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 1.45 by root, Tue Apr 24 12:19:01 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_2_APIS /* just guessing, you stupid idiots */
17
18#ifndef PREFER_1_1
19 #define PREFER_1_1 1
20#endif
21
22#if PREFER_1_1
23 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
24#endif
4 25
5#ifdef __APPLE__ 26#ifdef __APPLE__
6 #include <OpenCL/opencl.h> 27 #include <OpenCL/opencl.h>
7#else 28#else
8 #include <CL/opencl.h> 29 #include <CL/opencl.h>
30#endif
31
32#ifndef CL_VERSION_1_2
33 #undef PREFER_1_1
34 #define PREFER_1_1 1
9#endif 35#endif
10 36
11typedef cl_platform_id OpenCL__Platform; 37typedef cl_platform_id OpenCL__Platform;
12typedef cl_device_id OpenCL__Device; 38typedef cl_device_id OpenCL__Device;
13typedef cl_context OpenCL__Context; 39typedef cl_context OpenCL__Context;
31 57
32typedef SV *FUTURE; 58typedef SV *FUTURE;
33 59
34/*****************************************************************************/ 60/*****************************************************************************/
35 61
62// name must include a leading underscore
63// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
64// for perl. doh.
65static void *
66glsym (const char *name)
67{
68 void *fun = 0;
69
70 #if defined I_DLFCN && defined RTLD_DEFAULT
71 fun = dlsym (RTLD_DEFAULT, name + 1);
72 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
73
74 if (!fun)
75 {
76 static void *libgl;
77 static const char *glso[] = {
78 "libGL.so.1",
79 "libGL.so.3",
80 "libGL.so.4.0",
81 "libGL.so",
82 "/usr/lib/libGL.so",
83 "/usr/X11R6/lib/libGL.1.dylib"
84 };
85 int i;
86
87 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
88 {
89 libgl = dlopen (glso [i], RTLD_LAZY);
90 if (libgl)
91 break;
92 }
93
94 if (libgl)
95 {
96 fun = dlsym (libgl, name + 1);
97 if (!fun) fun = dlsym (libgl, name);
98 }
99 }
100 #endif
101
102 return fun;
103}
104
105/*****************************************************************************/
106
36/* up to two temporary buffers */ 107/* up to two temporary buffers */
37static void * 108static void *
38tmpbuf (size_t size) 109tmpbuf (size_t size)
39{ 110{
111 enum { buffers = 3 };
40 static int idx; 112 static int idx;
41 static void *buf [2]; 113 static void *buf [buffers];
42 static size_t len [2]; 114 static size_t len [buffers];
43 115
44 idx ^= 1; 116 idx = (idx + 1) % buffers;
45 117
46 if (len [idx] < size) 118 if (len [idx] < size)
47 { 119 {
48 free (buf [idx]); 120 free (buf [idx]);
49 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 121 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
131 croak ("%s: %s is not of type %s", func, svname, pkg); 203 croak ("%s: %s is not of type %s", func, svname, pkg);
132} 204}
133 205
134/*****************************************************************************/ 206/*****************************************************************************/
135 207
208static cl_context_properties *
209SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
210{
211 if (!sv || !SvOK (sv))
212 if (extra)
213 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
214 else
215 return 0;
216
217 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
218 {
219 AV *av = (AV *)SvRV (sv);
220 int i, len = av_len (av) + 1;
221 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
222 cl_context_properties *l = p;
223
224 if (len & 1)
225 croak ("%s: %s is not a property list (must be even number of elements)", func, svname);
226
227 while (extracount--)
228 *l++ = *extra++;
229
230 for (i = 0; i < len; i += 2)
231 {
232 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
233 SV *p_sv = *av_fetch (av, i + 1, 0);
234 cl_context_properties v = SvIV (p_sv); // code below can override
235
236 switch (t)
237 {
238 case CL_GLX_DISPLAY_KHR:
239 if (!SvOK (p_sv))
240 {
241 void *func = glsym ("_glXGetCurrentDisplay");
242 if (func)
243 v = (cl_context_properties)((void *(*)(void))func)();
244 }
245 break;
246
247 case CL_GL_CONTEXT_KHR:
248 if (!SvOK (p_sv))
249 {
250 void *func = glsym ("_glXGetCurrentContext");
251 if (func)
252 v = (cl_context_properties)((void *(*)(void))func)();
253 }
254 break;
255
256 default:
257 /* unknown property, treat as int */
258 break;
259 }
260
261 *l++ = t;
262 *l++ = v;
263 }
264
265 *l = 0;
266
267 return p;
268 }
269
270 croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname);
271}
272
