ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/OpenCL/OpenCL.xs
(Generate patch)

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.23 by root, Mon Apr 16 09:43:30 2012 UTC vs.
Revision 1.42 by root, Sat Apr 21 20:48:03 2012 UTC

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
5#ifdef I_DLFCN
6 #include <dlfcn.h>
7#endif
8
9// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
10// but then fail to define the api functions unless you ALSO define
11// this. This breaks 100% of the opencl 1.1 apps, for what reason?
12// after all, the functions are deprecated, not removed.
13// in addition, you cannot test for this in any future-proof way.
14// each time a new opencl version comes out, you need to make a new
15// release.
16#define CL_USE_DEPRECATED_OPENCL_1_1_APIS
17#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
4 18
5#ifdef __APPLE__ 19#ifdef __APPLE__
6 #include <OpenCL/opencl.h> 20 #include <OpenCL/opencl.h>
7#else 21#else
8 #include <CL/opencl.h> 22 #include <CL/opencl.h>
23#endif
24
25#if 0 // testing
26#undef CL_USE_DEPRECATED_OPENCL_1_1_APIS
27#undef CL_VERSION_1_2
9#endif 28#endif
10 29
11typedef cl_platform_id OpenCL__Platform; 30typedef cl_platform_id OpenCL__Platform;
12typedef cl_device_id OpenCL__Device; 31typedef cl_device_id OpenCL__Device;
13typedef cl_context OpenCL__Context; 32typedef cl_context OpenCL__Context;
31 50
32typedef SV *FUTURE; 51typedef SV *FUTURE;
33 52
34/*****************************************************************************/ 53/*****************************************************************************/
35 54
55// name must include a leading underscore
56// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
57// for perl. doh.
58static void *
59glsym (const char *name)
60{
61 void *fun = 0;
62
63 #if defined I_DLFCN && defined RTLD_DEFAULT
64 fun = dlsym (RTLD_DEFAULT, name + 1);
65 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
66
67 if (!fun)
68 {
69 static void *libgl;
70 static const char *glso[] = {
71 "libGL.so.1",
72 "libGL.so.3",
73 "libGL.so.4.0",
74 "libGL.so",
75 "/usr/lib/libGL.so",
76 "/usr/X11R6/lib/libGL.1.dylib"
77 };
78 int i;
79
80 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
81 {
82 libgl = dlopen (glso [i], RTLD_LAZY);
83 if (libgl)
84 break;
85 }
86
87 if (libgl)
88 {
89 fun = dlsym (libgl, name + 1);
90 if (!fun) fun = dlsym (libgl, name);
91 }
92 }
93 #endif
94
95 return fun;
96}
97
98/*****************************************************************************/
99
36/* up to two temporary buffers */ 100/* up to two temporary buffers */
37static void * 101static void *
38tmpbuf (size_t size) 102tmpbuf (size_t size)
39{ 103{
104 enum { buffers = 3 };
40 static int idx; 105 static int idx;
41 static void *buf [2]; 106 static void *buf [buffers];
42 static size_t len [2]; 107 static size_t len [buffers];
43 108
44 idx ^= 1; 109 idx = (idx + 1) % buffers;
45 110
46 if (len [idx] < size) 111 if (len [idx] < size)
47 { 112 {
48 free (buf [idx]); 113 free (buf [idx]);
49 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 114 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
131 croak ("%s: %s is not of type %s", func, svname, pkg); 196 croak ("%s: %s is not of type %s", func, svname, pkg);
132} 197}
133 198
134/*****************************************************************************/ 199/*****************************************************************************/
135 200
201static cl_context_properties *
202SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
203{
204 if (!sv || !SvOK (sv))
205 if (extra)
206 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
207 else
208 return 0;
209
210 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
211 {
212 AV *av = (AV *)SvRV (sv);
213 int i, len = av_len (av) + 1;
214 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
215 cl_context_properties *l = p;
216
217 if (len & 1)
218 croak ("%s: %s is not a property list (must be even number of elements)", func, svname);
219
220 while (extracount--)
221 *l++ = *extra++;
222
223 for (i = 0; i < len; i += 2)
224 {
225 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
226 SV *p_sv = *av_fetch (av, i + 1, 0);
227 cl_context_properties v = SvIV (p_sv); // code below can override
228
229 switch (t)
230 {
231 case CL_GLX_DISPLAY_KHR:
232 if (!SvOK (p_sv))
233 {
234 void *func = glsym ("_glXGetCurrentDisplay");
235 if (func)
236 v = (cl_context_properties)((void *(*)(void))func)();
237 }
238 break;
239
240 case CL_GL_CONTEXT_KHR:
241 if (!SvOK (p_sv))
242 {
243 void *func = glsym ("_glXGetCurrentContext");
244 if (func)
245 v = (cl_context_properties)((void *(*)(void))func)();
246 }
247 break;
248
