1 |
#include "EXTERN.h" |
2 |
#include "perl.h" |
3 |
#include "XSUB.h" |
4 |
|
5 |
#include <CL/opencl.h> |
6 |
|
7 |
typedef cl_platform_id OpenCL__Platform; |
8 |
typedef cl_device_id OpenCL__Device; |
9 |
typedef cl_context OpenCL__Context; |
10 |
typedef cl_command_queue OpenCL__Queue; |
11 |
typedef cl_mem OpenCL__Memory; |
12 |
typedef cl_mem OpenCL__Buffer; |
13 |
typedef cl_mem OpenCL__Image; |
14 |
typedef cl_mem OpenCL__Image2D; |
15 |
typedef cl_mem OpenCL__Image3D; |
16 |
typedef cl_mem OpenCL__Memory_ornull; |
17 |
typedef cl_mem OpenCL__Buffer_ornull; |
18 |
typedef cl_mem OpenCL__Image_ornull; |
19 |
typedef cl_mem OpenCL__Image2D_ornull; |
20 |
typedef cl_mem OpenCL__Image3D_ornull; |
21 |
typedef cl_sampler OpenCL__Sampler; |
22 |
typedef cl_program OpenCL__Program; |
23 |
typedef cl_kernel OpenCL__Kernel; |
24 |
typedef cl_event OpenCL__Event; |
25 |
typedef cl_event OpenCL__UserEvent; |
26 |
|
27 |
typedef SV *FUTURE; |
28 |
|
29 |
/*****************************************************************************/ |
30 |
|
31 |
/* up to two temporary buffers */ |
32 |
static void * |
33 |
tmpbuf (size_t size) |
34 |
{ |
35 |
static int idx; |
36 |
static void *buf [2]; |
37 |
static size_t len [2]; |
38 |
|
39 |
idx ^= 1; |
40 |
|
41 |
if (len [idx] < size) |
42 |
{ |
43 |
free (buf [idx]); |
44 |
len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
45 |
buf [idx] = malloc (len [idx]); |
46 |
} |
47 |
|
48 |
return buf [idx]; |
49 |
} |
50 |
|
51 |
/*****************************************************************************/ |
52 |
|
53 |
typedef struct |
54 |
{ |
55 |
IV iv; |
56 |
const char *name; |
57 |
#define const_iv(name) { (IV)CL_ ## name, # name }, |
58 |
} ivstr; |
59 |
|
60 |
static const char * |
61 |
iv2str (IV value, const ivstr *base, int count, const char *fallback) |
62 |
{ |
63 |
int i; |
64 |
static char strbuf [32]; |
65 |
|
66 |
for (i = count; i--; ) |
67 |
if (base [i].iv == value) |
68 |
return base [i].name; |
69 |
|
70 |
snprintf (strbuf, sizeof (strbuf), fallback, (int)value); |
71 |
|
72 |
return strbuf; |
73 |
} |
74 |
|
75 |
static const char * |
76 |
enum2str (cl_uint value) |
77 |
{ |
78 |
static const ivstr enumstr[] = { |
79 |
#include "enumstr.h" |
80 |
}; |
81 |
|
82 |
return iv2str (value, enumstr, sizeof (enumstr) / sizeof (enumstr [0]), "ENUM(0x%04x)"); |
83 |
} |
84 |
|
85 |
static const char * |
86 |
err2str (cl_int err) |
87 |
{ |
88 |
static const ivstr errstr[] = { |
89 |
#include "errstr.h" |
90 |
}; |
91 |
|
92 |
return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
93 |
} |
94 |
|
95 |
/*****************************************************************************/ |
96 |
|
97 |
static cl_int last_error; |
98 |
|
99 |
#define FAIL(name,err) \ |
100 |
croak ("cl" # name ": %s", err2str (last_error = err)); |
101 |
|
102 |
#define NEED_SUCCESS(name,args) \ |
103 |
do { \ |
104 |
cl_int res = cl ## name args; \ |
105 |
\ |
106 |
if (res) \ |
107 |
FAIL (name, res); \ |
108 |
} while (0) |
109 |
|
110 |
/*****************************************************************************/ |
111 |
|
112 |
#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
113 |
#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
114 |
|
115 |
static void * |
116 |
SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
117 |
{ |
118 |
if (SvROK (sv) && sv_derived_from (sv, pkg)) |
119 |
return (void *)SvIV (SvRV (sv)); |
120 |
|
121 |
croak ("%s: %s is not of type %s", func, svname, pkg); |
122 |
} |
123 |
|
124 |
/*****************************************************************************/ |
125 |
|
126 |
static cl_event * |
127 |
event_list (SV **items, int count) |
128 |
{ |
129 |
cl_event *list = tmpbuf (sizeof (cl_event) * count); |
130 |
|
131 |
while (count--) |
132 |
list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
133 |
|
134 |
return list; |
135 |
} |
136 |
|
137 |
#define EVENT_LIST(items,count) \ |
138 |
cl_uint event_list_count = (count); \ |
139 |
cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
140 |
|
141 |
#define INFO(class) \ |
142 |
{ \ |
143 |
size_t size; \ |
144 |
SV *sv; \ |
145 |
\ |
146 |
NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ |
147 |
sv = sv_2mortal (newSV (size)); \ |
148 |
SvUPGRADE (sv, SVt_PV); \ |
149 |
SvPOK_only (sv); \ |
150 |
SvCUR_set (sv, size); \ |
151 |
NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ |
152 |
XPUSHs (sv); \ |
153 |
} |
154 |
|
155 |
MODULE = OpenCL PACKAGE = OpenCL |
156 |
|
157 |
PROTOTYPES: ENABLE |
158 |
|
159 |
BOOT: |
160 |
{ |
161 |
HV *stash = gv_stashpv ("OpenCL", 1); |
162 |
static const ivstr *civ, const_iv[] = { |
163 |
{ sizeof (cl_char ), "SIZEOF_CHAR" }, |
164 |
{ sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
165 |
{ sizeof (cl_short ), "SIZEOF_SHORT" }, |
166 |
{ sizeof (cl_ushort), "SIZEOF_USHORT" }, |
167 |
{ sizeof (cl_int ), "SIZEOF_INT" }, |
168 |
{ sizeof (cl_uint ), "SIZEOF_UINT" }, |
169 |
{ sizeof (cl_long ), "SIZEOF_LONG" }, |
170 |
{ sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
171 |
{ sizeof (cl_half ), "SIZEOF_HALF" }, |
172 |
{ sizeof (cl_float ), "SIZEOF_FLOAT" }, |
173 |
{ sizeof (cl_double), "SIZEOF_DOUBLE" }, |
174 |
#include "constiv.h" |
175 |
}; |
176 |
for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
177 |
newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
178 |
} |
179 |
|
180 |
cl_int |
181 |
errno () |
182 |
CODE: |
183 |
errno = last_error; |
184 |
|
185 |
const char * |
186 |
err2str (cl_int err) |
187 |
|
188 |
const char * |
189 |
enum2str (cl_uint value) |
190 |
|
191 |
void |
192 |
platforms () |
193 |
PPCODE: |
194 |
{ |
195 |
cl_platform_id *list; |
196 |
cl_uint count; |
197 |
int i; |
198 |
|
199 |
NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
200 |
list = tmpbuf (sizeof (*list) * count); |
201 |
NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
202 |
|
203 |
EXTEND (SP, count); |
204 |
for (i = 0; i < count; ++i) |
205 |
PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
206 |
} |
207 |
|
208 |
void |
209 |
context_from_type (FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
210 |
PPCODE: |
211 |
{ |
212 |
cl_int res; |
213 |
cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
214 |
|
215 |
if (res) |
216 |
FAIL (CreateContextFromType, res); |
217 |
|
218 |
XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
219 |
} |
220 |
|
221 |
void |
222 |
wait_for_events (...) |
223 |
CODE: |
224 |
{ |
225 |
EVENT_LIST (0, items); |
226 |
NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
227 |
} |
228 |
|
229 |
PROTOTYPES: DISABLE |
230 |
|
231 |
MODULE = OpenCL PACKAGE = OpenCL::Platform |
232 |
|
233 |
void |
234 |
info (OpenCL::Platform this, cl_platform_info name) |
235 |
PPCODE: |
236 |
INFO (Platform) |
237 |
|
238 |
void |
239 |
devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) |
240 |
PPCODE: |
241 |
{ |
242 |
cl_device_id *list; |
243 |
cl_uint count; |
244 |
int i; |
245 |
|
246 |
NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); |
247 |
list = tmpbuf (sizeof (*list) * count); |
248 |
NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); |
249 |
|
250 |
EXTEND (SP, count); |
251 |
for (i = 0; i < count; ++i) |
252 |
PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
253 |
} |
254 |
|
255 |
void |
256 |
context_from_type (OpenCL::Platform this, FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
257 |
PPCODE: |
258 |
{ |
259 |
cl_int res; |
260 |
cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
261 |
cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
262 |
|
263 |
if (res) |
264 |
FAIL (CreateContextFromType, res); |
265 |
|
266 |
XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
267 |
} |
268 |
|
269 |
MODULE = OpenCL PACKAGE = OpenCL::Device |
270 |
|
271 |
void |
272 |
info (OpenCL::Device this, cl_device_info name) |
273 |
PPCODE: |
274 |
INFO (Device) |
275 |
|
276 |
void |
277 |
context (OpenCL::Device this, FUTURE properties = 0, FUTURE notify = 0) |
278 |
PPCODE: |
279 |
{ |
280 |
cl_int res; |
281 |
cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
282 |
|
283 |
if (res) |
284 |
FAIL (CreateContext, res); |
285 |
|
286 |
XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
287 |
} |
288 |
|
289 |
MODULE = OpenCL PACKAGE = OpenCL::Context |
290 |
|
291 |
void |
292 |
DESTROY (OpenCL::Context context) |
293 |
CODE: |
294 |
clReleaseContext (context); |
295 |
|
296 |
void |
297 |
info (OpenCL::Context this, cl_context_info name) |
298 |
PPCODE: |
299 |
INFO (Context) |
300 |
|
301 |
void |
302 |
queue (OpenCL::Context this, OpenCL::Device device, cl_command_queue_properties properties = 0) |
303 |
PPCODE: |
304 |
{ |
305 |
cl_int res; |
306 |
cl_command_queue queue = clCreateCommandQueue (this, device, properties, &res); |
307 |
|
308 |
if (res) |
309 |
FAIL (CreateCommandQueue, res); |
310 |
|
311 |
XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
312 |
} |
313 |
|
314 |
void |
315 |
user_event (OpenCL::Context this) |
316 |
PPCODE: |
317 |
{ |
318 |
cl_int res; |
319 |
cl_event ev = clCreateUserEvent (this, &res); |
320 |
|
321 |
if (res) |
322 |
FAIL (CreateUserevent, res); |
323 |
|
324 |
XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
325 |
} |
326 |
|
327 |
void |
328 |
buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) |
329 |
PPCODE: |
330 |
{ |
331 |
cl_int res; |
332 |
cl_mem mem; |
333 |
|
334 |
if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
335 |
croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
336 |
|
337 |
mem = clCreateBuffer (this, flags, len, 0, &res); |
338 |
|
339 |
if (res) |
340 |
FAIL (CreateBuffer, res); |
341 |
|
342 |
XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
343 |
} |
344 |
|
345 |
void |
346 |
buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) |
347 |
PPCODE: |
348 |
{ |
349 |
STRLEN len; |
350 |
char *ptr = SvPVbyte (data, len); |
351 |
cl_int res; |
352 |
cl_mem mem; |
353 |
|
354 |
if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
355 |
croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
356 |
|
357 |
mem = clCreateBuffer (this, flags, len, ptr, &res); |
358 |
|
359 |
if (res) |
360 |
FAIL (CreateBuffer, res); |
361 |
|
362 |
XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
363 |
} |
364 |
|
365 |
void |
366 |
image2d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, SV *data) |
367 |
PPCODE: |
368 |
{ |
369 |
STRLEN len; |
370 |
char *ptr = SvPVbyte (data, len); |
371 |
const cl_image_format format = { channel_order, channel_type }; |
372 |
cl_int res; |
373 |
cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); |
374 |
|
375 |
if (res) |
376 |
FAIL (CreateImage2D, res); |
377 |
|
378 |
XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
379 |
} |
380 |
|
381 |
void |
382 |
image3d (OpenCL::Context this, 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 slice_pitch, SV *data) |
383 |
PPCODE: |
384 |
{ |
385 |
STRLEN len; |
386 |
char *ptr = SvPVbyte (data, len); |
387 |
const cl_image_format format = { channel_order, channel_type }; |
388 |
cl_int res; |
389 |
cl_mem mem = clCreateImage3D (this, flags, &format, width, height, |
390 |
depth, len / (height * slice_pitch), slice_pitch, ptr, &res); |
391 |
|
392 |
if (res) |
393 |
FAIL (CreateImage3D, res); |
394 |
|
395 |
XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
396 |
} |
397 |
|
398 |
void |
399 |
supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) |
400 |
PPCODE: |
401 |
{ |
402 |
cl_uint count; |
403 |
cl_image_format *list; |
404 |
int i; |
405 |
|
406 |
NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); |
407 |
Newx (list, count, cl_image_format); |
408 |
NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); |
409 |
|
410 |
EXTEND (SP, count); |
411 |
for (i = 0; i < count; ++i) |
412 |
{ |
413 |
AV *av = newAV (); |
414 |
av_store (av, 1, newSVuv (list [i].image_channel_data_type)); |
415 |
av_store (av, 0, newSVuv (list [i].image_channel_order)); |
416 |
PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
417 |
} |
418 |
} |
419 |
|
420 |
void |
421 |
sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
422 |
PPCODE: |
423 |
{ |
424 |
cl_int res; |
425 |
cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); |
426 |
|
427 |
if (res) |
428 |
FAIL (CreateSampler, res); |
429 |
|
430 |
XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
431 |
} |
432 |
|
433 |
void |
434 |
program_with_source (OpenCL::Context this, SV *program) |
435 |
PPCODE: |
436 |
{ |
437 |
STRLEN len; |
438 |
size_t len2; |
439 |
const char *ptr = SvPVbyte (program, len); |
440 |
cl_int res; |
441 |
cl_program prog; |
442 |
|
443 |
len2 = len; |
444 |
prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); |
445 |
|
446 |
if (res) |
447 |
FAIL (CreateProgramWithSource, res); |
448 |
|
449 |
XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
450 |
} |
451 |
|
452 |
MODULE = OpenCL PACKAGE = OpenCL::Queue |
453 |
|
454 |
void |
455 |
DESTROY (OpenCL::Queue this) |
456 |
CODE: |
457 |
clReleaseCommandQueue (this); |
458 |
|
459 |
void |
460 |
info (OpenCL::Queue this, cl_command_queue_info name) |
461 |
PPCODE: |
462 |
INFO (CommandQueue) |
463 |
|
464 |
void |
465 |
enqueue_read_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
466 |
PPCODE: |
467 |
{ |
468 |
cl_event ev = 0; |
469 |
EVENT_LIST (6, items - 6); |
470 |
|
471 |
SvUPGRADE (data, SVt_PV); |
472 |
SvGROW (data, len); |
473 |
SvPOK_only (data); |
474 |
SvCUR_set (data, len); |
475 |
NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
476 |
|
477 |
if (ev) |
478 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
479 |
} |
480 |
|
481 |
void |
482 |
enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
483 |
PPCODE: |
484 |
{ |
485 |
cl_event ev = 0; |
486 |
STRLEN len; |
487 |
char *ptr = SvPVbyte (data, len); |
488 |
EVENT_LIST (5, items - 5); |
489 |
|
490 |
NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
491 |
|
492 |
if (ev) |
493 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
494 |
} |
495 |
|
496 |
void |
497 |
enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
498 |
PPCODE: |
499 |
{ |
500 |
cl_event ev = 0; |
501 |
EVENT_LIST (6, items - 6); |
502 |
|
503 |
NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
504 |
|
505 |
if (ev) |
506 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
507 |
} |
508 |
|
509 |
/*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueReadBufferRect.html */ |
510 |
/*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueWriteBufferRect.html */ |
511 |
|
512 |
void |
513 |
enqueue_read_image (OpenCL::Queue this, 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, ...) |
514 |
PPCODE: |
515 |
{ |
516 |
cl_event ev = 0; |
517 |
const size_t src_origin[3] = { src_x, src_y, src_z }; |
518 |
const size_t region[3] = { width, height, depth }; |
519 |
size_t len = row_pitch * slice_pitch * depth; |
520 |
EVENT_LIST (11, items - 11); |
521 |
|
522 |
SvUPGRADE (data, SVt_PV); |
523 |
SvGROW (data, len); |
524 |
SvPOK_only (data); |
525 |
SvCUR_set (data, len); |
526 |
NEED_SUCCESS (EnqueueReadImage, (this, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
527 |
|
528 |
if (ev) |
529 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
530 |
} |
531 |
|
532 |
void |
533 |
enqueue_write_image (OpenCL::Queue this, 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, SV *data, ...) |
534 |
PPCODE: |
535 |
{ |
536 |
cl_event ev = 0; |
537 |
const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
538 |
const size_t region[3] = { width, height, depth }; |
539 |
STRLEN len; |
540 |
char *ptr = SvPVbyte (data, len); |
541 |
size_t slice_pitch = len / (row_pitch * height); |
542 |
EVENT_LIST (11, items - 11); |
543 |
|
544 |
NEED_SUCCESS (EnqueueWriteImage, (this, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
545 |
|
546 |
if (ev) |
547 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
548 |
} |
549 |
|
550 |
void |
551 |
enqueue_copy_buffer_rect (OpenCL::Queue this, 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, ...) |
552 |
PPCODE: |
553 |
{ |
554 |
cl_event ev = 0; |
555 |
const size_t src_origin[3] = { src_x, src_y, src_z }; |
556 |
const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
557 |
const size_t region[3] = { width, height, depth }; |
558 |
EVENT_LIST (16, items - 16); |
559 |
|
560 |
NEED_SUCCESS (EnqueueCopyBufferRect, (this, 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)); |
561 |
|
562 |
if (ev) |
563 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
564 |
} |
565 |
|
566 |
void |
567 |
enqueue_copy_buffer_to_image (OpenCL::Queue this, 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, ...) |
568 |
PPCODE: |
569 |
{ |
570 |
cl_event ev = 0; |
571 |
const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
572 |
const size_t region[3] = { width, height, depth }; |
573 |
EVENT_LIST (10, items - 10); |
574 |
|
575 |
NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
576 |
|
577 |
if (ev) |
578 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
579 |
} |
580 |
|
581 |
void |
582 |
enqueue_copy_image (OpenCL::Queue this, OpenCL::Image 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, ...) |
583 |
PPCODE: |
584 |
{ |
585 |
cl_event ev = 0; |
586 |
const size_t src_origin[3] = { src_x, src_y, src_z }; |
587 |
const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
588 |
const size_t region[3] = { width, height, depth }; |
589 |
EVENT_LIST (12, items - 12); |
590 |
|
591 |
NEED_SUCCESS (EnqueueCopyImage, (this, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
592 |
|
593 |
if (ev) |
594 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
595 |
} |
596 |
|
597 |
void |
598 |
enqueue_copy_image_to_buffer (OpenCL::Queue this, 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, ...) |
599 |
PPCODE: |
600 |
{ |
601 |
cl_event ev = 0; |
602 |
const size_t src_origin[3] = { src_x, src_y, src_z }; |
603 |
const size_t region[3] = { width, height, depth }; |
604 |
EVENT_LIST (10, items - 10); |
605 |
|
606 |
NEED_SUCCESS (EnqueueCopyImageToBuffer, (this, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
607 |
|
608 |
if (ev) |
609 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
610 |
} |
611 |
|
612 |
void |
613 |
enqueue_task (OpenCL::Queue this, OpenCL::Kernel kernel, ...) |
614 |
PPCODE: |
615 |
{ |
616 |
cl_event ev = 0; |
617 |
EVENT_LIST (2, items - 2); |
618 |
|
619 |
NEED_SUCCESS (EnqueueTask, (this, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
620 |
|
621 |
if (ev) |
622 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
623 |
} |
624 |
|
625 |
void |
626 |
enqueue_nd_range_kernel (OpenCL::Queue this, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
627 |
PPCODE: |
628 |
{ |
629 |
cl_event ev = 0; |
630 |
size_t *gwo = 0, *gws, *lws = 0; |
631 |
int gws_len; |
632 |
size_t *lists; |
633 |
int i; |
634 |
EVENT_LIST (5, items - 5); |
635 |
|
636 |
if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
637 |
croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
638 |
|
639 |
gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
640 |
|
641 |
lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
642 |
|
643 |
gws = lists + gws_len * 0; |
644 |
for (i = 0; i < gws_len; ++i) |
645 |
gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
646 |
|
647 |
if (SvOK (global_work_offset)) |
648 |
{ |
649 |
if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
650 |
croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
651 |
|
652 |
if (AvFILLp (SvRV (global_work_size)) + 1 != gws_len) |
653 |
croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array of same size as global_work_size"); |
654 |
|
655 |
gwo = lists + gws_len * 1; |
656 |
for (i = 0; i < gws_len; ++i) |
657 |
gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
658 |
} |
659 |
|
660 |
if (SvOK (local_work_size)) |
661 |
{ |
662 |
if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
663 |
croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
664 |
|
665 |
if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
666 |
croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
667 |
|
668 |
lws = lists + gws_len * 2; |
669 |
for (i = 0; i < gws_len; ++i) |
670 |
lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
671 |
} |
672 |
|
673 |
NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
674 |
|
675 |
if (ev) |
676 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
677 |
} |
678 |
|
679 |
void |
680 |
enqueue_marker (OpenCL::Queue this) |
681 |
PPCODE: |
682 |
{ |
683 |
cl_event ev; |
684 |
NEED_SUCCESS (EnqueueMarker, (this, &ev)); |
685 |
XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
686 |
} |
687 |
|
688 |
void |
689 |
enqueue_wait_for_events (OpenCL::Queue this, ...) |
690 |
CODE: |
691 |
{ |
692 |
EVENT_LIST (1, items - 1); |
693 |
NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); |
694 |
} |
695 |
|
696 |
void |
697 |
enqueue_barrier (OpenCL::Queue this) |
698 |
CODE: |
699 |
NEED_SUCCESS (EnqueueBarrier, (this)); |
700 |
|
701 |
void |
702 |
flush (OpenCL::Queue this) |
703 |
CODE: |
704 |
NEED_SUCCESS (Flush, (this)); |
705 |
|
706 |
void |
707 |
finish (OpenCL::Queue this) |
708 |
CODE: |
709 |
NEED_SUCCESS (Finish, (this)); |
710 |
|
711 |
MODULE = OpenCL PACKAGE = OpenCL::Memory |
712 |
|
713 |
void |
714 |
DESTROY (OpenCL::Memory this) |
715 |
CODE: |
716 |
clReleaseMemObject (this); |
717 |
|
718 |
void |
719 |
info (OpenCL::Memory this, cl_mem_info name) |
720 |
PPCODE: |
721 |
INFO (MemObject) |
722 |
|
723 |
MODULE = OpenCL PACKAGE = OpenCL::Sampler |
724 |
|
725 |
void |
726 |
DESTROY (OpenCL::Sampler this) |
727 |
CODE: |
728 |
clReleaseSampler (this); |
729 |
|
730 |
void |
731 |
info (OpenCL::Sampler this, cl_sampler_info name) |
732 |
PPCODE: |
733 |
INFO (Sampler) |
734 |
|
735 |
MODULE = OpenCL PACKAGE = OpenCL::Program |
736 |
|
737 |
void |
738 |
DESTROY (OpenCL::Program this) |
739 |
CODE: |
740 |
clReleaseProgram (this); |
741 |
|
742 |
void |
743 |
info (OpenCL::Program this, cl_program_info name) |
744 |
PPCODE: |
745 |
INFO (Program) |
746 |
|
747 |
void |
748 |
build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) |
749 |
CODE: |
750 |
NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
751 |
|
752 |
void |
753 |
build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) |
754 |
PPCODE: |
755 |
{ |
756 |
size_t size; |
757 |
SV *sv; |
758 |
|
759 |
NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); |
760 |
sv = sv_2mortal (newSV (size)); |
761 |
SvUPGRADE (sv, SVt_PV); |
762 |
SvPOK_only (sv); |
763 |
SvCUR_set (sv, size); |
764 |
NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); |
765 |
XPUSHs (sv); |
766 |
} |
767 |
|
768 |
void |
769 |
kernel (OpenCL::Program program, SV *function) |
770 |
PPCODE: |
771 |
{ |
772 |
cl_int res; |
773 |
cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); |
774 |
|
775 |
if (res) |
776 |
FAIL (CreateKernel, res); |
777 |
|
778 |
XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
779 |
} |
780 |
|
781 |
MODULE = OpenCL PACKAGE = OpenCL::Kernel |
782 |
|
783 |
void |
784 |
DESTROY (OpenCL::Kernel this) |
785 |
CODE: |
786 |
clReleaseKernel (this); |
787 |
|
788 |
void |
789 |
info (OpenCL::Kernel this, cl_kernel_info name) |
790 |
PPCODE: |
791 |
INFO (Kernel) |
792 |
|
793 |
void |
794 |
set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) |
795 |
CODE: |
796 |
clSetKernelArg (this, idx, sizeof (value), &value); |
797 |
|
798 |
void |
799 |
set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value) |
800 |
CODE: |
801 |
clSetKernelArg (this, idx, sizeof (value), &value); |
802 |
|
803 |
void |
804 |
set_short (OpenCL::Kernel this, cl_uint idx, cl_short value) |
805 |
CODE: |
806 |
clSetKernelArg (this, idx, sizeof (value), &value); |
807 |
|
808 |
void |
809 |
set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) |
810 |
CODE: |
811 |
clSetKernelArg (this, idx, sizeof (value), &value); |
812 |
|
813 |
void |
814 |
set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) |
815 |
CODE: |
816 |
clSetKernelArg (this, idx, sizeof (value), &value); |
817 |
|
818 |
void |
819 |
set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value) |
820 |
CODE: |
821 |
clSetKernelArg (this, idx, sizeof (value), &value); |
822 |
|
823 |
void |
824 |
set_long (OpenCL::Kernel this, cl_uint idx, cl_long value) |
825 |
CODE: |
826 |
clSetKernelArg (this, idx, sizeof (value), &value); |
827 |
|
828 |
void |
829 |
set_ulong (OpenCL::Kernel this, cl_uint idx, cl_ulong value) |
830 |
CODE: |
831 |
clSetKernelArg (this, idx, sizeof (value), &value); |
832 |
|
833 |
void |
834 |
set_half (OpenCL::Kernel this, cl_uint idx, cl_half value) |
835 |
CODE: |
836 |
clSetKernelArg (this, idx, sizeof (value), &value); |
837 |
|
838 |
void |
839 |
set_float (OpenCL::Kernel this, cl_uint idx, cl_float value) |
840 |
CODE: |
841 |
clSetKernelArg (this, idx, sizeof (value), &value); |
842 |
|
843 |
void |
844 |
set_double (OpenCL::Kernel this, cl_uint idx, cl_double value) |
845 |
CODE: |
846 |
clSetKernelArg (this, idx, sizeof (value), &value); |
847 |
|
848 |
void |
849 |
set_memory (OpenCL::Kernel this, cl_uint idx, OpenCL::Memory_ornull value) |
850 |
CODE: |
851 |
clSetKernelArg (this, idx, sizeof (value), &value); |
852 |
|
853 |
void |
854 |
set_buffer (OpenCL::Kernel this, cl_uint idx, OpenCL::Buffer_ornull value) |
855 |
CODE: |
856 |
clSetKernelArg (this, idx, sizeof (value), &value); |
857 |
|
858 |
void |
859 |
set_image2d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image2D_ornull value) |
860 |
CODE: |
861 |
clSetKernelArg (this, idx, sizeof (value), &value); |
862 |
|
863 |
void |
864 |
set_image3d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image3D_ornull value) |
865 |
CODE: |
866 |
clSetKernelArg (this, idx, sizeof (value), &value); |
867 |
|
868 |
void |
869 |
set_sampler (OpenCL::Kernel this, cl_uint idx, OpenCL::Sampler value) |
870 |
CODE: |
871 |
clSetKernelArg (this, idx, sizeof (value), &value); |
872 |
|
873 |
void |
874 |
set_event (OpenCL::Kernel this, cl_uint idx, OpenCL::Event value) |
875 |
CODE: |
876 |
clSetKernelArg (this, idx, sizeof (value), &value); |
877 |
|
878 |
MODULE = OpenCL PACKAGE = OpenCL::Event |
879 |
|
880 |
void |
881 |
DESTROY (OpenCL::Event this) |
882 |
CODE: |
883 |
clReleaseEvent (this); |
884 |
|
885 |
void |
886 |
info (OpenCL::Event this, cl_event_info name) |
887 |
PPCODE: |
888 |
INFO (Event) |
889 |
|
890 |
void |
891 |
wait (OpenCL::Event this) |
892 |
CODE: |
893 |
clWaitForEvents (1, &this); |
894 |
|
895 |
MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
896 |
|
897 |
void |
898 |
set_status (OpenCL::UserEvent this, cl_int execution_status) |
899 |
CODE: |
900 |
clSetUserEventStatus (this, execution_status); |
901 |
|