); |
2791 | clSetKernelArg (self, idx, sizeof (value), &value); |
1627 | |
2792 | |
1628 | void |
2793 | void |
1629 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2794 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
1630 | CODE: |
2795 | CODE: |
1631 | clSetKernelArg (self, idx, sizeof (value), &value); |
2796 | clSetKernelArg (self, idx, size, 0); |
1632 | |
|
|
1633 | void |
|
|
1634 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
1635 | CODE: |
|
|
1636 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1637 | |
|
|
1638 | void |
|
|
1639 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1640 | CODE: |
|
|
1641 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1642 | |
|
|
1643 | void |
|
|
1644 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
1645 | CODE: |
|
|
1646 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1647 | |
2797 | |
1648 | void |
2798 | void |
1649 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
2799 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1650 | CODE: |
2800 | CODE: |
1651 | clSetKernelArg (self, idx, sizeof (value), &value); |
2801 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1687 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2837 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1688 | EXTEND (SP, 1); |
2838 | EXTEND (SP, 1); |
1689 | const int i = 0; |
2839 | const int i = 0; |
1690 | { |
2840 | { |
1691 | NEED_SUCCESS (RetainContext, (value [i])); |
2841 | NEED_SUCCESS (RetainContext, (value [i])); |
1692 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2842 | PUSH_CLOBJ (stash_context, value [i]); |
1693 | } |
2843 | } |
1694 | |
2844 | |
1695 | void |
2845 | void |
1696 | program (OpenCL::Kernel self) |
2846 | program (OpenCL::Kernel self) |
1697 | PPCODE: |
2847 | PPCODE: |
… | |
… | |
1699 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2849 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1700 | EXTEND (SP, 1); |
2850 | EXTEND (SP, 1); |
1701 | const int i = 0; |
2851 | const int i = 0; |
1702 | { |
2852 | { |
1703 | NEED_SUCCESS (RetainProgram, (value [i])); |
2853 | NEED_SUCCESS (RetainProgram, (value [i])); |
1704 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2854 | PUSH_CLOBJ (stash_program, value [i]); |
1705 | } |
2855 | } |
1706 | |
2856 | |
1707 | #END:kernel |
2857 | #END:kernel |
1708 | |
2858 | |
1709 | void |
2859 | void |
… | |
… | |
1756 | const int i = 0; |
2906 | const int i = 0; |
1757 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2907 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1758 | |
2908 | |
1759 | #END:kernel_work_group |
2909 | #END:kernel_work_group |
1760 | |
2910 | |
|
|
2911 | #if CL_VERSION_1_2 |
|
|
2912 | |
|
|
2913 | void |
|
|
2914 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2915 | PPCODE: |
|
|
2916 | size_t size; |
|
|
2917 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2918 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2919 | SvUPGRADE (sv, SVt_PV); |
|
|
2920 | SvPOK_only (sv); |
|
|
2921 | SvCUR_set (sv, size); |
|
|
2922 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2923 | XPUSHs (sv); |
|
|
2924 | |
|
|
2925 | #BEGIN:kernel_arg |
|
|
2926 | |
|
|
2927 | void |
|
|
2928 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2929 | PPCODE: |
|
|
2930 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2931 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2932 | EXTEND (SP, 1); |
|
|
2933 | const int i = 0; |
|
|
2934 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2935 | |
|
|
2936 | void |
|
|
2937 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2938 | PPCODE: |
|
|
2939 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2940 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2941 | EXTEND (SP, 1); |
|
|
2942 | const int i = 0; |
|
|
2943 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2944 | |
|
|
2945 | void |
|
|
2946 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2947 | ALIAS: |
|
|
2948 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2949 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2950 | PPCODE: |
|
|
2951 | size_t size; |
|
|
2952 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2953 | char *value = tmpbuf (size); |
|
|
2954 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2955 | EXTEND (SP, 1); |
|
|
2956 | const int i = 0; |
|
|
2957 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2958 | |
|
|
2959 | void |
|
|
2960 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2961 | PPCODE: |
|
|
2962 | cl_kernel_arg_type_qualifier value [1]; |
|
|
2963 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2964 | EXTEND (SP, 1); |
|
|
2965 | const int i = 0; |
|
|
2966 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2967 | |
|
|
2968 | #END:kernel_arg |
|
|
2969 | |
|
|
2970 | #endif |
