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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.29 by root, Thu Apr 19 13:51:36 2012 UTC vs.
Revision 1.48 by root, Tue Apr 24 14:24:42 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{
40 enum { buffers = 3 }; 111 enum { buffers = 3 };
41 static int idx; 112 static int idx;
42 static void *buf [buffers]; 113 static void *buf [buffers];
43 static size_t len [buffers]; 114 static size_t len [buffers];
44 115
45 idx = ++idx % buffers; 116 idx = (idx + 1) % buffers;
46 117
47 if (len [idx] < size) 118 if (len [idx] < size)
48 { 119 {
49 free (buf [idx]); 120 free (buf [idx]);
50 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 121 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
158 229
159 for (i = 0; i < len; i += 2) 230 for (i = 0; i < len; i += 2)
160 { 231 {
161 cl_context_properties t = SvIV (*av_fetch (av, i , 0)); 232 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
162 SV *p_sv = *av_fetch (av, i + 1, 0); 233 SV *p_sv = *av_fetch (av, i + 1, 0);
163 cl_context_properties v; 234 cl_context_properties v = SvIV (p_sv); // code below can override
164 235
165 switch (t) 236 switch (t)
166 { 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
167 default: 256 default:
168 /* unknown property, treat as int */ 257 /* unknown property, treat as int */
169 v = SvIV (p_sv);
170 break; 258 break;
171 } 259 }
172 260
173 *l++ = t; 261 *l++ = t;
174 *l++ = v; 262 *l++ = v;
191 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 279 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
192 return res; 280 return res;
193} 281}
194 282
195static cl_event * 283static cl_event *
196event_list (SV **items, int count) 284event_list (SV **items, cl_uint *rcount)
197{ 285{
286 cl_uint count = *rcount;
287
198 if (!count) 288 if (!count)
199 return 0; 289 return 0;
200 290
201 cl_event *list = tmpbuf (sizeof (cl_event) * count); 291 cl_event *list = tmpbuf (sizeof (cl_event) * count);
292 int i = 0;
202 293
203 while (count--) 294 do
295 {
296 --count;
297 if (SvOK (items [count]))
204 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);
205 301
302 *rcount = i;
303
206 return list; 304 return i ? list : 0;
207} 305}
208 306
209#define EVENT_LIST(items,count) \ 307#define EVENT_LIST(items,count) \
210 cl_uint event_list_count = (count); \ 308 cl_uint event_list_count = (count); \
211 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)
212 310
213#define INFO(class) \ 311#define INFO(class) \
214{ \ 312{ \
215 size_t size; \ 313 size_t size; \
216 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 314 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
248} 346}
249 347
250cl_int 348cl_int
251errno () 349errno ()
252 CODE: 350 CODE:
253 errno = res; 351 RETVAL = res;
352 OUTPUT:
353 RETVAL
254 354
255const char * 355const char *
256err2str (cl_int err) 356err2str (cl_int err)
257 357
258const char * 358const char *
277context_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)
278 PPCODE: 378 PPCODE:
279 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));
280 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 380 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
281 381
382#if 0
383
282void 384void
283context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) 385context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
284 PPCODE: 386 PPCODE:
285 /* der Gipfel der Kunst */ 387 /* der Gipfel der Kunst */
388
389#endif
286 390
287void 391void
288wait_for_events (...) 392wait_for_events (...)
289 CODE: 393 CODE:
290 EVENT_LIST (0, items); 394 EVENT_LIST (0, items);
296 400
297void 401void
298info (OpenCL::Platform self, cl_platform_info name) 402info (OpenCL::Platform self, cl_platform_info name)
299 PPCODE: 403 PPCODE:
300 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
301 412
302#BEGIN:platform 413#BEGIN:platform
303 414
304void 415void
305profile (OpenCL::Platform self) 416profile (OpenCL::Platform self)
405 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 516 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
406 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 517 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
407 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 518 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
408 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 519 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
409 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 520 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
410 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 521 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
411 PPCODE: 522 PPCODE:
412 cl_uint value [1]; 523 cl_uint value [1];
413 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 524 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
414 EXTEND (SP, 1); 525 EXTEND (SP, 1);
415 const int i = 0; 526 const int i = 0;
632 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)))
633 croak ("OpenCL::Context::buffer_sv: you 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?");
634 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));
635 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 746 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
636 747
748#if CL_VERSION_1_2
749
637void 750void
638image2d (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)
639 PPCODE: 752 PPCODE:
640 STRLEN len; 753 STRLEN len;
641 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 754 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
642 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 };
643 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 }
644 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 774 XPUSH_NEW_OBJ (klass, mem);
645 775
776#endif
777
646void 778void
647image3d (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)
648 PPCODE: 780 PPCODE:
649 STRLEN len; 781 STRLEN len;
650 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 782 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
651 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
652 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
653 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 804 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
654 805
655#if cl_apple_gl_sharing || cl_khr_gl_sharing 806#if cl_apple_gl_sharing || cl_khr_gl_sharing
656 807
657void 808void
659 PPCODE: 810 PPCODE:
660 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); 811 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
661 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 812 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
662 813
663void 814void
664gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
665 PPCODE:
666 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
667 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
668
669void
670gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
671 PPCODE:
672 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
673 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
674
675void
676gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) 815gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
677 PPCODE: 816 PPCODE:
678 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); 817 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
679 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 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);
680 862
681#endif 863#endif
682 864
683void 865void
684supported_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)
789 cl_event ev = 0; 971 cl_event ev = 0;
790 STRLEN len; 972 STRLEN len;
791 char *ptr = SvPVbyte (data, len); 973 char *ptr = SvPVbyte (data, len);
792 EVENT_LIST (5, items - 5); 974 EVENT_LIST (5, items - 5);
793 975
794 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));
795 977
796 if (ev) 978 if (ev)
797 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 (clGetImageInfo, (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
798 1023
799void 1024void
800enqueue_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, ...)
801 PPCODE: 1026 PPCODE:
802 cl_event ev = 0; 1027 cl_event ev = 0;
865 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];
866 1091
867 if (len < min_len) 1092 if (len < min_len)
868 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1093 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
869 1094
870 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));
871 1096
872 if (ev) 1097 if (ev)
873 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1098 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
874 1099
875void 1100void
930 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;
931 1156
932 if (len < min_len) 1157 if (len < min_len)
933 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1158 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
934 1159
935 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));
936 1161
937 if (ev) 1162 if (ev)
938 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1163 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
939 1164
940void 1165void
1022 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1247 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
1023 } 1248 }
1024 1249
1025 if (SvOK (local_work_size)) 1250 if (SvOK (local_work_size))
1026 { 1251 {
1027 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)
1028 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");
1029 1254
1030 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1255 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
1031 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");
1032 1257
1044 1269
1045void 1270void
1046enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) 1271enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1047 ALIAS: 1272 ALIAS:
1048 enqueue_release_gl_objects = 1 1273 enqueue_release_gl_objects = 1
1049 CODE: 1274 PPCODE:
1050 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 1275 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1051 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); 1276 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1052 1277
1053 cl_event ev = 0; 1278 cl_event ev = 0;
1054 EVENT_LIST (2, items - 2); 1279 EVENT_LIST (2, items - 2);
1069 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1294 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1070 1295
1071#endif 1296#endif
1072 1297
1073void 1298void
1074enqueue_marker (OpenCL::Queue self)
1075 PPCODE:
1076 cl_event ev;
1077 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1078 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1079
1080void
1081enqueue_wait_for_events (OpenCL::Queue self, ...) 1299enqueue_wait_for_events (OpenCL::Queue self, ...)
1082 CODE: 1300 CODE:
1083 EVENT_LIST (1, items - 1); 1301 EVENT_LIST (1, items - 1);
1302#if PREFER_1_1
1084 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
1085 1307
1086void 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 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); // also a barrier
1321#endif
1322#else
1323 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1324#endif
1325 if (ev)
1326 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1327
1328void
1087enqueue_barrier (OpenCL::Queue self) 1329enqueue_barrier (OpenCL::Queue self, ...)
