1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #include "ecb.h"//D |
|
|
6 | |
|
|
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
8 | #include "xthread.h" |
|
|
9 | #include "schmorp.h" |
|
|
10 | |
|
|
11 | #ifdef I_DLFCN |
|
|
12 | #include <dlfcn.h> |
|
|
13 | #endif |
|
|
14 | |
|
|
15 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
16 | // but then fail to define the api functions unless you ALSO define |
|
|
17 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
18 | // after all, the functions are deprecated, not removed. |
|
|
19 | // in addition, you cannot test for this in any future-proof way. |
|
|
20 | // each time a new opencl version comes out, you need to make a new |
|
|
21 | // release. |
|
|
22 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
23 | |
|
|
24 | #ifndef PREFER_1_1 |
|
|
25 | #define PREFER_1_1 1 |
|
|
26 | #endif |
|
|
27 | |
|
|
28 | #if PREFER_1_1 |
|
|
29 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
30 | #endif |
|
|
31 | |
5 | #ifdef __APPLE__ |
32 | #ifdef __APPLE__ |
6 | #include <OpenCL/opencl.h> |
33 | #define CLHDR(name) <OpenCL/name> |
7 | #else |
34 | #else |
8 | #include <CL/opencl.h> |
35 | #define CLHDR(name) <CL/name> |
9 | #endif |
36 | #endif |
|
|
37 | |
|
|
38 | #include CLHDR(opencl.h) |
|
|
39 | |
|
|
40 | #ifndef CL_VERSION_1_2 |
|
|
41 | #include CLHDR(cl_d3d9.h) |
|
|
42 | #endif |
|
|
43 | |
|
|
44 | #if _WIN32 |
|
|
45 | #include CLHDR(cl_d3d10.h) |
|
|
46 | #if CL_VERSION_1_2 |
|
|
47 | #include CLHDR<cl_d3d11.h> |
|
|
48 | #endif |
|
|
49 | #include CLHDR<cl_dx9_media_sharing.h.h> |
|
|
50 | #endif |
|
|
51 | |
|
|
52 | #ifndef CL_VERSION_1_2 |
|
|
53 | #undef PREFER_1_1 |
|
|
54 | #define PREFER_1_1 1 |
|
|
55 | #endif |
|
|
56 | |
|
|
57 | // make sure all constants we might use are actually defined |
|
|
58 | #include "default.h" |
10 | |
59 | |
11 | typedef cl_platform_id OpenCL__Platform; |
60 | typedef cl_platform_id OpenCL__Platform; |
12 | typedef cl_device_id OpenCL__Device; |
61 | typedef cl_device_id OpenCL__Device; |
|
|
62 | typedef cl_device_id OpenCL__SubDevice; |
13 | typedef cl_context OpenCL__Context; |
63 | typedef cl_context OpenCL__Context; |
14 | typedef cl_command_queue OpenCL__Queue; |
64 | typedef cl_command_queue OpenCL__Queue; |
15 | typedef cl_mem OpenCL__Memory; |
65 | typedef cl_mem OpenCL__Memory; |
16 | typedef cl_mem OpenCL__Buffer; |
66 | typedef cl_mem OpenCL__Buffer; |
17 | typedef cl_mem OpenCL__BufferObj; |
67 | typedef cl_mem OpenCL__BufferObj; |
18 | typedef cl_mem OpenCL__Image; |
68 | typedef cl_mem OpenCL__Image; |
19 | typedef cl_mem OpenCL__Image2D; |
|
|
20 | typedef cl_mem OpenCL__Image3D; |
|
|
21 | typedef cl_mem OpenCL__Memory_ornull; |
69 | typedef cl_mem OpenCL__Memory_ornull; |
22 | typedef cl_mem OpenCL__Buffer_ornull; |
70 | typedef cl_mem OpenCL__Buffer_ornull; |
23 | typedef cl_mem OpenCL__Image_ornull; |
71 | typedef cl_mem OpenCL__Image_ornull; |
24 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
25 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
26 | typedef cl_sampler OpenCL__Sampler; |
72 | typedef cl_sampler OpenCL__Sampler; |
27 | typedef cl_program OpenCL__Program; |
73 | typedef cl_program OpenCL__Program; |
28 | typedef cl_kernel OpenCL__Kernel; |
74 | typedef cl_kernel OpenCL__Kernel; |
29 | typedef cl_event OpenCL__Event; |
75 | typedef cl_event OpenCL__Event; |
30 | typedef cl_event OpenCL__UserEvent; |
76 | typedef cl_event OpenCL__UserEvent; |
31 | |
77 | |
32 | typedef SV *FUTURE; |
78 | typedef struct mapped * OpenCL__Mapped; |
|
|
79 | |
|
|
80 | static HV |
|
|
81 | *stash_platform, |
|
|
82 | *stash_device, |
|
|
83 | *stash_subdevice, |
|
|
84 | *stash_context, |
|
|
85 | *stash_queue, |
|
|
86 | *stash_program, |
|
|
87 | *stash_kernel, |
|
|
88 | *stash_sampler, |
|
|
89 | *stash_event, |
|
|
90 | *stash_userevent, |
|
|
91 | *stash_memory, |
|
|
92 | *stash_buffer, |
|
|
93 | *stash_bufferobj, |
|
|
94 | *stash_image, |
|
|
95 | *stash_image1d, |
|
|
96 | *stash_image1darray, |
|
|
97 | *stash_image1dbuffer, |
|
|
98 | *stash_image2d, |
|
|
99 | *stash_image2darray, |
|
|
100 | *stash_image3d, |
|
|
101 | *stash_mapped, |
|
|
102 | *stash_mappedbuffer, |
|
|
103 | *stash_mappedimage; |
33 | |
104 | |
34 | /*****************************************************************************/ |
105 | /*****************************************************************************/ |
35 | |
106 | |
|
|
107 | // name must include a leading underscore |
|
|
108 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
109 | // for perl. doh. |
|
|
110 | static void * |
|
|
111 | glsym (const char *name) |
|
|
112 | { |
|
|
113 | void *fun = 0; |
|
|
114 | |
|
|
115 | #if defined I_DLFCN && defined RTLD_DEFAULT |
|
|
116 | fun = dlsym (RTLD_DEFAULT, name + 1); |
|
|
117 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
|
|
118 | |
|
|
119 | if (!fun) |
|
|
120 | { |
|
|
121 | static void *libgl; |
|
|
122 | static const char *glso[] = { |
|
|
123 | "libGL.so.1", |
|
|
124 | "libGL.so.3", |
|
|
125 | "libGL.so.4.0", |
|
|
126 | "libGL.so", |
|
|
127 | "/usr/lib/libGL.so", |
|
|
128 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
129 | }; |
|
|
130 | int i; |
|
|
131 | |
|
|
132 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
133 | { |
|
|
134 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
135 | if (libgl) |
|
|
136 | break; |
|
|
137 | } |
|
|
138 | |
|
|
139 | if (libgl) |
|
|
140 | { |
|
|
141 | fun = dlsym (libgl, name + 1); |
|
|
142 | if (!fun) fun = dlsym (libgl, name); |
|
|
143 | } |
|
|
144 | } |
|
|
145 | #endif |
|
|
146 | |
|
|
147 | return fun; |
|
|
148 | } |
|
|
149 | |
|
|
150 | /*****************************************************************************/ |
|
|
151 | |
36 | /* up to two temporary buffers */ |
152 | /* up to two temporary buffers */ |
37 | static void * |
153 | static void * ecb_noinline |
38 | tmpbuf (size_t size) |
154 | tmpbuf (size_t size) |
39 | { |
155 | { |
40 | enum { buffers = 3 }; |
156 | enum { buffers = 4 }; |
41 | static int idx; |
157 | static int idx; |
42 | static void *buf [buffers]; |
158 | static void *buf [buffers]; |
43 | static size_t len [buffers]; |
159 | static size_t len [buffers]; |
44 | |
160 | |
45 | idx = ++idx % buffers; |
161 | idx = (idx + 1) % buffers; |
46 | |
162 | |
47 | if (len [idx] < size) |
163 | if (len [idx] < size) |
48 | { |
164 | { |
49 | free (buf [idx]); |
165 | free (buf [idx]); |
50 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
166 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
51 | buf [idx] = malloc (len [idx]); |
167 | buf [idx] = malloc (len [idx]); |
52 | } |
168 | } |
53 | |
169 | |
54 | return buf [idx]; |
170 | return buf [idx]; |
|
|
171 | } |
|
|
172 | |
|
|
173 | static const char * ecb_noinline |
|
|
174 | cv_get_name (CV *cv) |
|
|
175 | { |
|
|
176 | static char fullname [256]; |
|
|
177 | |
|
|
178 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
179 | |
|
|
180 | HV *stash = GvSTASH (gv); |
|
|
181 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
182 | const char *gvname = GvNAME (gv); |
|
|
183 | |
|
|
184 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
185 | return fullname; |
55 | } |
186 | } |
56 | |
187 | |
57 | /*****************************************************************************/ |
188 | /*****************************************************************************/ |
58 | |
189 | |
59 | typedef struct |
190 | typedef struct |
… | |
… | |
118 | if (res) \ |
249 | if (res) \ |
119 | FAIL (name); |
250 | FAIL (name); |
120 | |
251 | |
121 | /*****************************************************************************/ |
252 | /*****************************************************************************/ |
122 | |
253 | |
123 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
124 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
125 | |
|
|
126 | static void * |
254 | static SV * |
|
|
255 | new_clobj (HV *stash, IV id) |
|
|
256 | { |
|
|
257 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
258 | } |
|
|
259 | |
|
|
260 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
261 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
262 | |
|
|
263 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
264 | /* they can be upgraded at runtime to the array form */ |
|
|
265 | static void * ecb_noinline |
127 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
266 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
128 | { |
267 | { |
|
|
268 | // sv_derived_from is quite slow :( |
129 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
269 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
130 | return (void *)SvIV (SvRV (sv)); |
270 | return (void *)SvIV (SvRV (sv)); |
131 | |
271 | |
132 | croak ("%s: %s is not of type %s", func, svname, pkg); |
272 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
273 | } |
|
|
274 | |
|
|
275 | // the "no-inherit" version of the above |
|
|
276 | static void * ecb_noinline |
|
|
277 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
278 | { |
|
|
279 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
280 | return (void *)SvIV (SvRV (sv)); |
|
|
281 | |
|
|
282 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
133 | } |
283 | } |
134 | |
284 | |
135 | /*****************************************************************************/ |
285 | /*****************************************************************************/ |
136 | |
286 | |
137 | static cl_context_properties * |
287 | static cl_context_properties * ecb_noinline |
138 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
288 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
139 | { |
289 | { |
140 | if (!sv || !SvOK (sv)) |
290 | if (!sv || !SvOK (sv)) |
141 | if (extra) |
291 | if (extra) |
142 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
292 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
143 | else |
293 | else |
… | |
… | |
149 | int i, len = av_len (av) + 1; |
299 | int i, len = av_len (av) + 1; |
150 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
300 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
151 | cl_context_properties *l = p; |
301 | cl_context_properties *l = p; |
152 | |
302 | |
153 | if (len & 1) |
303 | if (len & 1) |
154 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
304 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
155 | |
305 | |
156 | while (extracount--) |
306 | while (extracount--) |
157 | *l++ = *extra++; |
307 | *l++ = *extra++; |
158 | |
308 | |
159 | for (i = 0; i < len; i += 2) |
309 | for (i = 0; i < len; i += 2) |
160 | { |
310 | { |
161 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
311 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
162 | SV *p_sv = *av_fetch (av, i + 1, 0); |
312 | SV *p_sv = *av_fetch (av, i + 1, 0); |
163 | cl_context_properties v; |
313 | cl_context_properties v = SvIV (p_sv); // code below can override |
164 | |
314 | |
165 | switch (t) |
315 | switch (t) |
166 | { |
316 | { |
|
|
317 | case CL_CONTEXT_PLATFORM: |
|
|
318 | if (SvROK (p_sv)) |
|
|
319 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
320 | break; |
|
|
321 | |
|
|
322 | case CL_GLX_DISPLAY_KHR: |
|
|
323 | if (!SvOK (p_sv)) |
|
|
324 | { |
|
|
325 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
326 | if (func) |
|
|
327 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
328 | } |
|
|
329 | break; |
|
|
330 | |
|
|
331 | case CL_GL_CONTEXT_KHR: |
|
|
332 | if (!SvOK (p_sv)) |
|
|
333 | { |
|
|
334 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
335 | if (func) |
|
|
336 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
337 | } |
|
|
338 | break; |
|
|
339 | |
167 | default: |
340 | default: |
168 | /* unknown property, treat as int */ |
341 | /* unknown property, treat as int */ |
169 | v = SvIV (p_sv); |
|
|
170 | break; |
342 | break; |
171 | } |
343 | } |
172 | |
344 | |
173 | *l++ = t; |
345 | *l++ = t; |
174 | *l++ = v; |
346 | *l++ = v; |
… | |
… | |
177 | *l = 0; |
349 | *l = 0; |
178 | |
350 | |
179 | return p; |
351 | return p; |
180 | } |
352 | } |
181 | |
353 | |
182 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
354 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
|
|
355 | } |
|
|
356 | |
|
|
357 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
|
|
358 | // is a pointer (and only on well-behaved systems). |
|
|
359 | static void * ecb_noinline |
|
|
360 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
361 | { |
|
|
362 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
363 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
364 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
|
|
365 | |
|
|
366 | AV *av = (AV *)SvRV (arg); |
|
|
367 | void **list = 0; |
|
|
368 | cl_uint count = av_len (av) + 1; |
|
|
369 | |
|
|
370 | if (count) |
|
|
371 | { |
|
|
372 | list = tmpbuf (sizeof (*list) * count); |
|
|
373 | int i; |
|
|
374 | for (i = 0; i < count; ++i) |
|
|
375 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
|
|
376 | } |
|
|
377 | |
|
|
378 | if (!count && !or_undef) |
|
|
379 | croak ("%s: '%s' must contain at least one %s object", |
|
|
380 | cv_get_name (cv), argname, klass); |
|
|
381 | |
|
|
382 | *rcount = count; |
|
|
383 | return (void *)list; |
183 | } |
384 | } |
184 | |
385 | |
185 | /*****************************************************************************/ |
386 | /*****************************************************************************/ |
|
|
387 | /* callback stuff */ |
|
|
388 | |
|
|
389 | /* default context callback, log to stderr */ |