273/*****************************************************************************/
274
136static size_t 275static size_t
137img_row_pitch (cl_mem img) 276img_row_pitch (cl_mem img)
138{ 277{
139 size_t res; 278 size_t res;
140 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 279 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
141 return res; 280 return res;
142} 281}
143 282
144static cl_event * 283static cl_event *
145event_list (SV **items, int count) 284event_list (SV **items, cl_uint *rcount)
146{ 285{
286 cl_uint count = *rcount;
287
147 if (!count) 288 if (!count)
148 return 0; 289 return 0;
149 290
150 cl_event *list = tmpbuf (sizeof (cl_event) * count); 291 cl_event *list = tmpbuf (sizeof (cl_event) * count);
292 int i = 0;
151 293
152 while (count--) 294 do
295 {
296 --count;
297 if (SvOK (items [count]))
153 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 298 list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
299 }
300 while (count);
154 301
302 *rcount = i;
303
155 return list; 304 return i ? list : 0;
156} 305}
157 306
158#define EVENT_LIST(items,count) \ 307#define EVENT_LIST(items,count) \
159 cl_uint event_list_count = (count); \ 308 cl_uint event_list_count = (count); \
160 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 309 cl_event *event_list_ptr = event_list (&ST (items), &event_list_count)
161 310
162#define INFO(class) \ 311#define INFO(class) \
163{ \ 312{ \
164 size_t size; \ 313 size_t size; \
165 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 314 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
175 324
176PROTOTYPES: ENABLE 325PROTOTYPES: ENABLE
177 326
178BOOT: 327BOOT:
179{ 328{
180 HV *stash = gv_stashpv ("OpenCL", 1); 329 HV *stash = gv_stashpv ("OpenCL", 1);
181 static const ivstr *civ, const_iv[] = { 330 static const ivstr *civ, const_iv[] = {
182 { sizeof (cl_char ), "SIZEOF_CHAR" }, 331 { sizeof (cl_char ), "SIZEOF_CHAR" },
183 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 332 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
184 { sizeof (cl_short ), "SIZEOF_SHORT" }, 333 { sizeof (cl_short ), "SIZEOF_SHORT" },
185 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 334 { sizeof (cl_ushort), "SIZEOF_USHORT" },
186 { sizeof (cl_int ), "SIZEOF_INT" }, 335 { sizeof (cl_int ), "SIZEOF_INT" },
187 { sizeof (cl_uint ), "SIZEOF_UINT" }, 336 { sizeof (cl_uint ), "SIZEOF_UINT" },
188 { sizeof (cl_long ), "SIZEOF_LONG" }, 337 { sizeof (cl_long ), "SIZEOF_LONG" },
189 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 338 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
190 { sizeof (cl_half ), "SIZEOF_HALF" }, 339 { sizeof (cl_half ), "SIZEOF_HALF" },
191 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 340 { sizeof (cl_float ), "SIZEOF_FLOAT" },
192 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 341 { sizeof (cl_double), "SIZEOF_DOUBLE" },
193#include "constiv.h" 342#include "constiv.h"
194 }; 343 };
195 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 344 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)); 345 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
197} 346}
198 347
199cl_int 348cl_int
200errno () 349errno ()
201 CODE: 350 CODE:
202 errno = res; 351 RETVAL = res;
352 OUTPUT:
353 RETVAL
203 354
204const char * 355const char *
205err2str (cl_int err) 356err2str (cl_int err)
206 357
207const char * 358const char *
221 EXTEND (SP, count); 372 EXTEND (SP, count);
222 for (i = 0; i < count; ++i) 373 for (i = 0; i < count; ++i)
223 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 374 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i]));
224 375
225void 376void
226context_from_type (FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 377context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
227 PPCODE: 378 PPCODE:
228 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (0, type, 0, 0, &res)); 379 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
229 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 380 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
230 381
382#if 0
383
231void 384void
232context (FUTURE properties, FUTURE devices, FUTURE notify = 0) 385context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
233 PPCODE: 386 PPCODE:
234 /* der Gipfel der Kunst */ 387 /* der Gipfel der Kunst */
388
389#endif
235 390
236void 391void
237wait_for_events (...) 392wait_for_events (...)
238 CODE: 393 CODE:
239 EVENT_LIST (0, items); 394 EVENT_LIST (0, items);
283 EXTEND (SP, count); 438 EXTEND (SP, count);
284 for (i = 0; i < count; ++i) 439 for (i = 0; i < count; ++i)
285 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 440 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i]));
286 441
287void 442void
288context (OpenCL::Platform self, FUTURE properties, SV *devices, FUTURE notify = 0) 443context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0)
289 PPCODE: 444 PPCODE:
290 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 445 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
291 croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); 446 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
292 447
293 AV *av = (AV *)SvRV (devices); 448 AV *av = (AV *)SvRV (devices);
294 cl_uint num_devices = av_len (av) + 1; 449 cl_uint num_devices = av_len (av) + 1;
295 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 450 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
296 int i; 451 int i;
297 452
298 for (i = num_devices; i--; ) 453 for (i = num_devices; i--; )
299 device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); 454 device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
300 455
301 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (0, num_devices, device_list, 0, 0, &res)); 456 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res));
302 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 457 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
303 458
304void 459void
305context_from_type (OpenCL::Platform self, FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 460context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
306 PPCODE: 461 PPCODE:
307 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self, 0 }; 462 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
463 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)); 464 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res));
309 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 465 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
310 466
311MODULE = OpenCL PACKAGE = OpenCL::Device 467MODULE = OpenCL PACKAGE = OpenCL::Device
312 468
353 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 509 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
354 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 510 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
355 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 511 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
356 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 512 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
357 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 513 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
358 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 514 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
359 PPCODE: 515 PPCODE:
360 cl_uint value [1]; 516 cl_uint value [1];
361 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 517 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
362 EXTEND (SP, 1); 518 EXTEND (SP, 1);
363 const int i = 0; 519 const int i = 0;
565 721
566void 722void
567buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) 723buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
568 PPCODE: 724 PPCODE:
569 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 725 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?"); 726 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
571 727
572 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); 728 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
573 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 729 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
574 730
575void 731void
576buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) 732buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
577 PPCODE: 733 PPCODE:
578 STRLEN len; 734 STRLEN len;
579 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 735 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
580 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 736 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?"); 737 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)); 738 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
583 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 739 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
584 740
741#if CL_VERSION_1_2
742
585void 743void
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) 744image (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: 745 PPCODE:
588 STRLEN len; 746 STRLEN len;
589 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 747 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
590 const cl_image_format format = { channel_order, channel_type }; 748 const cl_image_format format = { channel_order, channel_type };
749 const cl_image_desc desc = {
750 type,
751 width, height, depth,
752 array_size, row_pitch, slice_pitch,
753 num_mip_level, num_samples,
754 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
755 };
591 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); 756 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
757 char *klass = "OpenCL::Image";
758 switch (type)
759 {
760 case CL_MEM_OBJECT_IMAGE1D_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
761 case CL_MEM_OBJECT_IMAGE1D: klass = "OpenCL::Image1D"; break;
762 case CL_MEM_OBJECT_IMAGE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
763 case CL_MEM_OBJECT_IMAGE2D: klass = "OpenCL::Image2D"; break;
764 case CL_MEM_OBJECT_IMAGE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
765 case CL_MEM_OBJECT_IMAGE3D: klass = "OpenCL::Image3D"; break;
766 }
592 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 767 XPUSH_NEW_OBJ (klass, mem);
593 768
769#endif
770
594void 771void
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) 772image2d (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: 773 PPCODE:
597 STRLEN len; 774 STRLEN len;
598 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 775 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
599 const cl_image_format format = { channel_order, channel_type }; 776 const cl_image_format format = { channel_order, channel_type };
777#if PREFER_1_1
778 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
779#else
780 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
781 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
782#endif
783 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
784
785void
786image3d (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)
787 PPCODE:
788 STRLEN len;
789 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
790 const cl_image_format format = { channel_order, channel_type };
791#if PREFER_1_1
600 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 792 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
793#else
794 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
795 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
796#endif
601 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 797 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
798
799#if cl_apple_gl_sharing || cl_khr_gl_sharing
800
801void
802gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
803 PPCODE:
804 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
805 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
806
807void
808gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
809 PPCODE:
810 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
811 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
812
813#if CL_VERSION_1_2
814
815void
816gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
817 ALIAS:
818 PPCODE:
819 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
820 cl_gl_object_type type;
821 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
822 char *klass = "OpenCL::Memory";
823 switch (type)
824 {
825 case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
826 case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break;
827 case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
828 case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break;
829 case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
830 case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break;
831 }
832 XPUSH_NEW_OBJ (klass, mem);
833
834#endif
835
836void
837gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
838 PPCODE:
839#if PREFER_1_1
840 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
841#else
842 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
843#endif
844 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
845
846void
847gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
848 PPCODE:
849#if PREFER_1_1
850 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
851#else
852 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
853#endif
854 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
855
856#endif
602 857
603void 858void
604supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) 859supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
605 PPCODE: 860 PPCODE:
606{ 861{
709 cl_event ev = 0; 964 cl_event ev = 0;
710 STRLEN len; 965 STRLEN len;
711 char *ptr = SvPVbyte (data, len); 966 char *ptr = SvPVbyte (data, len);
712 EVENT_LIST (5, items - 5); 967 EVENT_LIST (5, items - 5);
713 968
714 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 969 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
715 970
716 if (ev) 971 if (ev)
717 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 972 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
718 973
719void 974void
785 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1040 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
786 1041
787 if (len < min_len) 1042 if (len < min_len)
788 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1043 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
789 1044
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)); 1045 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 1046
792 if (ev) 1047 if (ev)
793 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1048 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
794 1049
795void 1050void
850 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1105 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
851 1106
852 if (len < min_len) 1107 if (len < min_len)
853 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1108 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
854 1109
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)); 1110 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 1111
857 if (ev) 1112 if (ev)
858 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1113 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
859 1114
860void 1115void
942 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1197 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
943 } 1198 }
944 1199
945 if (SvOK (local_work_size)) 1200 if (SvOK (local_work_size))
946 { 1201 {
947 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1202 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"); 1203 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference");
949 1204
950 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1205 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"); 1206 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
952 1207
958 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1213 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
959 1214
960 if (ev) 1215 if (ev)
961 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1216 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
962 1217
1218#if cl_apple_gl_sharing || cl_khr_gl_sharing
1219
963void 1220void
964enqueue_marker (OpenCL::Queue self) 1221enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1222 ALIAS:
1223 enqueue_release_gl_objects = 1
965 PPCODE: 1224 PPCODE:
1225 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1226 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1227
966 cl_event ev; 1228 cl_event ev = 0;
967 NEED_SUCCESS (EnqueueMarker, (self, &ev)); 1229 EVENT_LIST (2, items - 2);
1230 AV *av = (AV *)SvRV (objects);
1231 cl_uint num_objects = av_len (av) + 1;
1232 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1233 int i;
1234
1235 for (i = num_objects; i--; )
1236 object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1237
1238 if (ix)
1239 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1240 else
1241 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1242
1243 if (ev)
968 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1244 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1245
1246#endif
1247
1248#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
969 1249
970void 1250void
971enqueue_wait_for_events (OpenCL::Queue self, ...) 1251enqueue_wait_for_events (OpenCL::Queue self, ...)
972 CODE: 1252 CODE:
973 EVENT_LIST (1, items - 1); 1253 EVENT_LIST (1, items - 1);
974 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 1254 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
975 1255
1256#endif
1257
976void 1258void
1259enqueue_marker (OpenCL::Queue self, ...)
1260 PPCODE:
1261 cl_event ev = 0;
1262 EVENT_LIST (1, items - 1);
1263#if PREFER_1_1
1264 if (event_list_count)
1265 croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1266 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1267#else
1268 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1269#endif
1270 if (ev)
1271 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1272
1273void
977enqueue_barrier (OpenCL::Queue self) 1274enqueue_barrier (OpenCL::Queue self, ...)