1088 CODE: 1330 PPCODE:
1331 cl_event ev = 0;
1332 EVENT_LIST (1, items - 1);
1333#if PREFER_1_1
1334 if (!event_list_count && GIMME_V == G_VOID)
1089 NEED_SUCCESS (EnqueueBarrier, (self)); 1335 NEED_SUCCESS (EnqueueBarrier, (self));
1336 else
1337#if CL_VERSION_1_2
1338 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1339#else
1340 {
1341 if (event_list_count)
1342 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1343
1344 if (GIMME_V != G_VOID)
1345 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1346 }
1347#endif
1348#else
1349 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1350#endif
1351 if (ev)
1352 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1090 1353
1091void 1354void
1092flush (OpenCL::Queue self) 1355flush (OpenCL::Queue self)
1093 CODE: 1356 CODE:
1094 NEED_SUCCESS (Flush, (self)); 1357 NEED_SUCCESS (Flush, (self));
1244void 1507void
1245gl_object_info (OpenCL::Memory self) 1508gl_object_info (OpenCL::Memory self)
1246 PPCODE: 1509 PPCODE:
1247 cl_gl_object_type type; 1510 cl_gl_object_type type;
1248 cl_GLuint name; 1511 cl_GLuint name;
1249 NEED_SUCCESS (clGetGLObjectInfo, (self, &type, &name)); 1512 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1250 EXTEND (SP, 2); 1513 EXTEND (SP, 2);
1251 PUSHs (sv_2mortal (newSVuv (type))); 1514 PUSHs (sv_2mortal (newSVuv (type)));
1252 PUSHs (sv_2mortal (newSVuv (name))); 1515 PUSHs (sv_2mortal (newSVuv (name)));
1253 1516
1254#endif 1517#endif
1299 1562
1300void 1563void
1301target (OpenCL::Image self) 1564target (OpenCL::Image self)
1302 PPCODE: 1565 PPCODE:
1303 cl_GLenum value [1]; 1566 cl_GLenum value [1];
1304 NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); 1567 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1305 EXTEND (SP, 1); 1568 EXTEND (SP, 1);
1306 const int i = 0; 1569 const int i = 0;
1307 PUSHs (sv_2mortal (newSVuv (value [i]))); 1570 PUSHs (sv_2mortal (newSVuv (value [i])));
1308 1571
1309void 1572void
1310gl_mipmap_level (OpenCL::Image self) 1573gl_mipmap_level (OpenCL::Image self)
1311 PPCODE: 1574 PPCODE:
1312 cl_GLint value [1]; 1575 cl_GLint value [1];
1313 NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); 1576 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1314 EXTEND (SP, 1); 1577 EXTEND (SP, 1);
1315 const int i = 0; 1578 const int i = 0;
1316 PUSHs (sv_2mortal (newSViv (value [i]))); 1579 PUSHs (sv_2mortal (newSViv (value [i])));
1317 1580
1318#END:gl_texture 1581#END:gl_texture
1439 PPCODE: 1702 PPCODE:
1440 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); 1703 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
1441 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 1704 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel);
1442 1705
1443void 1706void
1707kernels_in_program (OpenCL::Program program)
1708 PPCODE:
1709 cl_uint num_kernels;
1710 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
1711 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
1712 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
1713
1714 int i;
1715 EXTEND (SP, num_kernels);
1716 for (i = 0; i < num_kernels; ++i)
1717 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Kernel", kernels [i]));
1718
1719void
1444info (OpenCL::Program self, cl_program_info name) 1720info (OpenCL::Program self, cl_program_info name)
1445 PPCODE: 1721 PPCODE:
1446 INFO (Program) 1722 INFO (Program)
1447 1723
1448void 1724void
1464 { 1740 {
1465 SV *sv = sv_2mortal (newSV (sizes [i])); 1741 SV *sv = sv_2mortal (newSV (sizes [i]));
1466 SvUPGRADE (sv, SVt_PV); 1742 SvUPGRADE (sv, SVt_PV);
1467 SvPOK_only (sv); 1743 SvPOK_only (sv);
1468 SvCUR_set (sv, sizes [i]); 1744 SvCUR_set (sv, sizes [i]);
1469 ptrs [i] = SvPVX (sv); 1745 ptrs [i] = (void *)SvPVX (sv);
1470 PUSHs (sv); 1746 PUSHs (sv);
1471 } 1747 }
1472 1748
1473 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 1749 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1474 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1750 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1624set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1900set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1625 CODE: 1901 CODE:
1626 clSetKernelArg (self, idx, sizeof (value), &value); 1902 clSetKernelArg (self, idx, sizeof (value), &value);
1627 1903
1628void 1904void
1905set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1906 CODE:
1907 clSetKernelArg (self, idx, size, 0);
1908
1909void
1629set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1910set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1630 CODE: 1911 CODE:
1631 clSetKernelArg (self, idx, sizeof (value), &value); 1912 clSetKernelArg (self, idx, sizeof (value), &value);
1632 1913
1633void 1914void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines