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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.24 by root, Thu Apr 19 12:12:04 2012 UTC vs.
Revision 1.50 by root, Tue Apr 24 14:57:06 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>
9#endif 30#endif
10#if HAVE_OPENGL 31
11 #ifdef __APPLE__ 32#ifndef CL_VERSION_1_2
12 #include <OpenGL/gl.h> 33 #undef PREFER_1_1
13 #include <OpenCL/cl_gl.h> 34 #define PREFER_1_1 1
14 #else
15 #include <GL/gl.h>
16 #include <CL/cl_gl.h>
17 #endif
18#endif 35#endif
19 36
20typedef cl_platform_id OpenCL__Platform; 37typedef cl_platform_id OpenCL__Platform;
21typedef cl_device_id OpenCL__Device; 38typedef cl_device_id OpenCL__Device;
22typedef cl_context OpenCL__Context; 39typedef cl_context OpenCL__Context;
40 57
41typedef SV *FUTURE; 58typedef SV *FUTURE;
42 59
43/*****************************************************************************/ 60/*****************************************************************************/
44 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
45/* up to two temporary buffers */ 107/* up to two temporary buffers */
46static void * 108static void *
47tmpbuf (size_t size) 109tmpbuf (size_t size)
48{ 110{
49 enum { buffers = 3 }; 111 enum { buffers = 3 };
50 static int idx; 112 static int idx;
51 static void *buf [buffers]; 113 static void *buf [buffers];
52 static size_t len [buffers]; 114 static size_t len [buffers];
53 115
54 idx = ++idx % buffers; 116 idx = (idx + 1) % buffers;
55 117
56 if (len [idx] < size) 118 if (len [idx] < size)
57 { 119 {
58 free (buf [idx]); 120 free (buf [idx]);
59 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 121 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
153 return 0; 215 return 0;
154 216
155 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) 217 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
156 { 218 {
157 AV *av = (AV *)SvRV (sv); 219 AV *av = (AV *)SvRV (sv);
158 int i, len = av_len (av); 220 int i, len = av_len (av) + 1;
159 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); 221 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
160 cl_context_properties *l = p; 222 cl_context_properties *l = p;
161 223
162 if (len & 1) 224 if (len & 1)
163 croak ("%s: %s is not a property list (must be even number of elements)", func, svname); 225 croak ("%s: %s is not a property list (must be even number of elements)", func, svname);
165 while (extracount--) 227 while (extracount--)
166 *l++ = *extra++; 228 *l++ = *extra++;
167 229
168 for (i = 0; i < len; i += 2) 230 for (i = 0; i < len; i += 2)
169 { 231 {
170 cl_context_properties t = SvIV (*av_fetch (av, i, 0)); 232 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
171 cl_context_properties v; 233 SV *p_sv = *av_fetch (av, i + 1, 0);
234 cl_context_properties v = SvIV (p_sv); // code below can override
172 235
173 switch (t) 236 switch (t)
174 { 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
175 default: 256 default:
176 /* unknown property, treat as int */ 257 /* unknown property, treat as int */
177 v = SvIV (*av_fetch (av, i, 0));
178 break; 258 break;
179 } 259 }
180 260
181 *l++ = t; 261 *l++ = t;
182 *l++ = v; 262 *l++ = v;
199 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 279 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
200 return res; 280 return res;
201} 281}
202 282
203static cl_event * 283static cl_event *
204event_list (SV **items, int count) 284event_list (SV **items, cl_uint *rcount)
205{ 285{
286 cl_uint count = *rcount;
287
206 if (!count) 288 if (!count)
207 return 0; 289 return 0;
208 290
209 cl_event *list = tmpbuf (sizeof (cl_event) * count); 291 cl_event *list = tmpbuf (sizeof (cl_event) * count);
292 int i = 0;
210 293
211 while (count--) 294 do
295 {
296 --count;
297 if (SvOK (items [count]))
212 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);
213 301
302 *rcount = i;
303
214 return list; 304 return i ? list : 0;
215} 305}
216 306
217#define EVENT_LIST(items,count) \ 307#define EVENT_LIST(items,count) \
218 cl_uint event_list_count = (count); \ 308 cl_uint event_list_count = (count); \
219 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)
220 310
221#define INFO(class) \ 311#define INFO(class) \
222{ \ 312{ \
223 size_t size; \ 313 size_t size; \
224 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 314 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
247 { sizeof (cl_long ), "SIZEOF_LONG" }, 337 { sizeof (cl_long ), "SIZEOF_LONG" },
248 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 338 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
249 { sizeof (cl_half ), "SIZEOF_HALF" }, 339 { sizeof (cl_half ), "SIZEOF_HALF" },
250 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 340 { sizeof (cl_float ), "SIZEOF_FLOAT" },
251 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 341 { sizeof (cl_double), "SIZEOF_DOUBLE" },
252 { HAVE_OPENGL , "HAVE_OPENGL" },
253#include "constiv.h" 342#include "constiv.h"
254 }; 343 };
255 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--)
256 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 345 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
257} 346}
258 347
259cl_int 348cl_int
260errno () 349errno ()
261 CODE: 350 CODE:
262 errno = res; 351 RETVAL = res;
352 OUTPUT:
353 RETVAL
263 354
264const char * 355const char *
265err2str (cl_int err) 356err2str (cl_int err)
266 357
267const char * 358const char *
286context_from_type (cl_context_properties *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)
287 PPCODE: 378 PPCODE:
288 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); 379 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
289 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 380 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
290 381
382#if 0
383
291void 384void
292context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) 385context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
293 PPCODE: 386 PPCODE:
294 /* der Gipfel der Kunst */ 387 /* der Gipfel der Kunst */
388
389#endif
295 390
296void 391void
297wait_for_events (...) 392wait_for_events (...)
298 CODE: 393 CODE:
299 EVENT_LIST (0, items); 394 EVENT_LIST (0, items);
305 400
306void 401void
307info (OpenCL::Platform self, cl_platform_info name) 402info (OpenCL::Platform self, cl_platform_info name)
308 PPCODE: 403 PPCODE:
309 INFO (Platform) 404 INFO (Platform)
405
406void
407unload_compiler (OpenCL::Platform self)
408 CODE:
409#if CL_VERSION_1_2
410 clUnloadPlatformCompiler (self);
411#endif
310 412
311#BEGIN:platform 413#BEGIN:platform
312 414
313void 415void
314profile (OpenCL::Platform self) 416profile (OpenCL::Platform self)
346 448
347void 449void
348context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) 450context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0)
349 PPCODE: 451 PPCODE:
350 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 452 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
351 croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); 453 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
352 454
353 AV *av = (AV *)SvRV (devices); 455 AV *av = (AV *)SvRV (devices);
354 cl_uint num_devices = av_len (av) + 1; 456 cl_uint num_devices = av_len (av) + 1;
355 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 457 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
356 int i; 458 int i;
414 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 516 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
415 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 517 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
416 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 518 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
417 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 519 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
418 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 520 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
419 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 521 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
420 PPCODE: 522 PPCODE:
421 cl_uint value [1]; 523 cl_uint value [1];
422 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 524 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
423 EXTEND (SP, 1); 525 EXTEND (SP, 1);
424 const int i = 0; 526 const int i = 0;
626 728
627void 729void
628buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) 730buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
629 PPCODE: 731 PPCODE:
630 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 732 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
631 croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); 733 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
632 734
633 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); 735 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
634 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 736 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
635 737
636void 738void
637buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) 739buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
638 PPCODE: 740 PPCODE:
639 STRLEN len; 741 STRLEN len;
640 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 742 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
641 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 743 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
642 croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); 744 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
643 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); 745 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
644 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 746 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
645 747
748#if CL_VERSION_1_2
749
646void 750void
647image2d (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) 751image (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)
648 PPCODE: 752 PPCODE:
649 STRLEN len; 753 STRLEN len;
650 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 754 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
651 const cl_image_format format = { channel_order, channel_type }; 755 const cl_image_format format = { channel_order, channel_type };
756 const cl_image_desc desc = {
757 type,
758 width, height, depth,
759 array_size, row_pitch, slice_pitch,
760 num_mip_level, num_samples,
761 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
762 };
652 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); 763 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
764 char *klass = "OpenCL::Image";
765 switch (type)
766 {
767 case CL_MEM_OBJECT_IMAGE1D_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
768 case CL_MEM_OBJECT_IMAGE1D: klass = "OpenCL::Image1D"; break;
769 case CL_MEM_OBJECT_IMAGE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
770 case CL_MEM_OBJECT_IMAGE2D: klass = "OpenCL::Image2D"; break;
771 case CL_MEM_OBJECT_IMAGE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
772 case CL_MEM_OBJECT_IMAGE3D: klass = "OpenCL::Image3D"; break;
773 }
653 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 774 XPUSH_NEW_OBJ (klass, mem);
654 775
776#endif
777
655void 778void
656image3d (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) 779image2d (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)
657 PPCODE: 780 PPCODE:
658 STRLEN len; 781 STRLEN len;
659 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 782 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
660 const cl_image_format format = { channel_order, channel_type }; 783 const cl_image_format format = { channel_order, channel_type };
784#if PREFER_1_1
785 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
786#else
787 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
788 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
789#endif
790 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
791
792void
793image3d (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)
794 PPCODE:
795 STRLEN len;
796 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
797 const cl_image_format format = { channel_order, channel_type };
798#if PREFER_1_1
661 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 799 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
800#else
801 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
802 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
803#endif
662 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 804 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
805
806#if cl_apple_gl_sharing || cl_khr_gl_sharing
807
808void
809gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
810 PPCODE:
811 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
812 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
813
814void
815gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
816 PPCODE:
817 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
818 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
819
820#if CL_VERSION_1_2
821
822void
823gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
824 ALIAS:
825 PPCODE:
826 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
827 cl_gl_object_type type;
828 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
829 char *klass = "OpenCL::Memory";
830 switch (type)
831 {
832 case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break;
833 case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break;
834 case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break;
835 case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break;
836 case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break;
837 case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break;
838 }
839 XPUSH_NEW_OBJ (klass, mem);
840
841#endif
842
843void
844gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
845 PPCODE:
846#if PREFER_1_1
847 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
848#else
849 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
850#endif
851 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
852
853void
854gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
855 PPCODE:
856#if PREFER_1_1
857 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
858#else
859 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
860#endif
861 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
862
863#endif
663 864
664void 865void
665supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) 866supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
666 PPCODE: 867 PPCODE:
667{ 868{
770 cl_event ev = 0; 971 cl_event ev = 0;
771 STRLEN len; 972 STRLEN len;
772 char *ptr = SvPVbyte (data, len); 973 char *ptr = SvPVbyte (data, len);
773 EVENT_LIST (5, items - 5); 974 EVENT_LIST (5, items - 5);
774 975
775 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 976 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
776 977
777 if (ev) 978 if (ev)
778 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 979 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
980
981#if CL_VERSION_1_2
982
983void
984enqueue_fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
985 PPCODE:
986 cl_event ev = 0;
987 STRLEN len;
988 char *ptr = SvPVbyte (data, len);
989 EVENT_LIST (5, items - 5);
990
991 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
992
993 if (ev)
994 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
995
996void
997enqueue_fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...)
998 PPCODE:
999 cl_event ev = 0;
1000 STRLEN len;
1001 const size_t origin [3] = { x, y, z };
1002 const size_t region [3] = { width, height, depth };
1003 EVENT_LIST (12, items - 12);
1004
1005 const cl_float c_f [4] = { r, g, b, a };
1006 const cl_uint c_u [4] = { r, g, b, a };
1007 const cl_int c_s [4] = { r, g, b, a };
1008 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1009 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1010 cl_image_format format;
1011 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1012 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1013 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1014 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1015
1016 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type]],
1017 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1018
1019 if (ev)
1020 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1021
1022#endif
779 1023
780void 1024void
781enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1025enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
782 PPCODE: 1026 PPCODE:
783 cl_event ev = 0; 1027 cl_event ev = 0;
846 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1090 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
847 1091
848 if (len < min_len) 1092 if (len < min_len)
849 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1093 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
850 1094
851 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)); 1095 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));
852 1096
853 if (ev) 1097 if (ev)
854 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1098 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
855 1099
856void 1100void
911 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1155 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
912 1156
913 if (len < min_len) 1157 if (len < min_len)
914 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1158 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
915 1159
916 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)); 1160 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));
917 1161
918 if (ev) 1162 if (ev)
919 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1163 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
920 1164
921void 1165void
1003 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1247 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
1004 } 1248 }
1005 1249
1006 if (SvOK (local_work_size)) 1250 if (SvOK (local_work_size))
1007 { 1251 {
1008 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1252 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
1009 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); 1253 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference");
1010 1254
1011 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1255 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
1012 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); 1256 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
1013 1257
1019 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1263 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1020 1264
1021 if (ev) 1265 if (ev)
1022 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1266 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1023 1267
1268#if cl_apple_gl_sharing || cl_khr_gl_sharing
1269
1024void 1270void
1025enqueue_marker (OpenCL::Queue self) 1271enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1272 ALIAS:
1273 enqueue_release_gl_objects = 1
1026 PPCODE: 1274 PPCODE:
1275 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1276 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1277
1027 cl_event ev; 1278 cl_event ev = 0;
1028 NEED_SUCCESS (EnqueueMarker, (self, &ev)); 1279 EVENT_LIST (2, items - 2);
1280 AV *av = (AV *)SvRV (objects);
1281 cl_uint num_objects = av_len (av) + 1;
1282 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1283 int i;
1284
1285 for (i = num_objects; i--; )
1286 object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1287
1288 if (ix)
1289 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1290 else
1291 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1292
1293 if (ev)
1029 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1294 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1295
1296#endif
1030 1297
1031void 1298void
1032enqueue_wait_for_events (OpenCL::Queue self, ...) 1299enqueue_wait_for_events (OpenCL::Queue self, ...)
1033 CODE: 1300 CODE:
1034 EVENT_LIST (1, items - 1); 1301 EVENT_LIST (1, items - 1);
1302#if PREFER_1_1
1035 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 1303 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
1304#else
1305 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
1306#endif
1036 1307
1037void 1308void
1309enqueue_marker (OpenCL::Queue self, ...)
1310 PPCODE:
1311 cl_event ev = 0;
1312 EVENT_LIST (1, items - 1);
1313#if PREFER_1_1
1314 if (!event_list_count)
1315 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1316 else
1317#if CL_VERSION_1_2
1318 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1319#else
1320 {
1321 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
1322 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1323 }
1324#endif
1325#else
1326 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1327#endif
1328 if (ev)
1329 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1330
1331void
1038enqueue_barrier (OpenCL::Queue self) 1332enqueue_barrier (OpenCL::Queue self, ...)
1039 CODE: 1333 PPCODE:
1334 cl_event ev = 0;
1335 EVENT_LIST (1, items - 1);
1336#if PREFER_1_1
1337 if (!event_list_count && GIMME_V == G_VOID)
1040 NEED_SUCCESS (EnqueueBarrier, (self)); 1338 NEED_SUCCESS (EnqueueBarrier, (self));
1339 else
1340#if CL_VERSION_1_2
1341 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1342#else
1343 {
1344 if (event_list_count)
1345 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
1346
1347 if (GIMME_V != G_VOID)
1348 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1349 }
1350#endif
1351#else
1352 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1353#endif
1354 if (ev)
1355 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1041 1356
1042void 1357void
1043flush (OpenCL::Queue self) 1358flush (OpenCL::Queue self)
1044 CODE: 1359 CODE:
1045 NEED_SUCCESS (Flush, (self)); 1360 NEED_SUCCESS (Flush, (self));
1188 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); 1503 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i]));
1189 } 1504 }
1190 1505
1191#END:mem 1506#END:mem
1192 1507
1508#if cl_apple_gl_sharing || cl_khr_gl_sharing
1509
1510void
1511gl_object_info (OpenCL::Memory self)
1512 PPCODE:
1513 cl_gl_object_type type;
1514 cl_GLuint name;
1515 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1516 EXTEND (SP, 2);
1517 PUSHs (sv_2mortal (newSVuv (type)));
1518 PUSHs (sv_2mortal (newSVuv (name)));
1519
1520#endif
1521
1193MODULE = OpenCL PACKAGE = OpenCL::BufferObj 1522MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1194 1523
1195void 1524void
1196sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) 1525sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1197 PPCODE: 1526 PPCODE:
1207 1536
1208void 1537void
1209image_info (OpenCL::Image self, cl_image_info name) 1538image_info (OpenCL::Image self, cl_image_info name)
1210 PPCODE: 1539 PPCODE:
1211 INFO (Image) 1540 INFO (Image)
1541
1542void
1543format (OpenCL::Image self)
1544 PPCODE:
1545 cl_image_format format;
1546 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1547 EXTEND (SP, 2);
1548 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
1549 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
1212 1550
1213#BEGIN:image 1551#BEGIN:image
1214 1552
1215void 1553void
1216element_size (OpenCL::Image self) 1554element_size (OpenCL::Image self)
1228 const int i = 0; 1566 const int i = 0;
1229 PUSHs (sv_2mortal (newSVuv (value [i]))); 1567 PUSHs (sv_2mortal (newSVuv (value [i])));
1230 1568
1231#END:image 1569#END:image
1232 1570
1571#if cl_apple_gl_sharing || cl_khr_gl_sharing
1572
1573#BEGIN:gl_texture
1574
1575void
1576target (OpenCL::Image self)
1577 PPCODE:
1578 cl_GLenum value [1];
1579 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1580 EXTEND (SP, 1);
1581 const int i = 0;
1582 PUSHs (sv_2mortal (newSVuv (value [i])));
1583
1584void
1585gl_mipmap_level (OpenCL::Image self)
1586 PPCODE:
1587 cl_GLint value [1];
1588 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1589 EXTEND (SP, 1);
1590 const int i = 0;
1591 PUSHs (sv_2mortal (newSViv (value [i])));
1592
1593#END:gl_texture
1594
1595#endif
1596
1233MODULE = OpenCL PACKAGE = OpenCL::Sampler 1597MODULE = OpenCL PACKAGE = OpenCL::Sampler
1234 1598
1235void 1599void
1236DESTROY (OpenCL::Sampler self) 1600DESTROY (OpenCL::Sampler self)
1237 CODE: 1601 CODE:
1350 PPCODE: 1714 PPCODE:
1351 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); 1715 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
1352 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 1716 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel);
1353 1717
1354void 1718void
1719kernels_in_program (OpenCL::Program program)
1720 PPCODE:
1721 cl_uint num_kernels;
1722 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
1723 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
1724 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
1725
1726 int i;
1727 EXTEND (SP, num_kernels);
1728 for (i = 0; i < num_kernels; ++i)
1729 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Kernel", kernels [i]));
1730
1731void
1355info (OpenCL::Program self, cl_program_info name) 1732info (OpenCL::Program self, cl_program_info name)
1356 PPCODE: 1733 PPCODE:
1357 INFO (Program) 1734 INFO (Program)
1358 1735
1359void 1736void
1375 { 1752 {
1376 SV *sv = sv_2mortal (newSV (sizes [i])); 1753 SV *sv = sv_2mortal (newSV (sizes [i]));
1377 SvUPGRADE (sv, SVt_PV); 1754 SvUPGRADE (sv, SVt_PV);
1378 SvPOK_only (sv); 1755 SvPOK_only (sv);
1379 SvCUR_set (sv, sizes [i]); 1756 SvCUR_set (sv, sizes [i]);
1380 ptrs [i] = SvPVX (sv); 1757 ptrs [i] = (void *)SvPVX (sv);
1381 PUSHs (sv); 1758 PUSHs (sv);
1382 } 1759 }
1383 1760
1384 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 1761 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1385 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1762 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1535set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1912set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1536 CODE: 1913 CODE:
1537 clSetKernelArg (self, idx, sizeof (value), &value); 1914 clSetKernelArg (self, idx, sizeof (value), &value);
1538 1915
1539void 1916void
1917set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1918 CODE:
1919 clSetKernelArg (self, idx, size, 0);
1920
1921void
1540set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1922set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1541 CODE: 1923 CODE:
1542 clSetKernelArg (self, idx, sizeof (value), &value); 1924 clSetKernelArg (self, idx, sizeof (value), &value);
1543 1925
1544void 1926void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines