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