249 default:
250 /* unknown property, treat as int */
251 break;
252 }
253
254 *l++ = t;
255 *l++ = v;
256 }
257
258 *l = 0;
259
260 return p;
261 }
262
263 croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname);
264}
265
266/*****************************************************************************/
267
136static size_t 268static size_t
137img_row_pitch (cl_mem img) 269img_row_pitch (cl_mem img)
138{ 270{
139 size_t res; 271 size_t res;
140 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 272 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
141 return res; 273 return res;
142} 274}
143 275
144static cl_event * 276static cl_event *
145event_list (SV **items, int count) 277event_list (SV **items, cl_uint *rcount)
146{ 278{
279 cl_uint count = *rcount;
280
147 if (!count) 281 if (!count)
148 return 0; 282 return 0;
149 283
150 cl_event *list = tmpbuf (sizeof (cl_event) * count); 284 cl_event *list = tmpbuf (sizeof (cl_event) * count);
285 int i = 0;
151 286
152 while (count--) 287 do
288 {
289 --count;
290 if (SvOK (items [count]))
153 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 291 list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
292 }
293 while (count);
154 294
295 *rcount = i;
296
155 return list; 297 return i ? list : 0;
156} 298}
157 299
158#define EVENT_LIST(items,count) \ 300#define EVENT_LIST(items,count) \
159 cl_uint event_list_count = (count); \ 301 cl_uint event_list_count = (count); \
160 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 302 cl_event *event_list_ptr = event_list (&ST (items), &event_list_count)
161 303
162#define INFO(class) \ 304#define INFO(class) \
163{ \ 305{ \
164 size_t size; \ 306 size_t size; \
165 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 307 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
175 317
176PROTOTYPES: ENABLE 318PROTOTYPES: ENABLE
177 319
178BOOT: 320BOOT:
179{ 321{
180 HV *stash = gv_stashpv ("OpenCL", 1); 322 HV *stash = gv_stashpv ("OpenCL", 1);
181 static const ivstr *civ, const_iv[] = { 323 static const ivstr *civ, const_iv[] = {
182 { sizeof (cl_char ), "SIZEOF_CHAR" }, 324 { sizeof (cl_char ), "SIZEOF_CHAR" },
183 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 325 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
184 { sizeof (cl_short ), "SIZEOF_SHORT" }, 326 { sizeof (cl_short ), "SIZEOF_SHORT" },
185 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 327 { sizeof (cl_ushort), "SIZEOF_USHORT" },
186 { sizeof (cl_int ), "SIZEOF_INT" }, 328 { sizeof (cl_int ), "SIZEOF_INT" },
187 { sizeof (cl_uint ), "SIZEOF_UINT" }, 329 { sizeof (cl_uint ), "SIZEOF_UINT" },
188 { sizeof (cl_long ), "SIZEOF_LONG" }, 330 { sizeof (cl_long ), "SIZEOF_LONG" },
189 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 331 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
190 { sizeof (cl_half ), "SIZEOF_HALF" }, 332 { sizeof (cl_half ), "SIZEOF_HALF" },
191 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 333 { sizeof (cl_float ), "SIZEOF_FLOAT" },
192 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 334 { sizeof (cl_double), "SIZEOF_DOUBLE" },
193#include "constiv.h" 335#include "constiv.h"
194 }; 336 };
195 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 337 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
196 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 338 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
197} 339}
198 340
199cl_int 341cl_int
200errno () 342errno ()
201 CODE: 343 CODE:
202 errno = res; 344 RETVAL = res;
345 OUTPUT:
346 RETVAL
203 347
204const char * 348const char *
205err2str (cl_int err) 349err2str (cl_int err)
206 350
207const char * 351const char *
221 EXTEND (SP, count); 365 EXTEND (SP, count);
222 for (i = 0; i < count; ++i) 366 for (i = 0; i < count; ++i)
223 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 367 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i]));
224 368
225void 369void
226context_from_type (FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 370context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
227 PPCODE: 371 PPCODE:
228 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (0, type, 0, 0, &res)); 372 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
229 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 373 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
230 374
375#if 0
376
231void 377void
232context (FUTURE properties, FUTURE devices, FUTURE notify = 0) 378context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
233 PPCODE: 379 PPCODE:
234 /* der Gipfel der Kunst */ 380 /* der Gipfel der Kunst */
381
382#endif
235 383
236void 384void
237wait_for_events (...) 385wait_for_events (...)
238 CODE: 386 CODE:
239 EVENT_LIST (0, items); 387 EVENT_LIST (0, items);
283 EXTEND (SP, count); 431 EXTEND (SP, count);
284 for (i = 0; i < count; ++i) 432 for (i = 0; i < count; ++i)
285 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 433 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i]));
286 434
287void 435void
288context (OpenCL::Platform self, FUTURE properties, SV *devices, FUTURE notify = 0) 436context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0)
289 PPCODE: 437 PPCODE:
290 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 438 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
291 croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); 439 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
292 440
293 AV *av = (AV *)SvRV (devices); 441 AV *av = (AV *)SvRV (devices);
294 cl_uint num_devices = av_len (av) + 1; 442 cl_uint num_devices = av_len (av) + 1;
295 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 443 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
296 int i; 444 int i;
297 445
298 for (i = num_devices; i--; ) 446 for (i = num_devices; i--; )
299 device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); 447 device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
300 448
301 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (0, num_devices, device_list, 0, 0, &res)); 449 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res));
302 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 450 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
303 451
304void 452void
305context_from_type (OpenCL::Platform self, FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 453context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
306 PPCODE: 454 PPCODE:
307 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self, 0 }; 455 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
456 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2);
308 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); 457 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res));
309 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 458 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
310 459
311MODULE = OpenCL PACKAGE = OpenCL::Device 460MODULE = OpenCL PACKAGE = OpenCL::Device
312 461
353 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 502 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
354 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 503 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
355 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 504 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
356 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 505 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
357 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 506 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
358 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 507 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
359 PPCODE: 508 PPCODE:
360 cl_uint value [1]; 509 cl_uint value [1];
361 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 510 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
362 EXTEND (SP, 1); 511 EXTEND (SP, 1);
363 const int i = 0; 512 const int i = 0;
565 714
566void 715void
567buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) 716buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
568 PPCODE: 717 PPCODE:
569 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 718 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
570 croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); 719 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
571 720
572 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); 721 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
573 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 722 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
574 723
575void 724void
576buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) 725buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
577 PPCODE: 726 PPCODE:
578 STRLEN len; 727 STRLEN len;
579 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 728 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
580 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 729 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
581 croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); 730 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
582 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); 731 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
583 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 732 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
584 733
734#if CL_VERSION_1_2
735
585void 736void
586image2d (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) 737image (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, 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)
587 PPCODE: 738 PPCODE:
588 STRLEN len; 739 STRLEN len;
589 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 740 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
590 const cl_image_format format = { channel_order, channel_type }; 741 const cl_image_format format = { channel_order, channel_type };
742 const cl_image_desc desc = {
743 type,
744 width, height, depth,
745 array_size, row_pitch, slice_pitch,
746 num_mip_level, num_samples,
747 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
748 };
591 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); 749 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
750 char *klass = "OpenCL::Image";
751 switch (type)
752 {
753 case CL_MEM_OBJECT_IMAGE1D_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
754 case CL_MEM_OBJECT_IMAGE1D: klass = "OpenCL::Image1D"; break;
755 case CL_MEM_OBJECT_IMAGE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
756 case CL_MEM_OBJECT_IMAGE2D: klass = "OpenCL::Image2D"; break;
757 case CL_MEM_OBJECT_IMAGE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
758 case CL_MEM_OBJECT_IMAGE3D: klass = "OpenCL::Image3D"; break;
759 }
592 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 760 XPUSH_NEW_OBJ (klass, mem);
593 761
762#endif
763
764#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
765
594void 766void
595image3d (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) 767image2d (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)
596 PPCODE: 768 PPCODE:
597 STRLEN len; 769 STRLEN len;
598 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 770 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
599 const cl_image_format format = { channel_order, channel_type }; 771 const cl_image_format format = { channel_order, channel_type };
772#if CL_VERSION_1_2
773 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
774 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
775#else
776 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
777#endif
778 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
779
780void
781image3d (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)
782 PPCODE:
783 STRLEN len;
784 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
785 const cl_image_format format = { channel_order, channel_type };
786#if CL_VERSION_1_2
787 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
788 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
789#else
600 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 790 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
791#endif
601 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 792 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
793
794#endif
795
796#if cl_apple_gl_sharing || cl_khr_gl_sharing
797
798void
799gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
800 PPCODE:
801 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
802 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
803
804void
805gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
806 PPCODE:
807 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
808 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
809
810#if CL_VERSION_1_2
811
812void
813gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
814 PPCODE:
815 char *klass = "OpenCL::Memory";
816 cl_gl_object_type type;
817 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
818 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0));
819 switch (type)
820 {
821 case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
822 case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break;
823 case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
824 case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break;
825 case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
826 case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break;
827 }
828 XPUSH_NEW_OBJ (klass, mem);
829
830#endif
831
832#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
833
834void
835gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
836 PPCODE:
837 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
838 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
839
840void
841gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
842 PPCODE:
843 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
844 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
845
846#endif
847
848#endif
602 849
603void 850void
604supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) 851supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
605 PPCODE: 852 PPCODE:
606{ 853{
709 cl_event ev = 0; 956 cl_event ev = 0;
710 STRLEN len; 957 STRLEN len;
711 char *ptr = SvPVbyte (data, len); 958 char *ptr = SvPVbyte (data, len);
712 EVENT_LIST (5, items - 5); 959 EVENT_LIST (5, items - 5);
713 960
714 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 961 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
715 962
716 if (ev) 963 if (ev)
717 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 964 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
718 965
719void 966void
785 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1032 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
786 1033
787 if (len < min_len) 1034 if (len < min_len)
788 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1035 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
789 1036
790 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)); 1037 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));
791 1038
792 if (ev) 1039 if (ev)
793 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1040 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
794 1041
795void 1042void
850 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1097 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
851 1098
852 if (len < min_len) 1099 if (len < min_len)
853 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1100 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
854 1101
855 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)); 1102 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));
856 1103
857 if (ev) 1104 if (ev)
858 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1105 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
859 1106
860void 1107void
942 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1189 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
943 } 1190 }
944 1191
945 if (SvOK (local_work_size)) 1192 if (SvOK (local_work_size))
946 { 1193 {
947 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1194 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
948 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); 1195 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference");
949 1196
950 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1197 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
951 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); 1198 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
952 1199
958 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1205 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
959 1206
960 if (ev) 1207 if (ev)
961 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1208 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
962 1209
1210#if cl_apple_gl_sharing || cl_khr_gl_sharing
1211
963void 1212void
964enqueue_marker (OpenCL::Queue self) 1213enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1214 ALIAS:
1215 enqueue_release_gl_objects = 1
965 PPCODE: 1216 PPCODE:
1217 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1218 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1219
966 cl_event ev; 1220 cl_event ev = 0;
967 NEED_SUCCESS (EnqueueMarker, (self, &ev)); 1221 EVENT_LIST (2, items - 2);
1222 AV *av = (AV *)SvRV (objects);
1223 cl_uint num_objects = av_len (av) + 1;
1224 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1225 int i;
1226
1227 for (i = num_objects; i--; )
1228 object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1229
1230 if (ix)
1231 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1232 else
1233 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1234
1235 if (ev)
968 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1236 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1237
1238#endif
1239
1240#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
969 1241
970void 1242void
971enqueue_wait_for_events (OpenCL::Queue self, ...) 1243enqueue_wait_for_events (OpenCL::Queue self, ...)
972 CODE: 1244 CODE:
973 EVENT_LIST (1, items - 1); 1245 EVENT_LIST (1, items - 1);
974 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 1246 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
975 1247
1248#endif
1249
976void 1250void
1251enqueue_marker (OpenCL::Queue self, ...)
1252 PPCODE:
1253 cl_event ev = 0;
1254 EVENT_LIST (1, items - 1);
1255#if CL_VERSION_1_2
1256 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1257#else
1258 if (event_list_count)
1259 croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1260 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1261#endif
1262 if (ev)
1263 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1264
1265void
977enqueue_barrier (OpenCL::Queue self) 1266enqueue_barrier (OpenCL::Queue self, ...)
978 CODE: 1267 PPCODE:
1268 cl_event ev = 0;
1269 EVENT_LIST (1, items - 1);
1270#if CL_VERSION_1_2
1271 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev));
1272#else
1273 if (event_list_count)
1274 croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1275 if (GIMME_V != G_VOID)
1276 croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2");
979 NEED_SUCCESS (EnqueueBarrier, (self)); 1277 NEED_SUCCESS (EnqueueBarrier, (self));
1278#endif
1279 if (ev)
1280 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
980 1281
981void 1282void
982flush (OpenCL::Queue self) 1283flush (OpenCL::Queue self)
983 CODE: 1284 CODE:
984 NEED_SUCCESS (Flush, (self)); 1285 NEED_SUCCESS (Flush, (self));
1126 NEED_SUCCESS (RetainMemObject, (value [i])); 1427 NEED_SUCCESS (RetainMemObject, (value [i]));
1127 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); 1428 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i]));
1128 } 1429 }
1129 1430
1130#END:mem 1431#END:mem
1432
1433#if cl_apple_gl_sharing || cl_khr_gl_sharing
1434
1435void
1436gl_object_info (OpenCL::Memory self)
1437 PPCODE:
1438 cl_gl_object_type type;
1439 cl_GLuint name;
1440 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1441 EXTEND (SP, 2);
1442 PUSHs (sv_2mortal (newSVuv (type)));
1443 PUSHs (sv_2mortal (newSVuv (name)));
1444
1445#endif
1131 1446
1132MODULE = OpenCL PACKAGE = OpenCL::BufferObj 1447MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1133 1448
1134void 1449void
1135sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) 1450sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1167 const int i = 0; 1482 const int i = 0;
1168 PUSHs (sv_2mortal (newSVuv (value [i]))); 1483 PUSHs (sv_2mortal (newSVuv (value [i])));
1169 1484
1170#END:image 1485#END:image
1171 1486
1487#if cl_apple_gl_sharing || cl_khr_gl_sharing
1488
1489#BEGIN:gl_texture
1490
1491void
1492target (OpenCL::Image self)
1493 PPCODE:
1494 cl_GLenum value [1];
1495 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1496 EXTEND (SP, 1);
1497 const int i = 0;
1498 PUSHs (sv_2mortal (newSVuv (value [i])));
1499
1500void
1501gl_mipmap_level (OpenCL::Image self)
1502 PPCODE:
1503 cl_GLint value [1];
1504 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1505 EXTEND (SP, 1);
1506 const int i = 0;
1507 PUSHs (sv_2mortal (newSViv (value [i])));
1508
1509#END:gl_texture
1510
1511#endif
1512
1172MODULE = OpenCL PACKAGE = OpenCL::Sampler 1513MODULE = OpenCL PACKAGE = OpenCL::Sampler
1173 1514
1174void 1515void
1175DESTROY (OpenCL::Sampler self) 1516DESTROY (OpenCL::Sampler self)
1176 CODE: 1517 CODE:
1314 { 1655 {
1315 SV *sv = sv_2mortal (newSV (sizes [i])); 1656 SV *sv = sv_2mortal (newSV (sizes [i]));
1316 SvUPGRADE (sv, SVt_PV); 1657 SvUPGRADE (sv, SVt_PV);
1317 SvPOK_only (sv); 1658 SvPOK_only (sv);
1318 SvCUR_set (sv, sizes [i]); 1659 SvCUR_set (sv, sizes [i]);
1319 ptrs [i] = SvPVX (sv); 1660 ptrs [i] = (void *)SvPVX (sv);
1320 PUSHs (sv); 1661 PUSHs (sv);
1321 } 1662 }
1322 1663
1323 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 1664 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1324 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1665 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1474set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1815set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1475 CODE: 1816 CODE:
1476 clSetKernelArg (self, idx, sizeof (value), &value); 1817 clSetKernelArg (self, idx, sizeof (value), &value);
1477 1818
1478void 1819void
1820set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1821 CODE:
1822 clSetKernelArg (self, idx, size, 0);
1823
1824void
1479set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1825set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1480 CODE: 1826 CODE:
1481 clSetKernelArg (self, idx, sizeof (value), &value); 1827 clSetKernelArg (self, idx, sizeof (value), &value);
1482 1828
1483void 1829void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines