1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #include "ecb.h"//D |
|
|
6 | |
|
|
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
8 | #include "xthread.h" |
|
|
9 | #include "schmorp.h" |
|
|
10 | |
|
|
11 | #ifdef I_DLFCN |
|
|
12 | #include <dlfcn.h> |
|
|
13 | #endif |
|
|
14 | |
|
|
15 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
16 | // but then fail to define the api functions unless you ALSO define |
|
|
17 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
18 | // after all, the functions are deprecated, not removed. |
|
|
19 | // in addition, you cannot test for this in any future-proof way. |
|
|
20 | // each time a new opencl version comes out, you need to make a new |
|
|
21 | // release. |
|
|
22 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
23 | |
|
|
24 | #ifndef PREFER_1_1 |
|
|
25 | #define PREFER_1_1 1 |
|
|
26 | #endif |
|
|
27 | |
|
|
28 | #if PREFER_1_1 |
|
|
29 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
30 | #endif |
|
|
31 | |
5 | #ifdef __APPLE__ |
32 | #ifdef __APPLE__ |
6 | #include <OpenCL/opencl.h> |
33 | #define CLHDR(name) <OpenCL/name> |
7 | #else |
34 | #else |
8 | #include <CL/opencl.h> |
35 | #define CLHDR(name) <CL/name> |
9 | #endif |
36 | #endif |
|
|
37 | |
|
|
38 | #include CLHDR(opencl.h) |
|
|
39 | |
|
|
40 | #ifndef CL_VERSION_1_2 |
|
|
41 | #include CLHDR(cl_d3d9.h) |
|
|
42 | #endif |
|
|
43 | |
|
|
44 | #if _WIN32 |
|
|
45 | #include CLHDR(cl_d3d10.h) |
|
|
46 | #if CL_VERSION_1_2 |
|
|
47 | #include CLHDR<cl_d3d11.h> |
|
|
48 | #endif |
|
|
49 | #include CLHDR<cl_dx9_media_sharing.h.h> |
|
|
50 | #endif |
|
|
51 | |
|
|
52 | #ifndef CL_VERSION_1_2 |
|
|
53 | #undef PREFER_1_1 |
|
|
54 | #define PREFER_1_1 1 |
|
|
55 | #endif |
|
|
56 | |
|
|
57 | // make sure all constants we might use are actually defined |
|
|
58 | #include "default.h" |
10 | |
59 | |
11 | typedef cl_platform_id OpenCL__Platform; |
60 | typedef cl_platform_id OpenCL__Platform; |
12 | typedef cl_device_id OpenCL__Device; |
61 | typedef cl_device_id OpenCL__Device; |
|
|
62 | typedef cl_device_id OpenCL__SubDevice; |
13 | typedef cl_context OpenCL__Context; |
63 | typedef cl_context OpenCL__Context; |
14 | typedef cl_command_queue OpenCL__Queue; |
64 | typedef cl_command_queue OpenCL__Queue; |
15 | typedef cl_mem OpenCL__Memory; |
65 | typedef cl_mem OpenCL__Memory; |
16 | typedef cl_mem OpenCL__Buffer; |
66 | typedef cl_mem OpenCL__Buffer; |
17 | typedef cl_mem OpenCL__BufferObj; |
67 | typedef cl_mem OpenCL__BufferObj; |
18 | typedef cl_mem OpenCL__Image; |
68 | typedef cl_mem OpenCL__Image; |
19 | typedef cl_mem OpenCL__Image2D; |
|
|
20 | typedef cl_mem OpenCL__Image3D; |
|
|
21 | typedef cl_mem OpenCL__Memory_ornull; |
69 | typedef cl_mem OpenCL__Memory_ornull; |
22 | typedef cl_mem OpenCL__Buffer_ornull; |
70 | typedef cl_mem OpenCL__Buffer_ornull; |
23 | typedef cl_mem OpenCL__Image_ornull; |
71 | typedef cl_mem OpenCL__Image_ornull; |
24 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
25 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
26 | typedef cl_sampler OpenCL__Sampler; |
72 | typedef cl_sampler OpenCL__Sampler; |
27 | typedef cl_program OpenCL__Program; |
73 | typedef cl_program OpenCL__Program; |
28 | typedef cl_kernel OpenCL__Kernel; |
74 | typedef cl_kernel OpenCL__Kernel; |
29 | typedef cl_event OpenCL__Event; |
75 | typedef cl_event OpenCL__Event; |
30 | typedef cl_event OpenCL__UserEvent; |
76 | typedef cl_event OpenCL__UserEvent; |
31 | |
77 | |
32 | typedef SV *FUTURE; |
78 | typedef struct mapped * OpenCL__Mapped; |
|
|
79 | |
|
|
80 | static HV |
|
|
81 | *stash_platform, |
|
|
82 | *stash_device, |
|
|
83 | *stash_subdevice, |
|
|
84 | *stash_context, |
|
|
85 | *stash_queue, |
|
|
86 | *stash_program, |
|
|
87 | *stash_kernel, |
|
|
88 | *stash_sampler, |
|
|
89 | *stash_event, |
|
|
90 | *stash_userevent, |
|
|
91 | *stash_memory, |
|
|
92 | *stash_buffer, |
|
|
93 | *stash_bufferobj, |
|
|
94 | *stash_image, |
|
|
95 | *stash_image1d, |
|
|
96 | *stash_image1darray, |
|
|
97 | *stash_image1dbuffer, |
|
|
98 | *stash_image2d, |
|
|
99 | *stash_image2darray, |
|
|
100 | *stash_image3d, |
|
|
101 | *stash_mapped, |
|
|
102 | *stash_mappedbuffer, |
|
|
103 | *stash_mappedimage; |
33 | |
104 | |
34 | /*****************************************************************************/ |
105 | /*****************************************************************************/ |
35 | |
106 | |
|
|
107 | // name must include a leading underscore |
|
|
108 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
109 | // for perl. doh. |
|
|
110 | static void * |
|
|
111 | glsym (const char *name) |
|
|
112 | { |
|
|
113 | void *fun = 0; |
|
|
114 | |
|
|
115 | #if defined I_DLFCN && defined RTLD_DEFAULT |
|
|
116 | fun = dlsym (RTLD_DEFAULT, name + 1); |
|
|
117 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
|
|
118 | |
|
|
119 | if (!fun) |
|
|
120 | { |
|
|
121 | static void *libgl; |
|
|
122 | static const char *glso[] = { |
|
|
123 | "libGL.so.1", |
|
|
124 | "libGL.so.3", |
|
|
125 | "libGL.so.4.0", |
|
|
126 | "libGL.so", |
|
|
127 | "/usr/lib/libGL.so", |
|
|
128 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
129 | }; |
|
|
130 | int i; |
|
|
131 | |
|
|
132 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
133 | { |
|
|
134 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
135 | if (libgl) |
|
|
136 | break; |
|
|
137 | } |
|
|
138 | |
|
|
139 | if (libgl) |
|
|
140 | { |
|
|
141 | fun = dlsym (libgl, name + 1); |
|
|
142 | if (!fun) fun = dlsym (libgl, name); |
|
|
143 | } |
|
|
144 | } |
|
|
145 | #endif |
|
|
146 | |
|
|
147 | return fun; |
|
|
148 | } |
|
|
149 | |
|
|
150 | /*****************************************************************************/ |
|
|
151 | |
36 | /* up to two temporary buffers */ |
152 | /* up to two temporary buffers */ |
37 | static void * |
153 | static void * ecb_noinline |
38 | tmpbuf (size_t size) |
154 | tmpbuf (size_t size) |
39 | { |
155 | { |
40 | enum { buffers = 3 }; |
156 | enum { buffers = 4 }; |
41 | static int idx; |
157 | static int idx; |
42 | static void *buf [buffers]; |
158 | static void *buf [buffers]; |
43 | static size_t len [buffers]; |
159 | static size_t len [buffers]; |
44 | |
160 | |
45 | idx = ++idx % buffers; |
161 | idx = (idx + 1) % buffers; |
46 | |
162 | |
47 | if (len [idx] < size) |
163 | if (len [idx] < size) |
48 | { |
164 | { |
49 | free (buf [idx]); |
165 | free (buf [idx]); |
50 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
166 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
52 | } |
168 | } |
53 | |
169 | |
54 | return buf [idx]; |
170 | return buf [idx]; |
55 | } |
171 | } |
56 | |
172 | |
|
|
173 | static const char * ecb_noinline |
|
|
174 | cv_get_name (CV *cv) |
|
|
175 | { |
|
|
176 | static char fullname [256]; |
|
|
177 | |
|
|
178 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
179 | |
|
|
180 | HV *stash = GvSTASH (gv); |
|
|
181 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
182 | const char *gvname = GvNAME (gv); |
|
|
183 | |
|
|
184 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
185 | return fullname; |
|
|
186 | } |
|
|
187 | |
57 | /*****************************************************************************/ |
188 | /*****************************************************************************/ |
58 | |
189 | |
59 | typedef struct |
190 | typedef struct |
60 | { |
191 | { |
61 | IV iv; |
192 | IV iv; |
62 | const char *name; |
193 | const char *name; |
63 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
194 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
64 | } ivstr; |
195 | } ivstr; |
|
|
196 | |
|
|
197 | typedef struct |
|
|
198 | { |
|
|
199 | NV nv; |
|
|
200 | const char *name; |
|
|
201 | #define const_nv(name) { (NV)CL_ ## name, # name }, |
|
|
202 | } nvstr; |
65 | |
203 | |
66 | static const char * |
204 | static const char * |
67 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
205 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
68 | { |
206 | { |
69 | int i; |
207 | int i; |
… | |
… | |
118 | if (res) \ |
256 | if (res) \ |
119 | FAIL (name); |
257 | FAIL (name); |
120 | |
258 | |
121 | /*****************************************************************************/ |
259 | /*****************************************************************************/ |
122 | |
260 | |
123 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
124 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
125 | |
|
|
126 | static void * |
261 | static SV * |
|
|
262 | new_clobj (HV *stash, IV id) |
|
|
263 | { |
|
|
264 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
265 | } |
|
|
266 | |
|
|
267 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
268 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
269 | |
|
|
270 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
271 | /* they can be upgraded at runtime to the array form */ |
|
|
272 | static void * ecb_noinline |
127 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
273 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
128 | { |
274 | { |
|
|
275 | // sv_derived_from is quite slow :( |
129 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
276 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
130 | return (void *)SvIV (SvRV (sv)); |
277 | return (void *)SvIV (SvRV (sv)); |
131 | |
278 | |
132 | croak ("%s: %s is not of type %s", func, svname, pkg); |
279 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
280 | } |
|
|
281 | |
|
|
282 | // the "no-inherit" version of the above |
|
|
283 | static void * ecb_noinline |
|
|
284 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
285 | { |
|
|
286 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
287 | return (void *)SvIV (SvRV (sv)); |
|
|
288 | |
|
|
289 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
133 | } |
290 | } |
134 | |
291 | |
135 | /*****************************************************************************/ |
292 | /*****************************************************************************/ |
136 | |
293 | |
137 | static cl_context_properties * |
294 | static cl_context_properties * ecb_noinline |
138 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
295 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
139 | { |
296 | { |
140 | if (!sv || !SvOK (sv)) |
297 | if (!sv || !SvOK (sv)) |
141 | if (extra) |
298 | if (extra) |
142 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
299 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
143 | else |
300 | else |
144 | return 0; |
301 | return 0; |
145 | |
302 | |
146 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
303 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
147 | { |
304 | { |
148 | AV *av = (AV *)SvRV (sv); |
305 | AV *av = (AV *)SvRV (sv); |
149 | int i, len = av_len (av); |
306 | int i, len = av_len (av) + 1; |
150 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
307 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
151 | cl_context_properties *l = p; |
308 | cl_context_properties *l = p; |
152 | |
309 | |
153 | if (len & 1) |
310 | if (len & 1) |
154 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
311 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
155 | |
312 | |
156 | while (extracount--) |
313 | while (extracount--) |
157 | *l++ = *extra++; |
314 | *l++ = *extra++; |
158 | |
315 | |
159 | for (i = 0; i < len; i += 2) |
316 | for (i = 0; i < len; i += 2) |
160 | { |
317 | { |
161 | cl_context_properties t = SvIV (*av_fetch (av, i, 0)); |
318 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
162 | cl_context_properties v; |
319 | SV *p_sv = *av_fetch (av, i + 1, 0); |
|
|
320 | cl_context_properties v = SvIV (p_sv); // code below can override |
163 | |
321 | |
164 | switch (t) |
322 | switch (t) |
165 | { |
323 | { |
|
|
324 | case CL_CONTEXT_PLATFORM: |
|
|
325 | if (SvROK (p_sv)) |
|
|
326 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
327 | break; |
|
|
328 | |
|
|
329 | case CL_GLX_DISPLAY_KHR: |
|
|
330 | if (!SvOK (p_sv)) |
|
|
331 | { |
|
|
332 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
333 | if (func) |
|
|
334 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
335 | } |
|
|
336 | break; |
|
|
337 | |
|
|
338 | case CL_GL_CONTEXT_KHR: |
|
|
339 | if (!SvOK (p_sv)) |
|
|
340 | { |
|
|
341 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
342 | if (func) |
|
|
343 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
344 | } |
|
|
345 | break; |
|
|
346 | |
166 | default: |
347 | default: |
167 | /* unknown property, treat as int */ |
348 | /* unknown property, treat as int */ |
168 | v = SvIV (*av_fetch (av, i, 0)); |
|
|
169 | break; |
349 | break; |
170 | } |
350 | } |
171 | |
351 | |
172 | *l++ = t; |
352 | *l++ = t; |
173 | *l++ = v; |
353 | *l++ = v; |
… | |
… | |
176 | *l = 0; |
356 | *l = 0; |
177 | |
357 | |
178 | return p; |
358 | return p; |
179 | } |
359 | } |
180 | |
360 | |
181 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
361 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
|
|
362 | } |
|
|
363 | |
|
|
364 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
|
|
365 | // is a pointer (and only on well-behaved systems). |
|
|
366 | static void * ecb_noinline |
|
|
367 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
368 | { |
|
|
369 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
370 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
371 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
|
|
372 | |
|
|
373 | AV *av = (AV *)SvRV (arg); |
|
|
374 | void **list = 0; |
|
|
375 | cl_uint count = av_len (av) + 1; |
|
|
376 | |
|
|
377 | if (count) |
|
|
378 | { |
|
|
379 | list = tmpbuf (sizeof (*list) * count); |
|
|
380 | int i; |
|
|
381 | for (i = 0; i < count; ++i) |
|
|
382 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
|
|
383 | } |
|
|
384 | |
|
|
385 | if (!count && !or_undef) |
|
|
386 | croak ("%s: '%s' must contain at least one %s object", |
|
|
387 | cv_get_name (cv), argname, klass); |
|
|
388 | |
|
|
389 | *rcount = count; |
|
|
390 | return (void *)list; |
182 | } |
391 | } |
183 | |
392 | |
184 | /*****************************************************************************/ |
393 | /*****************************************************************************/ |
|
|
394 | /* callback stuff */ |
|
|
395 | |
|
|
396 | /* default context callback, log to stderr */ |
|
|
397 | static void CL_CALLBACK |
|
|
398 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
399 | { |
|
|
400 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
401 | } |
|
|
402 | |
|
|
403 | typedef struct |
|
|
404 | { |
|
|
405 | int free_cb; |
|
|
406 | void (*push)(void *data1, void *data2, void *data3); |
|
|
407 | } eq_vtbl; |
|
|
408 | |
|
|
409 | typedef struct eq_item |
|
|
410 | { |
|
|
411 | struct eq_item *next; |
|
|
412 | eq_vtbl *vtbl; |
|
|
413 | SV *cb; |
|
|
414 | void *data1, *data2, *data3; |
|
|
415 | } eq_item; |
|
|
416 | |
|
|
417 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
418 | static void *eq_signal_arg; |
|
|
419 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
420 | static eq_item *eq_head, *eq_tail; |
|
|
421 | |
|
|
422 | static void ecb_noinline |
|
|
423 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
424 | { |
|
|
425 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
426 | |
|
|
427 | item->next = 0; |
|
|
428 | item->vtbl = vtbl; |
|
|
429 | item->cb = cb; |
|
|
430 | item->data1 = data1; |
|
|
431 | item->data2 = data2; |
|
|
432 | item->data3 = data3; |
|
|
433 | |
|
|
434 | X_LOCK (eq_lock); |
|
|
435 | |
|
|
436 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
437 | eq_tail = item; |
|
|
438 | |
|
|
439 | X_UNLOCK (eq_lock); |
|
|
440 | |
|
|
441 | eq_signal_func (eq_signal_arg, 0); |
|
|
442 | } |
|
|
443 | |
|
|
444 | static eq_item * |
|
|
445 | eq_dec (void) |
|
|
446 | { |
|
|
447 | eq_item *res; |
|
|
448 | |
|
|
449 | X_LOCK (eq_lock); |
|
|
450 | |
|
|
451 | res = eq_head; |
|
|
452 | if (res) |
|
|
453 | eq_head = res->next; |
|
|
454 | |
|
|
455 | X_UNLOCK (eq_lock); |
|
|
456 | |
|
|
457 | return res; |
|
|
458 | } |
|
|
459 | |
|
|
460 | static void |
|
|
461 | eq_poll (void) |
|
|
462 | { |
|
|
463 | eq_item *item; |
|
|
464 | |
|
|
465 | while ((item = eq_dec ())) |
|
|
466 | { |
|
|
467 | ENTER; |
|
|
468 | SAVETMPS; |
|
|
469 | |
|
|
470 | dSP; |
|
|
471 | PUSHMARK (SP); |
|
|
472 | EXTEND (SP, 2); |
|
|
473 | |
|
|
474 | if (item->vtbl->free_cb) |
|
|
475 | sv_2mortal (item->cb); |
|
|
476 | |
|
|
477 | PUTBACK; |
|
|
478 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
479 | |
|
|
480 | SV *cb = item->cb; |
|
|
481 | free (item); |
|
|
482 | |
|
|
483 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
484 | |
|
|
485 | FREETMPS; |
|
|
486 | LEAVE; |
|
|
487 | } |
|
|
488 | } |
|
|
489 | |
|
|
490 | static void |
|
|
491 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
492 | { |
|
|
493 | eq_poll (); |
|
|
494 | } |
|
|
495 | |
|
|
496 | /*****************************************************************************/ |
|
|
497 | /* context notify */ |
|
|
498 | |
|
|
499 | static void ecb_noinline |
|
|
500 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
501 | { |
|
|
502 | dSP; |
|
|
503 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
504 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
505 | PUTBACK; |
|
|
506 | |
|
|
507 | free (data1); |
|
|
508 | free (data2); |
|
|
509 | } |
|
|
510 | |
|
|
511 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
512 | |
|
|
513 | static void CL_CALLBACK |
|
|
514 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
515 | { |
|
|
516 | void *pvt_copy = malloc (cb); |
|
|
517 | memcpy (pvt_copy, pvt, cb); |
|
|
518 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
519 | } |
|
|
520 | |
|
|
521 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
522 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
523 | void *user_data = 0; \ |
|
|
524 | \ |
|
|
525 | if (SvOK (notify)) \ |
|
|
526 | { \ |
|
|
527 | pfn_notify = eq_context_notify; \ |
|
|
528 | user_data = s_get_cv (notify); \ |
|
|
529 | } |
|
|
530 | |
|
|
531 | static SV * ecb_noinline |
|
|
532 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
533 | { |
|
|
534 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
535 | |
|
|
536 | if (user_data) |
|
|
537 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
538 | |
|
|
539 | return sv; |
|
|
540 | } |
|
|
541 | |
|
|
542 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
543 | |
|
|
544 | /*****************************************************************************/ |
|
|
545 | /* build/compile/link notify */ |
|
|
546 | |
|
|
547 | static void |
|
|
548 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
549 | { |
|
|
550 | dSP; |
|
|
551 | PUSH_CLOBJ (stash_program, data1); |
|
|
552 | PUTBACK; |
|
|
553 | } |
|
|
554 | |
|
|
555 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
556 | |
|
|
557 | static void CL_CALLBACK |
|
|
558 | eq_program_notify (cl_program program, void *user_data) |
|
|
559 | { |
|
|
560 | clRetainProgram (program); |
|
|
561 | |
|
|
562 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
563 | } |
|
|
564 | |
|
|
565 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
566 | |
|
|
567 | static program_callback ecb_noinline |
|
|
568 | make_program_callback (SV *notify, void **ruser_data) |
|
|
569 | { |
|
|
570 | if (SvOK (notify)) |
|
|
571 | { |
|
|
572 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
573 | return eq_program_notify; |
|
|
574 | } |
|
|
575 | else |
|
|
576 | { |
|
|
577 | *ruser_data = 0; |
|
|
578 | return 0; |
|
|
579 | } |
|
|
580 | } |
|
|
581 | |
|
|
582 | struct build_args |
|
|
583 | { |
|
|
584 | cl_program program; |
|
|
585 | char *options; |
|
|
586 | void *user_data; |
|
|
587 | cl_uint num_devices; |
|
|
588 | }; |
|
|
589 | |
|
|
590 | X_THREAD_PROC (build_program_thread) |
|
|
591 | { |
|
|
592 | struct build_args *arg = thr_arg; |
|
|
593 | |
|
|
594 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
595 | |
|
|
596 | if (arg->user_data) |
|
|
597 | eq_program_notify (arg->program, arg->user_data); |
|
|
598 | else |
|
|
599 | clReleaseProgram (arg->program); |
|
|
600 | |
|
|
601 | free (arg->options); |
|
|
602 | free (arg); |
|
|
603 | |
|
|
604 | return 0; |
|
|
605 | } |
|
|
606 | |
|
|
607 | static void |
|
|
608 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
609 | { |
|
|
610 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
611 | |
|
|
612 | arg->program = program; |
|
|
613 | arg->options = strdup (options); |
|
|
614 | arg->user_data = user_data; |
|
|
615 | arg->num_devices = num_devices; |
|
|
616 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
617 | |
|
|
618 | xthread_t id; |
|
|
619 | thread_create (&id, build_program_thread, arg); |
|
|
620 | } |
|
|
621 | |
|
|
622 | /*****************************************************************************/ |
|
|
623 | /* mem object destructor notify */ |
|
|
624 | |
|
|
625 | static void ecb_noinline |
|
|
626 | eq_destructor_push (void *data1, void *data2, void *data3) |
|
|
627 | { |
|
|
628 | } |
|
|
629 | |
|
|
630 | static eq_vtbl eq_destructor_vtbl = { 0, eq_destructor_push }; |
|
|
631 | |
|
|
632 | static void CL_CALLBACK |
|
|
633 | eq_destructor_notify (cl_mem memobj, void *user_data) |
|
|
634 | { |
|
|
635 | eq_enq (&eq_destructor_vtbl, (SV *)user_data, (void *)memobj, 0, 0); |
|
|
636 | } |
|
|
637 | |
|
|
638 | /*****************************************************************************/ |
|
|
639 | /* event objects */ |
|
|
640 | |
|
|
641 | static void |
|
|
642 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
643 | { |
|
|
644 | dSP; |
|
|
645 | PUSH_CLOBJ (stash_event, data1); |
|
|
646 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
647 | PUTBACK; |
|
|
648 | } |
|
|
649 | |
|
|
650 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
651 | |
|
|
652 | static void CL_CALLBACK |
|
|
653 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
654 | { |
|
|
655 | clRetainEvent (event); |
|
|
656 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
657 | } |
|
|
658 | |
|
|
659 | /*****************************************************************************/ |
|
|
660 | /* utilities for XS code */ |
185 | |
661 | |
186 | static size_t |
662 | static size_t |
187 | img_row_pitch (cl_mem img) |
663 | img_row_pitch (cl_mem img) |
188 | { |
664 | { |
189 | size_t res; |
665 | size_t res; |
190 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
666 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
191 | return res; |
667 | return res; |
192 | } |
668 | } |
193 | |
669 | |
194 | static cl_event * |
670 | static cl_event * ecb_noinline |
195 | event_list (SV **items, int count) |
671 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
196 | { |
672 | { |
197 | if (!count) |
673 | cl_uint count = *rcount; |
|
|
674 | |
|
|
675 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
676 | *rcount = count = 0; |
|
|
677 | |
|
|
678 | if (!count && !extra) |
198 | return 0; |
679 | return 0; |
199 | |
680 | |
200 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
681 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
682 | int i = 0; |
201 | |
683 | |
202 | while (count--) |
684 | while (count--) |
|
|
685 | if (SvOK (items [count])) |
203 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
686 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
204 | |
687 | |
|
|
688 | if (extra) |
|
|
689 | list [i++] = extra; |
|
|
690 | |
|
|
691 | *rcount = i; |
|
|
692 | |
205 | return list; |
693 | return i ? list : 0; |
206 | } |
694 | } |
207 | |
695 | |
208 | #define EVENT_LIST(items,count) \ |
696 | #define EVENT_LIST(skip) \ |
209 | cl_uint event_list_count = (count); \ |
697 | cl_uint event_list_count = items - (skip); \ |
210 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
698 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
211 | |
699 | |
212 | #define INFO(class) \ |
700 | #define INFO(class) \ |
213 | { \ |
701 | { \ |
214 | size_t size; \ |
702 | size_t size; \ |
215 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
703 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
219 | SvCUR_set (sv, size); \ |
707 | SvCUR_set (sv, size); \ |
220 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
708 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
221 | XPUSHs (sv); \ |
709 | XPUSHs (sv); \ |
222 | } |
710 | } |
223 | |
711 | |
|
|
712 | /*****************************************************************************/ |
|
|
713 | /* mapped_xxx */ |
|
|
714 | |
|
|
715 | static OpenCL__Mapped |
|
|
716 | SvMAPPED (SV *sv) |
|
|
717 | { |
|
|
718 | // no typechecking atm., keep your fingers crossed |
|
|
719 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
720 | } |
|
|
721 | |
|
|
722 | struct mapped |
|
|
723 | { |
|
|
724 | cl_command_queue queue; |
|
|
725 | cl_mem memobj; |
|
|
726 | void *ptr; |
|
|
727 | size_t cb; |
|
|
728 | cl_event event; |
|
|
729 | size_t row_pitch; |
|
|
730 | size_t slice_pitch; |
|
|
731 | }; |
|
|
732 | |
|
|
733 | static SV * |
|
|
734 | mapped_new (HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, void *ptr, size_t cb, cl_event ev, size_t row_pitch, size_t slice_pitch) |
|
|
735 | { |
|
|
736 | SV *data = newSV (0); |
|
|
737 | SvUPGRADE (data, SVt_PVMG); |
|
|
738 | |
|
|
739 | OpenCL__Mapped mapped; |
|
|
740 | New (0, mapped, 1, struct mapped); |
|
|
741 | |
|
|
742 | clRetainCommandQueue (queue); |
|
|
743 | |
|
|
744 | mapped->queue = queue; |
|
|
745 | mapped->memobj = memobj; |
|
|
746 | mapped->ptr = ptr; |
|
|
747 | mapped->cb = cb; |
|
|
748 | mapped->event = ev; |
|
|
749 | mapped->row_pitch = row_pitch; |
|
|
750 | mapped->slice_pitch = slice_pitch; |
|
|
751 | |
|
|
752 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
753 | |
|
|
754 | if (SvLEN (data)) |
|
|
755 | Safefree (data); |
|
|
756 | |
|
|
757 | SvPVX (data) = (char *)ptr; |
|
|
758 | SvCUR_set (data, cb); |
|
|
759 | SvLEN_set (data, 0); |
|
|
760 | SvPOK_only (data); |
|
|
761 | |
|
|
762 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
763 | |
|
|
764 | if (!(flags & CL_MAP_WRITE)) |
|
|
765 | SvREADONLY_on (data); |
|
|
766 | |
|
|
767 | return obj; |
|
|
768 | } |
|
|
769 | |
|
|
770 | static void |
|
|
771 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
772 | { |
|
|
773 | SV *data = SvRV (sv); |
|
|
774 | |
|
|
775 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
776 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
777 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
778 | else |
|
|
779 | { |
|
|
780 | SvREADONLY_off (data); |
|
|
781 | SvCUR_set (data, 0); |
|
|
782 | SvPVX (data) = 0; |
|
|
783 | SvOK_off (data); |
|
|
784 | } |
|
|
785 | |
|
|
786 | mapped->ptr = 0; |
|
|
787 | } |
|
|
788 | |
|
|
789 | static void |
|
|
790 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
791 | { |
|
|
792 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
793 | cl_event ev; |
|
|
794 | |
|
|
795 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
796 | |
|
|
797 | clReleaseEvent (mapped->event); |
|
|
798 | mapped->event = ev; |
|
|
799 | |
|
|
800 | mapped_detach (self, mapped); |
|
|
801 | } |
|
|
802 | |
|
|
803 | /*****************************************************************************/ |
|
|
804 | |
224 | MODULE = OpenCL PACKAGE = OpenCL |
805 | MODULE = OpenCL PACKAGE = OpenCL |
225 | |
806 | |
226 | PROTOTYPES: ENABLE |
807 | PROTOTYPES: ENABLE |
227 | |
808 | |
|
|
809 | void |
|
|
810 | poll () |
|
|
811 | CODE: |
|
|
812 | eq_poll (); |
|
|
813 | |
|
|
814 | void |
|
|
815 | _eq_initialise (IV func, IV arg) |
|
|
816 | CODE: |
|
|
817 | eq_signal_func = (void (*)(void *, int))func; |
|
|
818 | eq_signal_arg = (void*)arg; |
|
|
819 | |
228 | BOOT: |
820 | BOOT: |
229 | { |
821 | { |
230 | HV *stash = gv_stashpv ("OpenCL", 1); |