|
|
390 | static void CL_CALLBACK |
|
|
391 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
392 | { |
|
|
393 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
394 | } |
|
|
395 | |
|
|
396 | typedef struct |
|
|
397 | { |
|
|
398 | int free_cb; |
|
|
399 | void (*push)(void *data1, void *data2, void *data3); |
|
|
400 | } eq_vtbl; |
|
|
401 | |
|
|
402 | typedef struct eq_item |
|
|
403 | { |
|
|
404 | struct eq_item *next; |
|
|
405 | eq_vtbl *vtbl; |
|
|
406 | SV *cb; |
|
|
407 | void *data1, *data2, *data3; |
|
|
408 | } eq_item; |
|
|
409 | |
|
|
410 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
411 | static void *eq_signal_arg; |
|
|
412 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
413 | static eq_item *eq_head, *eq_tail; |
|
|
414 | |
|
|
415 | static void ecb_noinline |
|
|
416 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
417 | { |
|
|
418 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
419 | |
|
|
420 | item->next = 0; |
|
|
421 | item->vtbl = vtbl; |
|
|
422 | item->cb = cb; |
|
|
423 | item->data1 = data1; |
|
|
424 | item->data2 = data2; |
|
|
425 | item->data3 = data3; |
|
|
426 | |
|
|
427 | X_LOCK (eq_lock); |
|
|
428 | |
|
|
429 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
430 | eq_tail = item; |
|
|
431 | |
|
|
432 | X_UNLOCK (eq_lock); |
|
|
433 | |
|
|
434 | eq_signal_func (eq_signal_arg, 0); |
|
|
435 | } |
|
|
436 | |
|
|
437 | static eq_item * |
|
|
438 | eq_dec (void) |
|
|
439 | { |
|
|
440 | eq_item *res; |
|
|
441 | |
|
|
442 | X_LOCK (eq_lock); |
|
|
443 | |
|
|
444 | res = eq_head; |
|
|
445 | if (res) |
|
|
446 | eq_head = res->next; |
|
|
447 | |
|
|
448 | X_UNLOCK (eq_lock); |
|
|
449 | |
|
|
450 | return res; |
|
|
451 | } |
|
|
452 | |
|
|
453 | static void |
|
|
454 | eq_poll (void) |
|
|
455 | { |
|
|
456 | eq_item *item; |
|
|
457 | |
|
|
458 | while ((item = eq_dec ())) |
|
|
459 | { |
|
|
460 | ENTER; |
|
|
461 | SAVETMPS; |
|
|
462 | |
|
|
463 | dSP; |
|
|
464 | PUSHMARK (SP); |
|
|
465 | EXTEND (SP, 2); |
|
|
466 | |
|
|
467 | if (item->vtbl->free_cb) |
|
|
468 | sv_2mortal (item->cb); |
|
|
469 | |
|
|
470 | PUTBACK; |
|
|
471 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
472 | |
|
|
473 | SV *cb = item->cb; |
|
|
474 | free (item); |
|
|
475 | |
|
|
476 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
477 | |
|
|
478 | FREETMPS; |
|
|
479 | LEAVE; |
|
|
480 | } |
|
|
481 | } |
|
|
482 | |
|
|
483 | static void |
|
|
484 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
485 | { |
|
|
486 | eq_poll (); |
|
|
487 | } |
|
|
488 | |
|
|
489 | /*****************************************************************************/ |
|
|
490 | /* context notify */ |
|
|
491 | |
|
|
492 | static void ecb_noinline |
|
|
493 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
494 | { |
|
|
495 | dSP; |
|
|
496 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
497 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
498 | PUTBACK; |
|
|
499 | |
|
|
500 | free (data1); |
|
|
501 | free (data2); |
|
|
502 | } |
|
|
503 | |
|
|
504 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
505 | |
|
|
506 | static void CL_CALLBACK |
|
|
507 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
508 | { |
|
|
509 | void *pvt_copy = malloc (cb); |
|
|
510 | memcpy (pvt_copy, pvt, cb); |
|
|
511 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
512 | } |
|
|
513 | |
|
|
514 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
515 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
516 | void *user_data = 0; \ |
|
|
517 | \ |
|
|
518 | if (SvOK (notify)) \ |
|
|
519 | { \ |
|
|
520 | pfn_notify = eq_context_notify; \ |
|
|
521 | user_data = s_get_cv (notify); \ |
|
|
522 | } |
|
|
523 | |
|
|
524 | static SV * ecb_noinline |
|
|
525 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
526 | { |
|
|
527 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
528 | |
|
|
529 | if (user_data) |
|
|
530 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
531 | |
|
|
532 | return sv; |
|
|
533 | } |
|
|
534 | |
|
|
535 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
536 | |
|
|
537 | /*****************************************************************************/ |
|
|
538 | /* build/compile/link notify */ |
|
|
539 | |
|
|
540 | static void |
|
|
541 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
542 | { |
|
|
543 | dSP; |
|
|
544 | PUSH_CLOBJ (stash_program, data1); |
|
|
545 | PUTBACK; |
|
|
546 | } |
|
|
547 | |
|
|
548 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
549 | |
|
|
550 | static void CL_CALLBACK |
|
|
551 | eq_program_notify (cl_program program, void *user_data) |
|
|
552 | { |
|
|
553 | clRetainProgram (program); |
|
|
554 | |
|
|
555 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
556 | } |
|
|
557 | |
|
|
558 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
559 | |
|
|
560 | static program_callback ecb_noinline |
|
|
561 | make_program_callback (SV *notify, void **ruser_data) |
|
|
562 | { |
|
|
563 | if (SvOK (notify)) |
|
|
564 | { |
|
|
565 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
566 | return eq_program_notify; |
|
|
567 | } |
|
|
568 | else |
|
|
569 | { |
|
|
570 | *ruser_data = 0; |
|
|
571 | return 0; |
|
|
572 | } |
|
|
573 | } |
|
|
574 | |
|
|
575 | struct build_args |
|
|
576 | { |
|
|
577 | cl_program program; |
|
|
578 | char *options; |
|
|
579 | void *user_data; |
|
|
580 | cl_uint num_devices; |
|
|
581 | }; |
|
|
582 | |
|
|
583 | X_THREAD_PROC (build_program_thread) |
|
|
584 | { |
|
|
585 | struct build_args *arg = thr_arg; |
|
|
586 | |
|
|
587 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
588 | |
|
|
589 | if (arg->user_data) |
|
|
590 | eq_program_notify (arg->program, arg->user_data); |
|
|
591 | else |
|
|
592 | clReleaseProgram (arg->program); |
|
|
593 | |
|
|
594 | free (arg->options); |
|
|
595 | free (arg); |
|
|
596 | |
|
|
597 | return 0; |
|
|
598 | } |
|
|
599 | |
|
|
600 | static void |
|
|
601 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
602 | { |
|
|
603 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
604 | |
|
|
605 | arg->program = program; |
|
|
606 | arg->options = strdup (options); |
|
|
607 | arg->user_data = user_data; |
|
|
608 | arg->num_devices = num_devices; |
|
|
609 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
610 | |
|
|
611 | xthread_t id; |
|
|
612 | thread_create (&id, build_program_thread, arg); |
|
|
613 | } |
|
|
614 | |
|
|
615 | /*****************************************************************************/ |
|
|
616 | /* event objects */ |
|
|
617 | |
|
|
618 | static void |
|
|
619 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
620 | { |
|
|
621 | dSP; |
|
|
622 | PUSH_CLOBJ (stash_event, data1); |
|
|
623 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
624 | PUTBACK; |
|
|
625 | } |
|
|
626 | |
|
|
627 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
628 | |
|
|
629 | static void CL_CALLBACK |
|
|
630 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
631 | { |
|
|
632 | clRetainEvent (event); |
|
|
633 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
634 | } |
|
|
635 | |
|
|
636 | /*****************************************************************************/ |
|
|
637 | /* utilities for XS code */ |
186 | |
638 | |
187 | static size_t |
639 | static size_t |
188 | img_row_pitch (cl_mem img) |
640 | img_row_pitch (cl_mem img) |
189 | { |
641 | { |
190 | size_t res; |
642 | size_t res; |
191 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
643 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
192 | return res; |
644 | return res; |
193 | } |
645 | } |
194 | |
646 | |
195 | static cl_event * |
647 | static cl_event * ecb_noinline |
196 | event_list (SV **items, int count) |
648 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
197 | { |
649 | { |
198 | if (!count) |
650 | cl_uint count = *rcount; |
|
|
651 | |
|
|
652 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
653 | *rcount = count = 0; |
|
|
654 | |
|
|
655 | if (!count && !extra) |
199 | return 0; |
656 | return 0; |
200 | |
657 | |
201 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
658 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
659 | int i = 0; |
202 | |
660 | |
203 | while (count--) |
661 | while (count--) |
|
|
662 | if (SvOK (items [count])) |
204 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
663 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
205 | |
664 | |
|
|
665 | if (extra) |
|
|
666 | list [i++] = extra; |
|
|
667 | |
|
|
668 | *rcount = i; |
|
|
669 | |
206 | return list; |
670 | return i ? list : 0; |
207 | } |
671 | } |
208 | |
672 | |
209 | #define EVENT_LIST(items,count) \ |
673 | #define EVENT_LIST(skip) \ |
210 | cl_uint event_list_count = (count); \ |
674 | cl_uint event_list_count = items - (skip); \ |
211 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
675 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
212 | |
676 | |
213 | #define INFO(class) \ |
677 | #define INFO(class) \ |
214 | { \ |
678 | { \ |
215 | size_t size; \ |
679 | size_t size; \ |
216 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
680 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
220 | SvCUR_set (sv, size); \ |
684 | SvCUR_set (sv, size); \ |
221 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
685 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
222 | XPUSHs (sv); \ |
686 | XPUSHs (sv); \ |
223 | } |
687 | } |
224 | |
688 | |
|
|
689 | /*****************************************************************************/ |
|
|
690 | /* mapped_xxx */ |
|
|
691 | |
|
|
692 | static OpenCL__Mapped |
|
|
693 | SvMAPPED (SV *sv) |
|
|
694 | { |
|
|
695 | // no typechecking atm., keep your fingers crossed |
|
|
696 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
697 | } |
|
|
698 | |
|
|
699 | struct mapped |
|
|
700 | { |
|
|
701 | cl_command_queue queue; |
|
|
702 | cl_mem memobj; |
|
|
703 | void *ptr; |
|
|
704 | size_t cb; |
|
|
705 | cl_event event; |
|
|
706 | size_t row_pitch; |
|
|
707 | size_t slice_pitch; |
|
|
708 | }; |
|
|
709 | |
|
|
710 | static SV * |
|
|
711 | 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) |
|
|
712 | { |
|
|
713 | SV *data = newSV (0); |
|
|
714 | SvUPGRADE (data, SVt_PVMG); |
|
|
715 | |
|
|
716 | OpenCL__Mapped mapped; |
|
|
717 | New (0, mapped, 1, struct mapped); |
|
|
718 | |
|
|
719 | clRetainCommandQueue (queue); |
|
|
720 | |
|
|
721 | mapped->queue = queue; |
|
|
722 | mapped->memobj = memobj; |
|
|
723 | mapped->ptr = ptr; |
|
|
724 | mapped->cb = cb; |
|
|
725 | mapped->event = ev; |
|
|
726 | mapped->row_pitch = row_pitch; |
|
|
727 | mapped->slice_pitch = slice_pitch; |
|
|
728 | |
|
|
729 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
730 | |
|
|
731 | if (SvLEN (data)) |
|
|
732 | Safefree (data); |
|
|
733 | |
|
|
734 | SvPVX (data) = (char *)ptr; |
|
|
735 | SvCUR_set (data, cb); |
|
|
736 | SvLEN_set (data, 0); |
|
|
737 | SvPOK_only (data); |
|
|
738 | |
|
|
739 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
740 | |
|
|
741 | if (!(flags & CL_MAP_WRITE)) |
|
|
742 | SvREADONLY_on (data); |
|
|
743 | |
|
|
744 | return obj; |
|
|
745 | } |
|
|
746 | |
|
|
747 | static void |
|
|
748 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
749 | { |
|
|
750 | SV *data = SvRV (sv); |
|
|
751 | |
|
|
752 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
753 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
754 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
755 | else |
|
|
756 | { |
|
|
757 | SvREADONLY_off (data); |
|
|
758 | SvCUR_set (data, 0); |
|
|
759 | SvPVX (data) = 0; |
|
|
760 | SvOK_off (data); |
|
|
761 | } |
|
|
762 | |
|
|
763 | mapped->ptr = 0; |
|
|
764 | } |
|
|
765 | |
|
|
766 | static void |
|
|
767 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
768 | { |
|
|
769 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
770 | cl_event ev; |
|
|
771 | |
|
|
772 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
773 | |
|
|
774 | clReleaseEvent (mapped->event); |
|
|
775 | mapped->event = ev; |
|
|
776 | |
|
|
777 | mapped_detach (self, mapped); |
|
|
778 | } |
|
|
779 | |
|
|
780 | /*****************************************************************************/ |
|
|
781 | |
225 | MODULE = OpenCL PACKAGE = OpenCL |
782 | MODULE = OpenCL PACKAGE = OpenCL |
226 | |
783 | |
227 | PROTOTYPES: ENABLE |
784 | PROTOTYPES: ENABLE |
|
|
785 | |
|
|
786 | void |
|
|
787 | poll () |
|
|
788 | CODE: |
|
|
789 | eq_poll (); |
|
|
790 | |
|
|
791 | void |
|
|
792 | _eq_initialise (IV func, IV arg) |
|
|
793 | CODE: |
|
|
794 | eq_signal_func = (void (*)(void *, int))func; |
|
|
795 | eq_signal_arg = (void*)arg; |
228 | |
796 | |
229 | BOOT: |
797 | BOOT: |
230 | { |
798 | { |
231 | HV *stash = gv_stashpv ("OpenCL", 1); |
799 | HV *stash = gv_stashpv ("OpenCL", 1); |
232 | static const ivstr *civ, const_iv[] = { |
800 | static const ivstr *civ, const_iv[] = { |
… | |
… | |
239 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
807 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
240 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
808 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
241 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
809 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
242 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
810 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
243 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
811 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
812 | { PREFER_1_1 , "PREFER_1_1" }, |
244 | #include "constiv.h" |
813 | #include "constiv.h" |
245 | }; |
814 | }; |
|
|
815 | |
246 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
816 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
247 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
817 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
818 | |
|
|
819 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
820 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
821 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
822 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
823 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
824 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
825 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
826 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
827 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
828 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
829 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
830 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
831 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
832 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
833 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
834 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
835 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
836 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
837 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
838 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
839 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
840 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
841 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
842 | |
|
|
843 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
248 | } |
844 | } |
249 | |
845 | |
250 | cl_int |
846 | cl_int |
251 | errno () |
847 | errno () |
252 | CODE: |
848 | CODE: |
253 | errno = res; |
849 | RETVAL = res; |
|
|
850 | OUTPUT: |
|
|
851 | RETVAL |
254 | |
852 | |
255 | const char * |
853 | const char * |
256 | err2str (cl_int err) |
854 | err2str (cl_int err = res) |
257 | |
855 | |
258 | const char * |
856 | const char * |
259 | enum2str (cl_uint value) |
857 | enum2str (cl_uint value) |
260 | |
858 | |
261 | void |
859 | void |
… | |
… | |
269 | list = tmpbuf (sizeof (*list) * count); |
867 | list = tmpbuf (sizeof (*list) * count); |
270 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
868 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
271 | |
869 | |
272 | EXTEND (SP, count); |
870 | EXTEND (SP, count); |
273 | for (i = 0; i < count; ++i) |
871 | for (i = 0; i < count; ++i) |
274 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
872 | PUSH_CLOBJ (stash_platform, list [i]); |
275 | |
873 | |
276 | void |
874 | void |
277 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
875 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
278 | PPCODE: |
876 | PPCODE: |
|
|
877 | CONTEXT_NOTIFY_CALLBACK; |
279 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
878 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
280 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
879 | XPUSH_CLOBJ_CONTEXT; |
281 | |
880 | |
282 | void |
881 | void |
283 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
882 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
284 | PPCODE: |
883 | PPCODE: |
285 | /* der Gipfel der Kunst */ |
884 | cl_uint device_count; |
|
|
885 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
886 | |
|
|
887 | CONTEXT_NOTIFY_CALLBACK; |
|
|
888 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
889 | XPUSH_CLOBJ_CONTEXT; |
286 | |
890 | |
287 | void |
891 | void |
288 | wait_for_events (...) |
892 | wait_for_events (...) |
289 | CODE: |
893 | CODE: |
290 | EVENT_LIST (0, items); |
894 | EVENT_LIST (0); |
291 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
895 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
292 | |
896 | |
293 | PROTOTYPES: DISABLE |
897 | PROTOTYPES: DISABLE |
294 | |
898 | |
295 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
899 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
296 | |
900 | |
297 | void |
901 | void |
298 | info (OpenCL::Platform self, cl_platform_info name) |
902 | info (OpenCL::Platform self, cl_platform_info name) |
299 | PPCODE: |
903 | PPCODE: |
300 | INFO (Platform) |
904 | INFO (Platform) |
|
|
905 | |
|
|
906 | void |
|
|
907 | unload_compiler (OpenCL::Platform self) |
|
|
908 | CODE: |
|
|
909 | #if CL_VERSION_1_2 |
|
|
910 | clUnloadPlatformCompiler (self); |
|
|
911 | #endif |
301 | |
912 | |
302 | #BEGIN:platform |
913 | #BEGIN:platform |
303 | |
914 | |
304 | void |
915 | void |
305 | profile (OpenCL::Platform self) |
916 | profile (OpenCL::Platform self) |
… | |
… | |
331 | list = tmpbuf (sizeof (*list) * count); |
942 | list = tmpbuf (sizeof (*list) * count); |
332 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
943 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
333 | |
944 | |
334 | EXTEND (SP, count); |
945 | EXTEND (SP, count); |
335 | for (i = 0; i < count; ++i) |
946 | for (i = 0; i < count; ++i) |
336 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
947 | PUSH_CLOBJ (stash_device, list [i]); |
337 | |
948 | |
338 | void |
949 | void |
339 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
950 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
340 | PPCODE: |
951 | PPCODE: |
|
|
952 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
953 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
954 | |
|
|
955 | cl_uint device_count; |
|
|
956 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
957 | |
|
|
958 | CONTEXT_NOTIFY_CALLBACK; |
|
|
959 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
960 | XPUSH_CLOBJ_CONTEXT; |
|
|
961 | |
|
|
962 | void |
|
|
963 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
964 | PPCODE: |
|
|
965 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
966 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
967 | |
|
|
968 | CONTEXT_NOTIFY_CALLBACK; |
|
|
969 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
970 | XPUSH_CLOBJ_CONTEXT; |
|
|
971 | |
|
|
972 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
973 | |
|
|
974 | void |
|
|
975 | info (OpenCL::Device self, cl_device_info name) |
|
|
976 | PPCODE: |
|
|
977 | INFO (Device) |
|
|
978 | |
|
|
979 | #if CL_VERSION_1_2 |
|
|
980 | |
|
|
981 | void |
|
|
982 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
983 | PPCODE: |
341 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
984 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
342 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
985 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
343 | |
986 | |
344 | AV *av = (AV *)SvRV (devices); |
987 | properties = SvRV (properties); |
345 | cl_uint num_devices = av_len (av) + 1; |
988 | |
346 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
989 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
990 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
991 | |
347 | int i; |
992 | int i; |
|
|
993 | for (i = 0; i < count; ++i) |
|
|
994 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
348 | |
995 | |
349 | for (i = num_devices; i--; ) |
996 | props [count] = 0; |
350 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
351 | |
997 | |
352 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
998 | cl_uint num_devices; |
353 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
999 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
1000 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
1001 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
354 | |
1002 | |
355 | void |
1003 | EXTEND (SP, num_devices); |
356 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
1004 | for (i = 0; i < count; ++i) |
357 | PPCODE: |
1005 | PUSH_CLOBJ (stash_subdevice, list [i]); |
358 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
359 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
360 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
|
|
361 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
|
|
362 | |
1006 | |
363 | MODULE = OpenCL PACKAGE = OpenCL::Device |
1007 | #endif |
364 | |
|
|
365 | void |
|
|
366 | info (OpenCL::Device self, cl_device_info name) |
|
|
367 | PPCODE: |
|
|
368 | INFO (Device) |
|
|
369 | |
1008 | |
370 | #BEGIN:device |
1009 | #BEGIN:device |
371 | |
1010 | |
372 | void |
1011 | void |
373 | type (OpenCL::Device self) |
1012 | type (OpenCL::Device self) |
374 | PPCODE: |
1013 | PPCODE: |
375 | cl_device_type value [1]; |
1014 | cl_device_type value [1]; |
376 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
1015 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
377 | EXTEND (SP, 1); |
1016 | EXTEND (SP, 1); |
378 | const int i = 0; |
1017 | const int i = 0; |
379 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1018 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
380 | |
1019 | |
381 | void |
1020 | void |
382 | vendor_id (OpenCL::Device self) |
1021 | vendor_id (OpenCL::Device self) |
383 | ALIAS: |
1022 | ALIAS: |
384 | vendor_id = CL_DEVICE_VENDOR_ID |
1023 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
405 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
1044 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
406 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
1045 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
407 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
1046 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
408 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
1047 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
409 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
1048 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
410 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
1049 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
411 | PPCODE: |
1050 | PPCODE: |
412 | cl_uint value [1]; |
1051 | cl_uint value [1]; |
413 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1052 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
414 | EXTEND (SP, 1); |
1053 | EXTEND (SP, 1); |
415 | const int i = 0; |
1054 | const int i = 0; |
… | |
… | |
529 | PPCODE: |
1168 | PPCODE: |
530 | cl_command_queue_properties value [1]; |
1169 | cl_command_queue_properties value [1]; |
531 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1170 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
532 | EXTEND (SP, 1); |
1171 | EXTEND (SP, 1); |
533 | const int i = 0; |
1172 | const int i = 0; |
534 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1173 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
535 | |
1174 | |
536 | void |
1175 | void |
537 | platform (OpenCL::Device self) |
1176 | platform (OpenCL::Device self) |
538 | PPCODE: |
1177 | PPCODE: |
539 | cl_platform_id value [1]; |
1178 | cl_platform_id value [1]; |
540 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1179 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
541 | EXTEND (SP, 1); |
1180 | EXTEND (SP, 1); |
542 | const int i = 0; |
1181 | const int i = 0; |
543 | { |
1182 | { |
544 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1183 | PUSH_CLOBJ (stash_platform, value [i]); |
545 | } |
1184 | } |
546 | |
1185 | |
547 | void |
1186 | void |
548 | name (OpenCL::Device self) |
1187 | name (OpenCL::Device self) |
549 | ALIAS: |
1188 | ALIAS: |
… | |
… | |
568 | cl_device_id value [1]; |
1207 | cl_device_id value [1]; |
569 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1208 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
570 | EXTEND (SP, 1); |
1209 | EXTEND (SP, 1); |
571 | const int i = 0; |
1210 | const int i = 0; |
572 | { |
1211 | { |
573 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1212 | PUSH_CLOBJ (stash_device, value [i]); |
574 | } |
1213 | } |
575 | |
1214 | |
576 | void |
1215 | void |
577 | partition_types_ext (OpenCL::Device self) |
1216 | partition_types_ext (OpenCL::Device self) |
578 | ALIAS: |
1217 | ALIAS: |
… | |
… | |
589 | for (i = 0; i < n; ++i) |
1228 | for (i = 0; i < n; ++i) |
590 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1229 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
591 | |
1230 | |
592 | #END:device |
1231 | #END:device |
593 | |
1232 | |
|
|
1233 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1234 | |
|
|
1235 | #if CL_VERSION_1_2 |
|
|
1236 | |
|
|
1237 | void |
|
|
1238 | DESTROY (OpenCL::SubDevice self) |
|
|
1239 | CODE: |
|
|
1240 | clReleaseDevice (self); |
|
|
1241 | |
|
|
1242 | #endif |
|
|
1243 | |
594 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1244 | MODULE = OpenCL PACKAGE = OpenCL::Context |
595 | |
1245 | |
596 | void |
1246 | void |
597 | DESTROY (OpenCL::Context context) |
1247 | DESTROY (OpenCL::Context self) |
598 | CODE: |
1248 | CODE: |
599 | clReleaseContext (context); |
1249 | clReleaseContext (self); |
600 | |
1250 | |
601 | void |
1251 | void |
602 | info (OpenCL::Context self, cl_context_info name) |
1252 | info (OpenCL::Context self, cl_context_info name) |
603 | PPCODE: |
1253 | PPCODE: |
604 | INFO (Context) |
1254 | INFO (Context) |
605 | |
1255 | |
606 | void |
1256 | void |
607 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1257 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
608 | PPCODE: |
1258 | PPCODE: |
609 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1259 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
610 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1260 | XPUSH_CLOBJ (stash_queue, queue); |
611 | |
1261 | |
612 | void |
1262 | void |
613 | user_event (OpenCL::Context self) |
1263 | user_event (OpenCL::Context self) |
614 | PPCODE: |
1264 | PPCODE: |
615 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1265 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
616 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1266 | XPUSH_CLOBJ (stash_userevent, ev); |
617 | |
1267 | |
618 | void |
1268 | void |
619 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1269 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
620 | PPCODE: |
1270 | PPCODE: |
621 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1271 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
622 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1272 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
623 | |
1273 | |
624 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1274 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
625 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1275 | XPUSH_CLOBJ (stash_bufferobj, mem); |
626 | |
1276 | |
627 | void |
1277 | void |
628 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1278 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
629 | PPCODE: |
1279 | PPCODE: |
630 | STRLEN len; |
1280 | STRLEN len; |
631 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1281 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
632 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1282 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
633 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1283 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
634 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1284 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
635 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1285 | XPUSH_CLOBJ (stash_bufferobj, mem); |
636 | |
1286 | |
|
|
1287 | #if CL_VERSION_1_2 |
|
|
1288 | |
637 | void |
1289 | void |
638 | 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) |
1290 | 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) |
639 | PPCODE: |
1291 | PPCODE: |
640 | STRLEN len; |
1292 | STRLEN len; |
641 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1293 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
642 | const cl_image_format format = { channel_order, channel_type }; |
1294 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1295 | const cl_image_desc desc = { |
|
|
1296 | type, |
|
|
1297 | width, height, depth, |
|
|
1298 | array_size, row_pitch, slice_pitch, |
|
|
1299 | num_mip_level, num_samples, |
|
|
1300 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
|
|
1301 | }; |
643 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1302 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
644 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1303 | HV *stash = stash_image; |
|
|
1304 | switch (type) |
|
|
1305 | { |
|
|
1306 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1307 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1308 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1309 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1310 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1311 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1312 | } |
|
|
1313 | XPUSH_CLOBJ (stash, mem); |
645 | |
1314 | |
|
|
1315 | #endif |
|
|
1316 | |
646 | void |
1317 | void |
647 | 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) |
1318 | 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) |
648 | PPCODE: |
1319 | PPCODE: |
649 | STRLEN len; |
1320 | STRLEN len; |
650 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1321 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
651 | const cl_image_format format = { channel_order, channel_type }; |
1322 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1323 | #if PREFER_1_1 |
|
|
1324 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1325 | #else |
|
|
1326 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1327 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1328 | #endif |
|
|
1329 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1330 | |
|
|
1331 | void |
|
|
1332 | 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) |
|
|
1333 | PPCODE: |
|
|
1334 | STRLEN len; |
|
|
1335 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1336 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1337 | #if PREFER_1_1 |
652 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1338 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
653 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1339 | #else |
|
|
1340 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1341 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1342 | #endif |
|
|
1343 | XPUSH_CLOBJ (stash_image3d, mem); |
654 | |
1344 | |
655 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1345 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
656 | |
1346 | |
657 | void |
1347 | void |
658 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
1348 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
659 | PPCODE: |
1349 | PPCODE: |
660 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
1350 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
661 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1351 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1352 | |
|
|
1353 | void |
|
|
1354 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1355 | PPCODE: |
|
|
1356 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1357 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1358 | |
|
|
1359 | #if CL_VERSION_1_2 |
|
|
1360 | |
|
|
1361 | void |
|
|
1362 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1363 | ALIAS: |
|
|
1364 | PPCODE: |
|
|
1365 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1366 | cl_gl_object_type type; |
|
|
1367 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1368 | HV *stash = stash_memory; |
|
|
1369 | switch (type) |
|
|
1370 | { |
|
|
1371 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1372 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1373 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1374 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1375 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1376 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1377 | } |
|
|
1378 | XPUSH_CLOBJ (stash, mem); |
|
|
1379 | |
|
|
1380 | #endif |
662 | |
1381 | |
663 | void |
1382 | void |
664 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1383 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
665 | PPCODE: |
1384 | PPCODE: |
|
|
1385 | #if PREFER_1_1 |
666 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
1386 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
667 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1387 | #else |
|
|
1388 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1389 | #endif |
|
|
1390 | XPUSH_CLOBJ (stash_image2d, mem); |
668 | |
1391 | |
669 | void |
1392 | void |
670 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1393 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
671 | PPCODE: |
1394 | PPCODE: |
|
|
1395 | #if PREFER_1_1 |
672 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
1396 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
673 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1397 | #else |
674 | |
1398 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
675 | void |
1399 | #endif |
676 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
1400 | XPUSH_CLOBJ (stash_image3d, mem); |
677 | PPCODE: |
|
|
678 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
679 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
680 | |
1401 | |
681 | #endif |
1402 | #endif |
682 | |
1403 | |
683 | void |
1404 | void |
684 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1405 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
704 | |
1425 | |
705 | void |
1426 | void |
706 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1427 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
707 | PPCODE: |
1428 | PPCODE: |
708 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1429 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
709 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1430 | XPUSH_CLOBJ (stash_sampler, sampler); |
710 | |
1431 | |
711 | void |
1432 | void |
712 | program_with_source (OpenCL::Context self, SV *program) |
1433 | program_with_source (OpenCL::Context self, SV *program) |
713 | PPCODE: |
1434 | PPCODE: |
714 | STRLEN len; |
1435 | STRLEN len; |
715 | size_t len2; |
1436 | size_t len2; |
716 | const char *ptr = SvPVbyte (program, len); |
1437 | const char *ptr = SvPVbyte (program, len); |
717 | |
1438 | |
718 | len2 = len; |
1439 | len2 = len; |
719 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1440 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
720 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1441 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1442 | |
|
|
1443 | void |
|
|
1444 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1445 | PPCODE: |
|
|
1446 | cl_uint device_count; |
|
|
1447 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1448 | |
|
|
1449 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1450 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1451 | |
|
|
1452 | binaries = SvRV (binaries); |
|
|
1453 | |
|
|
1454 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1455 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1456 | |
|
|
1457 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1458 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1459 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1460 | |
|
|
1461 | int i; |
|
|
1462 | for (i = 0; i < device_count; ++i) |
|
|
1463 | { |
|
|
1464 | STRLEN len; |
|
|
1465 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1466 | length_list [i] = len; |
|
|
1467 | } |
|
|
1468 | |
|
|
1469 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1470 | length_list, binary_list, |
|
|
1471 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1472 | |
|
|
1473 | EXTEND (SP, 2); |
|
|
1474 | PUSH_CLOBJ (stash_program, prog); |
|
|
1475 | |
|
|
1476 | if (GIMME_V == G_ARRAY) |
|
|
1477 | { |
|
|
1478 | AV *av = newAV (); |
|
|
1479 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1480 | |
|
|
1481 | for (i = device_count; i--; ) |
|
|
1482 | av_store (av, i, newSViv (status_list [i])); |
|
|
1483 | } |
|
|
1484 | |
|
|
1485 | #if CL_VERSION_1_2 |
|
|
1486 | |
|
|
1487 | void |
|
|
1488 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1489 | PPCODE: |
|
|
1490 | cl_uint device_count; |
|
|
1491 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1492 | |
|
|
1493 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1494 | |
|
|
1495 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1496 | |
|
|
1497 | void |
|
|
1498 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1499 | CODE: |
|
|
1500 | cl_uint device_count = 0; |
|
|
1501 | cl_device_id *device_list = 0; |
|
|
1502 | |
|
|
1503 | if (SvOK (devices)) |
|
|
1504 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1505 | |
|
|
1506 | cl_uint program_count; |
|
|
1507 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1508 | |
|
|
1509 | void *user_data; |
|
|
1510 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1511 | |
|
|
1512 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1513 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1514 | |
|
|
1515 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1516 | |
|
|
1517 | #endif |
721 | |
1518 | |
722 | #BEGIN:context |
1519 | #BEGIN:context |
723 | |
1520 | |
724 | void |
1521 | void |
725 | reference_count (OpenCL::Context self) |
1522 | reference_count (OpenCL::Context self) |
… | |
… | |
742 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1539 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
743 | int i, n = size / sizeof (*value); |
1540 | int i, n = size / sizeof (*value); |
744 | EXTEND (SP, n); |
1541 | EXTEND (SP, n); |
745 | for (i = 0; i < n; ++i) |
1542 | for (i = 0; i < n; ++i) |
746 | { |
1543 | { |
747 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1544 | PUSH_CLOBJ (stash_device, value [i]); |
748 | } |
1545 | } |
749 | |
1546 | |
750 | void |
1547 | void |
751 | properties (OpenCL::Context self) |
1548 | properties (OpenCL::Context self) |
752 | PPCODE: |
1549 | PPCODE: |
… | |
… | |
767 | DESTROY (OpenCL::Queue self) |
1564 | DESTROY (OpenCL::Queue self) |
768 | CODE: |
1565 | CODE: |
769 | clReleaseCommandQueue (self); |
1566 | clReleaseCommandQueue (self); |
770 | |
1567 | |
771 | void |
1568 | void |
772 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1569 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1570 | ALIAS: |
|
|
1571 | enqueue_read_buffer = 0 |
773 | PPCODE: |
1572 | PPCODE: |
774 | cl_event ev = 0; |
|
|
775 | EVENT_LIST (6, items - 6); |
1573 | EVENT_LIST (6); |
776 | |
1574 | |
777 | SvUPGRADE (data, SVt_PV); |
1575 | SvUPGRADE (data, SVt_PV); |
778 | SvGROW (data, len); |
1576 | SvGROW (data, len); |
779 | SvPOK_only (data); |
1577 | SvPOK_only (data); |
780 | SvCUR_set (data, len); |
1578 | SvCUR_set (data, len); |
|
|
1579 | |
|
|
1580 | cl_event ev = 0; |
781 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1581 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
782 | |
1582 | |
783 | if (ev) |
1583 | if (ev) |
784 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1584 | XPUSH_CLOBJ (stash_event, ev); |
785 | |
1585 | |
786 | void |
1586 | void |
787 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1587 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1588 | ALIAS: |
|
|
1589 | enqueue_write_buffer = 0 |
788 | PPCODE: |
1590 | PPCODE: |
789 | cl_event ev = 0; |
1591 | EVENT_LIST (5); |
|
|
1592 | |
790 | STRLEN len; |
1593 | STRLEN len; |
791 | char *ptr = SvPVbyte (data, len); |
1594 | char *ptr = SvPVbyte (data, len); |
792 | EVENT_LIST (5, items - 5); |
|
|
793 | |
1595 | |
|
|
1596 | cl_event ev = 0; |
794 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1597 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
795 | |
1598 | |
796 | if (ev) |
1599 | if (ev) |
797 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1600 | XPUSH_CLOBJ (stash_event, ev); |
798 | |
1601 | |
|
|
1602 | #if CL_VERSION_1_2 |
|
|
1603 | |
799 | void |
1604 | void |
800 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1605 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1606 | ALIAS: |
|
|
1607 | enqueue_fill_buffer = 0 |
801 | PPCODE: |
1608 | PPCODE: |
|
|
1609 | EVENT_LIST (5); |
|
|
1610 | |
|
|
1611 | STRLEN len; |
|
|
1612 | char *ptr = SvPVbyte (data, len); |
|
|
1613 | |
802 | cl_event ev = 0; |
1614 | cl_event ev = 0; |
|
|
1615 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1616 | |
|
|
1617 | if (ev) |
|
|
1618 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1619 | |
|
|
1620 | void |
|
|
1621 | 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, ...) |
|
|
1622 | ALIAS: |
|
|
1623 | enqueue_fill_image = 0 |
|
|
1624 | PPCODE: |
|
|
1625 | EVENT_LIST (12); |
|
|
1626 | |
|
|
1627 | const size_t origin [3] = { x, y, z }; |
|
|
1628 | const size_t region [3] = { width, height, depth }; |
|
|
1629 | |
|
|
1630 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1631 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1632 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1633 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1634 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1635 | cl_image_format format; |
|
|
1636 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1637 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1638 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1639 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1640 | |
|
|
1641 | cl_event ev = 0; |
|
|
1642 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1643 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1644 | |
|
|
1645 | if (ev) |
|
|
1646 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1647 | |
|
|
1648 | #endif |
|
|
1649 | |
|
|
1650 | void |
|
|
1651 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1652 | ALIAS: |
|
|
1653 | enqueue_copy_buffer = 0 |
|
|
1654 | PPCODE: |
803 | EVENT_LIST (6, items - 6); |
1655 | EVENT_LIST (6); |
804 | |
1656 | |
|
|
1657 | cl_event ev = 0; |
805 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1658 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
806 | |
1659 | |
807 | if (ev) |
1660 | if (ev) |
808 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1661 | XPUSH_CLOBJ (stash_event, ev); |
809 | |
1662 | |
810 | void |
1663 | void |
811 | 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, ...) |
1664 | 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, ...) |
|
|
1665 | ALIAS: |
|
|
1666 | enqueue_read_buffer_rect = 0 |
812 | PPCODE: |
1667 | PPCODE: |
813 | cl_event ev = 0; |
1668 | EVENT_LIST (17); |
|
|
1669 | |
814 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1670 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
815 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1671 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
816 | const size_t region[3] = { width, height, depth }; |
1672 | const size_t region[3] = { width, height, depth }; |
817 | EVENT_LIST (17, items - 17); |
|
|
818 | |
1673 | |
819 | if (!buf_row_pitch) |
1674 | if (!buf_row_pitch) |
820 | buf_row_pitch = region [0]; |
1675 | buf_row_pitch = region [0]; |
821 | |
1676 | |
822 | if (!buf_slice_pitch) |
1677 | if (!buf_slice_pitch) |
… | |
… | |
832 | |
1687 | |
833 | SvUPGRADE (data, SVt_PV); |
1688 | SvUPGRADE (data, SVt_PV); |
834 | SvGROW (data, len); |
1689 | SvGROW (data, len); |
835 | SvPOK_only (data); |
1690 | SvPOK_only (data); |
836 | SvCUR_set (data, len); |
1691 | SvCUR_set (data, len); |
|
|
1692 | |
|
|
1693 | cl_event ev = 0; |
837 | 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)); |
1694 | 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)); |
838 | |
1695 | |
839 | if (ev) |
1696 | if (ev) |
840 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1697 | XPUSH_CLOBJ (stash_event, ev); |
841 | |
1698 | |
842 | void |
1699 | void |
843 | 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, ...) |
1700 | 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, ...) |
|
|
1701 | ALIAS: |
|
|
1702 | enqueue_write_buffer_rect = 0 |
844 | PPCODE: |
1703 | PPCODE: |
845 | cl_event ev = 0; |
1704 | EVENT_LIST (17); |
|
|
1705 | |
846 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1706 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
847 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1707 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
848 | const size_t region[3] = { width, height, depth }; |
1708 | const size_t region[3] = { width, height, depth }; |
849 | STRLEN len; |
1709 | STRLEN len; |
850 | char *ptr = SvPVbyte (data, len); |
1710 | char *ptr = SvPVbyte (data, len); |
851 | EVENT_LIST (17, items - 17); |
|
|
852 | |
1711 | |
853 | if (!buf_row_pitch) |
1712 | if (!buf_row_pitch) |
854 | buf_row_pitch = region [0]; |
1713 | buf_row_pitch = region [0]; |
855 | |
1714 | |
856 | if (!buf_slice_pitch) |
1715 | if (!buf_slice_pitch) |
… | |
… | |
865 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1724 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
866 | |
1725 | |
867 | if (len < min_len) |
1726 | if (len < min_len) |
868 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1727 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
869 | |
1728 | |
|
|
1729 | cl_event ev = 0; |
870 | 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)); |
1730 | 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)); |
871 | |
1731 | |
872 | if (ev) |
1732 | if (ev) |
873 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1733 | XPUSH_CLOBJ (stash_event, ev); |
874 | |
1734 | |
875 | void |
1735 | void |
876 | 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, ...) |
1736 | 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, ...) |
|
|
1737 | ALIAS: |
|
|
1738 | enqueue_copy_buffer_rect = 0 |
877 | PPCODE: |
1739 | PPCODE: |
878 | cl_event ev = 0; |
1740 | EVENT_LIST (16); |
|
|
1741 | |
879 | 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 }; |
880 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1743 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
881 | const size_t region[3] = { width, height, depth }; |
1744 | const size_t region[3] = { width, height, depth }; |
882 | EVENT_LIST (16, items - 16); |
|
|
883 | |
1745 | |
|
|
1746 | cl_event ev = 0; |
884 | 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)); |
1747 | 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)); |
885 | |
1748 | |
886 | if (ev) |
1749 | if (ev) |
887 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1750 | XPUSH_CLOBJ (stash_event, ev); |
888 | |
1751 | |
889 | void |
1752 | void |
890 | 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, ...) |
1753 | 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, ...) |
|
|
1754 | ALIAS: |
|
|
1755 | enqueue_read_image = 0 |
891 | PPCODE: |
1756 | PPCODE: |
892 | cl_event ev = 0; |
1757 | EVENT_LIST (12); |
|
|
1758 | |
893 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1759 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
894 | const size_t region[3] = { width, height, depth }; |
1760 | const size_t region[3] = { width, height, depth }; |
895 | EVENT_LIST (12, items - 12); |
|
|
896 | |
1761 | |
897 | if (!row_pitch) |
1762 | if (!row_pitch) |
898 | row_pitch = img_row_pitch (src); |
1763 | row_pitch = img_row_pitch (src); |
899 | |
1764 | |
900 | if (depth > 1 && !slice_pitch) |
1765 | if (depth > 1 && !slice_pitch) |
… | |
… | |
904 | |
1769 | |
905 | SvUPGRADE (data, SVt_PV); |
1770 | SvUPGRADE (data, SVt_PV); |
906 | SvGROW (data, len); |
1771 | SvGROW (data, len); |
907 | SvPOK_only (data); |
1772 | SvPOK_only (data); |
908 | SvCUR_set (data, len); |
1773 | SvCUR_set (data, len); |
|
|
1774 | |
|
|
1775 | cl_event ev = 0; |
909 | 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)); |
1776 | 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)); |
910 | |
1777 | |
911 | if (ev) |
1778 | if (ev) |
912 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1779 | XPUSH_CLOBJ (stash_event, ev); |
913 | |
1780 | |
914 | void |
1781 | void |
915 | 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, ...) |
1782 | 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, ...) |
|
|
1783 | ALIAS: |
|
|
1784 | enqueue_write_image = 0 |
916 | PPCODE: |
1785 | PPCODE: |
917 | cl_event ev = 0; |
1786 | EVENT_LIST (12); |
|
|
1787 | |
918 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1788 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
919 | const size_t region[3] = { width, height, depth }; |
1789 | const size_t region[3] = { width, height, depth }; |
920 | STRLEN len; |
1790 | STRLEN len; |
921 | char *ptr = SvPVbyte (data, len); |
1791 | char *ptr = SvPVbyte (data, len); |
922 | EVENT_LIST (12, items - 12); |
|
|
923 | |
1792 | |
924 | if (!row_pitch) |
1793 | if (!row_pitch) |
925 | row_pitch = img_row_pitch (dst); |
1794 | row_pitch = img_row_pitch (dst); |
926 | |
1795 | |
927 | if (depth > 1 && !slice_pitch) |
1796 | if (depth > 1 && !slice_pitch) |
… | |
… | |
930 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1799 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
931 | |
1800 | |
932 | if (len < min_len) |
1801 | if (len < min_len) |
933 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1802 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
934 | |
1803 | |
|
|
1804 | cl_event ev = 0; |
935 | 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)); |
1805 | 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)); |
936 | |
1806 | |
937 | if (ev) |
1807 | if (ev) |
938 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1808 | XPUSH_CLOBJ (stash_event, ev); |
939 | |
1809 | |
940 | void |
1810 | void |
941 | 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, ...) |
1811 | 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, ...) |
|
|
1812 | ALIAS: |
|
|
1813 | enqueue_copy_image = 0 |
942 | PPCODE: |
1814 | PPCODE: |
943 | cl_event ev = 0; |
1815 | EVENT_LIST (12); |
|
|
1816 | |
944 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1817 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
945 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1818 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
946 | const size_t region[3] = { width, height, depth }; |
1819 | const size_t region[3] = { width, height, depth }; |
947 | EVENT_LIST (12, items - 12); |
|
|
948 | |
1820 | |
|
|
1821 | cl_event ev = 0; |
949 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1822 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
950 | |
1823 | |
951 | if (ev) |
1824 | if (ev) |
952 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1825 | XPUSH_CLOBJ (stash_event, ev); |
953 | |
1826 | |
954 | void |
1827 | void |
955 | 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, ...) |
1828 | 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, ...) |
|
|
1829 | ALIAS: |
|
|
1830 | enqueue_copy_image_to_buffer = 0 |
956 | PPCODE: |
1831 | PPCODE: |
|
|
1832 | EVENT_LIST (10); |
|
|
1833 | |
|
|
1834 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1835 | const size_t region [3] = { width, height, depth }; |
|
|
1836 | |
957 | cl_event ev = 0; |
1837 | cl_event ev = 0; |
|
|
1838 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1839 | |
|
|
1840 | if (ev) |
|
|
1841 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1842 | |
|
|
1843 | void |
|
|
1844 | 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, ...) |
|
|
1845 | ALIAS: |
|
|
1846 | enqueue_copy_buffer_to_image = 0 |
|
|
1847 | PPCODE: |
|
|
1848 | EVENT_LIST (10); |
|
|
1849 | |
958 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1850 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1851 | const size_t region [3] = { width, height, depth }; |
|
|
1852 | |
|
|
1853 | cl_event ev = 0; |
|
|
1854 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1855 | |
|
|
1856 | if (ev) |
|
|
1857 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1858 | |
|
|
1859 | void |
|
|
1860 | 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, ...) |
|
|
1861 | ALIAS: |
|
|
1862 | enqueue_map_buffer = 0 |
|
|
1863 | PPCODE: |
|
|
1864 | EVENT_LIST (6); |
|
|
1865 | |
|
|
1866 | size_t cb = SvIV (cb_); |
|
|
1867 | |
|
|
1868 | if (!SvOK (cb_)) |
|
|
1869 | { |
|
|
1870 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1871 | cb -= offset; |
|
|
1872 | } |
|
|
1873 | |
|
|
1874 | cl_event ev; |
|
|
1875 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1876 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1877 | |
|
|
1878 | void |
|
|
1879 | 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, ...) |
|
|
1880 | ALIAS: |
|
|
1881 | enqueue_map_image = 0 |
|
|
1882 | PPCODE: |
|
|
1883 | size_t width = SvIV (width_); |
|
|
1884 | if (!SvOK (width_)) |
|
|
1885 | { |
|
|
1886 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1887 | width -= x; |
|
|
1888 | } |
|
|
1889 | |
|
|
1890 | size_t height = SvIV (width_); |
|
|
1891 | if (!SvOK (height_)) |
|
|
1892 | { |
|
|
1893 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1894 | height -= y; |
|
|
1895 | } |
|
|
1896 | |
|
|
1897 | size_t depth = SvIV (width_); |
|
|
1898 | if (!SvOK (depth_)) |
|
|
1899 | { |
|
|
1900 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1901 | depth -= z; |
|
|
1902 | |
|
|
1903 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1904 | if (!depth) |
|
|
1905 | depth = 1; |
|
|
1906 | } |
|
|
1907 | |
|
|
1908 | const size_t origin[3] = { x, y, z }; |
959 | const size_t region[3] = { width, height, depth }; |
1909 | const size_t region[3] = { width, height, depth }; |
|
|
1910 | size_t row_pitch, slice_pitch; |
960 | EVENT_LIST (10, items - 10); |
1911 | EVENT_LIST (10); |
961 | |
1912 | |
|
|
1913 | cl_event ev; |
|
|
1914 | 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)); |
|
|
1915 | |
|
|
1916 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1917 | : row_pitch ? row_pitch * region [1] |
|
|
1918 | : region [0]; |
|
|
1919 | |
|
|
1920 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1921 | |
|
|
1922 | void |
|
|
1923 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1924 | PPCODE: |
|
|
1925 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2); |
|
|
1926 | if (GIMME_V != G_VOID) |
|
|
1927 | { |
|
|
1928 | clRetainEvent (mapped->event); |
|
|
1929 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1930 | } |
|
|
1931 | |
|
|
1932 | void |
|
|
1933 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1934 | ALIAS: |
|
|
1935 | enqueue_task = 0 |
|
|
1936 | PPCODE: |
|
|
1937 | EVENT_LIST (2); |
|
|
1938 | |
|
|
1939 | cl_event ev = 0; |
962 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1940 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
963 | |
1941 | |
964 | if (ev) |
1942 | if (ev) |
965 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1943 | XPUSH_CLOBJ (stash_event, ev); |
966 | |
1944 | |
967 | void |
1945 | void |
968 | 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, ...) |
|
|
969 | PPCODE: |
|
|
970 | cl_event ev = 0; |
|
|
971 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
972 | const size_t region[3] = { width, height, depth }; |
|
|
973 | EVENT_LIST (10, items - 10); |
|
|
974 | |
|
|
975 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
976 | |
|
|
977 | if (ev) |
|
|
978 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
979 | |
|
|
980 | void |
|
|
981 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
982 | PPCODE: |
|
|
983 | cl_event ev = 0; |
|
|
984 | EVENT_LIST (2, items - 2); |
|
|
985 | |
|
|
986 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
987 | |
|
|
988 | if (ev) |
|
|
989 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
990 | |
|
|
991 | void |
|
|
992 | 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, ...) |
1946 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1947 | ALIAS: |
|
|
1948 | enqueue_nd_range_kernel = 0 |
993 | PPCODE: |
1949 | PPCODE: |
994 | cl_event ev = 0; |
1950 | EVENT_LIST (5); |
|
|
1951 | |
995 | size_t *gwo = 0, *gws, *lws = 0; |
1952 | size_t *gwo = 0, *gws, *lws = 0; |
996 | int gws_len; |
1953 | int gws_len; |
997 | size_t *lists; |
1954 | size_t *lists; |
998 | int i; |
1955 | int i; |
999 | EVENT_LIST (5, items - 5); |
|
|
1000 | |
1956 | |
1001 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1957 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1002 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1958 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1003 | |
1959 | |
1004 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1960 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1005 | |
1961 | |
1006 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1962 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1007 | |
1963 | |
1008 | gws = lists + gws_len * 0; |
1964 | gws = lists + gws_len * 0; |
1009 | for (i = 0; i < gws_len; ++i) |
1965 | for (i = 0; i < gws_len; ++i) |
|
|
1966 | { |
1010 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1967 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1968 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1969 | if (!gws [i]) |
|
|
1970 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1971 | } |
1011 | |
1972 | |
1012 | if (SvOK (global_work_offset)) |
1973 | if (SvOK (global_work_offset)) |
1013 | { |
1974 | { |
1014 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1975 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1015 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1976 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1022 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1983 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1023 | } |
1984 | } |
1024 | |
1985 | |
1025 | if (SvOK (local_work_size)) |
1986 | if (SvOK (local_work_size)) |
1026 | { |
1987 | { |
1027 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1988 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1028 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1989 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1029 | |
1990 | |
1030 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1991 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1031 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1992 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1032 | |
1993 | |
1033 | lws = lists + gws_len * 2; |
1994 | lws = lists + gws_len * 2; |
1034 | for (i = 0; i < gws_len; ++i) |
1995 | for (i = 0; i < gws_len; ++i) |
|
|
1996 | { |
1035 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
1997 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1998 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1999 | if (!lws [i]) |
|
|
2000 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
2001 | } |
1036 | } |
2002 | } |
1037 | |
2003 | |
|
|
2004 | cl_event ev = 0; |
1038 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2005 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1039 | |
2006 | |
1040 | if (ev) |
2007 | if (ev) |
1041 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2008 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2009 | |
|
|
2010 | #if CL_VERSION_1_2 |
|
|
2011 | |
|
|
2012 | void |
|
|
2013 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
2014 | ALIAS: |
|
|
2015 | enqueue_migrate_mem_objects = 0 |
|
|
2016 | PPCODE: |
|
|
2017 | EVENT_LIST (3); |
|
|
2018 | |
|
|
2019 | cl_uint obj_count; |
|
|
2020 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2021 | |
|
|
2022 | cl_event ev = 0; |
|
|
2023 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2024 | |
|
|
2025 | if (ev) |
|
|
2026 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2027 | |
|
|
2028 | #endif |
1042 | |
2029 | |
1043 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2030 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1044 | |
2031 | |
1045 | void |
2032 | void |
1046 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2033 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1047 | ALIAS: |
2034 | ALIAS: |
|
|
2035 | release_gl_objects = 1 |
|
|
2036 | enqueue_acquire_gl_objects = 0 |
1048 | enqueue_release_gl_objects = 1 |
2037 | enqueue_release_gl_objects = 1 |
1049 | CODE: |
2038 | PPCODE: |
1050 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
2039 | EVENT_LIST (2); |
1051 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
2040 | |
|
|
2041 | cl_uint obj_count; |
|
|
2042 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
1052 | |
2043 | |
1053 | cl_event ev = 0; |
2044 | cl_event ev = 0; |
1054 | EVENT_LIST (2, items - 2); |
|
|
1055 | AV *av = (AV *)SvRV (objects); |
|
|
1056 | cl_uint num_objects = av_len (av) + 1; |
|
|
1057 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1058 | int i; |
|
|
1059 | |
|
|
1060 | for (i = num_objects; i--; ) |
|
|
1061 | object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1062 | |
2045 | |
1063 | if (ix) |
2046 | if (ix) |
1064 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2047 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1065 | else |
2048 | else |
1066 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2049 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1067 | |
2050 | |
1068 | if (ev) |
2051 | if (ev) |
1069 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2052 | XPUSH_CLOBJ (stash_event, ev); |
1070 | |
2053 | |
1071 | #endif |
2054 | #endif |
1072 | |
2055 | |
1073 | void |
2056 | void |
1074 | enqueue_marker (OpenCL::Queue self) |
|
|
1075 | PPCODE: |
|
|
1076 | cl_event ev; |
|
|
1077 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1078 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1079 | |
|
|
1080 | void |
|
|
1081 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
2057 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2058 | ALIAS: |
|
|
2059 | enqueue_wait_for_events = 0 |
1082 | CODE: |
2060 | CODE: |
1083 | EVENT_LIST (1, items - 1); |
2061 | EVENT_LIST (1); |
|
|
2062 | #if PREFER_1_1 |
1084 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
2063 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2064 | #else |
|
|
2065 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2066 | #endif |
1085 | |
2067 | |
1086 | void |
2068 | void |
1087 | enqueue_barrier (OpenCL::Queue self) |
2069 | marker (OpenCL::Queue self, ...) |
|
|
2070 | ALIAS: |
|
|
2071 | enqueue_marker = 0 |
1088 | CODE: |
2072 | PPCODE: |
|
|
2073 | EVENT_LIST (1); |
|
|
2074 | cl_event ev = 0; |
|
|
2075 | #if PREFER_1_1 |
|
|
2076 | if (!event_list_count) |
|
|
2077 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2078 | else |
|
|
2079 | #if CL_VERSION_1_2 |
|
|
2080 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2081 | #else |
|
|
2082 | { |
|
|
2083 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2084 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2085 | } |
|
|
2086 | #endif |
|
|
2087 | #else |
|
|
2088 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2089 | #endif |
|
|
2090 | if (ev) |
|
|
2091 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2092 | |
|
|
2093 | void |
|
|
2094 | barrier (OpenCL::Queue self, ...) |
|
|
2095 | ALIAS: |
|
|
2096 | enqueue_barrier = 0 |
|
|
2097 | PPCODE: |
|
|
2098 | EVENT_LIST (1); |
|
|
2099 | cl_event ev = 0; |
|
|
2100 | #if PREFER_1_1 |
|
|
2101 | if (!event_list_count && GIMME_V == G_VOID) |
1089 | NEED_SUCCESS (EnqueueBarrier, (self)); |
2102 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2103 | else |
|
|
2104 | #if CL_VERSION_1_2 |
|
|
2105 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2106 | #else |
|
|
2107 | { |
|
|
2108 | if (event_list_count) |
|
|
2109 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2110 | |
|
|
2111 | if (GIMME_V != G_VOID) |
|
|
2112 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2113 | } |
|
|
2114 | #endif |
|
|
2115 | #else |
|
|
2116 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2117 | #endif |
|
|
2118 | if (ev) |
|
|
2119 | XPUSH_CLOBJ (stash_event, ev); |
1090 | |
2120 | |
1091 | void |
2121 | void |
1092 | flush (OpenCL::Queue self) |
2122 | flush (OpenCL::Queue self) |
1093 | CODE: |
2123 | CODE: |
1094 | NEED_SUCCESS (Flush, (self)); |
2124 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1112 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
2142 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1113 | EXTEND (SP, 1); |
2143 | EXTEND (SP, 1); |
1114 | const int i = 0; |
2144 | const int i = 0; |
1115 | { |
2145 | { |
1116 | NEED_SUCCESS (RetainContext, (value [i])); |
2146 | NEED_SUCCESS (RetainContext, (value [i])); |
1117 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2147 | PUSH_CLOBJ (stash_context, value [i]); |
1118 | } |
2148 | } |
1119 | |
2149 | |
1120 | void |
2150 | void |
1121 | device (OpenCL::Queue self) |
2151 | device (OpenCL::Queue self) |
1122 | PPCODE: |
2152 | PPCODE: |
1123 | cl_device_id value [1]; |
2153 | cl_device_id value [1]; |
1124 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
2154 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1125 | EXTEND (SP, 1); |
2155 | EXTEND (SP, 1); |
1126 | const int i = 0; |
2156 | const int i = 0; |
1127 | { |
2157 | { |
1128 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2158 | PUSH_CLOBJ (stash_device, value [i]); |
1129 | } |
2159 | } |
1130 | |
2160 | |
1131 | void |
2161 | void |
1132 | reference_count (OpenCL::Queue self) |
2162 | reference_count (OpenCL::Queue self) |
1133 | PPCODE: |
2163 | PPCODE: |
… | |
… | |
1142 | PPCODE: |
2172 | PPCODE: |
1143 | cl_command_queue_properties value [1]; |
2173 | cl_command_queue_properties value [1]; |
1144 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
2174 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1145 | EXTEND (SP, 1); |
2175 | EXTEND (SP, 1); |
1146 | const int i = 0; |
2176 | const int i = 0; |
1147 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2177 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1148 | |
2178 | |
1149 | #END:command_queue |
2179 | #END:command_queue |
1150 | |
2180 | |
1151 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2181 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1152 | |
2182 | |
… | |
… | |
1167 | PPCODE: |
2197 | PPCODE: |
1168 | cl_mem_object_type value [1]; |
2198 | cl_mem_object_type value [1]; |
1169 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
2199 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1170 | EXTEND (SP, 1); |
2200 | EXTEND (SP, 1); |
1171 | const int i = 0; |
2201 | const int i = 0; |
1172 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2202 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1173 | |
2203 | |
1174 | void |
2204 | void |
1175 | flags (OpenCL::Memory self) |
2205 | flags (OpenCL::Memory self) |
1176 | PPCODE: |
2206 | PPCODE: |
1177 | cl_mem_flags value [1]; |
2207 | cl_mem_flags value [1]; |
1178 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
2208 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1179 | EXTEND (SP, 1); |
2209 | EXTEND (SP, 1); |
1180 | const int i = 0; |
2210 | const int i = 0; |
1181 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2211 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1182 | |
2212 | |
1183 | void |
2213 | void |
1184 | size (OpenCL::Memory self) |
2214 | size (OpenCL::Memory self) |
1185 | ALIAS: |
2215 | ALIAS: |
1186 | size = CL_MEM_SIZE |
2216 | size = CL_MEM_SIZE |
… | |
… | |
1220 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
2250 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1221 | EXTEND (SP, 1); |
2251 | EXTEND (SP, 1); |
1222 | const int i = 0; |
2252 | const int i = 0; |
1223 | { |
2253 | { |
1224 | NEED_SUCCESS (RetainContext, (value [i])); |
2254 | NEED_SUCCESS (RetainContext, (value [i])); |
1225 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2255 | PUSH_CLOBJ (stash_context, value [i]); |
1226 | } |
2256 | } |
1227 | |
2257 | |
1228 | void |
2258 | void |
1229 | associated_memobject (OpenCL::Memory self) |
2259 | associated_memobject (OpenCL::Memory self) |
1230 | PPCODE: |
2260 | PPCODE: |
… | |
… | |
1232 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
2262 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1233 | EXTEND (SP, 1); |
2263 | EXTEND (SP, 1); |
1234 | const int i = 0; |
2264 | const int i = 0; |
1235 | { |
2265 | { |
1236 | NEED_SUCCESS (RetainMemObject, (value [i])); |
2266 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1237 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
2267 | PUSH_CLOBJ (stash_memory, value [i]); |
1238 | } |
2268 | } |
1239 | |
2269 | |
1240 | #END:mem |
2270 | #END:mem |
1241 | |
2271 | |
1242 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2272 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
… | |
… | |
1244 | void |
2274 | void |
1245 | gl_object_info (OpenCL::Memory self) |
2275 | gl_object_info (OpenCL::Memory self) |
1246 | PPCODE: |
2276 | PPCODE: |
1247 | cl_gl_object_type type; |
2277 | cl_gl_object_type type; |
1248 | cl_GLuint name; |
2278 | cl_GLuint name; |
1249 | NEED_SUCCESS (clGetGLObjectInfo, (self, &type, &name)); |
2279 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
1250 | EXTEND (SP, 2); |
2280 | EXTEND (SP, 2); |
1251 | PUSHs (sv_2mortal (newSVuv (type))); |
2281 | PUSHs (sv_2mortal (newSVuv (type))); |
1252 | PUSHs (sv_2mortal (newSVuv (name))); |
2282 | PUSHs (sv_2mortal (newSVuv (name))); |
1253 | |
2283 | |
1254 | #endif |
2284 | #endif |
… | |
… | |
1262 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
2292 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1263 | |
2293 | |
1264 | cl_buffer_region crdata = { origin, size }; |
2294 | cl_buffer_region crdata = { origin, size }; |
1265 | |
2295 | |
1266 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2296 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1267 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2297 | XPUSH_CLOBJ (stash_buffer, mem); |
1268 | |
2298 | |
1269 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2299 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1270 | |
2300 | |
1271 | void |
2301 | void |
1272 | image_info (OpenCL::Image self, cl_image_info name) |
2302 | image_info (OpenCL::Image self, cl_image_info name) |
1273 | PPCODE: |
2303 | PPCODE: |
1274 | INFO (Image) |
2304 | INFO (Image) |
|
|
2305 | |
|
|
2306 | void |
|
|
2307 | format (OpenCL::Image self) |
|
|
2308 | PPCODE: |
|
|
2309 | cl_image_format format; |
|
|
2310 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2311 | EXTEND (SP, 2); |
|
|
2312 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2313 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1275 | |
2314 | |
1276 | #BEGIN:image |
2315 | #BEGIN:image |
1277 | |
2316 | |
1278 | void |
2317 | void |
1279 | element_size (OpenCL::Image self) |
2318 | element_size (OpenCL::Image self) |
… | |
… | |
1299 | |
2338 | |
1300 | void |
2339 | void |
1301 | target (OpenCL::Image self) |
2340 | target (OpenCL::Image self) |
1302 | PPCODE: |
2341 | PPCODE: |
1303 | cl_GLenum value [1]; |
2342 | cl_GLenum value [1]; |
1304 | NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
2343 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
1305 | EXTEND (SP, 1); |
2344 | EXTEND (SP, 1); |
1306 | const int i = 0; |
2345 | const int i = 0; |
1307 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2346 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1308 | |
2347 | |
1309 | void |
2348 | void |
1310 | gl_mipmap_level (OpenCL::Image self) |
2349 | gl_mipmap_level (OpenCL::Image self) |
1311 | PPCODE: |
2350 | PPCODE: |
1312 | cl_GLint value [1]; |
2351 | cl_GLint value [1]; |
1313 | NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
2352 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
1314 | EXTEND (SP, 1); |
2353 | EXTEND (SP, 1); |
1315 | const int i = 0; |
2354 | const int i = 0; |
1316 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2355 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1317 | |
2356 | |
1318 | #END:gl_texture |
2357 | #END:gl_texture |
… | |
… | |
1349 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2388 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1350 | EXTEND (SP, 1); |
2389 | EXTEND (SP, 1); |
1351 | const int i = 0; |
2390 | const int i = 0; |
1352 | { |
2391 | { |
1353 | NEED_SUCCESS (RetainContext, (value [i])); |
2392 | NEED_SUCCESS (RetainContext, (value [i])); |
1354 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2393 | PUSH_CLOBJ (stash_context, value [i]); |
1355 | } |
2394 | } |
1356 | |
2395 | |
1357 | void |
2396 | void |
1358 | normalized_coords (OpenCL::Sampler self) |
2397 | normalized_coords (OpenCL::Sampler self) |
1359 | PPCODE: |
2398 | PPCODE: |
1360 | cl_addressing_mode value [1]; |
2399 | cl_addressing_mode value [1]; |
1361 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2400 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1362 | EXTEND (SP, 1); |
2401 | EXTEND (SP, 1); |
1363 | const int i = 0; |
2402 | const int i = 0; |
1364 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2403 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1365 | |
2404 | |
1366 | void |
2405 | void |
1367 | addressing_mode (OpenCL::Sampler self) |
2406 | addressing_mode (OpenCL::Sampler self) |
1368 | PPCODE: |
2407 | PPCODE: |
1369 | cl_filter_mode value [1]; |
2408 | cl_filter_mode value [1]; |
1370 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2409 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1371 | EXTEND (SP, 1); |
2410 | EXTEND (SP, 1); |
1372 | const int i = 0; |
2411 | const int i = 0; |
1373 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2412 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1374 | |
2413 | |
1375 | void |
2414 | void |
1376 | filter_mode (OpenCL::Sampler self) |
2415 | filter_mode (OpenCL::Sampler self) |
1377 | PPCODE: |
2416 | PPCODE: |
1378 | cl_bool value [1]; |
2417 | cl_bool value [1]; |
… | |
… | |
1389 | DESTROY (OpenCL::Program self) |
2428 | DESTROY (OpenCL::Program self) |
1390 | CODE: |
2429 | CODE: |
1391 | clReleaseProgram (self); |
2430 | clReleaseProgram (self); |
1392 | |
2431 | |
1393 | void |
2432 | void |
1394 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2433 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2434 | ALIAS: |
|
|
2435 | build_async = 1 |
1395 | CODE: |
2436 | CODE: |
|
|
2437 | cl_uint device_count = 0; |
|
|
2438 | cl_device_id *device_list = 0; |
|
|
2439 | |
|
|
2440 | if (SvOK (devices)) |
|
|
2441 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2442 | |
|
|
2443 | void *user_data; |
|
|
2444 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2445 | |
|
|
2446 | if (ix) |
|
|
2447 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2448 | else |
|
|
2449 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2450 | |
|
|
2451 | #if CL_VERSION_1_2 |
|
|
2452 | |
|
|
2453 | void |
|
|
2454 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2455 | CODE: |
|
|
2456 | cl_uint device_count = 0; |
|
|
2457 | cl_device_id *device_list = 0; |
|
|
2458 | |
|
|
2459 | if (SvOK (devices)) |
|
|
2460 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2461 | |
|
|
2462 | cl_uint header_count = 0; |
|
|
2463 | cl_program *header_list = 0; |
|
|
2464 | const char **header_name = 0; |
|
|
2465 | |
|
|
2466 | if (SvOK (headers)) |
|
|
2467 | { |
|
|
2468 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2469 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2470 | |
|
|
2471 | HV *hv = (HV *)SvRV (devices); |
|
|
2472 | |
|
|
2473 | header_count = hv_iterinit (hv); |
|
|
2474 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2475 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2476 | |
|
|
2477 | HE *he; |
|
|
2478 | int i = 0; |
|
|
2479 | while (he = hv_iternext (hv)) |
|
|
2480 | { |
|
|
2481 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2482 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2483 | ++i; |
|
|
2484 | } |
|
|
2485 | } |
|
|
2486 | |
|
|
2487 | void *user_data; |
|
|
2488 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2489 | |
1396 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2490 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2491 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2492 | |
|
|
2493 | #endif |
1397 | |
2494 | |
1398 | void |
2495 | void |
1399 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2496 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1400 | PPCODE: |
2497 | PPCODE: |
1401 | size_t size; |
2498 | size_t size; |
… | |
… | |
1430 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2527 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
1431 | EXTEND (SP, 1); |
2528 | EXTEND (SP, 1); |
1432 | const int i = 0; |
2529 | const int i = 0; |
1433 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
2530 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
1434 | |
2531 | |
|
|
2532 | void |
|
|
2533 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2534 | PPCODE: |
|
|
2535 | cl_program_binary_type value [1]; |
|
|
2536 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2537 | EXTEND (SP, 1); |
|
|
2538 | const int i = 0; |
|
|
2539 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
2540 | |
1435 | #END:program_build |
2541 | #END:program_build |
1436 | |
2542 | |
1437 | void |
2543 | void |
1438 | kernel (OpenCL::Program program, SV *function) |
2544 | kernel (OpenCL::Program program, SV *function) |
1439 | PPCODE: |
2545 | PPCODE: |
1440 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2546 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1441 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2547 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2548 | |
|
|
2549 | void |
|
|
2550 | kernels_in_program (OpenCL::Program program) |
|
|
2551 | PPCODE: |
|
|
2552 | cl_uint num_kernels; |
|
|
2553 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2554 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2555 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2556 | |
|
|
2557 | int i; |
|
|
2558 | EXTEND (SP, num_kernels); |
|
|
2559 | for (i = 0; i < num_kernels; ++i) |
|
|
2560 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1442 | |
2561 | |
1443 | void |
2562 | void |
1444 | info (OpenCL::Program self, cl_program_info name) |
2563 | info (OpenCL::Program self, cl_program_info name) |
1445 | PPCODE: |
2564 | PPCODE: |
1446 | INFO (Program) |
2565 | INFO (Program) |
… | |
… | |
1464 | { |
2583 | { |
1465 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2584 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1466 | SvUPGRADE (sv, SVt_PV); |
2585 | SvUPGRADE (sv, SVt_PV); |
1467 | SvPOK_only (sv); |
2586 | SvPOK_only (sv); |
1468 | SvCUR_set (sv, sizes [i]); |
2587 | SvCUR_set (sv, sizes [i]); |
1469 | ptrs [i] = SvPVX (sv); |
2588 | ptrs [i] = (void *)SvPVX (sv); |
1470 | PUSHs (sv); |
2589 | PUSHs (sv); |
1471 | } |
2590 | } |
1472 | |
2591 | |
1473 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2592 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1474 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2593 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1494 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2613 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1495 | EXTEND (SP, 1); |
2614 | EXTEND (SP, 1); |
1496 | const int i = 0; |
2615 | const int i = 0; |
1497 | { |
2616 | { |
1498 | NEED_SUCCESS (RetainContext, (value [i])); |
2617 | NEED_SUCCESS (RetainContext, (value [i])); |
1499 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2618 | PUSH_CLOBJ (stash_context, value [i]); |
1500 | } |
2619 | } |
1501 | |
2620 | |
1502 | void |
2621 | void |
1503 | devices (OpenCL::Program self) |
2622 | devices (OpenCL::Program self) |
1504 | PPCODE: |
2623 | PPCODE: |
… | |
… | |
1508 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2627 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1509 | int i, n = size / sizeof (*value); |
2628 | int i, n = size / sizeof (*value); |
1510 | EXTEND (SP, n); |
2629 | EXTEND (SP, n); |
1511 | for (i = 0; i < n; ++i) |
2630 | for (i = 0; i < n; ++i) |
1512 | { |
2631 | { |
1513 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2632 | PUSH_CLOBJ (stash_device, value [i]); |
1514 | } |
2633 | } |
1515 | |
2634 | |
1516 | void |
2635 | void |
1517 | source (OpenCL::Program self) |
2636 | source (OpenCL::Program self) |
1518 | PPCODE: |
2637 | PPCODE: |
… | |
… | |
1544 | DESTROY (OpenCL::Kernel self) |
2663 | DESTROY (OpenCL::Kernel self) |
1545 | CODE: |
2664 | CODE: |
1546 | clReleaseKernel (self); |
2665 | clReleaseKernel (self); |
1547 | |
2666 | |
1548 | void |
2667 | void |
|
|
2668 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2669 | CODE: |
|
|
2670 | int i; |
|
|
2671 | for (i = 2; ; ++i) |
|
|
2672 | { |
|
|
2673 | while (*format == ' ') |
|
|
2674 | ++format; |
|
|
2675 | |
|
|
2676 | char type = *format++; |
|
|
2677 | |
|
|
2678 | if (!type) |
|
|
2679 | break; |
|
|
2680 | |
|
|
2681 | if (i >= items) |
|
|
2682 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2683 | |
|
|
2684 | SV *sv = ST (i); |
|
|
2685 | |
|
|
2686 | union |
|
|
2687 | { |
|
|
2688 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2689 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2690 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2691 | cl_mem cm; |
|
|
2692 | cl_sampler ca; |
|
|
2693 | size_t cz; |
|
|
2694 | cl_event ce; |
|
|
2695 | } arg; |
|
|
2696 | size_t size; |
|
|
2697 | int nullarg = 0; |
|
|
2698 | |
|
|
2699 | switch (type) |
|
|
2700 | { |
|
|
2701 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2702 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2703 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2704 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2705 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2706 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2707 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2708 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2709 | |
|
|
2710 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2711 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2712 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2713 | |
|
|
2714 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2715 | |
|
|
2716 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2717 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2718 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2719 | |
|
|
2720 | default: |
|
|
2721 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2722 | } |
|
|
2723 | |
|
|
2724 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2725 | if (res) |
|
|
2726 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2727 | } |
|
|
2728 | |
|
|
2729 | if (i != items) |
|
|
2730 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2731 | |
|
|
2732 | void |
1549 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2733 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1550 | CODE: |
2734 | CODE: |
1551 | clSetKernelArg (self, idx, sizeof (value), &value); |
2735 | clSetKernelArg (self, idx, sizeof (value), &value); |
1552 | |
2736 | |
1553 | void |
2737 | void |
… | |
… | |
1601 | clSetKernelArg (self, idx, sizeof (value), &value); |
2785 | clSetKernelArg (self, idx, sizeof (value), &value); |
1602 | |
2786 | |
1603 | void |
2787 | void |
1604 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2788 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1605 | CODE: |
2789 | CODE: |
|
|
2790 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2791 | |
|
|
2792 | void |
|
|
2793 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2794 | CODE: |
|
|
2795 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2796 | |
|
|
2797 | void |
|
|
2798 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2799 | ALIAS: |
|
|
2800 | set_image2d = 0 |
|
|
2801 | set_image3d = 0 |
|
|
2802 | CODE: |
|
|
2803 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2804 | |
|
|
2805 | void |
|
|
2806 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2807 | CODE: |
1606 | clSetKernelArg (self, idx, sizeof (value), &value); |
2808 | clSetKernelArg (self, idx, sizeof (value), &value); |
1607 | |
2809 | |
1608 | void |
2810 | void |
1609 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2811 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
1610 | CODE: |
2812 | CODE: |
1611 | clSetKernelArg (self, idx, sizeof (value), &value); |
2813 | clSetKernelArg (self, idx, size, 0); |
1612 | |
|
|
1613 | void |
|
|
1614 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
1615 | CODE: |
|
|
1616 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1617 | |
|
|
1618 | void |
|
|
1619 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1620 | CODE: |
|
|
1621 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1622 | |
|
|
1623 | void |
|
|
1624 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
1625 | CODE: |
|
|
1626 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1627 | |
2814 | |
1628 | void |
2815 | void |
1629 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
2816 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1630 | CODE: |
2817 | CODE: |
1631 | clSetKernelArg (self, idx, sizeof (value), &value); |
2818 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1667 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2854 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1668 | EXTEND (SP, 1); |
2855 | EXTEND (SP, 1); |
1669 | const int i = 0; |
2856 | const int i = 0; |
1670 | { |
2857 | { |
1671 | NEED_SUCCESS (RetainContext, (value [i])); |
2858 | NEED_SUCCESS (RetainContext, (value [i])); |
1672 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2859 | PUSH_CLOBJ (stash_context, value [i]); |
1673 | } |
2860 | } |
1674 | |
2861 | |
1675 | void |
2862 | void |
1676 | program (OpenCL::Kernel self) |
2863 | program (OpenCL::Kernel self) |
1677 | PPCODE: |
2864 | PPCODE: |
… | |
… | |
1679 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2866 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1680 | EXTEND (SP, 1); |
2867 | EXTEND (SP, 1); |
1681 | const int i = 0; |
2868 | const int i = 0; |
1682 | { |
2869 | { |
1683 | NEED_SUCCESS (RetainProgram, (value [i])); |
2870 | NEED_SUCCESS (RetainProgram, (value [i])); |
1684 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2871 | PUSH_CLOBJ (stash_program, value [i]); |
1685 | } |
2872 | } |
1686 | |
2873 | |
1687 | #END:kernel |
2874 | #END:kernel |
1688 | |
2875 | |
1689 | void |
2876 | void |
… | |
… | |
1736 | const int i = 0; |
2923 | const int i = 0; |
1737 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2924 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1738 | |
2925 | |
1739 | #END:kernel_work_group |
2926 | #END:kernel_work_group |
1740 | |
2927 | |
|
|
2928 | #if CL_VERSION_1_2 |
|
|
2929 | |
|
|
2930 | void |
|
|
2931 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2932 | PPCODE: |
|
|
2933 | size_t size; |
|
|
2934 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2935 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2936 | SvUPGRADE (sv, SVt_PV); |
|
|
2937 | SvPOK_only (sv); |
|
|
2938 | SvCUR_set (sv, size); |
|
|
2939 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2940 | XPUSHs (sv); |
|
|
2941 | |
|
|
2942 | #BEGIN:kernel_arg |
|
|
2943 | |
|
|
2944 | void |
|
|
2945 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2946 | PPCODE: |
|
|
2947 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2948 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2949 | EXTEND (SP, 1); |
|
|
2950 | const int i = 0; |
|
|
2951 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2952 | |
|
|
2953 | void |
|
|
2954 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2955 | PPCODE: |
|
|
2956 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2957 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2958 | EXTEND (SP, 1); |
|
|
2959 | const int i = 0; |
|
|
2960 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2961 | |
|
|
2962 | void |
|
|
2963 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2964 | ALIAS: |
|
|
2965 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2966 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2967 | PPCODE: |
|
|
2968 | size_t size; |
|
|
2969 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2970 | char *value = tmpbuf (size); |
|
|
2971 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2972 | EXTEND (SP, 1); |
|
|
2973 | const int i = 0; |
|
|
2974 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2975 | |
|
|
2976 | void |
|
|
2977 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2978 | PPCODE: |
|
|
2979 | cl_kernel_arg_type_qualifier value [1]; |
|
|
2980 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2981 | EXTEND (SP, 1); |
|
|
2982 | const int i = 0; |
|
|
2983 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2984 | |
|
|
2985 | #END:kernel_arg |
|
|
2986 | |
|
|
2987 | #endif |
|
|
2988 | |
1741 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2989 | MODULE = OpenCL PACKAGE = OpenCL::Event |
1742 | |
2990 | |
1743 | void |
2991 | void |
1744 | DESTROY (OpenCL::Event self) |
2992 | DESTROY (OpenCL::Event self) |
1745 | CODE: |
2993 | CODE: |
… | |
… | |
1747 | |
2995 | |
1748 | void |
2996 | void |
1749 | wait (OpenCL::Event self) |
2997 | wait (OpenCL::Event self) |
1750 | CODE: |
2998 | CODE: |
1751 | clWaitForEvents (1, &self); |
2999 | clWaitForEvents (1, &self); |
|
|
3000 | |
|
|
3001 | void |
|
|
3002 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
3003 | CODE: |
|
|
3004 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
1752 | |
3005 | |
1753 | void |
3006 | void |
1754 | info (OpenCL::Event self, cl_event_info name) |
3007 | info (OpenCL::Event self, cl_event_info name) |
1755 | PPCODE: |
3008 | PPCODE: |
1756 | INFO (Event) |
3009 | INFO (Event) |
… | |
… | |
1764 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
3017 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1765 | EXTEND (SP, 1); |
3018 | EXTEND (SP, 1); |
1766 | const int i = 0; |
3019 | const int i = 0; |
1767 | { |
3020 | { |
1768 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
3021 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1769 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
3022 | PUSH_CLOBJ (stash_queue, value [i]); |
1770 | } |
3023 | } |
1771 | |
3024 | |
1772 | void |
3025 | void |
1773 | command_type (OpenCL::Event self) |
3026 | command_type (OpenCL::Event self) |
1774 | PPCODE: |
3027 | PPCODE: |
… | |
… | |
1797 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
3050 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1798 | EXTEND (SP, 1); |
3051 | EXTEND (SP, 1); |
1799 | const int i = 0; |
3052 | const int i = 0; |
1800 | { |
3053 | { |
1801 | NEED_SUCCESS (RetainContext, (value [i])); |
3054 | NEED_SUCCESS (RetainContext, (value [i])); |
1802 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
3055 | PUSH_CLOBJ (stash_context, value [i]); |
1803 | } |
3056 | } |
1804 | |
3057 | |
1805 | #END:event |
3058 | #END:event |
1806 | |
3059 | |
1807 | void |
3060 | void |
… | |
… | |
1832 | void |
3085 | void |
1833 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
3086 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1834 | CODE: |
3087 | CODE: |
1835 | clSetUserEventStatus (self, execution_status); |
3088 | clSetUserEventStatus (self, execution_status); |
1836 | |
3089 | |
|
|
3090 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3091 | |
|
|
3092 | void |
|
|
3093 | DESTROY (SV *self) |
|
|
3094 | CODE: |
|
|
3095 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3096 | |
|
|
3097 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3098 | mapped_detach (self, mapped); |
|
|
3099 | |
|
|
3100 | clReleaseCommandQueue (mapped->queue); |
|
|
3101 | clReleaseEvent (mapped->event); |
|
|
3102 | Safefree (mapped); |
|
|
3103 | |
|
|
3104 | void |
|
|
3105 | unmap (OpenCL::Mapped self, ...) |
|
|
3106 | CODE: |
|
|
3107 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3108 | |
|
|
3109 | bool |
|
|
3110 | mapped (OpenCL::Mapped self) |
|
|
3111 | CODE: |
|
|
3112 | RETVAL = !!self->ptr; |
|
|
3113 | OUTPUT: |
|
|
3114 | RETVAL |
|
|
3115 | |
|
|
3116 | void |
|
|
3117 | wait (OpenCL::Mapped self) |
|
|
3118 | PPCODE: |
|
|
3119 | if (self->event) |
|
|
3120 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3121 | |
|
|
3122 | void |
|
|
3123 | event (OpenCL::Mapped self) |
|
|
3124 | PPCODE: |
|
|
3125 | if (!self->event) |
|
|
3126 | XSRETURN_UNDEF; |
|
|
3127 | |
|
|
3128 | clRetainEvent (self->event); |
|
|
3129 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3130 | |
|
|
3131 | size_t |
|
|
3132 | size (OpenCL::Mapped self) |
|
|
3133 | CODE: |
|
|
3134 | RETVAL = self->cb; |
|
|
3135 | OUTPUT: |
|
|
3136 | RETVAL |
|
|
3137 | |
|
|
3138 | IV |
|
|
3139 | ptr (OpenCL::Mapped self) |
|
|
3140 | CODE: |
|
|
3141 | RETVAL = PTR2IV (self->ptr); |
|
|
3142 | OUTPUT: |
|
|
3143 | RETVAL |
|
|
3144 | |
|
|
3145 | void |
|
|
3146 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3147 | CODE: |
|
|
3148 | STRLEN len; |
|
|
3149 | const char *ptr = SvPVbyte (data, len); |
|
|
3150 | |
|
|
3151 | if (offset + len > self->cb) |
|
|
3152 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3153 | |
|
|
3154 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3155 | |
|
|
3156 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3157 | |
|
|
3158 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3159 | |
|
|
3160 | IV |
|
|
3161 | row_pitch (OpenCL::Mapped self) |
|
|
3162 | ALIAS: |
|
|
3163 | slice_pitch = 1 |
|
|
3164 | CODE: |
|
|
3165 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
3166 | OUTPUT: |
|
|
3167 | RETVAL |
|
|
3168 | |
|
|
3169 | |
|
|
3170 | |