|
|
1 | #include <stddef.h> |
|
|
2 | |
1 | #include "EXTERN.h" |
3 | #include "EXTERN.h" |
2 | #include "perl.h" |
4 | #include "perl.h" |
3 | #include "XSUB.h" |
5 | #include "XSUB.h" |
4 | |
6 | |
|
|
7 | #include "ecb.h"//D |
|
|
8 | |
|
|
9 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
10 | #include "xthread.h" |
|
|
11 | #include "schmorp.h" |
|
|
12 | |
|
|
13 | #ifdef I_DLFCN |
|
|
14 | #include <dlfcn.h> |
|
|
15 | #endif |
|
|
16 | |
|
|
17 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
18 | // but then fail to define the api functions unless you ALSO define |
|
|
19 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
20 | // after all, the functions are deprecated, not removed. |
|
|
21 | // in addition, you cannot test for this in any future-proof way. |
|
|
22 | // each time a new opencl version comes out, you need to make a new |
|
|
23 | // release. |
|
|
24 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
25 | |
|
|
26 | #ifndef PREFER_1_1 |
|
|
27 | #define PREFER_1_1 1 |
|
|
28 | #endif |
|
|
29 | |
|
|
30 | #if PREFER_1_1 |
|
|
31 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
32 | #endif |
|
|
33 | |
|
|
34 | #ifdef __APPLE__ |
|
|
35 | #define CLHDR(name) <OpenCL/name> |
|
|
36 | #else |
|
|
37 | #define CLHDR(name) <CL/name> |
|
|
38 | #endif |
|
|
39 | |
5 | #include <CL/opencl.h> |
40 | #include CLHDR(opencl.h) |
6 | |
41 | |
|
|
42 | #ifndef CL_VERSION_1_2 |
|
|
43 | #include CLHDR(cl_d3d9.h) |
|
|
44 | #endif |
|
|
45 | |
|
|
46 | #if _WIN32 |
|
|
47 | #include CLHDR(cl_d3d10.h) |
|
|
48 | #if CL_VERSION_1_2 |
|
|
49 | #include CLHDR<cl_d3d11.h> |
|
|
50 | #endif |
|
|
51 | #include CLHDR<cl_dx9_media_sharing.h.h> |
|
|
52 | #endif |
|
|
53 | |
|
|
54 | #ifndef CL_VERSION_1_2 |
|
|
55 | #undef PREFER_1_1 |
|
|
56 | #define PREFER_1_1 1 |
|
|
57 | #endif |
|
|
58 | |
|
|
59 | // make sure all constants we might use are actually defined |
|
|
60 | #include "default.h" |
|
|
61 | |
7 | typedef cl_platform_id OpenCL__Platform; |
62 | typedef cl_platform_id OpenCL__Platform; |
8 | typedef cl_device_id OpenCL__Device; |
63 | typedef cl_device_id OpenCL__Device; |
|
|
64 | typedef cl_device_id OpenCL__SubDevice; |
9 | typedef cl_context OpenCL__Context; |
65 | typedef cl_context OpenCL__Context; |
10 | typedef cl_command_queue OpenCL__Queue; |
66 | typedef cl_command_queue OpenCL__Queue; |
11 | typedef cl_mem OpenCL__Memory; |
67 | typedef cl_mem OpenCL__Memory; |
|
|
68 | typedef cl_mem OpenCL__Buffer; |
|
|
69 | typedef cl_mem OpenCL__BufferObj; |
|
|
70 | typedef cl_mem OpenCL__Image; |
|
|
71 | typedef cl_mem OpenCL__Memory_ornull; |
|
|
72 | typedef cl_mem OpenCL__Buffer_ornull; |
|
|
73 | typedef cl_mem OpenCL__Image_ornull; |
12 | typedef cl_sampler OpenCL__Sampler; |
74 | typedef cl_sampler OpenCL__Sampler; |
13 | typedef cl_program OpenCL__Program; |
75 | typedef cl_program OpenCL__Program; |
14 | typedef cl_kernel OpenCL__Kernel; |
76 | typedef cl_kernel OpenCL__Kernel; |
15 | typedef cl_event OpenCL__Event; |
77 | typedef cl_event OpenCL__Event; |
|
|
78 | typedef cl_event OpenCL__UserEvent; |
16 | |
79 | |
17 | static const struct { |
80 | typedef struct mapped * OpenCL__Mapped; |
|
|
81 | |
|
|
82 | static HV |
|
|
83 | *stash_platform, |
|
|
84 | *stash_device, |
|
|
85 | *stash_subdevice, |
|
|
86 | *stash_context, |
|
|
87 | *stash_queue, |
|
|
88 | *stash_program, |
|
|
89 | *stash_kernel, |
|
|
90 | *stash_sampler, |
|
|
91 | *stash_event, |
|
|
92 | *stash_userevent, |
|
|
93 | *stash_memory, |
|
|
94 | *stash_buffer, |
|
|
95 | *stash_bufferobj, |
|
|
96 | *stash_image, |
|
|
97 | *stash_image1d, |
|
|
98 | *stash_image1darray, |
|
|
99 | *stash_image1dbuffer, |
|
|
100 | *stash_image2d, |
|
|
101 | *stash_image2darray, |
|
|
102 | *stash_image3d, |
|
|
103 | *stash_mapped, |
|
|
104 | *stash_mappedbuffer, |
|
|
105 | *stash_mappedimage; |
|
|
106 | |
|
|
107 | /*****************************************************************************/ |
|
|
108 | |
|
|
109 | // name must include a leading underscore |
|
|
110 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
111 | // for perl. doh. |
|
|
112 | static void * |
|
|
113 | glsym (const char *name) |
|
|
114 | { |
|
|
115 | void *fun = 0; |
|
|
116 | |
|
|
117 | #if defined I_DLFCN && defined RTLD_DEFAULT |
|
|
118 | fun = dlsym (RTLD_DEFAULT, name + 1); |
|
|
119 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
|
|
120 | |
|
|
121 | if (!fun) |
|
|
122 | { |
|
|
123 | static void *libgl; |
|
|
124 | static const char *glso[] = { |
|
|
125 | "libGL.so.1", |
|
|
126 | "libGL.so.3", |
|
|
127 | "libGL.so.4.0", |
|
|
128 | "libGL.so", |
|
|
129 | "/usr/lib/libGL.so", |
|
|
130 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
131 | }; |
|
|
132 | int i; |
|
|
133 | |
|
|
134 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
135 | { |
|
|
136 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
137 | if (libgl) |
|
|
138 | break; |
|
|
139 | } |
|
|
140 | |
|
|
141 | if (libgl) |
|
|
142 | { |
|
|
143 | fun = dlsym (libgl, name + 1); |
|
|
144 | if (!fun) fun = dlsym (libgl, name); |
|
|
145 | } |
|
|
146 | } |
|
|
147 | #endif |
|
|
148 | |
|
|
149 | return fun; |
|
|
150 | } |
|
|
151 | |
|
|
152 | /*****************************************************************************/ |
|
|
153 | |
|
|
154 | /* up to two temporary buffers */ |
|
|
155 | static void * ecb_noinline |
|
|
156 | tmpbuf (size_t size) |
|
|
157 | { |
|
|
158 | enum { buffers = 4 }; |
|
|
159 | static int idx; |
|
|
160 | static void *buf [buffers]; |
|
|
161 | static size_t len [buffers]; |
|
|
162 | |
|
|
163 | idx = (idx + 1) % buffers; |
|
|
164 | |
|
|
165 | if (len [idx] < size) |
|
|
166 | { |
|
|
167 | free (buf [idx]); |
|
|
168 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
|
|
169 | buf [idx] = malloc (len [idx]); |
|
|
170 | } |
|
|
171 | |
|
|
172 | return buf [idx]; |
|
|
173 | } |
|
|
174 | |
|
|
175 | static const char * ecb_noinline |
|
|
176 | cv_get_name (CV *cv) |
|
|
177 | { |
|
|
178 | static char fullname [256]; |
|
|
179 | |
|
|
180 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
181 | |
|
|
182 | HV *stash = GvSTASH (gv); |
|
|
183 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
184 | const char *gvname = GvNAME (gv); |
|
|
185 | |
|
|
186 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
187 | return fullname; |
|
|
188 | } |
|
|
189 | |
|
|
190 | /*****************************************************************************/ |
|
|
191 | |
|
|
192 | typedef struct |
|
|
193 | { |
18 | IV iv; |
194 | IV iv; |
19 | const char *name; |
195 | const char *name; |
20 | } cl_error[] = { |
|
|
21 | #define def_error(name) { (IV)CL_ ## name, # name }, |
196 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
22 | #include "invalid.h" |
197 | } ivstr; |
23 | }; |
198 | |
|
|
199 | typedef struct |
|
|
200 | { |
|
|
201 | NV nv; |
|
|
202 | const char *name; |
|
|
203 | #define const_nv(name) { (NV)CL_ ## name, # name }, |
|
|
204 | } nvstr; |
24 | |
205 | |
25 | static const char * |
206 | static const char * |
26 | clstrerror (cl_int res) |
207 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
27 | { |
208 | { |
28 | int i; |
209 | int i; |
29 | static char numbuf [32]; |
210 | static char strbuf [32]; |
30 | |
211 | |
31 | for (i = sizeof (cl_error) / sizeof (cl_error [0]); i--; ) |
212 | for (i = count; i--; ) |
32 | if (cl_error [i].iv == res) |
213 | if (base [i].iv == value) |
33 | return cl_error [i].name; |
214 | return base [i].name; |
34 | |
215 | |
35 | snprintf (numbuf, sizeof (numbuf), "ERROR(%d)", res); |
216 | snprintf (strbuf, sizeof (strbuf), fallback, (int)value); |
36 | |
217 | |
37 | return numbuf; |
218 | return strbuf; |
38 | } |
219 | } |
39 | |
220 | |
|
|
221 | static const char * |
|
|
222 | enum2str (cl_uint value) |
|
|
223 | { |
|
|
224 | static const ivstr enumstr[] = { |
|
|
225 | #include "enumstr.h" |
|
|
226 | }; |
|
|
227 | |
|
|
228 | return iv2str (value, enumstr, sizeof (enumstr) / sizeof (enumstr [0]), "ENUM(0x%04x)"); |
|
|
229 | } |
|
|
230 | |
|
|
231 | static const char * |
|
|
232 | err2str (cl_int err) |
|
|
233 | { |
|
|
234 | static const ivstr errstr[] = { |
|
|
235 | #include "errstr.h" |
|
|
236 | }; |
|
|
237 | |
|
|
238 | return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
|
|
239 | } |
|
|
240 | |
|
|
241 | /*****************************************************************************/ |
|
|
242 | |
|
|
243 | static cl_int res; |
|
|
244 | |
40 | #define FAIL(name,res) \ |
245 | #define FAIL(name) \ |
41 | croak ("cl" # name ": %s", clstrerror (res)); |
246 | croak ("cl" # name ": %s", err2str (res)); |
42 | |
247 | |
43 | #define NEED_SUCCESS(name,args) \ |
248 | #define NEED_SUCCESS(name,args) \ |
44 | do { \ |
249 | do { \ |
45 | cl_int res = cl ## name args; \ |
250 | res = cl ## name args; \ |
46 | \ |
251 | \ |
47 | if (res) \ |
252 | if (res) \ |
48 | FAIL (name, res); \ |
253 | FAIL (name); \ |
49 | } while (0) |
254 | } while (0) |
50 | |
255 | |
51 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
256 | #define NEED_SUCCESS_ARG(retdecl, name, args) \ |
52 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
257 | retdecl = cl ## name args; \ |
|
|
258 | if (res) \ |
|
|
259 | FAIL (name); |
53 | |
260 | |
54 | /*TODO*/ |
261 | /*****************************************************************************/ |
55 | #define EVENT_LIST(items,count) cl_uint event_list_count = 0; cl_event *event_list_ptr = 0 |
262 | |
|
|
263 | static SV * |
|
|
264 | new_clobj (HV *stash, IV id) |
|
|
265 | { |
|
|
266 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
267 | } |
|
|
268 | |
|
|
269 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
270 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
271 | |
|
|
272 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
273 | /* they can be upgraded at runtime to the array form */ |
|
|
274 | static void * ecb_noinline |
|
|
275 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
|
|
276 | { |
|
|
277 | // sv_derived_from is quite slow :( |
|
|
278 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
279 | return (void *)SvIV (SvRV (sv)); |
|
|
280 | |
|
|
281 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
282 | } |
|
|
283 | |
|
|
284 | // the "no-inherit" version of the above |
|
|
285 | static void * ecb_noinline |
|
|
286 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
287 | { |
|
|
288 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
289 | return (void *)SvIV (SvRV (sv)); |
|
|
290 | |
|
|
291 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
|
|
292 | } |
|
|
293 | |
|
|
294 | /*****************************************************************************/ |
|
|
295 | |
|
|
296 | static cl_context_properties * ecb_noinline |
|
|
297 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
|
|
298 | { |
|
|
299 | if (!sv || !SvOK (sv)) |
|
|
300 | if (extra) |
|
|
301 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
|
|
302 | else |
|
|
303 | return 0; |
|
|
304 | |
|
|
305 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
|
|
306 | { |
|
|
307 | AV *av = (AV *)SvRV (sv); |
|
|
308 | int i, len = av_len (av) + 1; |
|
|
309 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
|
|
310 | cl_context_properties *l = p; |
|
|
311 | |
|
|
312 | if (len & 1) |
|
|
313 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
|
|
314 | |
|
|
315 | while (extracount--) |
|
|
316 | *l++ = *extra++; |
|
|
317 | |
|
|
318 | for (i = 0; i < len; i += 2) |
|
|
319 | { |
|
|
320 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
|
|
321 | SV *p_sv = *av_fetch (av, i + 1, 0); |
|
|
322 | cl_context_properties v = SvIV (p_sv); // code below can override |
|
|
323 | |
|
|
324 | switch (t) |
|
|
325 | { |
|
|
326 | case CL_CONTEXT_PLATFORM: |
|
|
327 | if (SvROK (p_sv)) |
|
|
328 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
329 | break; |
|
|
330 | |
|
|
331 | case CL_GLX_DISPLAY_KHR: |
|
|
332 | if (!SvOK (p_sv)) |
|
|
333 | { |
|
|
334 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
335 | if (func) |
|
|
336 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
337 | } |
|
|
338 | break; |
|
|
339 | |
|
|
340 | case CL_GL_CONTEXT_KHR: |
|
|
341 | if (!SvOK (p_sv)) |
|
|
342 | { |
|
|
343 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
344 | if (func) |
|
|
345 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
346 | } |
|
|
347 | break; |
|
|
348 | |
|
|
349 | default: |
|
|
350 | /* unknown property, treat as int */ |
|
|
351 | break; |
|
|
352 | } |
|
|
353 | |
|
|
354 | *l++ = t; |
|
|
355 | *l++ = v; |
|
|
356 | } |
|
|
357 | |
|
|
358 | *l = 0; |
|
|
359 | |
|
|
360 | return p; |
|
|
361 | } |
|
|
362 | |
|
|
363 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
|
|
364 | } |
|
|
365 | |
|
|
366 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
|
|
367 | // is a pointer (and only on well-behaved systems). |
|
|
368 | static void * ecb_noinline |
|
|
369 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
370 | { |
|
|
371 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
372 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
373 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
|
|
374 | |
|
|
375 | AV *av = (AV *)SvRV (arg); |
|
|
376 | void **list = 0; |
|
|
377 | cl_uint count = av_len (av) + 1; |
|
|
378 | |
|
|
379 | if (count) |
|
|
380 | { |
|
|
381 | list = tmpbuf (sizeof (*list) * count); |
|
|
382 | int i; |
|
|
383 | for (i = 0; i < count; ++i) |
|
|
384 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
|
|
385 | } |
|
|
386 | |
|
|
387 | if (!count && !or_undef) |
|
|
388 | croak ("%s: '%s' must contain at least one %s object", |
|
|
389 | cv_get_name (cv), argname, klass); |
|
|
390 | |
|
|
391 | *rcount = count; |
|
|
392 | return (void *)list; |
|
|
393 | } |
|
|
394 | |
|
|
395 | /*****************************************************************************/ |
|
|
396 | /* callback stuff */ |
|
|
397 | |
|
|
398 | /* default context callback, log to stderr */ |
|
|
399 | static void CL_CALLBACK |
|
|
400 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
401 | { |
|
|
402 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
403 | } |
|
|
404 | |
|
|
405 | typedef struct |
|
|
406 | { |
|
|
407 | int free_cb; |
|
|
408 | void (*push)(void *data1, void *data2, void *data3); |
|
|
409 | } eq_vtbl; |
|
|
410 | |
|
|
411 | typedef struct eq_item |
|
|
412 | { |
|
|
413 | struct eq_item *next; |
|
|
414 | eq_vtbl *vtbl; |
|
|
415 | SV *cb; |
|
|
416 | void *data1, *data2, *data3; |
|
|
417 | } eq_item; |
|
|
418 | |
|
|
419 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
420 | static void *eq_signal_arg; |
|
|
421 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
422 | static eq_item *eq_head, *eq_tail; |
|
|
423 | |
|
|
424 | static void ecb_noinline |
|
|
425 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
426 | { |
|
|
427 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
428 | |
|
|
429 | item->next = 0; |
|
|
430 | item->vtbl = vtbl; |
|
|
431 | item->cb = cb; |
|
|
432 | item->data1 = data1; |
|
|
433 | item->data2 = data2; |
|
|
434 | item->data3 = data3; |
|
|
435 | |
|
|
436 | X_LOCK (eq_lock); |
|
|
437 | |
|
|
438 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
439 | eq_tail = item; |
|
|
440 | |
|
|
441 | X_UNLOCK (eq_lock); |
|
|
442 | |
|
|
443 | eq_signal_func (eq_signal_arg, 0); |
|
|
444 | } |
|
|
445 | |
|
|
446 | static eq_item * |
|
|
447 | eq_dec (void) |
|
|
448 | { |
|
|
449 | eq_item *res; |
|
|
450 | |
|
|
451 | X_LOCK (eq_lock); |
|
|
452 | |
|
|
453 | res = eq_head; |
|
|
454 | if (res) |
|
|
455 | eq_head = res->next; |
|
|
456 | |
|
|
457 | X_UNLOCK (eq_lock); |
|
|
458 | |
|
|
459 | return res; |
|
|
460 | } |
|
|
461 | |
|
|
462 | static void |
|
|
463 | eq_poll (void) |
|
|
464 | { |
|
|
465 | eq_item *item; |
|
|
466 | |
|
|
467 | while ((item = eq_dec ())) |
|
|
468 | { |
|
|
469 | ENTER; |
|
|
470 | SAVETMPS; |
|
|
471 | |
|
|
472 | dSP; |
|
|
473 | PUSHMARK (SP); |
|
|
474 | EXTEND (SP, 2); |
|
|
475 | |
|
|
476 | if (item->vtbl->free_cb) |
|
|
477 | sv_2mortal (item->cb); |
|
|
478 | |
|
|
479 | PUTBACK; |
|
|
480 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
481 | |
|
|
482 | SV *cb = item->cb; |
|
|
483 | free (item); |
|
|
484 | |
|
|
485 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
486 | |
|
|
487 | FREETMPS; |
|
|
488 | LEAVE; |
|
|
489 | } |
|
|
490 | } |
|
|
491 | |
|
|
492 | static void |
|
|
493 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
494 | { |
|
|
495 | eq_poll (); |
|
|
496 | } |
|
|
497 | |
|
|
498 | /*****************************************************************************/ |
|
|
499 | /* context notify */ |
|
|
500 | |
|
|
501 | static void ecb_noinline |
|
|
502 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
503 | { |
|
|
504 | dSP; |
|
|
505 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
506 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
507 | PUTBACK; |
|
|
508 | |
|
|
509 | free (data1); |
|
|
510 | free (data2); |
|
|
511 | } |
|
|
512 | |
|
|
513 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
514 | |
|
|
515 | static void CL_CALLBACK |
|
|
516 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
517 | { |
|
|
518 | void *pvt_copy = malloc (cb); |
|
|
519 | memcpy (pvt_copy, pvt, cb); |
|
|
520 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
521 | } |
|
|
522 | |
|
|
523 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
524 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
525 | void *user_data = 0; \ |
|
|
526 | \ |
|
|
527 | if (SvOK (notify)) \ |
|
|
528 | { \ |
|
|
529 | pfn_notify = eq_context_notify; \ |
|
|
530 | user_data = s_get_cv (notify); \ |
|
|
531 | } |
|
|
532 | |
|
|
533 | static SV * ecb_noinline |
|
|
534 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
535 | { |
|
|
536 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
537 | |
|
|
538 | if (user_data) |
|
|
539 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
540 | |
|
|
541 | return sv; |
|
|
542 | } |
|
|
543 | |
|
|
544 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
545 | |
|
|
546 | /*****************************************************************************/ |
|
|
547 | /* build/compile/link notify */ |
|
|
548 | |
|
|
549 | static void |
|
|
550 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
551 | { |
|
|
552 | dSP; |
|
|
553 | PUSH_CLOBJ (stash_program, data1); |
|
|
554 | PUTBACK; |
|
|
555 | } |
|
|
556 | |
|
|
557 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
558 | |
|
|
559 | static void CL_CALLBACK |
|
|
560 | eq_program_notify (cl_program program, void *user_data) |
|
|
561 | { |
|
|
562 | clRetainProgram (program); |
|
|
563 | |
|
|
564 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
565 | } |
|
|
566 | |
|
|
567 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
568 | |
|
|
569 | static program_callback ecb_noinline |
|
|
570 | make_program_callback (SV *notify, void **ruser_data) |
|
|
571 | { |
|
|
572 | if (SvOK (notify)) |
|
|
573 | { |
|
|
574 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
575 | return eq_program_notify; |
|
|
576 | } |
|
|
577 | else |
|
|
578 | { |
|
|
579 | *ruser_data = 0; |
|
|
580 | return 0; |
|
|
581 | } |
|
|
582 | } |
|
|
583 | |
|
|
584 | struct build_args |
|
|
585 | { |
|
|
586 | cl_program program; |
|
|
587 | char *options; |
|
|
588 | void *user_data; |
|
|
589 | cl_uint num_devices; |
|
|
590 | }; |
|
|
591 | |
|
|
592 | X_THREAD_PROC (build_program_thread) |
|
|
593 | { |
|
|
594 | struct build_args *arg = thr_arg; |
|
|
595 | |
|
|
596 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
597 | |
|
|
598 | if (arg->user_data) |
|
|
599 | eq_program_notify (arg->program, arg->user_data); |
|
|
600 | else |
|
|
601 | clReleaseProgram (arg->program); |
|
|
602 | |
|
|
603 | free (arg->options); |
|
|
604 | free (arg); |
|
|
605 | |
|
|
606 | return 0; |
|
|
607 | } |
|
|
608 | |
|
|
609 | static void |
|
|
610 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
611 | { |
|
|
612 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
613 | |
|
|
614 | arg->program = program; |
|
|
615 | arg->options = strdup (options); |
|
|
616 | arg->user_data = user_data; |
|
|
617 | arg->num_devices = num_devices; |
|
|
618 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
619 | |
|
|
620 | xthread_t id; |
|
|
621 | thread_create (&id, build_program_thread, arg); |
|
|
622 | } |
|
|
623 | |
|
|
624 | /*****************************************************************************/ |
|
|
625 | /* mem object destructor notify */ |
|
|
626 | |
|
|
627 | static void ecb_noinline |
|
|
628 | eq_destructor_push (void *data1, void *data2, void *data3) |
|
|
629 | { |
|
|
630 | } |
|
|
631 | |
|
|
632 | static eq_vtbl eq_destructor_vtbl = { 0, eq_destructor_push }; |
|
|
633 | |
|
|
634 | static void CL_CALLBACK |
|
|
635 | eq_destructor_notify (cl_mem memobj, void *user_data) |
|
|
636 | { |
|
|
637 | eq_enq (&eq_destructor_vtbl, (SV *)user_data, (void *)memobj, 0, 0); |
|
|
638 | } |
|
|
639 | |
|
|
640 | /*****************************************************************************/ |
|
|
641 | /* event objects */ |
|
|
642 | |
|
|
643 | static void |
|
|
644 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
645 | { |
|
|
646 | dSP; |
|
|
647 | PUSH_CLOBJ (stash_event, data1); |
|
|
648 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
649 | PUTBACK; |
|
|
650 | } |
|
|
651 | |
|
|
652 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
653 | |
|
|
654 | static void CL_CALLBACK |
|
|
655 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
656 | { |
|
|
657 | clRetainEvent (event); |
|
|
658 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
659 | } |
|
|
660 | |
|
|
661 | /*****************************************************************************/ |
|
|
662 | /* utilities for XS code */ |
|
|
663 | |
|
|
664 | static size_t |
|
|
665 | img_row_pitch (cl_mem img) |
|
|
666 | { |
|
|
667 | size_t res; |
|
|
668 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
|
|
669 | return res; |
|
|
670 | } |
|
|
671 | |
|
|
672 | static cl_event * ecb_noinline |
|
|
673 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
|
|
674 | { |
|
|
675 | cl_uint count = *rcount; |
|
|
676 | |
|
|
677 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
678 | *rcount = count = 0; |
|
|
679 | |
|
|
680 | if (!count && !extra) |
|
|
681 | return 0; |
|
|
682 | |
|
|
683 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
684 | int i = 0; |
|
|
685 | |
|
|
686 | while (count--) |
|
|
687 | if (SvOK (items [count])) |
|
|
688 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
|
|
689 | |
|
|
690 | if (extra) |
|
|
691 | list [i++] = extra; |
|
|
692 | |
|
|
693 | *rcount = i; |
|
|
694 | |
|
|
695 | return i ? list : 0; |
|
|
696 | } |
|
|
697 | |
|
|
698 | #define EVENT_LIST(skip) \ |
|
|
699 | cl_uint event_list_count = items - (skip); \ |
|
|
700 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
56 | |
701 | |
57 | #define INFO(class) \ |
702 | #define INFO(class) \ |
58 | { \ |
703 | { \ |
59 | size_t size; \ |
704 | size_t size; \ |
60 | SV *sv; \ |
|
|
61 | \ |
|
|
62 | NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ |
705 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
63 | sv = sv_2mortal (newSV (size)); \ |
706 | SV *sv = sv_2mortal (newSV (size)); \ |
64 | SvUPGRADE (sv, SVt_PV); \ |
707 | SvUPGRADE (sv, SVt_PV); \ |
65 | SvPOK_only (sv); \ |
708 | SvPOK_only (sv); \ |
66 | SvCUR_set (sv, size); \ |
709 | SvCUR_set (sv, size); \ |
67 | NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ |
710 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
68 | XPUSHs (sv); \ |
711 | XPUSHs (sv); \ |
69 | } |
712 | } |
70 | |
713 | |
|
|
714 | /*****************************************************************************/ |
|
|
715 | /* mapped_xxx */ |
|
|
716 | |
|
|
717 | static OpenCL__Mapped |
|
|
718 | SvMAPPED (SV *sv) |
|
|
719 | { |
|
|
720 | // no typechecking atm., keep your fingers crossed |
|
|
721 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
722 | } |
|
|
723 | |
|
|
724 | struct mapped |
|
|
725 | { |
|
|
726 | cl_command_queue queue; |
|
|
727 | cl_mem memobj; |
|
|
728 | void *ptr; |
|
|
729 | size_t cb; |
|
|
730 | cl_event event; |
|
|
731 | size_t row_pitch; |
|
|
732 | size_t slice_pitch; |
|
|
733 | |
|
|
734 | size_t element_size; |
|
|
735 | size_t width, height, depth; |
|
|
736 | }; |
|
|
737 | |
|
|
738 | static SV * |
|
|
739 | mapped_new ( |
|
|
740 | HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, |
|
|
741 | void *ptr, size_t cb, cl_event ev, |
|
|
742 | size_t row_pitch, size_t slice_pitch, size_t element_size, |
|
|
743 | size_t width, size_t height, size_t depth |
|
|
744 | ) |
|
|
745 | { |
|
|
746 | SV *data = newSV (0); |
|
|
747 | SvUPGRADE (data, SVt_PVMG); |
|
|
748 | |
|
|
749 | OpenCL__Mapped mapped; |
|
|
750 | New (0, mapped, 1, struct mapped); |
|
|
751 | |
|
|
752 | clRetainCommandQueue (queue); |
|
|
753 | |
|
|
754 | mapped->queue = queue; |
|
|
755 | mapped->memobj = memobj; |
|
|
756 | mapped->ptr = ptr; |
|
|
757 | mapped->cb = cb; |
|
|
758 | mapped->event = ev; |
|
|
759 | mapped->row_pitch = row_pitch; |
|
|
760 | mapped->slice_pitch = slice_pitch; |
|
|
761 | |
|
|
762 | mapped->element_size = element_size; |
|
|
763 | mapped->width = width; |
|
|
764 | mapped->height = height; |
|
|
765 | mapped->depth = depth; |
|
|
766 | |
|
|
767 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
768 | |
|
|
769 | if (SvLEN (data)) |
|
|
770 | Safefree (data); |
|
|
771 | |
|
|
772 | SvPVX (data) = (char *)ptr; |
|
|
773 | SvCUR_set (data, cb); |
|
|
774 | SvLEN_set (data, 0); |
|
|
775 | SvPOK_only (data); |
|
|
776 | |
|
|
777 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
778 | |
|
|
779 | if (!(flags & CL_MAP_WRITE)) |
|
|
780 | SvREADONLY_on (data); |
|
|
781 | |
|
|
782 | return obj; |
|
|
783 | } |
|
|
784 | |
|
|
785 | static void |
|
|
786 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
787 | { |
|
|
788 | SV *data = SvRV (sv); |
|
|
789 | |
|
|
790 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
791 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
792 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
793 | else |
|
|
794 | { |
|
|
795 | SvREADONLY_off (data); |
|
|
796 | SvCUR_set (data, 0); |
|
|
797 | SvPVX (data) = 0; |
|
|
798 | SvOK_off (data); |
|
|
799 | } |
|
|
800 | |
|
|
801 | mapped->ptr = 0; |
|
|
802 | } |
|
|
803 | |
|
|
804 | static void |
|
|
805 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
806 | { |
|
|
807 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
808 | cl_event ev; |
|
|
809 | |
|
|
810 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
811 | |
|
|
812 | clReleaseEvent (mapped->event); |
|
|
813 | mapped->event = ev; |
|
|
814 | |
|
|
815 | mapped_detach (self, mapped); |
|
|
816 | } |
|
|
817 | |
|
|
818 | static size_t |
|
|
819 | mapped_element_size (OpenCL__Mapped self) |
|
|
820 | { |
|
|
821 | if (!self->element_size) |
|
|
822 | clGetImageInfo (self->memobj, CL_IMAGE_ELEMENT_SIZE, sizeof (self->element_size), &self->element_size, 0); |
|
|
823 | |
|
|
824 | return self->element_size; |
|
|
825 | } |
|
|
826 | |
|
|
827 | /*****************************************************************************/ |
|
|
828 | |
71 | MODULE = OpenCL PACKAGE = OpenCL |
829 | MODULE = OpenCL PACKAGE = OpenCL |
72 | |
830 | |
73 | PROTOTYPES: ENABLE |
831 | PROTOTYPES: ENABLE |
74 | |
832 | |
|
|
833 | void |
|
|
834 | poll () |
|
|
835 | CODE: |
|
|
836 | eq_poll (); |
|
|
837 | |
|
|
838 | void |
|
|
839 | _eq_initialise (IV func, IV arg) |
|
|
840 | CODE: |
|
|
841 | eq_signal_func = (void (*)(void *, int))func; |
|
|
842 | eq_signal_arg = (void*)arg; |
|
|
843 | |
75 | BOOT: |
844 | BOOT: |
76 | { |
845 | { |
77 | HV *stash = gv_stashpv ("OpenCL", 1); |
846 | HV *stash = gv_stashpv ("OpenCL", 1); |
78 | static const struct { |
847 | |
79 | const char *name; |
848 | static const ivstr *civ, const_iv[] = { |
80 | IV iv; |
849 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
81 | } *civ, const_iv[] = { |
850 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
82 | #define const_iv(name) { # name, (IV)CL_ ## name }, |
851 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
|
|
852 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
|
|
853 | { sizeof (cl_int ), "SIZEOF_INT" }, |
|
|
854 | { sizeof (cl_uint ), "SIZEOF_UINT" }, |
|
|
855 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
|
|
856 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
|
|
857 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
|
|
858 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
|
|
859 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
860 | { PREFER_1_1 , "PREFER_1_1" }, |
83 | #include "constiv.h" |
861 | #include "constiv.h" |
84 | }; |
862 | }; |
|
|
863 | |
85 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
864 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
86 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
865 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
866 | |
|
|
867 | static const nvstr *cnv, const_nv[] = { |
|
|
868 | #include "constnv.h" |
|
|
869 | }; |
|
|
870 | |
|
|
871 | for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--) |
|
|
872 | newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv)); |
|
|
873 | |
|
|
874 | newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call |
|
|
875 | |
|
|
876 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
877 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
878 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
879 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
880 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
881 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
882 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
883 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
884 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
885 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
886 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
887 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
888 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
889 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
890 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
891 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
892 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
893 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
894 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
895 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
896 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
897 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
898 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
899 | |
|
|
900 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
87 | } |
901 | } |
|
|
902 | |
|
|
903 | cl_int |
|
|
904 | errno () |
|
|
905 | CODE: |
|
|
906 | RETVAL = res; |
|
|
907 | OUTPUT: |
|
|
908 | RETVAL |
|
|
909 | |
|
|
910 | const char * |
|
|
911 | err2str (cl_int err = res) |
|
|
912 | |
|
|
913 | const char * |
|
|
914 | enum2str (cl_uint value) |
88 | |
915 | |
89 | void |
916 | void |
90 | platforms () |
917 | platforms () |
91 | PPCODE: |
918 | PPCODE: |
92 | { |
|
|
93 | cl_platform_id *list; |
919 | cl_platform_id *list; |
94 | cl_uint count; |
920 | cl_uint count; |
95 | int i; |
921 | int i; |
96 | |
922 | |
97 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
923 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
98 | Newx (list, count, cl_platform_id); |
924 | list = tmpbuf (sizeof (*list) * count); |
99 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
925 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
100 | |
926 | |
101 | EXTEND (SP, count); |
927 | EXTEND (SP, count); |
102 | for (i = 0; i < count; ++i) |
928 | for (i = 0; i < count; ++i) |
103 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
929 | PUSH_CLOBJ (stash_platform, list [i]); |
104 | |
930 | |
105 | Safefree (list); |
|
|
106 | } |
|
|
107 | |
|
|
108 | void |
931 | void |
109 | context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
932 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
110 | PPCODE: |
933 | PPCODE: |
111 | { |
934 | CONTEXT_NOTIFY_CALLBACK; |
112 | cl_int res; |
935 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
113 | cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
936 | XPUSH_CLOBJ_CONTEXT; |
114 | |
937 | |
115 | if (res) |
938 | void |
116 | FAIL (CreateContextFromType, res); |
939 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
|
|
940 | PPCODE: |
|
|
941 | cl_uint device_count; |
|
|
942 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
117 | |
943 | |
118 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
944 | CONTEXT_NOTIFY_CALLBACK; |
119 | } |
945 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
946 | XPUSH_CLOBJ_CONTEXT; |
120 | |
947 | |
121 | void |
948 | void |
122 | wait_for_events (...) |
949 | wait_for_events (...) |
123 | CODE: |
950 | CODE: |
124 | { |
|
|
125 | EVENT_LIST (0, items); |
951 | EVENT_LIST (0); |
126 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
952 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
127 | } |
|
|
128 | |
953 | |
129 | PROTOTYPES: DISABLE |
954 | PROTOTYPES: DISABLE |
130 | |
955 | |
131 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
956 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
132 | |
957 | |
133 | void |
958 | void |
134 | info (OpenCL::Platform this, cl_platform_info name) |
959 | info (OpenCL::Platform self, cl_platform_info name) |
135 | PPCODE: |
960 | PPCODE: |
136 | INFO (Platform) |
961 | INFO (Platform) |
137 | |
962 | |
138 | void |
963 | void |
|
|
964 | unload_compiler (OpenCL::Platform self) |
|
|
965 | CODE: |
|
|
966 | #if CL_VERSION_1_2 |
|
|
967 | clUnloadPlatformCompiler (self); |
|
|
968 | #endif |
|
|
969 | |
|
|
970 | #BEGIN:platform |
|
|
971 | |
|
|
972 | void |
|
|
973 | profile (OpenCL::Platform self) |
|
|
974 | ALIAS: |
|
|
975 | profile = CL_PLATFORM_PROFILE |
|
|
976 | version = CL_PLATFORM_VERSION |
|
|
977 | name = CL_PLATFORM_NAME |
|
|
978 | vendor = CL_PLATFORM_VENDOR |
|
|
979 | extensions = CL_PLATFORM_EXTENSIONS |
|
|
980 | PPCODE: |
|
|
981 | size_t size; |
|
|
982 | NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size)); |
|
|
983 | char *value = tmpbuf (size); |
|
|
984 | NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0)); |
|
|
985 | EXTEND (SP, 1); |
|
|
986 | const int i = 0; |
|
|
987 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
988 | |
|
|
989 | #END:platform |
|
|
990 | |
|
|
991 | void |
139 | devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) |
992 | devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL) |
140 | PPCODE: |
993 | PPCODE: |
141 | { |
|
|
142 | cl_device_id *list; |
994 | cl_device_id *list; |
143 | cl_uint count; |
995 | cl_uint count; |
144 | int i; |
996 | int i; |
145 | |
997 | |
146 | NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); |
998 | NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count)); |
147 | Newx (list, count, cl_device_id); |
999 | list = tmpbuf (sizeof (*list) * count); |
148 | NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); |
1000 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
149 | |
1001 | |
150 | EXTEND (SP, count); |
1002 | EXTEND (SP, count); |
151 | for (i = 0; i < count; ++i) |
1003 | for (i = 0; i < count; ++i) |
152 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
1004 | PUSH_CLOBJ (stash_device, list [i]); |
153 | |
1005 | |
154 | Safefree (list); |
|
|
155 | } |
|
|
156 | |
|
|
157 | void |
1006 | void |
158 | context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
1007 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
159 | PPCODE: |
1008 | PPCODE: |
160 | { |
|
|
161 | cl_int res; |
|
|
162 | cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
1009 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
163 | cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
1010 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
164 | |
1011 | |
165 | if (res) |
1012 | cl_uint device_count; |
166 | FAIL (CreateContextFromType, res); |
1013 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
167 | |
1014 | |
168 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
1015 | CONTEXT_NOTIFY_CALLBACK; |
169 | } |
1016 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
1017 | XPUSH_CLOBJ_CONTEXT; |
170 | |
1018 | |
171 | void |
1019 | void |
172 | unload_compiler () |
1020 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
173 | CODE: |
1021 | PPCODE: |
174 | NEED_SUCCESS (UnloadCompiler, ()); |
1022 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
1023 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
1024 | |
|
|
1025 | CONTEXT_NOTIFY_CALLBACK; |
|
|
1026 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
1027 | XPUSH_CLOBJ_CONTEXT; |
175 | |
1028 | |
176 | MODULE = OpenCL PACKAGE = OpenCL::Device |
1029 | MODULE = OpenCL PACKAGE = OpenCL::Device |
177 | |
1030 | |
178 | void |
1031 | void |
179 | info (OpenCL::Device this, cl_device_info name) |
1032 | info (OpenCL::Device self, cl_device_info name) |
180 | PPCODE: |
1033 | PPCODE: |
181 | INFO (Device) |
1034 | INFO (Device) |
182 | |
1035 | |
|
|
1036 | #if CL_VERSION_1_2 |
|
|
1037 | |
183 | void |
1038 | void |
184 | context_simple (OpenCL::Device this) |
1039 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
1040 | PPCODE: |
|
|
1041 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
|
|
1042 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
|
|
1043 | |
|
|
1044 | properties = SvRV (properties); |
|
|
1045 | |
|
|
1046 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
1047 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
1048 | |
|
|
1049 | int i; |
|
|
1050 | for (i = 0; i < count; ++i) |
|
|
1051 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
|
|
1052 | |
|
|
1053 | props [count] = 0; |
|
|
1054 | |
|
|
1055 | cl_uint num_devices; |
|
|
1056 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
1057 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
1058 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
|
|
1059 | |
|
|
1060 | EXTEND (SP, num_devices); |
|
|
1061 | for (i = 0; i < count; ++i) |
|
|
1062 | PUSH_CLOBJ (stash_subdevice, list [i]); |
|
|
1063 | |
|
|
1064 | #endif |
|
|
1065 | |
|
|
1066 | #BEGIN:device |
|
|
1067 | |
|
|
1068 | void |
|
|
1069 | type (OpenCL::Device self) |
|
|
1070 | ALIAS: |
|
|
1071 | type = CL_DEVICE_TYPE |
|
|
1072 | address_bits = CL_DEVICE_ADDRESS_BITS |
|
|
1073 | max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE |
|
|
1074 | single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG |
|
|
1075 | global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE |
|
|
1076 | global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE |
|
|
1077 | max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE |
|
|
1078 | local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE |
|
|
1079 | execution_capabilities = CL_DEVICE_EXECUTION_CAPABILITIES |
|
|
1080 | properties = CL_DEVICE_QUEUE_PROPERTIES |
|
|
1081 | double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG |
|
|
1082 | half_fp_config = CL_DEVICE_HALF_FP_CONFIG |
185 | PPCODE: |
1083 | PPCODE: |
186 | { |
1084 | cl_ulong value [1]; |
187 | cl_int res; |
1085 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
188 | cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
1086 | EXTEND (SP, 1); |
|
|
1087 | const int i = 0; |
|
|
1088 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
189 | |
1089 | |
190 | if (res) |
1090 | void |
191 | FAIL (CreateContext, res); |
1091 | vendor_id (OpenCL::Device self) |
|
|
1092 | ALIAS: |
|
|
1093 | vendor_id = CL_DEVICE_VENDOR_ID |
|
|
1094 | max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS |
|
|
1095 | max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS |
|
|
1096 | preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR |
|
|
1097 | preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT |
|
|
1098 | preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT |
|
|
1099 | preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG |
|
|
1100 | preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT |
|
|
1101 | preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE |
|
|
1102 | max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY |
|
|
1103 | max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS |
|
|
1104 | max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS |
|
|
1105 | image_support = CL_DEVICE_IMAGE_SUPPORT |
|
|
1106 | max_samplers = CL_DEVICE_MAX_SAMPLERS |
|
|
1107 | mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN |
|
|
1108 | min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE |
|
|
1109 | global_mem_cache_type = CL_DEVICE_GLOBAL_MEM_CACHE_TYPE |
|
|
1110 | global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
|
|
1111 | max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS |
|
|
1112 | local_mem_type = CL_DEVICE_LOCAL_MEM_TYPE |
|
|
1113 | preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF |
|
|
1114 | native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR |
|
|
1115 | native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT |
|
|
1116 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
|
|
1117 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
|
|
1118 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
|
|
1119 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
|
|
1120 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
|
|
1121 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
|
|
1122 | PPCODE: |
|
|
1123 | cl_uint value [1]; |
|
|
1124 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1125 | EXTEND (SP, 1); |
|
|
1126 | const int i = 0; |
|
|
1127 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
192 | |
1128 | |
193 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
1129 | void |
194 | } |
1130 | max_work_group_size (OpenCL::Device self) |
|
|
1131 | ALIAS: |
|
|
1132 | max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE |
|
|
1133 | image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH |
|
|
1134 | image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT |
|
|
1135 | image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH |
|
|
1136 | image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT |
|
|
1137 | image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH |
|
|
1138 | max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE |
|
|
1139 | profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION |
|
|
1140 | PPCODE: |
|
|
1141 | size_t value [1]; |
|
|
1142 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1143 | EXTEND (SP, 1); |
|
|
1144 | const int i = 0; |
|
|
1145 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1146 | |
|
|
1147 | void |
|
|
1148 | max_work_item_sizes (OpenCL::Device self) |
|
|
1149 | PPCODE: |
|
|
1150 | size_t size; |
|
|
1151 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size)); |
|
|
1152 | size_t *value = tmpbuf (size); |
|
|
1153 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0)); |
|
|
1154 | int i, n = size / sizeof (*value); |
|
|
1155 | EXTEND (SP, n); |
|
|
1156 | for (i = 0; i < n; ++i) |
|
|
1157 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1158 | |
|
|
1159 | void |
|
|
1160 | error_correction_support (OpenCL::Device self) |
|
|
1161 | ALIAS: |
|
|
1162 | error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT |
|
|
1163 | endian_little = CL_DEVICE_ENDIAN_LITTLE |
|
|
1164 | available = CL_DEVICE_AVAILABLE |
|
|
1165 | compiler_available = CL_DEVICE_COMPILER_AVAILABLE |
|
|
1166 | host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY |
|
|
1167 | PPCODE: |
|
|
1168 | cl_uint value [1]; |
|
|
1169 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1170 | EXTEND (SP, 1); |
|
|
1171 | const int i = 0; |
|
|
1172 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
1173 | |
|
|
1174 | void |
|
|
1175 | platform (OpenCL::Device self) |
|
|
1176 | PPCODE: |
|
|
1177 | cl_platform_id value [1]; |
|
|
1178 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
|
|
1179 | EXTEND (SP, 1); |
|
|
1180 | const int i = 0; |
|
|
1181 | PUSH_CLOBJ (stash_platform, value [i]); |
|
|
1182 | |
|
|
1183 | void |
|
|
1184 | name (OpenCL::Device self) |
|
|
1185 | ALIAS: |
|
|
1186 | name = CL_DEVICE_NAME |
|
|
1187 | vendor = CL_DEVICE_VENDOR |
|
|
1188 | driver_version = CL_DRIVER_VERSION |
|
|
1189 | profile = CL_DEVICE_PROFILE |
|
|
1190 | version = CL_DEVICE_VERSION |
|
|
1191 | extensions = CL_DEVICE_EXTENSIONS |
|
|
1192 | PPCODE: |
|
|
1193 | size_t size; |
|
|
1194 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1195 | char *value = tmpbuf (size); |
|
|
1196 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1197 | EXTEND (SP, 1); |
|
|
1198 | const int i = 0; |
|
|
1199 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
1200 | |
|
|
1201 | void |
|
|
1202 | parent_device_ext (OpenCL::Device self) |
|
|
1203 | PPCODE: |
|
|
1204 | cl_device_id value [1]; |
|
|
1205 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
|
|
1206 | EXTEND (SP, 1); |
|
|
1207 | const int i = 0; |
|
|
1208 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1209 | |
|
|
1210 | void |
|
|
1211 | partition_types_ext (OpenCL::Device self) |
|
|
1212 | ALIAS: |
|
|
1213 | partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT |
|
|
1214 | affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT |
|
|
1215 | partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT |
|
|
1216 | PPCODE: |
|
|
1217 | size_t size; |
|
|
1218 | NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size)); |
|
|
1219 | cl_device_partition_property_ext *value = tmpbuf (size); |
|
|
1220 | NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0)); |
|
|
1221 | int i, n = size / sizeof (*value); |
|
|
1222 | EXTEND (SP, n); |
|
|
1223 | for (i = 0; i < n; ++i) |
|
|
1224 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1225 | |
|
|
1226 | #END:device |
|
|
1227 | |
|
|
1228 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1229 | |
|
|
1230 | #if CL_VERSION_1_2 |
|
|
1231 | |
|
|
1232 | void |
|
|
1233 | DESTROY (OpenCL::SubDevice self) |
|
|
1234 | CODE: |
|
|
1235 | clReleaseDevice (self); |
|
|
1236 | |
|
|
1237 | #endif |
195 | |
1238 | |
196 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1239 | MODULE = OpenCL PACKAGE = OpenCL::Context |
197 | |
1240 | |
198 | void |
1241 | void |
199 | DESTROY (OpenCL::Context context) |
1242 | DESTROY (OpenCL::Context self) |
200 | CODE: |
1243 | CODE: |
201 | clReleaseContext (context); |
1244 | clReleaseContext (self); |
202 | |
1245 | |
203 | void |
1246 | void |
204 | info (OpenCL::Context this, cl_context_info name) |
1247 | info (OpenCL::Context self, cl_context_info name) |
205 | PPCODE: |
1248 | PPCODE: |
206 | INFO (Context) |
1249 | INFO (Context) |
207 | |
1250 | |
208 | void |
1251 | void |
209 | command_queue_simple (OpenCL::Context this, OpenCL::Device device) |
1252 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
210 | PPCODE: |
1253 | PPCODE: |
211 | { |
1254 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
212 | cl_int res; |
1255 | XPUSH_CLOBJ (stash_queue, queue); |
213 | cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res); |
|
|
214 | |
1256 | |
215 | if (res) |
|
|
216 | FAIL (CreateCommandQueue, res); |
|
|
217 | |
|
|
218 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
|
|
219 | } |
|
|
220 | |
|
|
221 | void |
1257 | void |
|
|
1258 | user_event (OpenCL::Context self) |
|
|
1259 | PPCODE: |
|
|
1260 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
|
|
1261 | XPUSH_CLOBJ (stash_userevent, ev); |
|
|
1262 | |
|
|
1263 | void |
222 | buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) |
1264 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
223 | PPCODE: |
1265 | PPCODE: |
224 | { |
1266 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
225 | cl_int res; |
1267 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
226 | cl_mem mem = clCreateBuffer (this, flags, len, 0, &res); |
1268 | |
|
|
1269 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
|
|
1270 | XPUSH_CLOBJ (stash_bufferobj, mem); |
227 | |
1271 | |
228 | if (res) |
|
|
229 | FAIL (CreateBuffer, res); |
|
|
230 | |
|
|
231 | XPUSH_NEW_OBJ ("OpenCL::Memory", mem); |
|
|
232 | } |
|
|
233 | |
|
|
234 | void |
1272 | void |
235 | buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) |
1273 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
236 | PPCODE: |
1274 | PPCODE: |
237 | { |
|
|
238 | STRLEN len; |
1275 | STRLEN len; |
239 | char *ptr = SvPVbyte (data, len); |
1276 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
240 | cl_int res; |
1277 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
241 | cl_mem mem = clCreateBuffer (this, flags, len, ptr, &res); |
1278 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
|
|
1279 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
|
|
1280 | XPUSH_CLOBJ (stash_bufferobj, mem); |
242 | |
1281 | |
243 | if (res) |
1282 | #if CL_VERSION_1_2 |
244 | FAIL (CreateBuffer, res); |
|
|
245 | |
1283 | |
246 | XPUSH_NEW_OBJ ("OpenCL::Memory", mem); |
|
|
247 | } |
|
|
248 | |
|
|
249 | void |
1284 | void |
|
|
1285 | 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) |
|
|
1286 | PPCODE: |
|
|
1287 | STRLEN len; |
|
|
1288 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1289 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1290 | const cl_image_desc desc = { |
|
|
1291 | type, |
|
|
1292 | width, height, depth, |
|
|
1293 | array_size, row_pitch, slice_pitch, |
|
|
1294 | num_mip_level, num_samples, |
|
|
1295 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
|
|
1296 | }; |
|
|
1297 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1298 | HV *stash = stash_image; |
|
|
1299 | switch (type) |
|
|
1300 | { |
|
|
1301 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1302 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1303 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1304 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1305 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1306 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1307 | } |
|
|
1308 | XPUSH_CLOBJ (stash, mem); |
|
|
1309 | |
|
|
1310 | #endif |
|
|
1311 | |
|
|
1312 | void |
|
|
1313 | 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) |
|
|
1314 | PPCODE: |
|
|
1315 | STRLEN len; |
|
|
1316 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1317 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1318 | #if PREFER_1_1 |
|
|
1319 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1320 | #else |
|
|
1321 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1322 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1323 | #endif |
|
|
1324 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1325 | |
|
|
1326 | void |
|
|
1327 | 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) |
|
|
1328 | PPCODE: |
|
|
1329 | STRLEN len; |
|
|
1330 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1331 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1332 | #if PREFER_1_1 |
|
|
1333 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
|
|
1334 | #else |
|
|
1335 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1336 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1337 | #endif |
|
|
1338 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1339 | |
|
|
1340 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1341 | |
|
|
1342 | void |
|
|
1343 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
|
|
1344 | PPCODE: |
|
|
1345 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
|
|
1346 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1347 | |
|
|
1348 | void |
|
|
1349 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1350 | PPCODE: |
|
|
1351 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1352 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1353 | |
|
|
1354 | #if CL_VERSION_1_2 |
|
|
1355 | |
|
|
1356 | void |
|
|
1357 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1358 | ALIAS: |
|
|
1359 | PPCODE: |
|
|
1360 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1361 | cl_gl_object_type type; |
|
|
1362 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1363 | HV *stash = stash_memory; |
|
|
1364 | switch (type) |
|
|
1365 | { |
|
|
1366 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1367 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1368 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1369 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1370 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1371 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1372 | } |
|
|
1373 | XPUSH_CLOBJ (stash, mem); |
|
|
1374 | |
|
|
1375 | #endif |
|
|
1376 | |
|
|
1377 | void |
|
|
1378 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1379 | PPCODE: |
|
|
1380 | #if PREFER_1_1 |
|
|
1381 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
1382 | #else |
|
|
1383 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1384 | #endif |
|
|
1385 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1386 | |
|
|
1387 | void |
|
|
1388 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1389 | PPCODE: |
|
|
1390 | #if PREFER_1_1 |
|
|
1391 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
1392 | #else |
|
|
1393 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1394 | #endif |
|
|
1395 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1396 | |
|
|
1397 | #endif |
|
|
1398 | |
|
|
1399 | void |
|
|
1400 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
|
|
1401 | PPCODE: |
|
|
1402 | { |
|
|
1403 | cl_uint count; |
|
|
1404 | cl_image_format *list; |
|
|
1405 | int i; |
|
|
1406 | |
|
|
1407 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count)); |
|
|
1408 | Newx (list, count, cl_image_format); |
|
|
1409 | NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0)); |
|
|
1410 | |
|
|
1411 | EXTEND (SP, count); |
|
|
1412 | for (i = 0; i < count; ++i) |
|
|
1413 | { |
|
|
1414 | AV *av = newAV (); |
|
|
1415 | av_store (av, 1, newSVuv (list [i].image_channel_data_type)); |
|
|
1416 | av_store (av, 0, newSVuv (list [i].image_channel_order)); |
|
|
1417 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1418 | } |
|
|
1419 | } |
|
|
1420 | |
|
|
1421 | void |
250 | sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1422 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
251 | PPCODE: |
1423 | PPCODE: |
252 | { |
|
|
253 | cl_int res; |
|
|
254 | cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); |
1424 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
255 | |
|
|
256 | if (res) |
|
|
257 | FAIL (CreateSampler, res); |
|
|
258 | |
|
|
259 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1425 | XPUSH_CLOBJ (stash_sampler, sampler); |
260 | } |
|
|
261 | |
1426 | |
262 | void |
1427 | void |
263 | program_with_source (OpenCL::Context this, SV *program) |
1428 | program_with_source (OpenCL::Context self, SV *program) |
264 | PPCODE: |
1429 | PPCODE: |
265 | { |
|
|
266 | STRLEN len; |
1430 | STRLEN len; |
267 | size_t len2; |
1431 | size_t len2; |
268 | const char *ptr = SvPVbyte (program, len); |
1432 | const char *ptr = SvPVbyte (program, len); |
269 | cl_int res; |
|
|
270 | cl_program prog; |
|
|
271 | |
1433 | |
272 | len2 = len; |
1434 | len2 = len; |
273 | prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); |
1435 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
274 | |
|
|
275 | if (res) |
|
|
276 | FAIL (CreateProgramWithSource, res); |
|
|
277 | |
|
|
278 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1436 | XPUSH_CLOBJ (stash_program, prog); |
279 | } |
1437 | |
|
|
1438 | void |
|
|
1439 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1440 | PPCODE: |
|
|
1441 | cl_uint device_count; |
|
|
1442 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1443 | |
|
|
1444 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1445 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1446 | |
|
|
1447 | binaries = SvRV (binaries); |
|
|
1448 | |
|
|
1449 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1450 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1451 | |
|
|
1452 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1453 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1454 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1455 | |
|
|
1456 | int i; |
|
|
1457 | for (i = 0; i < device_count; ++i) |
|
|
1458 | { |
|
|
1459 | STRLEN len; |
|
|
1460 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1461 | length_list [i] = len; |
|
|
1462 | } |
|
|
1463 | |
|
|
1464 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1465 | length_list, binary_list, |
|
|
1466 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1467 | |
|
|
1468 | EXTEND (SP, 2); |
|
|
1469 | PUSH_CLOBJ (stash_program, prog); |
|
|
1470 | |
|
|
1471 | if (GIMME_V == G_ARRAY) |
|
|
1472 | { |
|
|
1473 | AV *av = newAV (); |
|
|
1474 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1475 | |
|
|
1476 | for (i = device_count; i--; ) |
|
|
1477 | av_store (av, i, newSViv (status_list [i])); |
|
|
1478 | } |
|
|
1479 | |
|
|
1480 | #if CL_VERSION_1_2 |
|
|
1481 | |
|
|
1482 | void |
|
|
1483 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1484 | PPCODE: |
|
|
1485 | cl_uint device_count; |
|
|
1486 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1487 | |
|
|
1488 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1489 | |
|
|
1490 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1491 | |
|
|
1492 | void |
|
|
1493 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1494 | CODE: |
|
|
1495 | cl_uint device_count = 0; |
|
|
1496 | cl_device_id *device_list = 0; |
|
|
1497 | |
|
|
1498 | if (SvOK (devices)) |
|
|
1499 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1500 | |
|
|
1501 | cl_uint program_count; |
|
|
1502 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1503 | |
|
|
1504 | void *user_data; |
|
|
1505 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1506 | |
|
|
1507 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1508 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1509 | |
|
|
1510 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1511 | |
|
|
1512 | #endif |
|
|
1513 | |
|
|
1514 | #BEGIN:context |
|
|
1515 | |
|
|
1516 | void |
|
|
1517 | reference_count (OpenCL::Context self) |
|
|
1518 | ALIAS: |
|
|
1519 | reference_count = CL_CONTEXT_REFERENCE_COUNT |
|
|
1520 | num_devices = CL_CONTEXT_NUM_DEVICES |
|
|
1521 | PPCODE: |
|
|
1522 | cl_uint value [1]; |
|
|
1523 | NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1524 | EXTEND (SP, 1); |
|
|
1525 | const int i = 0; |
|
|
1526 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1527 | |
|
|
1528 | void |
|
|
1529 | devices (OpenCL::Context self) |
|
|
1530 | PPCODE: |
|
|
1531 | size_t size; |
|
|
1532 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size)); |
|
|
1533 | cl_device_id *value = tmpbuf (size); |
|
|
1534 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
|
|
1535 | int i, n = size / sizeof (*value); |
|
|
1536 | EXTEND (SP, n); |
|
|
1537 | for (i = 0; i < n; ++i) |
|
|
1538 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
1539 | |
|
|
1540 | void |
|
|
1541 | properties (OpenCL::Context self) |
|
|
1542 | PPCODE: |
|
|
1543 | size_t size; |
|
|
1544 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size)); |
|
|
1545 | cl_context_properties *value = tmpbuf (size); |
|
|
1546 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0)); |
|
|
1547 | int i, n = size / sizeof (*value); |
|
|
1548 | EXTEND (SP, n); |
|
|
1549 | for (i = 0; i < n; ++i) |
|
|
1550 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
1551 | |
|
|
1552 | #END:context |
280 | |
1553 | |
281 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
1554 | MODULE = OpenCL PACKAGE = OpenCL::Queue |
282 | |
1555 | |
283 | void |
1556 | void |
284 | DESTROY (OpenCL::Queue this) |
1557 | DESTROY (OpenCL::Queue self) |
285 | CODE: |
1558 | CODE: |
286 | clReleaseCommandQueue (this); |
1559 | clReleaseCommandQueue (self); |
287 | |
1560 | |
288 | void |
1561 | void |
289 | info (OpenCL::Queue this, cl_command_queue_info name) |
|
|
290 | PPCODE: |
|
|
291 | INFO (CommandQueue) |
|
|
292 | |
|
|
293 | void |
|
|
294 | enqueue_read_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1562 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1563 | ALIAS: |
|
|
1564 | enqueue_read_buffer = 0 |
295 | PPCODE: |
1565 | PPCODE: |
296 | { |
|
|
297 | cl_event ev = 0; |
|
|
298 | EVENT_LIST (6, items - 6); |
1566 | EVENT_LIST (6); |
299 | |
1567 | |
300 | SvUPGRADE (data, SVt_PV); |
1568 | SvUPGRADE (data, SVt_PV); |
301 | SvGROW (data, len); |
1569 | SvGROW (data, len); |
302 | SvPOK_only (data); |
1570 | SvPOK_only (data); |
303 | SvCUR_set (data, len); |
1571 | SvCUR_set (data, len); |
|
|
1572 | |
|
|
1573 | cl_event ev = 0; |
304 | NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1574 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
305 | |
1575 | |
306 | if (ev) |
1576 | if (ev) |
307 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1577 | XPUSH_CLOBJ (stash_event, ev); |
308 | } |
|
|
309 | |
1578 | |
310 | void |
1579 | void |
311 | enqueue_write_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, SV *data, ...) |
1580 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1581 | ALIAS: |
|
|
1582 | enqueue_write_buffer = 0 |
312 | PPCODE: |
1583 | PPCODE: |
313 | { |
1584 | EVENT_LIST (5); |
314 | cl_event ev = 0; |
1585 | |
315 | STRLEN len; |
1586 | STRLEN len; |
316 | char *ptr = SvPVbyte (data, len); |
1587 | char *ptr = SvPVbyte (data, len); |
317 | EVENT_LIST (5, items - 5); |
|
|
318 | |
1588 | |
|
|
1589 | cl_event ev = 0; |
319 | NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1590 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
320 | |
1591 | |
321 | if (ev) |
1592 | if (ev) |
322 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1593 | XPUSH_CLOBJ (stash_event, ev); |
323 | } |
|
|
324 | |
1594 | |
|
|
1595 | #if CL_VERSION_1_2 |
|
|
1596 | |
325 | void |
1597 | void |
326 | enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Memory src, OpenCL::Memory dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1598 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1599 | ALIAS: |
|
|
1600 | enqueue_fill_buffer = 0 |
327 | PPCODE: |
1601 | PPCODE: |
328 | { |
1602 | EVENT_LIST (5); |
|
|
1603 | |
|
|
1604 | STRLEN len; |
|
|
1605 | char *ptr = SvPVbyte (data, len); |
|
|
1606 | |
329 | cl_event ev = 0; |
1607 | cl_event ev = 0; |
330 | EVENT_LIST (6, items - 6); |
|
|
331 | |
|
|
332 | NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1608 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
333 | |
1609 | |
334 | if (ev) |
1610 | if (ev) |
335 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1611 | XPUSH_CLOBJ (stash_event, ev); |
336 | } |
|
|
337 | |
1612 | |
338 | void |
1613 | void |
339 | enqueue_marker (OpenCL::Queue this) |
1614 | 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, ...) |
|
|
1615 | ALIAS: |
|
|
1616 | enqueue_fill_image = 0 |
340 | PPCODE: |
1617 | PPCODE: |
341 | { |
1618 | EVENT_LIST (12); |
|
|
1619 | |
|
|
1620 | const size_t origin [3] = { x, y, z }; |
|
|
1621 | const size_t region [3] = { width, height, depth }; |
|
|
1622 | |
|
|
1623 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1624 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1625 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1626 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1627 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1628 | cl_image_format format; |
|
|
1629 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1630 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1631 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1632 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1633 | |
|
|
1634 | cl_event ev = 0; |
|
|
1635 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1636 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1637 | |
|
|
1638 | if (ev) |
|
|
1639 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1640 | |
|
|
1641 | #endif |
|
|
1642 | |
|
|
1643 | void |
|
|
1644 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1645 | ALIAS: |
|
|
1646 | enqueue_copy_buffer = 0 |
|
|
1647 | PPCODE: |
|
|
1648 | EVENT_LIST (6); |
|
|
1649 | |
|
|
1650 | cl_event ev = 0; |
|
|
1651 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1652 | |
|
|
1653 | if (ev) |
|
|
1654 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1655 | |
|
|
1656 | void |
|
|
1657 | 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, ...) |
|
|
1658 | ALIAS: |
|
|
1659 | enqueue_read_buffer_rect = 0 |
|
|
1660 | PPCODE: |
|
|
1661 | EVENT_LIST (17); |
|
|
1662 | |
|
|
1663 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
|
|
1664 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
|
|
1665 | const size_t region[3] = { width, height, depth }; |
|
|
1666 | |
|
|
1667 | if (!buf_row_pitch) |
|
|
1668 | buf_row_pitch = region [0]; |
|
|
1669 | |
|
|
1670 | if (!buf_slice_pitch) |
|
|
1671 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1672 | |
|
|
1673 | if (!host_row_pitch) |
|
|
1674 | host_row_pitch = region [0]; |
|
|
1675 | |
|
|
1676 | if (!host_slice_pitch) |
|
|
1677 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1678 | |
|
|
1679 | size_t len = host_row_pitch * host_slice_pitch * region [2]; |
|
|
1680 | |
|
|
1681 | SvUPGRADE (data, SVt_PV); |
|
|
1682 | SvGROW (data, len); |
|
|
1683 | SvPOK_only (data); |
|
|
1684 | SvCUR_set (data, len); |
|
|
1685 | |
|
|
1686 | cl_event ev = 0; |
|
|
1687 | 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)); |
|
|
1688 | |
|
|
1689 | if (ev) |
|
|
1690 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1691 | |
|
|
1692 | void |
|
|
1693 | 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, ...) |
|
|
1694 | ALIAS: |
|
|
1695 | enqueue_write_buffer_rect = 0 |
|
|
1696 | PPCODE: |
|
|
1697 | EVENT_LIST (17); |
|
|
1698 | |
|
|
1699 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
|
|
1700 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
|
|
1701 | const size_t region[3] = { width, height, depth }; |
|
|
1702 | STRLEN len; |
|
|
1703 | char *ptr = SvPVbyte (data, len); |
|
|
1704 | |
|
|
1705 | if (!buf_row_pitch) |
|
|
1706 | buf_row_pitch = region [0]; |
|
|
1707 | |
|
|
1708 | if (!buf_slice_pitch) |
|
|
1709 | buf_slice_pitch = region [1] * buf_row_pitch; |
|
|
1710 | |
|
|
1711 | if (!host_row_pitch) |
|
|
1712 | host_row_pitch = region [0]; |
|
|
1713 | |
|
|
1714 | if (!host_slice_pitch) |
|
|
1715 | host_slice_pitch = region [1] * host_row_pitch; |
|
|
1716 | |
|
|
1717 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
|
|
1718 | |
|
|
1719 | if (len < min_len) |
|
|
1720 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1721 | |
|
|
1722 | cl_event ev = 0; |
|
|
1723 | 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)); |
|
|
1724 | |
|
|
1725 | if (ev) |
|
|
1726 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1727 | |
|
|
1728 | void |
|
|
1729 | 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, ...) |
|
|
1730 | ALIAS: |
|
|
1731 | enqueue_copy_buffer_rect = 0 |
|
|
1732 | PPCODE: |
|
|
1733 | EVENT_LIST (16); |
|
|
1734 | |
|
|
1735 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1736 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1737 | const size_t region[3] = { width, height, depth }; |
|
|
1738 | |
|
|
1739 | cl_event ev = 0; |
|
|
1740 | 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)); |
|
|
1741 | |
|
|
1742 | if (ev) |
|
|
1743 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1744 | |
|
|
1745 | void |
|
|
1746 | 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, ...) |
|
|
1747 | ALIAS: |
|
|
1748 | enqueue_read_image = 0 |
|
|
1749 | PPCODE: |
|
|
1750 | EVENT_LIST (12); |
|
|
1751 | |
|
|
1752 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1753 | const size_t region[3] = { width, height, depth }; |
|
|
1754 | |
|
|
1755 | if (!row_pitch) |
|
|
1756 | row_pitch = img_row_pitch (src); |
|
|
1757 | |
|
|
1758 | if (depth > 1 && !slice_pitch) |
|
|
1759 | slice_pitch = row_pitch * height; |
|
|
1760 | |
|
|
1761 | size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1762 | |
|
|
1763 | SvUPGRADE (data, SVt_PV); |
|
|
1764 | SvGROW (data, len); |
|
|
1765 | SvPOK_only (data); |
|
|
1766 | SvCUR_set (data, len); |
|
|
1767 | |
|
|
1768 | cl_event ev = 0; |
|
|
1769 | 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)); |
|
|
1770 | |
|
|
1771 | if (ev) |
|
|
1772 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1773 | |
|
|
1774 | void |
|
|
1775 | 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, ...) |
|
|
1776 | ALIAS: |
|
|
1777 | enqueue_write_image = 0 |
|
|
1778 | PPCODE: |
|
|
1779 | EVENT_LIST (12); |
|
|
1780 | |
|
|
1781 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1782 | const size_t region[3] = { width, height, depth }; |
|
|
1783 | STRLEN len; |
|
|
1784 | char *ptr = SvPVbyte (data, len); |
|
|
1785 | |
|
|
1786 | if (!row_pitch) |
|
|
1787 | row_pitch = img_row_pitch (dst); |
|
|
1788 | |
|
|
1789 | if (depth > 1 && !slice_pitch) |
|
|
1790 | slice_pitch = row_pitch * height; |
|
|
1791 | |
|
|
1792 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
|
|
1793 | |
|
|
1794 | if (len < min_len) |
|
|
1795 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
|
|
1796 | |
|
|
1797 | cl_event ev = 0; |
|
|
1798 | 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)); |
|
|
1799 | |
|
|
1800 | if (ev) |
|
|
1801 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1802 | |
|
|
1803 | void |
|
|
1804 | 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, ...) |
|
|
1805 | ALIAS: |
|
|
1806 | enqueue_copy_image = 0 |
|
|
1807 | PPCODE: |
|
|
1808 | EVENT_LIST (12); |
|
|
1809 | |
|
|
1810 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1811 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1812 | const size_t region[3] = { width, height, depth }; |
|
|
1813 | |
|
|
1814 | cl_event ev = 0; |
|
|
1815 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1816 | |
|
|
1817 | if (ev) |
|
|
1818 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1819 | |
|
|
1820 | void |
|
|
1821 | 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, ...) |
|
|
1822 | ALIAS: |
|
|
1823 | enqueue_copy_image_to_buffer = 0 |
|
|
1824 | PPCODE: |
|
|
1825 | EVENT_LIST (10); |
|
|
1826 | |
|
|
1827 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1828 | const size_t region [3] = { width, height, depth }; |
|
|
1829 | |
|
|
1830 | cl_event ev = 0; |
|
|
1831 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1832 | |
|
|
1833 | if (ev) |
|
|
1834 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1835 | |
|
|
1836 | void |
|
|
1837 | 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, ...) |
|
|
1838 | ALIAS: |
|
|
1839 | enqueue_copy_buffer_to_image = 0 |
|
|
1840 | PPCODE: |
|
|
1841 | EVENT_LIST (10); |
|
|
1842 | |
|
|
1843 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1844 | const size_t region [3] = { width, height, depth }; |
|
|
1845 | |
|
|
1846 | cl_event ev = 0; |
|
|
1847 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1848 | |
|
|
1849 | if (ev) |
|
|
1850 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1851 | |
|
|
1852 | void |
|
|
1853 | 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, ...) |
|
|
1854 | ALIAS: |
|
|
1855 | enqueue_map_buffer = 0 |
|
|
1856 | PPCODE: |
|
|
1857 | EVENT_LIST (6); |
|
|
1858 | |
|
|
1859 | size_t cb = SvIV (cb_); |
|
|
1860 | |
|
|
1861 | if (!SvOK (cb_)) |
|
|
1862 | { |
|
|
1863 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1864 | cb -= offset; |
|
|
1865 | } |
|
|
1866 | |
342 | cl_event ev; |
1867 | cl_event ev; |
343 | NEED_SUCCESS (EnqueueMarker, (this, &ev)); |
1868 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
344 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1869 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0, 1, cb, 1, 1)); |
345 | } |
|
|
346 | |
1870 | |
347 | void |
1871 | void |
348 | enqueue_wait_for_events (OpenCL::Queue this, ...) |
1872 | 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, ...) |
|
|
1873 | ALIAS: |
|
|
1874 | enqueue_map_image = 0 |
|
|
1875 | PPCODE: |
|
|
1876 | size_t width = SvIV (width_); |
|
|
1877 | if (!SvOK (width_)) |
|
|
1878 | { |
|
|
1879 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1880 | width -= x; |
|
|
1881 | } |
|
|
1882 | |
|
|
1883 | size_t height = SvIV (width_); |
|
|
1884 | if (!SvOK (height_)) |
|
|
1885 | { |
|
|
1886 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1887 | height -= y; |
|
|
1888 | |
|
|
1889 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1890 | if (!height) |
|
|
1891 | height = 1; |
|
|
1892 | } |
|
|
1893 | |
|
|
1894 | size_t depth = SvIV (width_); |
|
|
1895 | if (!SvOK (depth_)) |
|
|
1896 | { |
|
|
1897 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1898 | depth -= z; |
|
|
1899 | |
|
|
1900 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1901 | if (!depth) |
|
|
1902 | depth = 1; |
|
|
1903 | } |
|
|
1904 | |
|
|
1905 | const size_t origin[3] = { x, y, z }; |
|
|
1906 | const size_t region[3] = { width, height, depth }; |
|
|
1907 | size_t row_pitch, slice_pitch; |
|
|
1908 | EVENT_LIST (10); |
|
|
1909 | |
|
|
1910 | cl_event ev; |
|
|
1911 | 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)); |
|
|
1912 | |
|
|
1913 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1914 | : row_pitch ? row_pitch * region [1] |
|
|
1915 | : region [0]; |
|
|
1916 | |
|
|
1917 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch, 0, width, height, depth)); |
|
|
1918 | |
|
|
1919 | void |
|
|
1920 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
349 | CODE: |
1921 | PPCODE: |
350 | { |
1922 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2); |
|
|
1923 | if (GIMME_V != G_VOID) |
|
|
1924 | { |
|
|
1925 | clRetainEvent (mapped->event); |
|
|
1926 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1927 | } |
|
|
1928 | |
|
|
1929 | void |
|
|
1930 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1931 | ALIAS: |
|
|
1932 | enqueue_task = 0 |
|
|
1933 | PPCODE: |
|
|
1934 | EVENT_LIST (2); |
|
|
1935 | |
|
|
1936 | cl_event ev = 0; |
|
|
1937 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1938 | |
|
|
1939 | if (ev) |
|
|
1940 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1941 | |
|
|
1942 | void |
|
|
1943 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1944 | ALIAS: |
|
|
1945 | enqueue_nd_range_kernel = 0 |
|
|
1946 | PPCODE: |
|
|
1947 | EVENT_LIST (5); |
|
|
1948 | |
|
|
1949 | size_t *gwo = 0, *gws, *lws = 0; |
|
|
1950 | int gws_len; |
|
|
1951 | size_t *lists; |
|
|
1952 | int i; |
|
|
1953 | |
|
|
1954 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
|
|
1955 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
|
|
1956 | |
|
|
1957 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
|
|
1958 | |
|
|
1959 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
|
|
1960 | |
|
|
1961 | gws = lists + gws_len * 0; |
|
|
1962 | for (i = 0; i < gws_len; ++i) |
|
|
1963 | { |
|
|
1964 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1965 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1966 | if (!gws [i]) |
|
|
1967 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1968 | } |
|
|
1969 | |
|
|
1970 | if (SvOK (global_work_offset)) |
|
|
1971 | { |
|
|
1972 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
|
|
1973 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
|
|
1974 | |
|
|
1975 | if (AvFILLp (SvRV (global_work_size)) + 1 != gws_len) |
|
|
1976 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array of same size as global_work_size"); |
|
|
1977 | |
|
|
1978 | gwo = lists + gws_len * 1; |
|
|
1979 | for (i = 0; i < gws_len; ++i) |
|
|
1980 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
|
|
1981 | } |
|
|
1982 | |
|
|
1983 | if (SvOK (local_work_size)) |
|
|
1984 | { |
|
|
1985 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
|
|
1986 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
|
|
1987 | |
|
|
1988 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
|
|
1989 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
|
|
1990 | |
|
|
1991 | lws = lists + gws_len * 2; |
|
|
1992 | for (i = 0; i < gws_len; ++i) |
|
|
1993 | { |
|
|
1994 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1995 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1996 | if (!lws [i]) |
|
|
1997 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1998 | } |
|
|
1999 | } |
|
|
2000 | |
|
|
2001 | cl_event ev = 0; |
|
|
2002 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2003 | |
|
|
2004 | if (ev) |
|
|
2005 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2006 | |
|
|
2007 | #if CL_VERSION_1_2 |
|
|
2008 | |
|
|
2009 | void |
|
|
2010 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
2011 | ALIAS: |
|
|
2012 | enqueue_migrate_mem_objects = 0 |
|
|
2013 | PPCODE: |
|
|
2014 | EVENT_LIST (3); |
|
|
2015 | |
|
|
2016 | cl_uint obj_count; |
|
|
2017 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2018 | |
|
|
2019 | cl_event ev = 0; |
|
|
2020 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2021 | |
|
|
2022 | if (ev) |
|
|
2023 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2024 | |
|
|
2025 | #endif |
|
|
2026 | |
|
|
2027 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2028 | |
|
|
2029 | void |
|
|
2030 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
2031 | ALIAS: |
|
|
2032 | release_gl_objects = 1 |
|
|
2033 | enqueue_acquire_gl_objects = 0 |
|
|
2034 | enqueue_release_gl_objects = 1 |
|
|
2035 | PPCODE: |
|
|
2036 | EVENT_LIST (2); |
|
|
2037 | |
|
|
2038 | cl_uint obj_count; |
|
|
2039 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2040 | |
|
|
2041 | cl_event ev = 0; |
|
|
2042 | |
|
|
2043 | if (ix) |
|
|
2044 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2045 | else |
|
|
2046 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2047 | |
|
|
2048 | if (ev) |
|
|
2049 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2050 | |
|
|
2051 | #endif |
|
|
2052 | |
|
|
2053 | void |
|
|
2054 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2055 | ALIAS: |
|
|
2056 | enqueue_wait_for_events = 0 |
|
|
2057 | CODE: |
351 | EVENT_LIST (1, items - 1); |
2058 | EVENT_LIST (1); |
|
|
2059 | #if PREFER_1_1 |
352 | NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); |
2060 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
353 | } |
2061 | #else |
|
|
2062 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2063 | #endif |
354 | |
2064 | |
355 | void |
2065 | void |
356 | enqueue_barrier (OpenCL::Queue this) |
2066 | marker (OpenCL::Queue self, ...) |
|
|
2067 | ALIAS: |
|
|
2068 | enqueue_marker = 0 |
357 | CODE: |
2069 | PPCODE: |
|
|
2070 | EVENT_LIST (1); |
|
|
2071 | cl_event ev = 0; |
|
|
2072 | #if PREFER_1_1 |
|
|
2073 | if (!event_list_count) |
|
|
2074 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2075 | else |
|
|
2076 | #if CL_VERSION_1_2 |
|
|
2077 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2078 | #else |
|
|
2079 | { |
|
|
2080 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2081 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2082 | } |
|
|
2083 | #endif |
|
|
2084 | #else |
|
|
2085 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2086 | #endif |
|
|
2087 | if (ev) |
|
|
2088 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2089 | |
|
|
2090 | void |
|
|
2091 | barrier (OpenCL::Queue self, ...) |
|
|
2092 | ALIAS: |
|
|
2093 | enqueue_barrier = 0 |
|
|
2094 | PPCODE: |
|
|
2095 | EVENT_LIST (1); |
|
|
2096 | cl_event ev = 0; |
|
|
2097 | #if PREFER_1_1 |
|
|
2098 | if (!event_list_count && GIMME_V == G_VOID) |
358 | NEED_SUCCESS (EnqueueBarrier, (this)); |
2099 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2100 | else |
|
|
2101 | #if CL_VERSION_1_2 |
|
|
2102 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2103 | #else |
|
|
2104 | { |
|
|
2105 | if (event_list_count) |
|
|
2106 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2107 | |
|
|
2108 | if (GIMME_V != G_VOID) |
|
|
2109 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2110 | } |
|
|
2111 | #endif |
|
|
2112 | #else |
|
|
2113 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2114 | #endif |
|
|
2115 | if (ev) |
|
|
2116 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2117 | |
|
|
2118 | void |
|
|
2119 | flush (OpenCL::Queue self) |
|
|
2120 | CODE: |
|
|
2121 | NEED_SUCCESS (Flush, (self)); |
|
|
2122 | |
|
|
2123 | void |
|
|
2124 | finish (OpenCL::Queue self) |
|
|
2125 | CODE: |
|
|
2126 | NEED_SUCCESS (Finish, (self)); |
|
|
2127 | |
|
|
2128 | void |
|
|
2129 | info (OpenCL::Queue self, cl_command_queue_info name) |
|
|
2130 | PPCODE: |
|
|
2131 | INFO (CommandQueue) |
|
|
2132 | |
|
|
2133 | #BEGIN:command_queue |
|
|
2134 | |
|
|
2135 | void |
|
|
2136 | context (OpenCL::Queue self) |
|
|
2137 | PPCODE: |
|
|
2138 | cl_context value [1]; |
|
|
2139 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
|
|
2140 | EXTEND (SP, 1); |
|
|
2141 | const int i = 0; |
|
|
2142 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2143 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2144 | |
|
|
2145 | void |
|
|
2146 | device (OpenCL::Queue self) |
|
|
2147 | PPCODE: |
|
|
2148 | cl_device_id value [1]; |
|
|
2149 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
|
|
2150 | EXTEND (SP, 1); |
|
|
2151 | const int i = 0; |
|
|
2152 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2153 | |
|
|
2154 | void |
|
|
2155 | reference_count (OpenCL::Queue self) |
|
|
2156 | PPCODE: |
|
|
2157 | cl_uint value [1]; |
|
|
2158 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0)); |
|
|
2159 | EXTEND (SP, 1); |
|
|
2160 | const int i = 0; |
|
|
2161 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2162 | |
|
|
2163 | void |
|
|
2164 | properties (OpenCL::Queue self) |
|
|
2165 | PPCODE: |
|
|
2166 | cl_ulong value [1]; |
|
|
2167 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
2168 | EXTEND (SP, 1); |
|
|
2169 | const int i = 0; |
|
|
2170 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2171 | |
|
|
2172 | #END:command_queue |
359 | |
2173 | |
360 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2174 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
361 | |
2175 | |
362 | void |
2176 | void |
363 | DESTROY (OpenCL::Memory this) |
2177 | DESTROY (OpenCL::Memory self) |
364 | CODE: |
2178 | CODE: |
365 | clReleaseMemObject (this); |
2179 | clReleaseMemObject (self); |
366 | |
2180 | |
367 | void |
2181 | void |
368 | info (OpenCL::Memory this, cl_mem_info name) |
2182 | info (OpenCL::Memory self, cl_mem_info name) |
369 | PPCODE: |
2183 | PPCODE: |
370 | INFO (MemObject) |
2184 | INFO (MemObject) |
371 | |
2185 | |
|
|
2186 | void |
|
|
2187 | destructor_callback (OpenCL::Memory self, SV *notify) |
|
|
2188 | PPCODE: |
|
|
2189 | clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify))); |
|
|
2190 | |
|
|
2191 | #BEGIN:mem |
|
|
2192 | |
|
|
2193 | void |
|
|
2194 | type (OpenCL::Memory self) |
|
|
2195 | ALIAS: |
|
|
2196 | type = CL_MEM_TYPE |
|
|
2197 | map_count = CL_MEM_MAP_COUNT |
|
|
2198 | reference_count = CL_MEM_REFERENCE_COUNT |
|
|
2199 | PPCODE: |
|
|
2200 | cl_uint value [1]; |
|
|
2201 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2202 | EXTEND (SP, 1); |
|
|
2203 | const int i = 0; |
|
|
2204 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2205 | |
|
|
2206 | void |
|
|
2207 | flags (OpenCL::Memory self) |
|
|
2208 | PPCODE: |
|
|
2209 | cl_ulong value [1]; |
|
|
2210 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
|
|
2211 | EXTEND (SP, 1); |
|
|
2212 | const int i = 0; |
|
|
2213 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2214 | |
|
|
2215 | void |
|
|
2216 | size (OpenCL::Memory self) |
|
|
2217 | ALIAS: |
|
|
2218 | size = CL_MEM_SIZE |
|
|
2219 | offset = CL_MEM_OFFSET |
|
|
2220 | PPCODE: |
|
|
2221 | size_t value [1]; |
|
|
2222 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2223 | EXTEND (SP, 1); |
|
|
2224 | const int i = 0; |
|
|
2225 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2226 | |
|
|
2227 | void |
|
|
2228 | host_ptr (OpenCL::Memory self) |
|
|
2229 | PPCODE: |
|
|
2230 | void * value [1]; |
|
|
2231 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0)); |
|
|
2232 | EXTEND (SP, 1); |
|
|
2233 | const int i = 0; |
|
|
2234 | PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i]))); |
|
|
2235 | |
|
|
2236 | void |
|
|
2237 | context (OpenCL::Memory self) |
|
|
2238 | PPCODE: |
|
|
2239 | cl_context value [1]; |
|
|
2240 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
|
|
2241 | EXTEND (SP, 1); |
|
|
2242 | const int i = 0; |
|
|
2243 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2244 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2245 | |
|
|
2246 | void |
|
|
2247 | associated_memobject (OpenCL::Memory self) |
|
|
2248 | PPCODE: |
|
|
2249 | cl_mem value [1]; |
|
|
2250 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
|
|
2251 | EXTEND (SP, 1); |
|
|
2252 | const int i = 0; |
|
|
2253 | NEED_SUCCESS (RetainMemObject, (value [i])); |
|
|
2254 | PUSH_CLOBJ (stash_memory, value [i]); |
|
|
2255 | |
|
|
2256 | #END:mem |
|
|
2257 | |
|
|
2258 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2259 | |
|
|
2260 | void |
|
|
2261 | gl_object_info (OpenCL::Memory self) |
|
|
2262 | PPCODE: |
|
|
2263 | cl_gl_object_type type; |
|
|
2264 | cl_GLuint name; |
|
|
2265 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
|
|
2266 | EXTEND (SP, 2); |
|
|
2267 | PUSHs (sv_2mortal (newSVuv (type))); |
|
|
2268 | PUSHs (sv_2mortal (newSVuv (name))); |
|
|
2269 | |
|
|
2270 | #endif |
|
|
2271 | |
|
|
2272 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
|
|
2273 | |
|
|
2274 | void |
|
|
2275 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
|
|
2276 | PPCODE: |
|
|
2277 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR)) |
|
|
2278 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
|
|
2279 | |
|
|
2280 | cl_buffer_region crdata = { origin, size }; |
|
|
2281 | |
|
|
2282 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
|
|
2283 | XPUSH_CLOBJ (stash_buffer, mem); |
|
|
2284 | |
|
|
2285 | MODULE = OpenCL PACKAGE = OpenCL::Image |
|
|
2286 | |
|
|
2287 | void |
|
|
2288 | image_info (OpenCL::Image self, cl_image_info name) |
|
|
2289 | PPCODE: |
|
|
2290 | INFO (Image) |
|
|
2291 | |
|
|
2292 | void |
|
|
2293 | format (OpenCL::Image self) |
|
|
2294 | PPCODE: |
|
|
2295 | cl_image_format format; |
|
|
2296 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2297 | EXTEND (SP, 2); |
|
|
2298 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2299 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
|
|
2300 | |
|
|
2301 | #BEGIN:image |
|
|
2302 | |
|
|
2303 | void |
|
|
2304 | element_size (OpenCL::Image self) |
|
|
2305 | ALIAS: |
|
|
2306 | element_size = CL_IMAGE_ELEMENT_SIZE |
|
|
2307 | row_pitch = CL_IMAGE_ROW_PITCH |
|
|
2308 | slice_pitch = CL_IMAGE_SLICE_PITCH |
|
|
2309 | width = CL_IMAGE_WIDTH |
|
|
2310 | height = CL_IMAGE_HEIGHT |
|
|
2311 | depth = CL_IMAGE_DEPTH |
|
|
2312 | PPCODE: |
|
|
2313 | size_t value [1]; |
|
|
2314 | NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2315 | EXTEND (SP, 1); |
|
|
2316 | const int i = 0; |
|
|
2317 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2318 | |
|
|
2319 | #END:image |
|
|
2320 | |
|
|
2321 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2322 | |
|
|
2323 | #BEGIN:gl_texture |
|
|
2324 | |
|
|
2325 | void |
|
|
2326 | target (OpenCL::Image self) |
|
|
2327 | PPCODE: |
|
|
2328 | cl_GLenum value [1]; |
|
|
2329 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
|
|
2330 | EXTEND (SP, 1); |
|
|
2331 | const int i = 0; |
|
|
2332 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2333 | |
|
|
2334 | void |
|
|
2335 | gl_mipmap_level (OpenCL::Image self) |
|
|
2336 | PPCODE: |
|
|
2337 | cl_GLint value [1]; |
|
|
2338 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
|
|
2339 | EXTEND (SP, 1); |
|
|
2340 | const int i = 0; |
|
|
2341 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2342 | |
|
|
2343 | #END:gl_texture |
|
|
2344 | |
|
|
2345 | #endif |
|
|
2346 | |
372 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
2347 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
373 | |
2348 | |
374 | void |
2349 | void |
375 | DESTROY (OpenCL::Sampler this) |
2350 | DESTROY (OpenCL::Sampler self) |
376 | CODE: |
2351 | CODE: |
377 | clReleaseSampler (this); |
2352 | clReleaseSampler (self); |
378 | |
2353 | |
379 | void |
2354 | void |
380 | info (OpenCL::Sampler this, cl_sampler_info name) |
2355 | info (OpenCL::Sampler self, cl_sampler_info name) |
381 | PPCODE: |
2356 | PPCODE: |
382 | INFO (Sampler) |
2357 | INFO (Sampler) |
383 | |
2358 | |
|
|
2359 | #BEGIN:sampler |
|
|
2360 | |
|
|
2361 | void |
|
|
2362 | reference_count (OpenCL::Sampler self) |
|
|
2363 | ALIAS: |
|
|
2364 | reference_count = CL_SAMPLER_REFERENCE_COUNT |
|
|
2365 | normalized_coords = CL_SAMPLER_NORMALIZED_COORDS |
|
|
2366 | addressing_mode = CL_SAMPLER_ADDRESSING_MODE |
|
|
2367 | PPCODE: |
|
|
2368 | cl_uint value [1]; |
|
|
2369 | NEED_SUCCESS (GetSamplerInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2370 | EXTEND (SP, 1); |
|
|
2371 | const int i = 0; |
|
|
2372 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2373 | |
|
|
2374 | void |
|
|
2375 | context (OpenCL::Sampler self) |
|
|
2376 | PPCODE: |
|
|
2377 | cl_context value [1]; |
|
|
2378 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
|
|
2379 | EXTEND (SP, 1); |
|
|
2380 | const int i = 0; |
|
|
2381 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2382 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2383 | |
|
|
2384 | void |
|
|
2385 | filter_mode (OpenCL::Sampler self) |
|
|
2386 | PPCODE: |
|
|
2387 | cl_uint value [1]; |
|
|
2388 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0)); |
|
|
2389 | EXTEND (SP, 1); |
|
|
2390 | const int i = 0; |
|
|
2391 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
|
|
2392 | |
|
|
2393 | #END:sampler |
|
|
2394 | |
384 | MODULE = OpenCL PACKAGE = OpenCL::Program |
2395 | MODULE = OpenCL PACKAGE = OpenCL::Program |
385 | |
2396 | |
386 | void |
2397 | void |
387 | DESTROY (OpenCL::Program this) |
2398 | DESTROY (OpenCL::Program self) |
388 | CODE: |
2399 | CODE: |
389 | clReleaseProgram (this); |
2400 | clReleaseProgram (self); |
390 | |
2401 | |
391 | void |
2402 | void |
392 | info (OpenCL::Program this, cl_program_info name) |
2403 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2404 | ALIAS: |
|
|
2405 | build_async = 1 |
393 | PPCODE: |
2406 | CODE: |
394 | INFO (Program) |
2407 | cl_uint device_count = 0; |
|
|
2408 | cl_device_id *device_list = 0; |
395 | |
2409 | |
|
|
2410 | if (SvOK (devices)) |
|
|
2411 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2412 | |
|
|
2413 | void *user_data; |
|
|
2414 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2415 | |
|
|
2416 | if (ix) |
|
|
2417 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2418 | else |
|
|
2419 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2420 | |
|
|
2421 | #if CL_VERSION_1_2 |
|
|
2422 | |
396 | void |
2423 | void |
397 | build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) |
2424 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
398 | CODE: |
2425 | CODE: |
|
|
2426 | cl_uint device_count = 0; |
|
|
2427 | cl_device_id *device_list = 0; |
|
|
2428 | |
|
|
2429 | if (SvOK (devices)) |
|
|
2430 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2431 | |
|
|
2432 | cl_uint header_count = 0; |
|
|
2433 | cl_program *header_list = 0; |
|
|
2434 | const char **header_name = 0; |
|
|
2435 | |
|
|
2436 | if (SvOK (headers)) |
|
|
2437 | { |
|
|
2438 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2439 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2440 | |
|
|
2441 | HV *hv = (HV *)SvRV (devices); |
|
|
2442 | |
|
|
2443 | header_count = hv_iterinit (hv); |
|
|
2444 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2445 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2446 | |
|
|
2447 | HE *he; |
|
|
2448 | int i = 0; |
|
|
2449 | while (he = hv_iternext (hv)) |
|
|
2450 | { |
|
|
2451 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2452 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2453 | ++i; |
|
|
2454 | } |
|
|
2455 | } |
|
|
2456 | |
|
|
2457 | void *user_data; |
|
|
2458 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2459 | |
399 | NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2460 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2461 | header_count, header_list, header_name, pfn_notify, user_data)); |
400 | |
2462 | |
|
|
2463 | #endif |
|
|
2464 | |
401 | void |
2465 | void |
402 | build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) |
2466 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
403 | PPCODE: |
2467 | PPCODE: |
404 | { |
|
|
405 | size_t size; |
2468 | size_t size; |
406 | SV *sv; |
|
|
407 | |
|
|
408 | NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); |
2469 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size)); |
409 | sv = sv_2mortal (newSV (size)); |
2470 | SV *sv = sv_2mortal (newSV (size)); |
410 | SvUPGRADE (sv, SVt_PV); |
2471 | SvUPGRADE (sv, SVt_PV); |
411 | SvPOK_only (sv); |
2472 | SvPOK_only (sv); |
412 | SvCUR_set (sv, size); |
2473 | SvCUR_set (sv, size); |
413 | NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); |
2474 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0)); |
414 | XPUSHs (sv); |
2475 | XPUSHs (sv); |
415 | } |
2476 | |
|
|
2477 | #BEGIN:program_build |
|
|
2478 | |
|
|
2479 | void |
|
|
2480 | build_status (OpenCL::Program self, OpenCL::Device device) |
|
|
2481 | PPCODE: |
|
|
2482 | cl_int value [1]; |
|
|
2483 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0)); |
|
|
2484 | EXTEND (SP, 1); |
|
|
2485 | const int i = 0; |
|
|
2486 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2487 | |
|
|
2488 | void |
|
|
2489 | build_options (OpenCL::Program self, OpenCL::Device device) |
|
|
2490 | ALIAS: |
|
|
2491 | build_options = CL_PROGRAM_BUILD_OPTIONS |
|
|
2492 | build_log = CL_PROGRAM_BUILD_LOG |
|
|
2493 | PPCODE: |
|
|
2494 | size_t size; |
|
|
2495 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size)); |
|
|
2496 | char *value = tmpbuf (size); |
|
|
2497 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
|
|
2498 | EXTEND (SP, 1); |
|
|
2499 | const int i = 0; |
|
|
2500 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2501 | |
|
|
2502 | void |
|
|
2503 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2504 | PPCODE: |
|
|
2505 | cl_uint value [1]; |
|
|
2506 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2507 | EXTEND (SP, 1); |
|
|
2508 | const int i = 0; |
|
|
2509 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
2510 | |
|
|
2511 | #END:program_build |
416 | |
2512 | |
417 | void |
2513 | void |
418 | kernel (OpenCL::Program program, SV *function) |
2514 | kernel (OpenCL::Program program, SV *function) |
419 | PPCODE: |
2515 | PPCODE: |
420 | { |
|
|
421 | cl_int res; |
|
|
422 | cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); |
2516 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
|
|
2517 | XPUSH_CLOBJ (stash_kernel, kernel); |
423 | |
2518 | |
424 | if (res) |
2519 | void |
425 | FAIL (CreateKernel, res); |
2520 | kernels_in_program (OpenCL::Program program) |
|
|
2521 | PPCODE: |
|
|
2522 | cl_uint num_kernels; |
|
|
2523 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2524 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2525 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
426 | |
2526 | |
427 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2527 | int i; |
428 | } |
2528 | EXTEND (SP, num_kernels); |
|
|
2529 | for (i = 0; i < num_kernels; ++i) |
|
|
2530 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
|
|
2531 | |
|
|
2532 | void |
|
|
2533 | info (OpenCL::Program self, cl_program_info name) |
|
|
2534 | PPCODE: |
|
|
2535 | INFO (Program) |
|
|
2536 | |
|
|
2537 | void |
|
|
2538 | binaries (OpenCL::Program self) |
|
|
2539 | PPCODE: |
|
|
2540 | cl_uint n, i; |
|
|
2541 | size_t size; |
|
|
2542 | |
|
|
2543 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0)); |
|
|
2544 | if (!n) XSRETURN_EMPTY; |
|
|
2545 | |
|
|
2546 | size_t *sizes = tmpbuf (sizeof (*sizes) * n); |
|
|
2547 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size)); |
|
|
2548 | if (size != sizeof (*sizes) * n) XSRETURN_EMPTY; |
|
|
2549 | unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n); |
|
|
2550 | |
|
|
2551 | EXTEND (SP, n); |
|
|
2552 | for (i = 0; i < n; ++i) |
|
|
2553 | { |
|
|
2554 | SV *sv = sv_2mortal (newSV (sizes [i])); |
|
|
2555 | SvUPGRADE (sv, SVt_PV); |
|
|
2556 | SvPOK_only (sv); |
|
|
2557 | SvCUR_set (sv, sizes [i]); |
|
|
2558 | ptrs [i] = (void *)SvPVX (sv); |
|
|
2559 | PUSHs (sv); |
|
|
2560 | } |
|
|
2561 | |
|
|
2562 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
|
|
2563 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
|
|
2564 | |
|
|
2565 | #BEGIN:program |
|
|
2566 | |
|
|
2567 | void |
|
|
2568 | reference_count (OpenCL::Program self) |
|
|
2569 | ALIAS: |
|
|
2570 | reference_count = CL_PROGRAM_REFERENCE_COUNT |
|
|
2571 | num_devices = CL_PROGRAM_NUM_DEVICES |
|
|
2572 | PPCODE: |
|
|
2573 | cl_uint value [1]; |
|
|
2574 | NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2575 | EXTEND (SP, 1); |
|
|
2576 | const int i = 0; |
|
|
2577 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2578 | |
|
|
2579 | void |
|
|
2580 | context (OpenCL::Program self) |
|
|
2581 | PPCODE: |
|
|
2582 | cl_context value [1]; |
|
|
2583 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
|
|
2584 | EXTEND (SP, 1); |
|
|
2585 | const int i = 0; |
|
|
2586 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2587 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2588 | |
|
|
2589 | void |
|
|
2590 | devices (OpenCL::Program self) |
|
|
2591 | PPCODE: |
|
|
2592 | size_t size; |
|
|
2593 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size)); |
|
|
2594 | cl_device_id *value = tmpbuf (size); |
|
|
2595 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
|
|
2596 | int i, n = size / sizeof (*value); |
|
|
2597 | EXTEND (SP, n); |
|
|
2598 | for (i = 0; i < n; ++i) |
|
|
2599 | PUSH_CLOBJ (stash_device, value [i]); |
|
|
2600 | |
|
|
2601 | void |
|
|
2602 | source (OpenCL::Program self) |
|
|
2603 | PPCODE: |
|
|
2604 | size_t size; |
|
|
2605 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size)); |
|
|
2606 | char *value = tmpbuf (size); |
|
|
2607 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0)); |
|
|
2608 | EXTEND (SP, 1); |
|
|
2609 | const int i = 0; |
|
|
2610 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2611 | |
|
|
2612 | void |
|
|
2613 | binary_sizes (OpenCL::Program self) |
|
|
2614 | PPCODE: |
|
|
2615 | size_t size; |
|
|
2616 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size)); |
|
|
2617 | size_t *value = tmpbuf (size); |
|
|
2618 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0)); |
|
|
2619 | int i, n = size / sizeof (*value); |
|
|
2620 | EXTEND (SP, n); |
|
|
2621 | for (i = 0; i < n; ++i) |
|
|
2622 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2623 | |
|
|
2624 | #END:program |
429 | |
2625 | |
430 | MODULE = OpenCL PACKAGE = OpenCL::Kernel |
2626 | MODULE = OpenCL PACKAGE = OpenCL::Kernel |
431 | |
2627 | |
432 | void |
2628 | void |
433 | DESTROY (OpenCL::Kernel this) |
2629 | DESTROY (OpenCL::Kernel self) |
434 | CODE: |
2630 | CODE: |
435 | clReleaseKernel (this); |
2631 | clReleaseKernel (self); |
436 | |
2632 | |
437 | void |
2633 | void |
|
|
2634 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2635 | CODE: |
|
|
2636 | int i; |
|
|
2637 | for (i = 2; ; ++i) |
|
|
2638 | { |
|
|
2639 | while (*format == ' ') |
|
|
2640 | ++format; |
|
|
2641 | |
|
|
2642 | char type = *format++; |
|
|
2643 | |
|
|
2644 | if (!type) |
|
|
2645 | break; |
|
|
2646 | |
|
|
2647 | if (i >= items) |
|
|
2648 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2649 | |
|
|
2650 | SV *sv = ST (i); |
|
|
2651 | |
|
|
2652 | union |
|
|
2653 | { |
|
|
2654 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2655 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2656 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2657 | cl_mem cm; |
|
|
2658 | cl_sampler ca; |
|
|
2659 | size_t cz; |
|
|
2660 | cl_event ce; |
|
|
2661 | } arg; |
|
|
2662 | size_t size; |
|
|
2663 | int nullarg = 0; |
|
|
2664 | |
|
|
2665 | switch (type) |
|
|
2666 | { |
|
|
2667 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2668 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2669 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2670 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2671 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2672 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2673 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2674 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2675 | |
|
|
2676 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2677 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2678 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2679 | |
|
|
2680 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2681 | |
|
|
2682 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2683 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2684 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2685 | |
|
|
2686 | default: |
|
|
2687 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2688 | } |
|
|
2689 | |
|
|
2690 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2691 | if (res) |
|
|
2692 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2693 | } |
|
|
2694 | |
|
|
2695 | if (i != items) |
|
|
2696 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2697 | |
|
|
2698 | void |
|
|
2699 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
|
|
2700 | CODE: |
|
|
2701 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2702 | |
|
|
2703 | void |
|
|
2704 | set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value) |
|
|
2705 | CODE: |
|
|
2706 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2707 | |
|
|
2708 | void |
|
|
2709 | set_short (OpenCL::Kernel self, cl_uint idx, cl_short value) |
|
|
2710 | CODE: |
|
|
2711 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2712 | |
|
|
2713 | void |
|
|
2714 | set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value) |
|
|
2715 | CODE: |
|
|
2716 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2717 | |
|
|
2718 | void |
|
|
2719 | set_int (OpenCL::Kernel self, cl_uint idx, cl_int value) |
|
|
2720 | CODE: |
|
|
2721 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2722 | |
|
|
2723 | void |
|
|
2724 | set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value) |
|
|
2725 | CODE: |
|
|
2726 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2727 | |
|
|
2728 | void |
|
|
2729 | set_long (OpenCL::Kernel self, cl_uint idx, cl_long value) |
|
|
2730 | CODE: |
|
|
2731 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2732 | |
|
|
2733 | void |
|
|
2734 | set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value) |
|
|
2735 | CODE: |
|
|
2736 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2737 | |
|
|
2738 | void |
|
|
2739 | set_half (OpenCL::Kernel self, cl_uint idx, cl_half value) |
|
|
2740 | CODE: |
|
|
2741 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2742 | |
|
|
2743 | void |
|
|
2744 | set_float (OpenCL::Kernel self, cl_uint idx, cl_float value) |
|
|
2745 | CODE: |
|
|
2746 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2747 | |
|
|
2748 | void |
|
|
2749 | set_double (OpenCL::Kernel self, cl_uint idx, cl_double value) |
|
|
2750 | CODE: |
|
|
2751 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2752 | |
|
|
2753 | void |
|
|
2754 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
|
|
2755 | CODE: |
|
|
2756 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2757 | |
|
|
2758 | void |
|
|
2759 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2760 | CODE: |
|
|
2761 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2762 | |
|
|
2763 | void |
|
|
2764 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2765 | CODE: |
|
|
2766 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2767 | |
|
|
2768 | void |
|
|
2769 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2770 | CODE: |
|
|
2771 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2772 | |
|
|
2773 | void |
|
|
2774 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
|
|
2775 | CODE: |
|
|
2776 | clSetKernelArg (self, idx, size, 0); |
|
|
2777 | |
|
|
2778 | void |
|
|
2779 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
|
|
2780 | CODE: |
|
|
2781 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
2782 | |
|
|
2783 | void |
438 | info (OpenCL::Kernel this, cl_kernel_info name) |
2784 | info (OpenCL::Kernel self, cl_kernel_info name) |
439 | PPCODE: |
2785 | PPCODE: |
440 | INFO (Kernel) |
2786 | INFO (Kernel) |
441 | |
2787 | |
|
|
2788 | #BEGIN:kernel |
|
|
2789 | |
442 | void |
2790 | void |
443 | set_bool (OpenCL::Kernel this, cl_uint idx, cl_bool value) |
2791 | function_name (OpenCL::Kernel self) |
|
|
2792 | PPCODE: |
|
|
2793 | size_t size; |
|
|
2794 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size)); |
|
|
2795 | char *value = tmpbuf (size); |
|
|
2796 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0)); |
|
|
2797 | EXTEND (SP, 1); |
|
|
2798 | const int i = 0; |
|
|
2799 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2800 | |
|
|
2801 | void |
|
|
2802 | num_args (OpenCL::Kernel self) |
|
|
2803 | ALIAS: |
|
|
2804 | num_args = CL_KERNEL_NUM_ARGS |
|
|
2805 | reference_count = CL_KERNEL_REFERENCE_COUNT |
|
|
2806 | PPCODE: |
|
|
2807 | cl_uint value [1]; |
|
|
2808 | NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2809 | EXTEND (SP, 1); |
|
|
2810 | const int i = 0; |
|
|
2811 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2812 | |
|
|
2813 | void |
|
|
2814 | context (OpenCL::Kernel self) |
|
|
2815 | PPCODE: |
|
|
2816 | cl_context value [1]; |
|
|
2817 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
|
|
2818 | EXTEND (SP, 1); |
|
|
2819 | const int i = 0; |
|
|
2820 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2821 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2822 | |
|
|
2823 | void |
|
|
2824 | program (OpenCL::Kernel self) |
|
|
2825 | PPCODE: |
|
|
2826 | cl_program value [1]; |
|
|
2827 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
|
|
2828 | EXTEND (SP, 1); |
|
|
2829 | const int i = 0; |
|
|
2830 | NEED_SUCCESS (RetainProgram, (value [i])); |
|
|
2831 | PUSH_CLOBJ (stash_program, value [i]); |
|
|
2832 | |
|
|
2833 | #END:kernel |
|
|
2834 | |
|
|
2835 | void |
|
|
2836 | work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name) |
444 | CODE: |
2837 | PPCODE: |
445 | clKernelSetArg (this, idx, sizeof (value), &value); |
2838 | size_t size; |
|
|
2839 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size)); |
|
|
2840 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2841 | SvUPGRADE (sv, SVt_PV); |
|
|
2842 | SvPOK_only (sv); |
|
|
2843 | SvCUR_set (sv, size); |
|
|
2844 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0)); |
|
|
2845 | XPUSHs (sv); |
|
|
2846 | |
|
|
2847 | #BEGIN:kernel_work_group |
|
|
2848 | |
|
|
2849 | void |
|
|
2850 | work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2851 | ALIAS: |
|
|
2852 | work_group_size = CL_KERNEL_WORK_GROUP_SIZE |
|
|
2853 | preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE |
|
|
2854 | PPCODE: |
|
|
2855 | size_t value [1]; |
|
|
2856 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
|
|
2857 | EXTEND (SP, 1); |
|
|
2858 | const int i = 0; |
|
|
2859 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2860 | |
|
|
2861 | void |
|
|
2862 | compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2863 | PPCODE: |
|
|
2864 | size_t size; |
|
|
2865 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size)); |
|
|
2866 | size_t *value = tmpbuf (size); |
|
|
2867 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0)); |
|
|
2868 | int i, n = size / sizeof (*value); |
|
|
2869 | EXTEND (SP, n); |
|
|
2870 | for (i = 0; i < n; ++i) |
|
|
2871 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2872 | |
|
|
2873 | void |
|
|
2874 | local_mem_size (OpenCL::Kernel self, OpenCL::Device device) |
|
|
2875 | ALIAS: |
|
|
2876 | local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE |
|
|
2877 | private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE |
|
|
2878 | PPCODE: |
|
|
2879 | cl_ulong value [1]; |
|
|
2880 | NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0)); |
|
|
2881 | EXTEND (SP, 1); |
|
|
2882 | const int i = 0; |
|
|
2883 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2884 | |
|
|
2885 | #END:kernel_work_group |
|
|
2886 | |
|
|
2887 | #if CL_VERSION_1_2 |
|
|
2888 | |
|
|
2889 | void |
|
|
2890 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2891 | PPCODE: |
|
|
2892 | size_t size; |
|
|
2893 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2894 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2895 | SvUPGRADE (sv, SVt_PV); |
|
|
2896 | SvPOK_only (sv); |
|
|
2897 | SvCUR_set (sv, size); |
|
|
2898 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2899 | XPUSHs (sv); |
|
|
2900 | |
|
|
2901 | #BEGIN:kernel_arg |
|
|
2902 | |
|
|
2903 | void |
|
|
2904 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2905 | ALIAS: |
|
|
2906 | arg_address_qualifier = CL_KERNEL_ARG_ADDRESS_QUALIFIER |
|
|
2907 | arg_access_qualifier = CL_KERNEL_ARG_ACCESS_QUALIFIER |
|
|
2908 | PPCODE: |
|
|
2909 | cl_uint value [1]; |
|
|
2910 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, sizeof (value), value, 0)); |
|
|
2911 | EXTEND (SP, 1); |
|
|
2912 | const int i = 0; |
|
|
2913 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2914 | |
|
|
2915 | void |
|
|
2916 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2917 | ALIAS: |
|
|
2918 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2919 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2920 | PPCODE: |
|
|
2921 | size_t size; |
|
|
2922 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2923 | char *value = tmpbuf (size); |
|
|
2924 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2925 | EXTEND (SP, 1); |
|
|
2926 | const int i = 0; |
|
|
2927 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2928 | |
|
|
2929 | void |
|
|
2930 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2931 | PPCODE: |
|
|
2932 | cl_ulong value [1]; |
|
|
2933 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2934 | EXTEND (SP, 1); |
|
|
2935 | const int i = 0; |
|
|
2936 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2937 | |
|
|
2938 | #END:kernel_arg |
|
|
2939 | |
|
|
2940 | #endif |
446 | |
2941 | |
447 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2942 | MODULE = OpenCL PACKAGE = OpenCL::Event |
448 | |
2943 | |
449 | void |
2944 | void |
450 | DESTROY (OpenCL::Event this) |
2945 | DESTROY (OpenCL::Event self) |
451 | CODE: |
2946 | CODE: |
452 | clReleaseEvent (this); |
2947 | clReleaseEvent (self); |
453 | |
2948 | |
454 | void |
2949 | void |
|
|
2950 | wait (OpenCL::Event self) |
|
|
2951 | CODE: |
|
|
2952 | clWaitForEvents (1, &self); |
|
|
2953 | |
|
|
2954 | void |
|
|
2955 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2956 | CODE: |
|
|
2957 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
|
|
2958 | |
|
|
2959 | void |
455 | info (OpenCL::Event this, cl_event_info name) |
2960 | info (OpenCL::Event self, cl_event_info name) |
456 | PPCODE: |
2961 | PPCODE: |
457 | INFO (Event) |
2962 | INFO (Event) |
458 | |
2963 | |
|
|
2964 | #BEGIN:event |
|
|
2965 | |
459 | void |
2966 | void |
460 | wait (OpenCL::Event this) |
2967 | command_queue (OpenCL::Event self) |
|
|
2968 | PPCODE: |
|
|
2969 | cl_command_queue value [1]; |
|
|
2970 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
|
|
2971 | EXTEND (SP, 1); |
|
|
2972 | const int i = 0; |
|
|
2973 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
|
|
2974 | PUSH_CLOBJ (stash_queue, value [i]); |
|
|
2975 | |
|
|
2976 | void |
|
|
2977 | command_type (OpenCL::Event self) |
|
|
2978 | ALIAS: |
|
|
2979 | command_type = CL_EVENT_COMMAND_TYPE |
|
|
2980 | reference_count = CL_EVENT_REFERENCE_COUNT |
|
|
2981 | command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS |
|
|
2982 | PPCODE: |
|
|
2983 | cl_uint value [1]; |
|
|
2984 | NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0)); |
|
|
2985 | EXTEND (SP, 1); |
|
|
2986 | const int i = 0; |
|
|
2987 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2988 | |
|
|
2989 | void |
|
|
2990 | context (OpenCL::Event self) |
|
|
2991 | PPCODE: |
|
|
2992 | cl_context value [1]; |
|
|
2993 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
|
|
2994 | EXTEND (SP, 1); |
|
|
2995 | const int i = 0; |
|
|
2996 | NEED_SUCCESS (RetainContext, (value [i])); |
|
|
2997 | PUSH_CLOBJ (stash_context, value [i]); |
|
|
2998 | |
|
|
2999 | #END:event |
|
|
3000 | |
|
|
3001 | void |
|
|
3002 | profiling_info (OpenCL::Event self, cl_profiling_info name) |
461 | CODE: |
3003 | PPCODE: |
462 | clWaitForEvents (1, &this); |
3004 | INFO (EventProfiling) |
463 | |
3005 | |
|
|
3006 | #BEGIN:profiling |
|
|
3007 | |
|
|
3008 | void |
|
|
3009 | profiling_command_queued (OpenCL::Event self) |
|
|
3010 | ALIAS: |
|
|
3011 | profiling_command_queued = CL_PROFILING_COMMAND_QUEUED |
|
|
3012 | profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT |
|
|
3013 | profiling_command_start = CL_PROFILING_COMMAND_START |
|
|
3014 | profiling_command_end = CL_PROFILING_COMMAND_END |
|
|
3015 | PPCODE: |
|
|
3016 | cl_ulong value [1]; |
|
|
3017 | NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0)); |
|
|
3018 | EXTEND (SP, 1); |
|
|
3019 | const int i = 0; |
|
|
3020 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3021 | |
|
|
3022 | #END:profiling |
|
|
3023 | |
|
|
3024 | MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
|
|
3025 | |
|
|
3026 | void |
|
|
3027 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
|
|
3028 | CODE: |
|
|
3029 | clSetUserEventStatus (self, execution_status); |
|
|
3030 | |
|
|
3031 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3032 | |
|
|
3033 | void |
|
|
3034 | DESTROY (SV *self) |
|
|
3035 | CODE: |
|
|
3036 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3037 | |
|
|
3038 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3039 | mapped_detach (self, mapped); |
|
|
3040 | |
|
|
3041 | clReleaseCommandQueue (mapped->queue); |
|
|
3042 | clReleaseEvent (mapped->event); |
|
|
3043 | Safefree (mapped); |
|
|
3044 | |
|
|
3045 | void |
|
|
3046 | unmap (OpenCL::Mapped self, ...) |
|
|
3047 | CODE: |
|
|
3048 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3049 | |
|
|
3050 | bool |
|
|
3051 | mapped (OpenCL::Mapped self) |
|
|
3052 | CODE: |
|
|
3053 | RETVAL = !!self->ptr; |
|
|
3054 | OUTPUT: |
|
|
3055 | RETVAL |
|
|
3056 | |
|
|
3057 | void |
|
|
3058 | wait (OpenCL::Mapped self) |
|
|
3059 | PPCODE: |
|
|
3060 | if (self->event) |
|
|
3061 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3062 | |
|
|
3063 | void |
|
|
3064 | event (OpenCL::Mapped self) |
|
|
3065 | PPCODE: |
|
|
3066 | if (!self->event) |
|
|
3067 | XSRETURN_UNDEF; |
|
|
3068 | |
|
|
3069 | clRetainEvent (self->event); |
|
|
3070 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3071 | |
|
|
3072 | #define MAPPED_OFFSET_CB offsetof (struct mapped, cb) |
|
|
3073 | #define MAPPED_OFFSET_ROW_PITCH offsetof (struct mapped, row_pitch) |
|
|
3074 | #define MAPPED_OFFSET_SLICE_PITCH offsetof (struct mapped, slice_pitch) |
|
|
3075 | #define MAPPED_OFFSET_WIDTH offsetof (struct mapped, width) |
|
|
3076 | #define MAPPED_OFFSET_HEIGHT offsetof (struct mapped, height) |
|
|
3077 | #define MAPPED_OFFSET_DEPTH offsetof (struct mapped, depth) |
|
|
3078 | |
|
|
3079 | IV |
|
|
3080 | size (OpenCL::Mapped self) |
|
|
3081 | ALIAS: |
|
|
3082 | size = MAPPED_OFFSET_CB |
|
|
3083 | row_pitch = MAPPED_OFFSET_ROW_PITCH |
|
|
3084 | slice_pitch = MAPPED_OFFSET_SLICE_PITCH |
|
|
3085 | width = MAPPED_OFFSET_WIDTH |
|
|
3086 | height = MAPPED_OFFSET_HEIGHT |
|
|
3087 | depth = MAPPED_OFFSET_DEPTH |
|
|
3088 | CODE: |
|
|
3089 | RETVAL = *(size_t *)((char *)self + ix); |
|
|
3090 | OUTPUT: |
|
|
3091 | RETVAL |
|
|
3092 | |
|
|
3093 | IV |
|
|
3094 | ptr (OpenCL::Mapped self) |
|
|
3095 | CODE: |
|
|
3096 | RETVAL = PTR2IV (self->ptr); |
|
|
3097 | OUTPUT: |
|
|
3098 | RETVAL |
|
|
3099 | |
|
|
3100 | void |
|
|
3101 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3102 | CODE: |
|
|
3103 | STRLEN len; |
|
|
3104 | const char *ptr = SvPVbyte (data, len); |
|
|
3105 | |
|
|
3106 | if (offset + len > self->cb) |
|
|
3107 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3108 | |
|
|
3109 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3110 | |
|
|
3111 | void |
|
|
3112 | get_row (OpenCL::Mapped self, size_t count, size_t x = 0, size_t y = 0, size_t z = 0) |
|
|
3113 | PPCODE: |
|
|
3114 | if (!SvOK (ST (1))) |
|
|
3115 | count = self->width - x; |
|
|
3116 | |
|
|
3117 | if (x + count > self->width) |
|
|
3118 | croak ("OpenCL::Mapped::get: x + count crosses a row boundary"); |
|
|
3119 | |
|
|
3120 | if (y >= self->height) |
|
|
3121 | croak ("OpenCL::Mapped::get: y coordinate out of bounds"); |
|
|
3122 | |
|
|
3123 | if (z >= self->depth) |
|
|
3124 | croak ("OpenCL::Mapped::get: z coordinate out of bounds"); |
|
|
3125 | |
|
|
3126 | size_t element = mapped_element_size (self); |
|
|
3127 | |
|
|
3128 | count *= element; |
|
|
3129 | x *= element; |
|
|
3130 | |
|
|
3131 | char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch; |
|
|
3132 | XPUSHs (sv_2mortal (newSVpvn (ptr, count))); |
|
|
3133 | |
|
|
3134 | void |
|
|
3135 | set_row (OpenCL::Mapped self, SV *data, size_t x = 0, size_t y = 0, size_t z = 0) |
|
|
3136 | PPCODE: |
|
|
3137 | STRLEN count; |
|
|
3138 | char *dataptr = SvPVbyte (data, count); |
|
|
3139 | size_t element = mapped_element_size (self); |
|
|
3140 | |
|
|
3141 | x *= element; |
|
|
3142 | |
|
|
3143 | if (x + count > self->width * element) |
|
|
3144 | croak ("OpenCL::Mapped::set: x + data size crosses a row boundary"); |
|
|
3145 | |
|
|
3146 | if (y >= self->height) |
|
|
3147 | croak ("OpenCL::Mapped::set: y coordinate out of bounds"); |
|
|
3148 | |
|
|
3149 | if (z >= self->depth) |
|
|
3150 | croak ("OpenCL::Mapped::set: z coordinate out of bounds"); |
|
|
3151 | |
|
|
3152 | char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch; |
|
|
3153 | memcpy (ptr, dataptr, count); |
|
|
3154 | |
|
|
3155 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3156 | |
|
|
3157 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3158 | |
|
|
3159 | IV |
|
|
3160 | element_size (OpenCL::Mapped self) |
|
|
3161 | CODE: |
|
|
3162 | RETVAL = mapped_element_size (self); |
|
|
3163 | OUTPUT: |
|
|
3164 | RETVAL |
|
|
3165 | |