|
|
2971 | |
1761 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2972 | MODULE = OpenCL PACKAGE = OpenCL::Event |
1762 | |
2973 | |
1763 | void |
2974 | void |
1764 | DESTROY (OpenCL::Event self) |
2975 | DESTROY (OpenCL::Event self) |
1765 | CODE: |
2976 | CODE: |
… | |
… | |
1767 | |
2978 | |
1768 | void |
2979 | void |
1769 | wait (OpenCL::Event self) |
2980 | wait (OpenCL::Event self) |
1770 | CODE: |
2981 | CODE: |
1771 | clWaitForEvents (1, &self); |
2982 | clWaitForEvents (1, &self); |
|
|
2983 | |
|
|
2984 | void |
|
|
2985 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2986 | CODE: |
|
|
2987 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
1772 | |
2988 | |
1773 | void |
2989 | void |
1774 | info (OpenCL::Event self, cl_event_info name) |
2990 | info (OpenCL::Event self, cl_event_info name) |
1775 | PPCODE: |
2991 | PPCODE: |
1776 | INFO (Event) |
2992 | INFO (Event) |
… | |
… | |
1784 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
3000 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1785 | EXTEND (SP, 1); |
3001 | EXTEND (SP, 1); |
1786 | const int i = 0; |
3002 | const int i = 0; |
1787 | { |
3003 | { |
1788 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
3004 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1789 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
3005 | PUSH_CLOBJ (stash_queue, value [i]); |
1790 | } |
3006 | } |
1791 | |
3007 | |
1792 | void |
3008 | void |
1793 | command_type (OpenCL::Event self) |
3009 | command_type (OpenCL::Event self) |
1794 | PPCODE: |
3010 | PPCODE: |
… | |
… | |
1817 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
3033 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1818 | EXTEND (SP, 1); |
3034 | EXTEND (SP, 1); |
1819 | const int i = 0; |
3035 | const int i = 0; |
1820 | { |
3036 | { |
1821 | NEED_SUCCESS (RetainContext, (value [i])); |
3037 | NEED_SUCCESS (RetainContext, (value [i])); |
1822 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
3038 | PUSH_CLOBJ (stash_context, value [i]); |
1823 | } |
3039 | } |
1824 | |
3040 | |
1825 | #END:event |
3041 | #END:event |
1826 | |
3042 | |
1827 | void |
3043 | void |
… | |
… | |
1852 | void |
3068 | void |
1853 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
3069 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1854 | CODE: |
3070 | CODE: |
1855 | clSetUserEventStatus (self, execution_status); |
3071 | clSetUserEventStatus (self, execution_status); |
1856 | |
3072 | |
|
|
3073 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3074 | |
|
|
3075 | void |
|
|
3076 | DESTROY (SV *self) |
|
|
3077 | CODE: |
|
|
3078 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3079 | |
|
|
3080 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3081 | mapped_detach (self, mapped); |
|
|
3082 | |
|
|
3083 | clReleaseCommandQueue (mapped->queue); |
|
|
3084 | clReleaseEvent (mapped->event); |
|
|
3085 | Safefree (mapped); |
|
|
3086 | |
|
|
3087 | void |
|
|
3088 | unmap (OpenCL::Mapped self, ...) |
|
|
3089 | CODE: |
|
|
3090 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3091 | |
|
|
3092 | bool |
|
|
3093 | mapped (OpenCL::Mapped self) |
|
|
3094 | CODE: |
|
|
3095 | RETVAL = !!self->ptr; |
|
|
3096 | OUTPUT: |
|
|
3097 | RETVAL |
|
|
3098 | |
|
|
3099 | void |
|
|
3100 | wait (OpenCL::Mapped self) |
|
|
3101 | PPCODE: |
|
|
3102 | if (self->event) |
|
|
3103 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3104 | |
|
|
3105 | void |
|
|
3106 | event (OpenCL::Mapped self) |
|
|
3107 | PPCODE: |
|
|
3108 | if (!self->event) |
|
|
3109 | XSRETURN_UNDEF; |
|
|
3110 | |
|
|
3111 | clRetainEvent (self->event); |
|
|
3112 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3113 | |
|
|
3114 | size_t |
|
|
3115 | size (OpenCL::Mapped self) |
|
|
3116 | CODE: |
|
|
3117 | RETVAL = self->cb; |
|
|
3118 | OUTPUT: |
|
|
3119 | RETVAL |
|
|
3120 | |
|
|
3121 | IV |
|
|
3122 | ptr (OpenCL::Mapped self) |
|
|
3123 | CODE: |
|
|
3124 | RETVAL = PTR2IV (self->ptr); |
|
|
3125 | OUTPUT: |
|
|
3126 | RETVAL |
|
|
3127 | |
|
|
3128 | void |
|
|
3129 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3130 | CODE: |
|
|
3131 | STRLEN len; |
|
|
3132 | const char *ptr = SvPVbyte (data, len); |
|
|
3133 | |
|
|
3134 | if (offset + len > self->cb) |
|
|
3135 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3136 | |
|
|
3137 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3138 | |
|
|
3139 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3140 | |
|
|
3141 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3142 | |
|
|
3143 | IV |
|
|
3144 | row_pitch (OpenCL::Mapped self) |
|
|
3145 | ALIAS: |
|
|
3146 | slice_pitch = 1 |
|
|
3147 | CODE: |
|
|
3148 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
3149 | OUTPUT: |
|
|
3150 | RETVAL |
|
|
3151 | |
|
|
3152 | |
|
|
3153 | |