822 | HV *stash = gv_stashpv ("OpenCL", 1); |
|
|
823 | |
231 | static const ivstr *civ, const_iv[] = { |
824 | static const ivstr *civ, const_iv[] = { |
232 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
825 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
233 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
826 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
234 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
827 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
235 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
828 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
… | |
… | |
238 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
831 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
239 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
832 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
240 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
833 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
241 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
834 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
242 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
835 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
836 | { PREFER_1_1 , "PREFER_1_1" }, |
243 | #include "constiv.h" |
837 | #include "constiv.h" |
244 | }; |
838 | }; |
|
|
839 | |
245 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
840 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
246 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
841 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
842 | |
|
|
843 | static const nvstr *cnv, const_nv[] = { |
|
|
844 | #include "constnv.h" |
|
|
845 | }; |
|
|
846 | |
|
|
847 | for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--) |
|
|
848 | newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv)); |
|
|
849 | |
|
|
850 | newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call |
|
|
851 | |
|
|
852 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
853 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
854 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
855 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
856 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
857 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
858 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
859 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
860 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
861 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
862 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
863 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
864 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
865 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
866 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
867 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
868 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
869 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
870 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
871 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
872 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
873 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
874 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
875 | |
|
|
876 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
247 | } |
877 | } |
248 | |
878 | |
249 | cl_int |
879 | cl_int |
250 | errno () |
880 | errno () |
251 | CODE: |
881 | CODE: |
252 | errno = res; |
882 | RETVAL = res; |
|
|
883 | OUTPUT: |
|
|
884 | RETVAL |
253 | |
885 | |
254 | const char * |
886 | const char * |
255 | err2str (cl_int err) |
887 | err2str (cl_int err = res) |
256 | |
888 | |
257 | const char * |
889 | const char * |
258 | enum2str (cl_uint value) |
890 | enum2str (cl_uint value) |
259 | |
891 | |
260 | void |
892 | void |
… | |
… | |
268 | list = tmpbuf (sizeof (*list) * count); |
900 | list = tmpbuf (sizeof (*list) * count); |
269 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
901 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
270 | |
902 | |
271 | EXTEND (SP, count); |
903 | EXTEND (SP, count); |
272 | for (i = 0; i < count; ++i) |
904 | for (i = 0; i < count; ++i) |
273 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
905 | PUSH_CLOBJ (stash_platform, list [i]); |
274 | |
906 | |
275 | void |
907 | void |
276 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
908 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
277 | PPCODE: |
909 | PPCODE: |
|
|
910 | CONTEXT_NOTIFY_CALLBACK; |
278 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
911 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
279 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
912 | XPUSH_CLOBJ_CONTEXT; |
280 | |
913 | |
281 | void |
914 | void |
282 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
915 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
283 | PPCODE: |
916 | PPCODE: |
284 | /* der Gipfel der Kunst */ |
917 | cl_uint device_count; |
|
|
918 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
919 | |
|
|
920 | CONTEXT_NOTIFY_CALLBACK; |
|
|
921 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
922 | XPUSH_CLOBJ_CONTEXT; |
285 | |
923 | |
286 | void |
924 | void |
287 | wait_for_events (...) |
925 | wait_for_events (...) |
288 | CODE: |
926 | CODE: |
289 | EVENT_LIST (0, items); |
927 | EVENT_LIST (0); |
290 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
928 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
291 | |
929 | |
292 | PROTOTYPES: DISABLE |
930 | PROTOTYPES: DISABLE |
293 | |
931 | |
294 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
932 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
295 | |
933 | |
296 | void |
934 | void |
297 | info (OpenCL::Platform self, cl_platform_info name) |
935 | info (OpenCL::Platform self, cl_platform_info name) |
298 | PPCODE: |
936 | PPCODE: |
299 | INFO (Platform) |
937 | INFO (Platform) |
|
|
938 | |
|
|
939 | void |
|
|
940 | unload_compiler (OpenCL::Platform self) |
|
|
941 | CODE: |
|
|
942 | #if CL_VERSION_1_2 |
|
|
943 | clUnloadPlatformCompiler (self); |
|
|
944 | #endif |
300 | |
945 | |
301 | #BEGIN:platform |
946 | #BEGIN:platform |
302 | |
947 | |
303 | void |
948 | void |
304 | profile (OpenCL::Platform self) |
949 | profile (OpenCL::Platform self) |
… | |
… | |
330 | list = tmpbuf (sizeof (*list) * count); |
975 | list = tmpbuf (sizeof (*list) * count); |
331 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
976 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
332 | |
977 | |
333 | EXTEND (SP, count); |
978 | EXTEND (SP, count); |
334 | for (i = 0; i < count; ++i) |
979 | for (i = 0; i < count; ++i) |
335 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
980 | PUSH_CLOBJ (stash_device, list [i]); |
336 | |
981 | |
337 | void |
982 | void |
338 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
983 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
339 | PPCODE: |
984 | PPCODE: |
|
|
985 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
986 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
987 | |
|
|
988 | cl_uint device_count; |
|
|
989 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
990 | |
|
|
991 | CONTEXT_NOTIFY_CALLBACK; |
|
|
992 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
993 | XPUSH_CLOBJ_CONTEXT; |
|
|
994 | |
|
|
995 | void |
|
|
996 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
997 | PPCODE: |
|
|
998 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
999 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
1000 | |
|
|
1001 | CONTEXT_NOTIFY_CALLBACK; |
|
|
1002 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
1003 | XPUSH_CLOBJ_CONTEXT; |
|
|
1004 | |
|
|
1005 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
1006 | |
|
|
1007 | void |
|
|
1008 | info (OpenCL::Device self, cl_device_info name) |
|
|
1009 | PPCODE: |
|
|
1010 | INFO (Device) |
|
|
1011 | |
|
|
1012 | #if CL_VERSION_1_2 |
|
|
1013 | |
|
|
1014 | void |
|
|
1015 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
1016 | PPCODE: |
340 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
1017 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
341 | croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); |
1018 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
342 | |
1019 | |
343 | AV *av = (AV *)SvRV (devices); |
1020 | properties = SvRV (properties); |
344 | cl_uint num_devices = av_len (av) + 1; |
1021 | |
345 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
1022 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
1023 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
1024 | |
346 | int i; |
1025 | int i; |
|
|
1026 | for (i = 0; i < count; ++i) |
|
|
1027 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
347 | |
1028 | |
348 | for (i = num_devices; i--; ) |
1029 | props [count] = 0; |
349 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
350 | |
1030 | |
351 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
1031 | cl_uint num_devices; |
352 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
1032 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
1033 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
1034 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
353 | |
1035 | |
354 | void |
1036 | EXTEND (SP, num_devices); |
355 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
1037 | for (i = 0; i < count; ++i) |
356 | PPCODE: |
1038 | PUSH_CLOBJ (stash_subdevice, list [i]); |
357 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
358 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
359 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
|
|
360 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
|
|
361 | |
1039 | |
362 | MODULE = OpenCL PACKAGE = OpenCL::Device |
1040 | #endif |
363 | |
|
|
364 | void |
|
|
365 | info (OpenCL::Device self, cl_device_info name) |
|
|
366 | PPCODE: |
|
|
367 | INFO (Device) |
|
|
368 | |
1041 | |
369 | #BEGIN:device |
1042 | #BEGIN:device |
370 | |
1043 | |
371 | void |
1044 | void |
372 | type (OpenCL::Device self) |
1045 | type (OpenCL::Device self) |
373 | PPCODE: |
1046 | PPCODE: |
374 | cl_device_type value [1]; |
1047 | cl_device_type value [1]; |
375 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
1048 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
376 | EXTEND (SP, 1); |
1049 | EXTEND (SP, 1); |
377 | const int i = 0; |
1050 | const int i = 0; |
378 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1051 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
379 | |
1052 | |
380 | void |
1053 | void |
381 | vendor_id (OpenCL::Device self) |
1054 | vendor_id (OpenCL::Device self) |
382 | ALIAS: |
1055 | ALIAS: |
383 | vendor_id = CL_DEVICE_VENDOR_ID |
1056 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
404 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
1077 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
405 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
1078 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
406 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
1079 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
407 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
1080 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
408 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
1081 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
409 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
1082 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
410 | PPCODE: |
1083 | PPCODE: |
411 | cl_uint value [1]; |
1084 | cl_uint value [1]; |
412 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1085 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
413 | EXTEND (SP, 1); |
1086 | EXTEND (SP, 1); |
414 | const int i = 0; |
1087 | const int i = 0; |
… | |
… | |
528 | PPCODE: |
1201 | PPCODE: |
529 | cl_command_queue_properties value [1]; |
1202 | cl_command_queue_properties value [1]; |
530 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1203 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
531 | EXTEND (SP, 1); |
1204 | EXTEND (SP, 1); |
532 | const int i = 0; |
1205 | const int i = 0; |
533 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1206 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
534 | |
1207 | |
535 | void |
1208 | void |
536 | platform (OpenCL::Device self) |
1209 | platform (OpenCL::Device self) |
537 | PPCODE: |
1210 | PPCODE: |
538 | cl_platform_id value [1]; |
1211 | cl_platform_id value [1]; |
539 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1212 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
540 | EXTEND (SP, 1); |
1213 | EXTEND (SP, 1); |
541 | const int i = 0; |
1214 | const int i = 0; |
542 | { |
1215 | { |
543 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1216 | PUSH_CLOBJ (stash_platform, value [i]); |
544 | } |
1217 | } |
545 | |
1218 | |
546 | void |
1219 | void |
547 | name (OpenCL::Device self) |
1220 | name (OpenCL::Device self) |
548 | ALIAS: |
1221 | ALIAS: |
… | |
… | |
567 | cl_device_id value [1]; |
1240 | cl_device_id value [1]; |
568 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1241 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
569 | EXTEND (SP, 1); |
1242 | EXTEND (SP, 1); |
570 | const int i = 0; |
1243 | const int i = 0; |
571 | { |
1244 | { |
572 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1245 | PUSH_CLOBJ (stash_device, value [i]); |
573 | } |
1246 | } |
574 | |
1247 | |
575 | void |
1248 | void |
576 | partition_types_ext (OpenCL::Device self) |
1249 | partition_types_ext (OpenCL::Device self) |
577 | ALIAS: |
1250 | ALIAS: |
… | |
… | |
588 | for (i = 0; i < n; ++i) |
1261 | for (i = 0; i < n; ++i) |
589 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1262 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
590 | |
1263 | |
591 | #END:device |
1264 | #END:device |
592 | |
1265 | |
|
|
1266 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1267 | |
|
|
1268 | #if CL_VERSION_1_2 |
|
|
1269 | |
|
|
1270 | void |
|
|
1271 | DESTROY (OpenCL::SubDevice self) |
|
|
1272 | CODE: |
|
|
1273 | clReleaseDevice (self); |
|
|
1274 | |
|
|
1275 | #endif |
|
|
1276 | |
593 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1277 | MODULE = OpenCL PACKAGE = OpenCL::Context |
594 | |
1278 | |
595 | void |
1279 | void |
596 | DESTROY (OpenCL::Context context) |
1280 | DESTROY (OpenCL::Context self) |
597 | CODE: |
1281 | CODE: |
598 | clReleaseContext (context); |
1282 | clReleaseContext (self); |
599 | |
1283 | |
600 | void |
1284 | void |
601 | info (OpenCL::Context self, cl_context_info name) |
1285 | info (OpenCL::Context self, cl_context_info name) |
602 | PPCODE: |
1286 | PPCODE: |
603 | INFO (Context) |
1287 | INFO (Context) |
604 | |
1288 | |
605 | void |
1289 | void |
606 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1290 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
607 | PPCODE: |
1291 | PPCODE: |
608 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1292 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
609 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1293 | XPUSH_CLOBJ (stash_queue, queue); |
610 | |
1294 | |
611 | void |
1295 | void |
612 | user_event (OpenCL::Context self) |
1296 | user_event (OpenCL::Context self) |
613 | PPCODE: |
1297 | PPCODE: |
614 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1298 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
615 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1299 | XPUSH_CLOBJ (stash_userevent, ev); |
616 | |
1300 | |
617 | void |
1301 | void |
618 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1302 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
619 | PPCODE: |
1303 | PPCODE: |
620 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1304 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
621 | croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1305 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
622 | |
1306 | |
623 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1307 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
624 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1308 | XPUSH_CLOBJ (stash_bufferobj, mem); |
625 | |
1309 | |
626 | void |
1310 | void |
627 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1311 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
628 | PPCODE: |
1312 | PPCODE: |
629 | STRLEN len; |
1313 | STRLEN len; |
630 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1314 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
631 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1315 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
632 | croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1316 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
633 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1317 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
634 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1318 | XPUSH_CLOBJ (stash_bufferobj, mem); |
635 | |
1319 | |
|
|
1320 | #if CL_VERSION_1_2 |
|
|
1321 | |
636 | void |
1322 | void |
637 | 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) |
1323 | 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) |
638 | PPCODE: |
1324 | PPCODE: |
639 | STRLEN len; |
1325 | STRLEN len; |
640 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1326 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
641 | const cl_image_format format = { channel_order, channel_type }; |
1327 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1328 | const cl_image_desc desc = { |
|
|
1329 | type, |
|
|
1330 | width, height, depth, |
|
|
1331 | array_size, row_pitch, slice_pitch, |
|
|
1332 | num_mip_level, num_samples, |
|
|
1333 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
|
|
1334 | }; |
642 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1335 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
643 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1336 | HV *stash = stash_image; |
|
|
1337 | switch (type) |
|
|
1338 | { |
|
|
1339 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1340 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1341 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1342 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1343 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1344 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1345 | } |
|
|
1346 | XPUSH_CLOBJ (stash, mem); |
644 | |
1347 | |
|
|
1348 | #endif |
|
|
1349 | |
645 | void |
1350 | void |
646 | 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) |
1351 | 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) |
647 | PPCODE: |
1352 | PPCODE: |
648 | STRLEN len; |
1353 | STRLEN len; |
649 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1354 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
650 | const cl_image_format format = { channel_order, channel_type }; |
1355 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1356 | #if PREFER_1_1 |
|
|
1357 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1358 | #else |
|
|
1359 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1360 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1361 | #endif |
|
|
1362 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1363 | |
|
|
1364 | void |
|
|
1365 | 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) |
|
|
1366 | PPCODE: |
|
|
1367 | STRLEN len; |
|
|
1368 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1369 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1370 | #if PREFER_1_1 |
651 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1371 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
652 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1372 | #else |
|
|
1373 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1374 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1375 | #endif |
|
|
1376 | XPUSH_CLOBJ (stash_image3d, mem); |
653 | |
1377 | |
654 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1378 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
655 | |
1379 | |
656 | void |
1380 | void |
657 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
1381 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
658 | PPCODE: |
1382 | PPCODE: |
659 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
1383 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
660 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1384 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1385 | |
|
|
1386 | void |
|
|
1387 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1388 | PPCODE: |
|
|
1389 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1390 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1391 | |
|
|
1392 | #if CL_VERSION_1_2 |
|
|
1393 | |
|
|
1394 | void |
|
|
1395 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1396 | ALIAS: |
|
|
1397 | PPCODE: |
|
|
1398 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1399 | cl_gl_object_type type; |
|
|
1400 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1401 | HV *stash = stash_memory; |
|
|
1402 | switch (type) |
|
|
1403 | { |
|
|
1404 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1405 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1406 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1407 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1408 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1409 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1410 | } |
|
|
1411 | XPUSH_CLOBJ (stash, mem); |
|
|
1412 | |
|
|
1413 | #endif |
661 | |
1414 | |
662 | void |
1415 | void |
663 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1416 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
664 | PPCODE: |
1417 | PPCODE: |
|
|
1418 | #if PREFER_1_1 |
665 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
1419 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
666 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1420 | #else |
|
|
1421 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1422 | #endif |
|
|
1423 | XPUSH_CLOBJ (stash_image2d, mem); |
667 | |
1424 | |
668 | void |
1425 | void |
669 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1426 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
670 | PPCODE: |
1427 | PPCODE: |
|
|
1428 | #if PREFER_1_1 |
671 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
1429 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
672 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1430 | #else |
673 | |
1431 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
674 | void |
1432 | #endif |
675 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
1433 | XPUSH_CLOBJ (stash_image3d, mem); |
676 | PPCODE: |
|
|
677 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderBuffer, (self, flags, renderbuffer, &res)); |
|
|
678 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
679 | |
1434 | |
680 | #endif |
1435 | #endif |
681 | |
1436 | |
682 | void |
1437 | void |
683 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1438 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
703 | |
1458 | |
704 | void |
1459 | void |
705 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1460 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
706 | PPCODE: |
1461 | PPCODE: |
707 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1462 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
708 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1463 | XPUSH_CLOBJ (stash_sampler, sampler); |
709 | |
1464 | |
710 | void |
1465 | void |
711 | program_with_source (OpenCL::Context self, SV *program) |
1466 | program_with_source (OpenCL::Context self, SV *program) |
712 | PPCODE: |
1467 | PPCODE: |
713 | STRLEN len; |
1468 | STRLEN len; |
714 | size_t len2; |
1469 | size_t len2; |
715 | const char *ptr = SvPVbyte (program, len); |
1470 | const char *ptr = SvPVbyte (program, len); |
716 | |
1471 | |
717 | len2 = len; |
1472 | len2 = len; |
718 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1473 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
719 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1474 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1475 | |
|
|
1476 | void |
|
|
1477 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1478 | PPCODE: |
|
|
1479 | cl_uint device_count; |
|
|
1480 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1481 | |
|
|
1482 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1483 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1484 | |
|
|
1485 | binaries = SvRV (binaries); |
|
|
1486 | |
|
|
1487 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1488 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1489 | |
|
|
1490 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1491 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1492 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1493 | |
|
|
1494 | int i; |
|
|
1495 | for (i = 0; i < device_count; ++i) |
|
|
1496 | { |
|
|
1497 | STRLEN len; |
|
|
1498 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1499 | length_list [i] = len; |
|
|
1500 | } |
|
|
1501 | |
|
|
1502 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1503 | length_list, binary_list, |
|
|
1504 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1505 | |
|
|
1506 | EXTEND (SP, 2); |
|
|
1507 | PUSH_CLOBJ (stash_program, prog); |
|
|
1508 | |
|
|
1509 | if (GIMME_V == G_ARRAY) |
|
|
1510 | { |
|
|
1511 | AV *av = newAV (); |
|
|
1512 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1513 | |
|
|
1514 | for (i = device_count; i--; ) |
|
|
1515 | av_store (av, i, newSViv (status_list [i])); |
|
|
1516 | } |
|
|
1517 | |
|
|
1518 | #if CL_VERSION_1_2 |
|
|
1519 | |
|
|
1520 | void |
|
|
1521 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1522 | PPCODE: |
|
|
1523 | cl_uint device_count; |
|
|
1524 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1525 | |
|
|
1526 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1527 | |
|
|
1528 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1529 | |
|
|
1530 | void |
|
|
1531 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1532 | CODE: |
|
|
1533 | cl_uint device_count = 0; |
|
|
1534 | cl_device_id *device_list = 0; |
|
|
1535 | |
|
|
1536 | if (SvOK (devices)) |
|
|
1537 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1538 | |
|
|
1539 | cl_uint program_count; |
|
|
1540 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1541 | |
|
|
1542 | void *user_data; |
|
|
1543 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1544 | |
|
|
1545 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1546 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1547 | |
|
|
1548 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1549 | |
|
|
1550 | #endif |
720 | |
1551 | |
721 | #BEGIN:context |
1552 | #BEGIN:context |
722 | |
1553 | |
723 | void |
1554 | void |
724 | reference_count (OpenCL::Context self) |
1555 | reference_count (OpenCL::Context self) |
… | |
… | |
741 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1572 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
742 | int i, n = size / sizeof (*value); |
1573 | int i, n = size / sizeof (*value); |
743 | EXTEND (SP, n); |
1574 | EXTEND (SP, n); |
744 | for (i = 0; i < n; ++i) |
1575 | for (i = 0; i < n; ++i) |
745 | { |
1576 | { |
746 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1577 | PUSH_CLOBJ (stash_device, value [i]); |
747 | } |
1578 | } |
748 | |
1579 | |
749 | void |
1580 | void |
750 | properties (OpenCL::Context self) |
1581 | properties (OpenCL::Context self) |
751 | PPCODE: |
1582 | PPCODE: |
… | |
… | |
766 | DESTROY (OpenCL::Queue self) |
1597 | DESTROY (OpenCL::Queue self) |
767 | CODE: |
1598 | CODE: |
768 | clReleaseCommandQueue (self); |
1599 | clReleaseCommandQueue (self); |
769 | |
1600 | |
770 | void |
1601 | void |
771 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1602 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1603 | ALIAS: |
|
|
1604 | enqueue_read_buffer = 0 |
772 | PPCODE: |
1605 | PPCODE: |
773 | cl_event ev = 0; |
|
|
774 | EVENT_LIST (6, items - 6); |
1606 | EVENT_LIST (6); |
775 | |
1607 | |
776 | SvUPGRADE (data, SVt_PV); |
1608 | SvUPGRADE (data, SVt_PV); |
777 | SvGROW (data, len); |
1609 | SvGROW (data, len); |
778 | SvPOK_only (data); |
1610 | SvPOK_only (data); |
779 | SvCUR_set (data, len); |
1611 | SvCUR_set (data, len); |
|
|
1612 | |
|
|
1613 | cl_event ev = 0; |
780 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1614 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
781 | |
1615 | |
782 | if (ev) |
1616 | if (ev) |
783 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1617 | XPUSH_CLOBJ (stash_event, ev); |
784 | |
1618 | |
785 | void |
1619 | void |
786 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1620 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1621 | ALIAS: |
|
|
1622 | enqueue_write_buffer = 0 |
787 | PPCODE: |
1623 | PPCODE: |
788 | cl_event ev = 0; |
1624 | EVENT_LIST (5); |
|
|
1625 | |
789 | STRLEN len; |
1626 | STRLEN len; |
790 | char *ptr = SvPVbyte (data, len); |
1627 | char *ptr = SvPVbyte (data, len); |
791 | EVENT_LIST (5, items - 5); |
|
|
792 | |
1628 | |
|
|
1629 | cl_event ev = 0; |
793 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1630 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
794 | |
1631 | |
795 | if (ev) |
1632 | if (ev) |
796 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1633 | XPUSH_CLOBJ (stash_event, ev); |
797 | |
1634 | |
|
|
1635 | #if CL_VERSION_1_2 |
|
|
1636 | |
798 | void |
1637 | void |
799 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1638 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1639 | ALIAS: |
|
|
1640 | enqueue_fill_buffer = 0 |
800 | PPCODE: |
1641 | PPCODE: |
|
|
1642 | EVENT_LIST (5); |
|
|
1643 | |
|
|
1644 | STRLEN len; |
|
|
1645 | char *ptr = SvPVbyte (data, len); |
|
|
1646 | |
801 | cl_event ev = 0; |
1647 | cl_event ev = 0; |
|
|
1648 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1649 | |
|
|
1650 | if (ev) |
|
|
1651 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1652 | |
|
|
1653 | void |
|
|
1654 | 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, ...) |
|
|
1655 | ALIAS: |
|
|
1656 | enqueue_fill_image = 0 |
|
|
1657 | PPCODE: |
|
|
1658 | EVENT_LIST (12); |
|
|
1659 | |
|
|
1660 | const size_t origin [3] = { x, y, z }; |
|
|
1661 | const size_t region [3] = { width, height, depth }; |
|
|
1662 | |
|
|
1663 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1664 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1665 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1666 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1667 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1668 | cl_image_format format; |
|
|
1669 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1670 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1671 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1672 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1673 | |
|
|
1674 | cl_event ev = 0; |
|
|
1675 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1676 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1677 | |
|
|
1678 | if (ev) |
|
|
1679 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1680 | |
|
|
1681 | #endif |
|
|
1682 | |
|
|
1683 | void |
|
|
1684 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1685 | ALIAS: |
|
|
1686 | enqueue_copy_buffer = 0 |
|
|
1687 | PPCODE: |
802 | EVENT_LIST (6, items - 6); |
1688 | EVENT_LIST (6); |
803 | |
1689 | |
|
|
1690 | cl_event ev = 0; |
804 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1691 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
805 | |
1692 | |
806 | if (ev) |
1693 | if (ev) |
807 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1694 | XPUSH_CLOBJ (stash_event, ev); |
808 | |
1695 | |
809 | void |
1696 | void |
810 | enqueue_read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1697 | 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, ...) |
|
|
1698 | ALIAS: |
|
|
1699 | enqueue_read_buffer_rect = 0 |
811 | PPCODE: |
1700 | PPCODE: |
812 | cl_event ev = 0; |
1701 | EVENT_LIST (17); |
|
|
1702 | |
813 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1703 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
814 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1704 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
815 | const size_t region[3] = { width, height, depth }; |
1705 | const size_t region[3] = { width, height, depth }; |
816 | EVENT_LIST (17, items - 17); |
|
|
817 | |
1706 | |
818 | if (!buf_row_pitch) |
1707 | if (!buf_row_pitch) |
819 | buf_row_pitch = region [0]; |
1708 | buf_row_pitch = region [0]; |
820 | |
1709 | |
821 | if (!buf_slice_pitch) |
1710 | if (!buf_slice_pitch) |
… | |
… | |
831 | |
1720 | |
832 | SvUPGRADE (data, SVt_PV); |
1721 | SvUPGRADE (data, SVt_PV); |
833 | SvGROW (data, len); |
1722 | SvGROW (data, len); |
834 | SvPOK_only (data); |
1723 | SvPOK_only (data); |
835 | SvCUR_set (data, len); |
1724 | SvCUR_set (data, len); |
|
|
1725 | |
|
|
1726 | cl_event ev = 0; |
836 | 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)); |
1727 | 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)); |
837 | |
1728 | |
838 | if (ev) |
1729 | if (ev) |
839 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1730 | XPUSH_CLOBJ (stash_event, ev); |
840 | |
1731 | |
841 | void |
1732 | void |
842 | enqueue_write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1733 | 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, ...) |
|
|
1734 | ALIAS: |
|
|
1735 | enqueue_write_buffer_rect = 0 |
843 | PPCODE: |
1736 | PPCODE: |
844 | cl_event ev = 0; |
1737 | EVENT_LIST (17); |
|
|
1738 | |
845 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1739 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
846 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1740 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
847 | const size_t region[3] = { width, height, depth }; |
1741 | const size_t region[3] = { width, height, depth }; |
848 | STRLEN len; |
1742 | STRLEN len; |
849 | char *ptr = SvPVbyte (data, len); |
1743 | char *ptr = SvPVbyte (data, len); |
850 | EVENT_LIST (17, items - 17); |
|
|
851 | |
1744 | |
852 | if (!buf_row_pitch) |
1745 | if (!buf_row_pitch) |
853 | buf_row_pitch = region [0]; |
1746 | buf_row_pitch = region [0]; |
854 | |
1747 | |
855 | if (!buf_slice_pitch) |
1748 | if (!buf_slice_pitch) |
… | |
… | |
864 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1757 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
865 | |
1758 | |
866 | if (len < min_len) |
1759 | if (len < min_len) |
867 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1760 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
868 | |
1761 | |
|
|
1762 | cl_event ev = 0; |
869 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1763 | 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)); |
870 | |
1764 | |
871 | if (ev) |
1765 | if (ev) |
872 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1766 | XPUSH_CLOBJ (stash_event, ev); |
873 | |
1767 | |
874 | void |
1768 | void |
875 | enqueue_copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
1769 | 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, ...) |
|
|
1770 | ALIAS: |
|
|
1771 | enqueue_copy_buffer_rect = 0 |
876 | PPCODE: |
1772 | PPCODE: |
877 | cl_event ev = 0; |
1773 | EVENT_LIST (16); |
|
|
1774 | |
878 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1775 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
879 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1776 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
880 | const size_t region[3] = { width, height, depth }; |
1777 | const size_t region[3] = { width, height, depth }; |
881 | EVENT_LIST (16, items - 16); |
|
|
882 | |
1778 | |
|
|
1779 | cl_event ev = 0; |
883 | 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)); |
1780 | 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)); |
884 | |
1781 | |
885 | if (ev) |
1782 | if (ev) |
886 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1783 | XPUSH_CLOBJ (stash_event, ev); |
887 | |
1784 | |
888 | void |
1785 | void |
889 | enqueue_read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1786 | 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, ...) |
|
|
1787 | ALIAS: |
|
|
1788 | enqueue_read_image = 0 |
890 | PPCODE: |
1789 | PPCODE: |
891 | cl_event ev = 0; |
1790 | EVENT_LIST (12); |
|
|
1791 | |
892 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1792 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
893 | const size_t region[3] = { width, height, depth }; |
1793 | const size_t region[3] = { width, height, depth }; |
894 | EVENT_LIST (12, items - 12); |
|
|
895 | |
1794 | |
896 | if (!row_pitch) |
1795 | if (!row_pitch) |
897 | row_pitch = img_row_pitch (src); |
1796 | row_pitch = img_row_pitch (src); |
898 | |
1797 | |
899 | if (depth > 1 && !slice_pitch) |
1798 | if (depth > 1 && !slice_pitch) |
… | |
… | |
903 | |
1802 | |
904 | SvUPGRADE (data, SVt_PV); |
1803 | SvUPGRADE (data, SVt_PV); |
905 | SvGROW (data, len); |
1804 | SvGROW (data, len); |
906 | SvPOK_only (data); |
1805 | SvPOK_only (data); |
907 | SvCUR_set (data, len); |
1806 | SvCUR_set (data, len); |
|
|
1807 | |
|
|
1808 | cl_event ev = 0; |
908 | 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)); |
1809 | 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)); |
909 | |
1810 | |
910 | if (ev) |
1811 | if (ev) |
911 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1812 | XPUSH_CLOBJ (stash_event, ev); |
912 | |
1813 | |
913 | void |
1814 | void |
914 | enqueue_write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1815 | 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, ...) |
|
|
1816 | ALIAS: |
|
|
1817 | enqueue_write_image = 0 |
915 | PPCODE: |
1818 | PPCODE: |
916 | cl_event ev = 0; |
1819 | EVENT_LIST (12); |
|
|
1820 | |
917 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1821 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
918 | const size_t region[3] = { width, height, depth }; |
1822 | const size_t region[3] = { width, height, depth }; |
919 | STRLEN len; |
1823 | STRLEN len; |
920 | char *ptr = SvPVbyte (data, len); |
1824 | char *ptr = SvPVbyte (data, len); |
921 | EVENT_LIST (12, items - 12); |
|
|
922 | |
1825 | |
923 | if (!row_pitch) |
1826 | if (!row_pitch) |
924 | row_pitch = img_row_pitch (dst); |
1827 | row_pitch = img_row_pitch (dst); |
925 | |
1828 | |
926 | if (depth > 1 && !slice_pitch) |
1829 | if (depth > 1 && !slice_pitch) |
… | |
… | |
929 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1832 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
930 | |
1833 | |
931 | if (len < min_len) |
1834 | if (len < min_len) |
932 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1835 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
933 | |
1836 | |
|
|
1837 | cl_event ev = 0; |
934 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1838 | 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)); |
935 | |
1839 | |
936 | if (ev) |
1840 | if (ev) |
937 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1841 | XPUSH_CLOBJ (stash_event, ev); |
938 | |
1842 | |
939 | void |
1843 | void |
940 | enqueue_copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
1844 | 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, ...) |
|
|
1845 | ALIAS: |
|
|
1846 | enqueue_copy_image = 0 |
941 | PPCODE: |
1847 | PPCODE: |
942 | cl_event ev = 0; |
1848 | EVENT_LIST (12); |
|
|
1849 | |
943 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1850 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
944 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1851 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
945 | const size_t region[3] = { width, height, depth }; |
1852 | const size_t region[3] = { width, height, depth }; |
946 | EVENT_LIST (12, items - 12); |
|
|
947 | |
1853 | |
|
|
1854 | cl_event ev = 0; |
948 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1855 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
949 | |
1856 | |
950 | if (ev) |
1857 | if (ev) |
951 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1858 | XPUSH_CLOBJ (stash_event, ev); |
952 | |
1859 | |
953 | void |
1860 | void |
954 | enqueue_copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
1861 | 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, ...) |
|
|
1862 | ALIAS: |
|
|
1863 | enqueue_copy_image_to_buffer = 0 |
955 | PPCODE: |
1864 | PPCODE: |
|
|
1865 | EVENT_LIST (10); |
|
|
1866 | |
|
|
1867 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1868 | const size_t region [3] = { width, height, depth }; |
|
|
1869 | |
956 | cl_event ev = 0; |
1870 | cl_event ev = 0; |
|
|
1871 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1872 | |
|
|
1873 | if (ev) |
|
|
1874 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1875 | |
|
|
1876 | void |
|
|
1877 | 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, ...) |
|
|
1878 | ALIAS: |
|
|
1879 | enqueue_copy_buffer_to_image = 0 |
|
|
1880 | PPCODE: |
|
|
1881 | EVENT_LIST (10); |
|
|
1882 | |
957 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1883 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1884 | const size_t region [3] = { width, height, depth }; |
|
|
1885 | |
|
|
1886 | cl_event ev = 0; |
|
|
1887 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1888 | |
|
|
1889 | if (ev) |
|
|
1890 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1891 | |
|
|
1892 | void |
|
|
1893 | 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, ...) |
|
|
1894 | ALIAS: |
|
|
1895 | enqueue_map_buffer = 0 |
|
|
1896 | PPCODE: |
|
|
1897 | EVENT_LIST (6); |
|
|
1898 | |
|
|
1899 | size_t cb = SvIV (cb_); |
|
|
1900 | |
|
|
1901 | if (!SvOK (cb_)) |
|
|
1902 | { |
|
|
1903 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1904 | cb -= offset; |
|
|
1905 | } |
|
|
1906 | |
|
|
1907 | cl_event ev; |
|
|
1908 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1909 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1910 | |
|
|
1911 | void |
|
|
1912 | 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, ...) |
|
|
1913 | ALIAS: |
|
|
1914 | enqueue_map_image = 0 |
|
|
1915 | PPCODE: |
|
|
1916 | size_t width = SvIV (width_); |
|
|
1917 | if (!SvOK (width_)) |
|
|
1918 | { |
|
|
1919 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1920 | width -= x; |
|
|
1921 | } |
|
|
1922 | |
|
|
1923 | size_t height = SvIV (width_); |
|
|
1924 | if (!SvOK (height_)) |
|
|
1925 | { |
|
|
1926 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1927 | height -= y; |
|
|
1928 | } |
|
|
1929 | |
|
|
1930 | size_t depth = SvIV (width_); |
|
|
1931 | if (!SvOK (depth_)) |
|
|
1932 | { |
|
|
1933 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1934 | depth -= z; |
|
|
1935 | |
|
|
1936 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1937 | if (!depth) |
|
|
1938 | depth = 1; |
|
|
1939 | } |
|
|
1940 | |
|
|
1941 | const size_t origin[3] = { x, y, z }; |
958 | const size_t region[3] = { width, height, depth }; |
1942 | const size_t region[3] = { width, height, depth }; |
|
|
1943 | size_t row_pitch, slice_pitch; |
959 | EVENT_LIST (10, items - 10); |
1944 | EVENT_LIST (10); |
960 | |
1945 | |
|
|
1946 | cl_event ev; |
|
|
1947 | 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)); |
|
|
1948 | |
|
|
1949 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1950 | : row_pitch ? row_pitch * region [1] |
|
|
1951 | : region [0]; |
|
|
1952 | |
|
|
1953 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1954 | |
|
|
1955 | void |
|
|
1956 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1957 | PPCODE: |
|
|
1958 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2); |
|
|
1959 | if (GIMME_V != G_VOID) |
|
|
1960 | { |
|
|
1961 | clRetainEvent (mapped->event); |
|
|
1962 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1963 | } |
|
|
1964 | |
|
|
1965 | void |
|
|
1966 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1967 | ALIAS: |
|
|
1968 | enqueue_task = 0 |
|
|
1969 | PPCODE: |
|
|
1970 | EVENT_LIST (2); |
|
|
1971 | |
|
|
1972 | cl_event ev = 0; |
961 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1973 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
962 | |
1974 | |
963 | if (ev) |
1975 | if (ev) |
964 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1976 | XPUSH_CLOBJ (stash_event, ev); |
965 | |
1977 | |
966 | void |
1978 | void |
967 | enqueue_copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
968 | PPCODE: |
|
|
969 | cl_event ev = 0; |
|
|
970 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
971 | const size_t region[3] = { width, height, depth }; |
|
|
972 | EVENT_LIST (10, items - 10); |
|
|
973 | |
|
|
974 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
975 | |
|
|
976 | if (ev) |
|
|
977 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
978 | |
|
|
979 | void |
|
|
980 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
981 | PPCODE: |
|
|
982 | cl_event ev = 0; |
|
|
983 | EVENT_LIST (2, items - 2); |
|
|
984 | |
|
|
985 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
986 | |
|
|
987 | if (ev) |
|
|
988 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
989 | |
|
|
990 | void |
|
|
991 | enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1979 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1980 | ALIAS: |
|
|
1981 | enqueue_nd_range_kernel = 0 |
992 | PPCODE: |
1982 | PPCODE: |
993 | cl_event ev = 0; |
1983 | EVENT_LIST (5); |
|
|
1984 | |
994 | size_t *gwo = 0, *gws, *lws = 0; |
1985 | size_t *gwo = 0, *gws, *lws = 0; |
995 | int gws_len; |
1986 | int gws_len; |
996 | size_t *lists; |
1987 | size_t *lists; |
997 | int i; |
1988 | int i; |
998 | EVENT_LIST (5, items - 5); |
|
|
999 | |
1989 | |
1000 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1990 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1001 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1991 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1002 | |
1992 | |
1003 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1993 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1004 | |
1994 | |
1005 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1995 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1006 | |
1996 | |
1007 | gws = lists + gws_len * 0; |
1997 | gws = lists + gws_len * 0; |
1008 | for (i = 0; i < gws_len; ++i) |
1998 | for (i = 0; i < gws_len; ++i) |
|
|
1999 | { |
1009 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
2000 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
2001 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
2002 | if (!gws [i]) |
|
|
2003 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
2004 | } |
1010 | |
2005 | |
1011 | if (SvOK (global_work_offset)) |
2006 | if (SvOK (global_work_offset)) |
1012 | { |
2007 | { |
1013 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
2008 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1014 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
2009 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1021 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
2016 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1022 | } |
2017 | } |
1023 | |
2018 | |
1024 | if (SvOK (local_work_size)) |
2019 | if (SvOK (local_work_size)) |
1025 | { |
2020 | { |
1026 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
2021 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1027 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
2022 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1028 | |
2023 | |
1029 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
2024 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1030 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
2025 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1031 | |
2026 | |
1032 | lws = lists + gws_len * 2; |
2027 | lws = lists + gws_len * 2; |
1033 | for (i = 0; i < gws_len; ++i) |
2028 | for (i = 0; i < gws_len; ++i) |
|
|
2029 | { |
1034 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
2030 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
2031 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
2032 | if (!lws [i]) |
|
|
2033 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
2034 | } |
1035 | } |
2035 | } |
1036 | |
2036 | |
|
|
2037 | cl_event ev = 0; |
1037 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2038 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1038 | |
2039 | |
1039 | if (ev) |
2040 | if (ev) |
1040 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2041 | XPUSH_CLOBJ (stash_event, ev); |
1041 | |
2042 | |
1042 | #if 0 |
2043 | #if CL_VERSION_1_2 |
|
|
2044 | |
|
|
2045 | void |
|
|
2046 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
2047 | ALIAS: |
|
|
2048 | enqueue_migrate_mem_objects = 0 |
|
|
2049 | PPCODE: |
|
|
2050 | EVENT_LIST (3); |
|
|
2051 | |
|
|
2052 | cl_uint obj_count; |
|
|
2053 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2054 | |
|
|
2055 | cl_event ev = 0; |
|
|
2056 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2057 | |
|
|
2058 | if (ev) |
|
|
2059 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2060 | |
|
|
2061 | #endif |
|
|
2062 | |
1043 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2063 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1044 | |
2064 | |
1045 | void |
2065 | void |
1046 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2066 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
2067 | ALIAS: |
|
|
2068 | release_gl_objects = 1 |
|
|
2069 | enqueue_acquire_gl_objects = 0 |
|
|
2070 | enqueue_release_gl_objects = 1 |
1047 | CODE: |
2071 | PPCODE: |
|
|
2072 | EVENT_LIST (2); |
|
|
2073 | |
|
|
2074 | cl_uint obj_count; |
|
|
2075 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2076 | |
1048 | cl_event ev = 0; |
2077 | cl_event ev = 0; |
1049 | EVENT_LIST (2, items - 2); |
|
|
1050 | |
2078 | |
|
|
2079 | if (ix) |
|
|
2080 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2081 | else |
1051 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, ..., event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2082 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1052 | |
2083 | |
1053 | if (ev) |
2084 | if (ev) |
1054 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2085 | XPUSH_CLOBJ (stash_event, ev); |
1055 | |
2086 | |
|
|
2087 | #endif |
|
|
2088 | |
1056 | void |
2089 | void |
1057 | enqueue_release_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2090 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2091 | ALIAS: |
|
|
2092 | enqueue_wait_for_events = 0 |
|
|
2093 | CODE: |
|
|
2094 | EVENT_LIST (1); |
|
|
2095 | #if PREFER_1_1 |
|
|
2096 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2097 | #else |
|
|
2098 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2099 | #endif |
|
|
2100 | |
|
|
2101 | void |
|
|
2102 | marker (OpenCL::Queue self, ...) |
|
|
2103 | ALIAS: |
|
|
2104 | enqueue_marker = 0 |
1058 | CODE: |
2105 | PPCODE: |
|
|
2106 | EVENT_LIST (1); |
1059 | cl_event ev = 0; |
2107 | cl_event ev = 0; |
1060 | EVENT_LIST (2, items - 2); |
2108 | #if PREFER_1_1 |
1061 | |
2109 | if (!event_list_count) |
|
|
2110 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2111 | else |
|
|
2112 | #if CL_VERSION_1_2 |
|
|
2113 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2114 | #else |
|
|
2115 | { |
|
|
2116 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2117 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2118 | } |
|
|
2119 | #endif |
|
|
2120 | #else |
1062 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, ..., event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2121 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1063 | |
2122 | #endif |
1064 | if (ev) |
2123 | if (ev) |
1065 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2124 | XPUSH_CLOBJ (stash_event, ev); |
1066 | |
2125 | |
|
|
2126 | void |
|
|
2127 | barrier (OpenCL::Queue self, ...) |
|
|
2128 | ALIAS: |
|
|
2129 | enqueue_barrier = 0 |
|
|
2130 | PPCODE: |
|
|
2131 | EVENT_LIST (1); |
|
|
2132 | cl_event ev = 0; |
|
|
2133 | #if PREFER_1_1 |
|
|
2134 | if (!event_list_count && GIMME_V == G_VOID) |
|
|
2135 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2136 | else |
|
|
2137 | #if CL_VERSION_1_2 |
|
|
2138 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2139 | #else |
|
|
2140 | { |
|
|
2141 | if (event_list_count) |
|
|
2142 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2143 | |
|
|
2144 | if (GIMME_V != G_VOID) |
|
|
2145 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2146 | } |
1067 | #endif |
2147 | #endif |
|
|
2148 | #else |
|
|
2149 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1068 | #endif |
2150 | #endif |
1069 | |
2151 | if (ev) |
1070 | void |
2152 | XPUSH_CLOBJ (stash_event, ev); |
1071 | enqueue_marker (OpenCL::Queue self) |
|
|
1072 | PPCODE: |
|
|
1073 | cl_event ev; |
|
|
1074 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1075 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1076 | |
|
|
1077 | void |
|
|
1078 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
|
|
1079 | CODE: |
|
|
1080 | EVENT_LIST (1, items - 1); |
|
|
1081 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1082 | |
|
|
1083 | void |
|
|
1084 | enqueue_barrier (OpenCL::Queue self) |
|
|
1085 | CODE: |
|
|
1086 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1087 | |
2153 | |
1088 | void |
2154 | void |
1089 | flush (OpenCL::Queue self) |
2155 | flush (OpenCL::Queue self) |
1090 | CODE: |
2156 | CODE: |
1091 | NEED_SUCCESS (Flush, (self)); |
2157 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1109 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
2175 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1110 | EXTEND (SP, 1); |
2176 | EXTEND (SP, 1); |
1111 | const int i = 0; |
2177 | const int i = 0; |
1112 | { |
2178 | { |
1113 | NEED_SUCCESS (RetainContext, (value [i])); |
2179 | NEED_SUCCESS (RetainContext, (value [i])); |
1114 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2180 | PUSH_CLOBJ (stash_context, value [i]); |
1115 | } |
2181 | } |
1116 | |
2182 | |
1117 | void |
2183 | void |
1118 | device (OpenCL::Queue self) |
2184 | device (OpenCL::Queue self) |
1119 | PPCODE: |
2185 | PPCODE: |
1120 | cl_device_id value [1]; |
2186 | cl_device_id value [1]; |
1121 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
2187 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1122 | EXTEND (SP, 1); |
2188 | EXTEND (SP, 1); |
1123 | const int i = 0; |
2189 | const int i = 0; |
1124 | { |
2190 | { |
1125 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2191 | PUSH_CLOBJ (stash_device, value [i]); |
1126 | } |
2192 | } |
1127 | |
2193 | |
1128 | void |
2194 | void |
1129 | reference_count (OpenCL::Queue self) |
2195 | reference_count (OpenCL::Queue self) |
1130 | PPCODE: |
2196 | PPCODE: |
… | |
… | |
1139 | PPCODE: |
2205 | PPCODE: |
1140 | cl_command_queue_properties value [1]; |
2206 | cl_command_queue_properties value [1]; |
1141 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
2207 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1142 | EXTEND (SP, 1); |
2208 | EXTEND (SP, 1); |
1143 | const int i = 0; |
2209 | const int i = 0; |
1144 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2210 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1145 | |
2211 | |
1146 | #END:command_queue |
2212 | #END:command_queue |
1147 | |
2213 | |
1148 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2214 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1149 | |
2215 | |
… | |
… | |
1154 | |
2220 | |
1155 | void |
2221 | void |
1156 | info (OpenCL::Memory self, cl_mem_info name) |
2222 | info (OpenCL::Memory self, cl_mem_info name) |
1157 | PPCODE: |
2223 | PPCODE: |
1158 | INFO (MemObject) |
2224 | INFO (MemObject) |
|
|
2225 | |
|
|
2226 | void |
|
|
2227 | destructor_callback (OpenCL::Memory self, SV *notify) |
|
|
2228 | PPCODE: |
|
|
2229 | clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify))); |
1159 | |
2230 | |
1160 | #BEGIN:mem |
2231 | #BEGIN:mem |
1161 | |
2232 | |
1162 | void |
2233 | void |
1163 | type (OpenCL::Memory self) |
2234 | type (OpenCL::Memory self) |
1164 | PPCODE: |
2235 | PPCODE: |
1165 | cl_mem_object_type value [1]; |
2236 | cl_mem_object_type value [1]; |
1166 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
2237 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1167 | EXTEND (SP, 1); |
2238 | EXTEND (SP, 1); |
1168 | const int i = 0; |
2239 | const int i = 0; |
1169 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2240 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1170 | |
2241 | |
1171 | void |
2242 | void |
1172 | flags (OpenCL::Memory self) |
2243 | flags (OpenCL::Memory self) |
1173 | PPCODE: |
2244 | PPCODE: |
1174 | cl_mem_flags value [1]; |
2245 | cl_mem_flags value [1]; |
1175 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
2246 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1176 | EXTEND (SP, 1); |
2247 | EXTEND (SP, 1); |
1177 | const int i = 0; |
2248 | const int i = 0; |
1178 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2249 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1179 | |
2250 | |
1180 | void |
2251 | void |
1181 | size (OpenCL::Memory self) |
2252 | size (OpenCL::Memory self) |
1182 | ALIAS: |
2253 | ALIAS: |
1183 | size = CL_MEM_SIZE |
2254 | size = CL_MEM_SIZE |
… | |
… | |
1217 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
2288 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1218 | EXTEND (SP, 1); |
2289 | EXTEND (SP, 1); |
1219 | const int i = 0; |
2290 | const int i = 0; |
1220 | { |
2291 | { |
1221 | NEED_SUCCESS (RetainContext, (value [i])); |
2292 | NEED_SUCCESS (RetainContext, (value [i])); |
1222 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2293 | PUSH_CLOBJ (stash_context, value [i]); |
1223 | } |
2294 | } |
1224 | |
2295 | |
1225 | void |
2296 | void |
1226 | associated_memobject (OpenCL::Memory self) |
2297 | associated_memobject (OpenCL::Memory self) |
1227 | PPCODE: |
2298 | PPCODE: |
… | |
… | |
1229 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
2300 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1230 | EXTEND (SP, 1); |
2301 | EXTEND (SP, 1); |
1231 | const int i = 0; |
2302 | const int i = 0; |
1232 | { |
2303 | { |
1233 | NEED_SUCCESS (RetainMemObject, (value [i])); |
2304 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1234 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
2305 | PUSH_CLOBJ (stash_memory, value [i]); |
1235 | } |
2306 | } |
1236 | |
2307 | |
1237 | #END:mem |
2308 | #END:mem |
|
|
2309 | |
|
|
2310 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2311 | |
|
|
2312 | void |
|
|
2313 | gl_object_info (OpenCL::Memory self) |
|
|
2314 | PPCODE: |
|
|
2315 | cl_gl_object_type type; |
|
|
2316 | cl_GLuint name; |
|
|
2317 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
|
|
2318 | EXTEND (SP, 2); |
|
|
2319 | PUSHs (sv_2mortal (newSVuv (type))); |
|
|
2320 | PUSHs (sv_2mortal (newSVuv (name))); |
|
|
2321 | |
|
|
2322 | #endif |
1238 | |
2323 | |
1239 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
2324 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
1240 | |
2325 | |
1241 | void |
2326 | void |
1242 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
2327 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
… | |
… | |
1245 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
2330 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1246 | |
2331 | |
1247 | cl_buffer_region crdata = { origin, size }; |
2332 | cl_buffer_region crdata = { origin, size }; |
1248 | |
2333 | |
1249 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2334 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1250 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2335 | XPUSH_CLOBJ (stash_buffer, mem); |
1251 | |
2336 | |
1252 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2337 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1253 | |
2338 | |
1254 | void |
2339 | void |
1255 | image_info (OpenCL::Image self, cl_image_info name) |
2340 | image_info (OpenCL::Image self, cl_image_info name) |
1256 | PPCODE: |
2341 | PPCODE: |
1257 | INFO (Image) |
2342 | INFO (Image) |
|
|
2343 | |
|
|
2344 | void |
|
|
2345 | format (OpenCL::Image self) |
|
|
2346 | PPCODE: |
|
|
2347 | cl_image_format format; |
|
|
2348 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2349 | EXTEND (SP, 2); |
|
|
2350 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2351 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1258 | |
2352 | |
1259 | #BEGIN:image |
2353 | #BEGIN:image |
1260 | |
2354 | |
1261 | void |
2355 | void |
1262 | element_size (OpenCL::Image self) |
2356 | element_size (OpenCL::Image self) |
… | |
… | |
1274 | const int i = 0; |
2368 | const int i = 0; |
1275 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2369 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1276 | |
2370 | |
1277 | #END:image |
2371 | #END:image |
1278 | |
2372 | |
|
|
2373 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2374 | |
|
|
2375 | #BEGIN:gl_texture |
|
|
2376 | |
|
|
2377 | void |
|
|
2378 | target (OpenCL::Image self) |
|
|
2379 | PPCODE: |
|
|
2380 | cl_GLenum value [1]; |
|
|
2381 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
|
|
2382 | EXTEND (SP, 1); |
|
|
2383 | const int i = 0; |
|
|
2384 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2385 | |
|
|
2386 | void |
|
|
2387 | gl_mipmap_level (OpenCL::Image self) |
|
|
2388 | PPCODE: |
|
|
2389 | cl_GLint value [1]; |
|
|
2390 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
|
|
2391 | EXTEND (SP, 1); |
|
|
2392 | const int i = 0; |
|
|
2393 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2394 | |
|
|
2395 | #END:gl_texture |
|
|
2396 | |
|
|
2397 | #endif |
|
|
2398 | |
1279 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
2399 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
1280 | |
2400 | |
1281 | void |
2401 | void |
1282 | DESTROY (OpenCL::Sampler self) |
2402 | DESTROY (OpenCL::Sampler self) |
1283 | CODE: |
2403 | CODE: |
… | |
… | |
1306 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2426 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1307 | EXTEND (SP, 1); |
2427 | EXTEND (SP, 1); |
1308 | const int i = 0; |
2428 | const int i = 0; |
1309 | { |
2429 | { |
1310 | NEED_SUCCESS (RetainContext, (value [i])); |
2430 | NEED_SUCCESS (RetainContext, (value [i])); |
1311 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2431 | PUSH_CLOBJ (stash_context, value [i]); |
1312 | } |
2432 | } |
1313 | |
2433 | |
1314 | void |
2434 | void |
1315 | normalized_coords (OpenCL::Sampler self) |
2435 | normalized_coords (OpenCL::Sampler self) |
1316 | PPCODE: |
2436 | PPCODE: |
1317 | cl_addressing_mode value [1]; |
2437 | cl_addressing_mode value [1]; |
1318 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2438 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1319 | EXTEND (SP, 1); |
2439 | EXTEND (SP, 1); |
1320 | const int i = 0; |
2440 | const int i = 0; |
1321 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2441 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1322 | |
2442 | |
1323 | void |
2443 | void |
1324 | addressing_mode (OpenCL::Sampler self) |
2444 | addressing_mode (OpenCL::Sampler self) |
1325 | PPCODE: |
2445 | PPCODE: |
1326 | cl_filter_mode value [1]; |
2446 | cl_filter_mode value [1]; |
1327 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2447 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1328 | EXTEND (SP, 1); |
2448 | EXTEND (SP, 1); |
1329 | const int i = 0; |
2449 | const int i = 0; |
1330 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2450 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1331 | |
2451 | |
1332 | void |
2452 | void |
1333 | filter_mode (OpenCL::Sampler self) |
2453 | filter_mode (OpenCL::Sampler self) |
1334 | PPCODE: |
2454 | PPCODE: |
1335 | cl_bool value [1]; |
2455 | cl_bool value [1]; |
… | |
… | |
1346 | DESTROY (OpenCL::Program self) |
2466 | DESTROY (OpenCL::Program self) |
1347 | CODE: |
2467 | CODE: |
1348 | clReleaseProgram (self); |
2468 | clReleaseProgram (self); |
1349 | |
2469 | |
1350 | void |
2470 | void |
1351 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2471 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2472 | ALIAS: |
|
|
2473 | build_async = 1 |
1352 | CODE: |
2474 | CODE: |
|
|
2475 | cl_uint device_count = 0; |
|
|
2476 | cl_device_id *device_list = 0; |
|
|
2477 | |
|
|
2478 | if (SvOK (devices)) |
|
|
2479 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2480 | |
|
|
2481 | void *user_data; |
|
|
2482 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2483 | |
|
|
2484 | if (ix) |
|
|
2485 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2486 | else |
|
|
2487 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2488 | |
|
|
2489 | #if CL_VERSION_1_2 |
|
|
2490 | |
|
|
2491 | void |
|
|
2492 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2493 | CODE: |
|
|
2494 | cl_uint device_count = 0; |
|
|
2495 | cl_device_id *device_list = 0; |
|
|
2496 | |
|
|
2497 | if (SvOK (devices)) |
|
|
2498 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2499 | |
|
|
2500 | cl_uint header_count = 0; |
|
|
2501 | cl_program *header_list = 0; |
|
|
2502 | const char **header_name = 0; |
|
|
2503 | |
|
|
2504 | if (SvOK (headers)) |
|
|
2505 | { |
|
|
2506 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2507 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2508 | |
|
|
2509 | HV *hv = (HV *)SvRV (devices); |
|
|
2510 | |
|
|
2511 | header_count = hv_iterinit (hv); |
|
|
2512 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2513 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2514 | |
|
|
2515 | HE *he; |
|
|
2516 | int i = 0; |
|
|
2517 | while (he = hv_iternext (hv)) |
|
|
2518 | { |
|
|
2519 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2520 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2521 | ++i; |
|
|
2522 | } |
|
|
2523 | } |
|
|
2524 | |
|
|
2525 | void *user_data; |
|
|
2526 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2527 | |
1353 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2528 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2529 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2530 | |
|
|
2531 | #endif |
1354 | |
2532 | |
1355 | void |
2533 | void |
1356 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2534 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1357 | PPCODE: |
2535 | PPCODE: |
1358 | size_t size; |
2536 | size_t size; |
… | |
… | |
1387 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2565 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
1388 | EXTEND (SP, 1); |
2566 | EXTEND (SP, 1); |
1389 | const int i = 0; |
2567 | const int i = 0; |
1390 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
2568 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
1391 | |
2569 | |
|
|
2570 | void |
|
|
2571 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2572 | PPCODE: |
|
|
2573 | cl_program_binary_type value [1]; |
|
|
2574 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2575 | EXTEND (SP, 1); |
|
|
2576 | const int i = 0; |
|
|
2577 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
2578 | |
1392 | #END:program_build |
2579 | #END:program_build |
1393 | |
2580 | |
1394 | void |
2581 | void |
1395 | kernel (OpenCL::Program program, SV *function) |
2582 | kernel (OpenCL::Program program, SV *function) |
1396 | PPCODE: |
2583 | PPCODE: |
1397 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2584 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1398 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2585 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2586 | |
|
|
2587 | void |
|
|
2588 | kernels_in_program (OpenCL::Program program) |
|
|
2589 | PPCODE: |
|
|
2590 | cl_uint num_kernels; |
|
|
2591 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2592 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2593 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2594 | |
|
|
2595 | int i; |
|
|
2596 | EXTEND (SP, num_kernels); |
|
|
2597 | for (i = 0; i < num_kernels; ++i) |
|
|
2598 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1399 | |
2599 | |
1400 | void |
2600 | void |
1401 | info (OpenCL::Program self, cl_program_info name) |
2601 | info (OpenCL::Program self, cl_program_info name) |
1402 | PPCODE: |
2602 | PPCODE: |
1403 | INFO (Program) |
2603 | INFO (Program) |
… | |
… | |
1421 | { |
2621 | { |
1422 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2622 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1423 | SvUPGRADE (sv, SVt_PV); |
2623 | SvUPGRADE (sv, SVt_PV); |
1424 | SvPOK_only (sv); |
2624 | SvPOK_only (sv); |
1425 | SvCUR_set (sv, sizes [i]); |
2625 | SvCUR_set (sv, sizes [i]); |
1426 | ptrs [i] = SvPVX (sv); |
2626 | ptrs [i] = (void *)SvPVX (sv); |
1427 | PUSHs (sv); |
2627 | PUSHs (sv); |
1428 | } |
2628 | } |
1429 | |
2629 | |
1430 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2630 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1431 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2631 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1451 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2651 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1452 | EXTEND (SP, 1); |
2652 | EXTEND (SP, 1); |
1453 | const int i = 0; |
2653 | const int i = 0; |
1454 | { |
2654 | { |
1455 | NEED_SUCCESS (RetainContext, (value [i])); |
2655 | NEED_SUCCESS (RetainContext, (value [i])); |
1456 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2656 | PUSH_CLOBJ (stash_context, value [i]); |
1457 | } |
2657 | } |
1458 | |
2658 | |
1459 | void |
2659 | void |
1460 | devices (OpenCL::Program self) |
2660 | devices (OpenCL::Program self) |
1461 | PPCODE: |
2661 | PPCODE: |
… | |
… | |
1465 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2665 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1466 | int i, n = size / sizeof (*value); |
2666 | int i, n = size / sizeof (*value); |
1467 | EXTEND (SP, n); |
2667 | EXTEND (SP, n); |
1468 | for (i = 0; i < n; ++i) |
2668 | for (i = 0; i < n; ++i) |
1469 | { |
2669 | { |
1470 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2670 | PUSH_CLOBJ (stash_device, value [i]); |
1471 | } |
2671 | } |
1472 | |
2672 | |
1473 | void |
2673 | void |
1474 | source (OpenCL::Program self) |
2674 | source (OpenCL::Program self) |
1475 | PPCODE: |
2675 | PPCODE: |
… | |
… | |
1501 | DESTROY (OpenCL::Kernel self) |
2701 | DESTROY (OpenCL::Kernel self) |
1502 | CODE: |
2702 | CODE: |
1503 | clReleaseKernel (self); |
2703 | clReleaseKernel (self); |
1504 | |
2704 | |
1505 | void |
2705 | void |
|
|
2706 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2707 | CODE: |
|
|
2708 | int i; |
|
|
2709 | for (i = 2; ; ++i) |
|
|
2710 | { |
|
|
2711 | while (*format == ' ') |
|
|
2712 | ++format; |
|
|
2713 | |
|
|
2714 | char type = *format++; |
|
|
2715 | |
|
|
2716 | if (!type) |
|
|
2717 | break; |
|
|
2718 | |
|
|
2719 | if (i >= items) |
|
|
2720 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2721 | |
|
|
2722 | SV *sv = ST (i); |
|
|
2723 | |
|
|
2724 | union |
|
|
2725 | { |
|
|
2726 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2727 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2728 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2729 | cl_mem cm; |
|
|
2730 | cl_sampler ca; |
|
|
2731 | size_t cz; |
|
|
2732 | cl_event ce; |
|
|
2733 | } arg; |
|
|
2734 | size_t size; |
|
|
2735 | int nullarg = 0; |
|
|
2736 | |
|
|
2737 | switch (type) |
|
|
2738 | { |
|
|
2739 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2740 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2741 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2742 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2743 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2744 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2745 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2746 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2747 | |
|
|
2748 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2749 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2750 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2751 | |
|
|
2752 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2753 | |
|
|
2754 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2755 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2756 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2757 | |
|
|
2758 | default: |
|
|
2759 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2760 | } |
|
|
2761 | |
|
|
2762 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2763 | if (res) |
|
|
2764 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2765 | } |
|
|
2766 | |
|
|
2767 | if (i != items) |
|
|
2768 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2769 | |
|
|
2770 | void |
1506 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2771 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1507 | CODE: |
2772 | CODE: |
1508 | clSetKernelArg (self, idx, sizeof (value), &value); |
2773 | clSetKernelArg (self, idx, sizeof (value), &value); |
1509 | |
2774 | |
1510 | void |
2775 | void |
… | |
… | |
1558 | clSetKernelArg (self, idx, sizeof (value), &value); |
2823 | clSetKernelArg (self, idx, sizeof (value), &value); |
1559 | |
2824 | |
1560 | void |
2825 | void |
1561 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2826 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1562 | CODE: |
2827 | CODE: |
|
|
2828 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2829 | |
|
|
2830 | void |
|
|
2831 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2832 | CODE: |
|
|
2833 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2834 | |
|
|
2835 | void |
|
|
2836 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2837 | CODE: |
|
|
2838 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2839 | |
|
|
2840 | void |
|
|
2841 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2842 | CODE: |
1563 | clSetKernelArg (self, idx, sizeof (value), &value); |
2843 | clSetKernelArg (self, idx, sizeof (value), &value); |
1564 | |
2844 | |
1565 | void |
2845 | void |
1566 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2846 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
1567 | CODE: |
2847 | CODE: |
1568 | clSetKernelArg (self, idx, sizeof (value), &value); |
2848 | clSetKernelArg (self, idx, size, 0); |
1569 | |
|
|
1570 | void |
|
|
1571 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
1572 | CODE: |
|
|
1573 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1574 | |
|
|
1575 | void |
|
|
1576 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1577 | CODE: |
|
|
1578 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1579 | |
|
|
1580 | void |
|
|
1581 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
1582 | CODE: |
|
|
1583 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1584 | |
2849 | |
1585 | void |
2850 | void |
1586 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
2851 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1587 | CODE: |
2852 | CODE: |
1588 | clSetKernelArg (self, idx, sizeof (value), &value); |
2853 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1624 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2889 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1625 | EXTEND (SP, 1); |
2890 | EXTEND (SP, 1); |
1626 | const int i = 0; |
2891 | const int i = 0; |
1627 | { |
2892 | { |
1628 | NEED_SUCCESS (RetainContext, (value [i])); |
2893 | NEED_SUCCESS (RetainContext, (value [i])); |
1629 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2894 | PUSH_CLOBJ (stash_context, value [i]); |
1630 | } |
2895 | } |
1631 | |
2896 | |
1632 | void |
2897 | void |
1633 | program (OpenCL::Kernel self) |
2898 | program (OpenCL::Kernel self) |
1634 | PPCODE: |
2899 | PPCODE: |
… | |
… | |
1636 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2901 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1637 | EXTEND (SP, 1); |
2902 | EXTEND (SP, 1); |
1638 | const int i = 0; |
2903 | const int i = 0; |
1639 | { |
2904 | { |
1640 | NEED_SUCCESS (RetainProgram, (value [i])); |
2905 | NEED_SUCCESS (RetainProgram, (value [i])); |
1641 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2906 | PUSH_CLOBJ (stash_program, value [i]); |
1642 | } |
2907 | } |
1643 | |
2908 | |
1644 | #END:kernel |
2909 | #END:kernel |
1645 | |
2910 | |
1646 | void |
2911 | void |
… | |
… | |
1693 | const int i = 0; |
2958 | const int i = 0; |
1694 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2959 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1695 | |
2960 | |
1696 | #END:kernel_work_group |
2961 | #END:kernel_work_group |
1697 | |
2962 | |
|
|
2963 | #if CL_VERSION_1_2 |
|
|
2964 | |
|
|
2965 | void |
|
|
2966 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2967 | PPCODE: |
|
|
2968 | size_t size; |
|
|
2969 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2970 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2971 | SvUPGRADE (sv, SVt_PV); |
|
|
2972 | SvPOK_only (sv); |
|
|
2973 | SvCUR_set (sv, size); |
|
|
2974 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2975 | XPUSHs (sv); |
|
|
2976 | |
|
|
2977 | #BEGIN:kernel_arg |
|
|
2978 | |
|
|
2979 | void |
|
|
2980 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2981 | PPCODE: |
|
|
2982 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2983 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2984 | EXTEND (SP, 1); |
|
|
2985 | const int i = 0; |
|
|
2986 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2987 | |
|
|
2988 | void |
|
|
2989 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2990 | PPCODE: |
|
|
2991 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2992 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2993 | EXTEND (SP, 1); |
|
|
2994 | const int i = 0; |
|
|
2995 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2996 | |
|
|
2997 | void |
|
|
2998 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2999 | ALIAS: |
|
|
3000 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
3001 | arg_name = CL_KERNEL_ARG_NAME |
|
|
3002 | PPCODE: |
|
|
3003 | size_t size; |
|
|
3004 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
3005 | char *value = tmpbuf (size); |
|
|
3006 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
3007 | EXTEND (SP, 1); |
|
|
3008 | const int i = 0; |
|
|
3009 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
3010 | |
|
|
3011 | void |
|
|
3012 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
3013 | PPCODE: |
|
|
3014 | cl_kernel_arg_type_qualifier value [1]; |
|
|
3015 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
3016 | EXTEND (SP, 1); |
|
|
3017 | const int i = 0; |
|
|
3018 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3019 | |
|
|
3020 | #END:kernel_arg |
|
|
3021 | |
|
|
3022 | #endif |
|
|
3023 | |
1698 | MODULE = OpenCL PACKAGE = OpenCL::Event |
3024 | MODULE = OpenCL PACKAGE = OpenCL::Event |
1699 | |
3025 | |
1700 | void |
3026 | void |
1701 | DESTROY (OpenCL::Event self) |
3027 | DESTROY (OpenCL::Event self) |
1702 | CODE: |
3028 | CODE: |
… | |
… | |
1704 | |
3030 | |
1705 | void |
3031 | void |
1706 | wait (OpenCL::Event self) |
3032 | wait (OpenCL::Event self) |
1707 | CODE: |
3033 | CODE: |
1708 | clWaitForEvents (1, &self); |
3034 | clWaitForEvents (1, &self); |
|
|
3035 | |
|
|
3036 | void |
|
|
3037 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
3038 | CODE: |
|
|
3039 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
1709 | |
3040 | |
1710 | void |
3041 | void |
1711 | info (OpenCL::Event self, cl_event_info name) |
3042 | info (OpenCL::Event self, cl_event_info name) |
1712 | PPCODE: |
3043 | PPCODE: |
1713 | INFO (Event) |
3044 | INFO (Event) |
… | |
… | |
1721 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
3052 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1722 | EXTEND (SP, 1); |
3053 | EXTEND (SP, 1); |
1723 | const int i = 0; |
3054 | const int i = 0; |
1724 | { |
3055 | { |
1725 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
3056 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1726 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
3057 | PUSH_CLOBJ (stash_queue, value [i]); |
1727 | } |
3058 | } |
1728 | |
3059 | |
1729 | void |
3060 | void |
1730 | command_type (OpenCL::Event self) |
3061 | command_type (OpenCL::Event self) |
1731 | PPCODE: |
3062 | PPCODE: |
… | |
… | |
1754 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
3085 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1755 | EXTEND (SP, 1); |
3086 | EXTEND (SP, 1); |
1756 | const int i = 0; |
3087 | const int i = 0; |
1757 | { |
3088 | { |
1758 | NEED_SUCCESS (RetainContext, (value [i])); |
3089 | NEED_SUCCESS (RetainContext, (value [i])); |
1759 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
3090 | PUSH_CLOBJ (stash_context, value [i]); |
1760 | } |
3091 | } |
1761 | |
3092 | |
1762 | #END:event |
3093 | #END:event |
1763 | |
3094 | |
1764 | void |
3095 | void |
… | |
… | |
1789 | void |
3120 | void |
1790 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
3121 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1791 | CODE: |
3122 | CODE: |
1792 | clSetUserEventStatus (self, execution_status); |
3123 | clSetUserEventStatus (self, execution_status); |
1793 | |
3124 | |
|
|
3125 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3126 | |
|
|
3127 | void |
|
|
3128 | DESTROY (SV *self) |
|
|
3129 | CODE: |
|
|
3130 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3131 | |
|
|
3132 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3133 | mapped_detach (self, mapped); |
|
|
3134 | |
|
|
3135 | clReleaseCommandQueue (mapped->queue); |
|
|
3136 | clReleaseEvent (mapped->event); |
|
|
3137 | Safefree (mapped); |
|
|
3138 | |
|
|
3139 | void |
|
|
3140 | unmap (OpenCL::Mapped self, ...) |
|
|
3141 | CODE: |
|
|
3142 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3143 | |
|
|
3144 | bool |
|
|
3145 | mapped (OpenCL::Mapped self) |
|
|
3146 | CODE: |
|
|
3147 | RETVAL = !!self->ptr; |
|
|
3148 | OUTPUT: |
|
|
3149 | RETVAL |
|
|
3150 | |
|
|
3151 | void |
|
|
3152 | wait (OpenCL::Mapped self) |
|
|
3153 | PPCODE: |
|
|
3154 | if (self->event) |
|
|
3155 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3156 | |
|
|
3157 | void |
|
|
3158 | event (OpenCL::Mapped self) |
|
|
3159 | PPCODE: |
|
|
3160 | if (!self->event) |
|
|
3161 | XSRETURN_UNDEF; |
|
|
3162 | |
|
|
3163 | clRetainEvent (self->event); |
|
|
3164 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3165 | |
|
|
3166 | size_t |
|
|
3167 | size (OpenCL::Mapped self) |
|
|
3168 | CODE: |
|
|
3169 | RETVAL = self->cb; |
|
|
3170 | OUTPUT: |
|
|
3171 | RETVAL |
|
|
3172 | |
|
|
3173 | IV |
|
|
3174 | ptr (OpenCL::Mapped self) |
|
|
3175 | CODE: |
|
|
3176 | RETVAL = PTR2IV (self->ptr); |
|
|
3177 | OUTPUT: |
|
|
3178 | RETVAL |
|
|
3179 | |
|
|
3180 | void |
|
|
3181 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3182 | CODE: |
|
|
3183 | STRLEN len; |
|
|
3184 | const char *ptr = SvPVbyte (data, len); |
|
|
3185 | |
|
|
3186 | if (offset + len > self->cb) |
|
|
3187 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3188 | |
|
|
3189 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3190 | |
|
|
3191 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3192 | |
|
|
3193 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3194 | |
|
|
3195 | IV |
|
|
3196 | row_pitch (OpenCL::Mapped self) |
|
|
3197 | ALIAS: |
|
|
3198 | slice_pitch = 1 |
|
|
3199 | CODE: |
|
|
3200 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
3201 | OUTPUT: |
|
|
3202 | RETVAL |
|
|
3203 | |
|
|
3204 | |
|
|
3205 | |