978 CODE: 1275 PPCODE:
1276 cl_event ev = 0;
1277 EVENT_LIST (1, items - 1);
1278#if PREFER_1_1
1279 if (event_list_count)
1280 croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1281 if (GIMME_V != G_VOID)
1282 croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2");
979 NEED_SUCCESS (EnqueueBarrier, (self)); 1283 NEED_SUCCESS (EnqueueBarrier, (self));
1284#else
1285 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev));
1286#endif
1287 if (ev)
1288 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
980 1289
981void 1290void
982flush (OpenCL::Queue self) 1291flush (OpenCL::Queue self)
983 CODE: 1292 CODE:
984 NEED_SUCCESS (Flush, (self)); 1293 NEED_SUCCESS (Flush, (self));
1126 NEED_SUCCESS (RetainMemObject, (value [i])); 1435 NEED_SUCCESS (RetainMemObject, (value [i]));
1127 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); 1436 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i]));
1128 } 1437 }
1129 1438
1130#END:mem 1439#END:mem
1440
1441#if cl_apple_gl_sharing || cl_khr_gl_sharing
1442
1443void
1444gl_object_info (OpenCL::Memory self)
1445 PPCODE:
1446 cl_gl_object_type type;
1447 cl_GLuint name;
1448 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1449 EXTEND (SP, 2);
1450 PUSHs (sv_2mortal (newSVuv (type)));
1451 PUSHs (sv_2mortal (newSVuv (name)));
1452
1453#endif
1131 1454
1132MODULE = OpenCL PACKAGE = OpenCL::BufferObj 1455MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1133 1456
1134void 1457void
1135sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) 1458sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1167 const int i = 0; 1490 const int i = 0;
1168 PUSHs (sv_2mortal (newSVuv (value [i]))); 1491 PUSHs (sv_2mortal (newSVuv (value [i])));
1169 1492
1170#END:image 1493#END:image
1171 1494
1495#if cl_apple_gl_sharing || cl_khr_gl_sharing
1496
1497#BEGIN:gl_texture
1498
1499void
1500target (OpenCL::Image self)
1501 PPCODE:
1502 cl_GLenum value [1];
1503 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1504 EXTEND (SP, 1);
1505 const int i = 0;
1506 PUSHs (sv_2mortal (newSVuv (value [i])));
1507
1508void
1509gl_mipmap_level (OpenCL::Image self)
1510 PPCODE:
1511 cl_GLint value [1];
1512 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1513 EXTEND (SP, 1);
1514 const int i = 0;
1515 PUSHs (sv_2mortal (newSViv (value [i])));
1516
1517#END:gl_texture
1518
1519#endif
1520
1172MODULE = OpenCL PACKAGE = OpenCL::Sampler 1521MODULE = OpenCL PACKAGE = OpenCL::Sampler
1173 1522
1174void 1523void
1175DESTROY (OpenCL::Sampler self) 1524DESTROY (OpenCL::Sampler self)
1176 CODE: 1525 CODE:
1314 { 1663 {
1315 SV *sv = sv_2mortal (newSV (sizes [i])); 1664 SV *sv = sv_2mortal (newSV (sizes [i]));
1316 SvUPGRADE (sv, SVt_PV); 1665 SvUPGRADE (sv, SVt_PV);
1317 SvPOK_only (sv); 1666 SvPOK_only (sv);
1318 SvCUR_set (sv, sizes [i]); 1667 SvCUR_set (sv, sizes [i]);
1319 ptrs [i] = SvPVX (sv); 1668 ptrs [i] = (void *)SvPVX (sv);
1320 PUSHs (sv); 1669 PUSHs (sv);
1321 } 1670 }
1322 1671
1323 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 1672 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1324 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1673 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1474set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1823set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1475 CODE: 1824 CODE:
1476 clSetKernelArg (self, idx, sizeof (value), &value); 1825 clSetKernelArg (self, idx, sizeof (value), &value);
1477 1826
1478void 1827void
1828set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1829 CODE:
1830 clSetKernelArg (self, idx, size, 0);
1831
1832void
1479set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1833set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1480 CODE: 1834 CODE:
1481 clSetKernelArg (self, idx, sizeof (value), &value); 1835 clSetKernelArg (self, idx, sizeof (value), &value);
1482 1836
1483void 1837void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines