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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.27 by root, Thu Apr 19 13:18:31 2012 UTC vs.
Revision 1.44 by root, Mon Apr 23 13:50:12 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_!
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;
144 return 0; 215 return 0;
145 216
146 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) 217 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
147 { 218 {
148 AV *av = (AV *)SvRV (sv); 219 AV *av = (AV *)SvRV (sv);
149 int i, len = av_len (av); 220 int i, len = av_len (av) + 1;
150 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));
151 cl_context_properties *l = p; 222 cl_context_properties *l = p;
152 223
153 if (len & 1) 224 if (len & 1)
154 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);
156 while (extracount--) 227 while (extracount--)
157 *l++ = *extra++; 228 *l++ = *extra++;
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 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
163 235
164 switch (t) 236 switch (t)
165 { 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
166 default: 256 default:
167 /* unknown property, treat as int */ 257 /* unknown property, treat as int */
168 v = SvIV (*av_fetch (av, i, 0));
169 break; 258 break;
170 } 259 }
171 260
172 *l++ = t; 261 *l++ = t;
173 *l++ = v; 262 *l++ = v;
190 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 279 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
191 return res; 280 return res;
192} 281}
193 282
194static cl_event * 283static cl_event *
195event_list (SV **items, int count) 284event_list (SV **items, cl_uint *rcount)
196{ 285{
286 cl_uint count = *rcount;
287
197 if (!count) 288 if (!count)
198 return 0; 289 return 0;
199 290
200 cl_event *list = tmpbuf (sizeof (cl_event) * count); 291 cl_event *list = tmpbuf (sizeof (cl_event) * count);
292 int i = 0;
201 293
202 while (count--) 294 do
295 {
296 --count;
297 if (SvOK (items [count]))
203 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);
204 301
302 *rcount = i;
303
205 return list; 304 return i ? list : 0;
206} 305}
207 306
208#define EVENT_LIST(items,count) \ 307#define EVENT_LIST(items,count) \
209 cl_uint event_list_count = (count); \ 308 cl_uint event_list_count = (count); \
210 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)
211 310
212#define INFO(class) \ 311#define INFO(class) \
213{ \ 312{ \
214 size_t size; \ 313 size_t size; \
215 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 314 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
247} 346}
248 347
249cl_int 348cl_int
250errno () 349errno ()
251 CODE: 350 CODE:
252 errno = res; 351 RETVAL = res;
352 OUTPUT:
353 RETVAL
253 354
254const char * 355const char *
255err2str (cl_int err) 356err2str (cl_int err)
256 357
257const char * 358const char *
276context_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)
277 PPCODE: 378 PPCODE:
278 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));
279 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 380 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
280 381
382#if 0
383
281void 384void
282context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) 385context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
283 PPCODE: 386 PPCODE:
284 /* der Gipfel der Kunst */ 387 /* der Gipfel der Kunst */
388
389#endif
285 390
286void 391void
287wait_for_events (...) 392wait_for_events (...)
288 CODE: 393 CODE:
289 EVENT_LIST (0, items); 394 EVENT_LIST (0, items);
404 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 509 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
405 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 510 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
406 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 511 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
407 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 512 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
408 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 513 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
409 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 514 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
410 PPCODE: 515 PPCODE:
411 cl_uint value [1]; 516 cl_uint value [1];
412 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 517 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
413 EXTEND (SP, 1); 518 EXTEND (SP, 1);
414 const int i = 0; 519 const int i = 0;
631 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)))
632 croak ("OpenCL::Context::buffer_sv: you 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?");
633 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));
634 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 739 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
635 740
741#if CL_VERSION_1_2
742
636void 743void
637image2d (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)
638 PPCODE: 745 PPCODE:
639 STRLEN len; 746 STRLEN len;
640 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 747 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
641 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 };
642 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 }
643 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 767 XPUSH_NEW_OBJ (klass, mem);
644 768
769#endif
770
645void 771void
646image3d (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)
647 PPCODE: 773 PPCODE:
648 STRLEN len; 774 STRLEN len;
649 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 775 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
650 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
651 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
652 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 797 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
653 798
654#if cl_apple_gl_sharing || cl_khr_gl_sharing 799#if cl_apple_gl_sharing || cl_khr_gl_sharing
655 800
656void 801void
658 PPCODE: 803 PPCODE:
659 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); 804 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
660 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 805 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
661 806
662void 807void
663gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
664 PPCODE:
665 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
666 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
667
668void
669gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
670 PPCODE:
671 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
672 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
673
674void
675gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) 808gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
676 PPCODE: 809 PPCODE:
677 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); 810 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
678 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 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);
679 855
680#endif 856#endif
681 857
682void 858void
683supported_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)
788 cl_event ev = 0; 964 cl_event ev = 0;
789 STRLEN len; 965 STRLEN len;
790 char *ptr = SvPVbyte (data, len); 966 char *ptr = SvPVbyte (data, len);
791 EVENT_LIST (5, items - 5); 967 EVENT_LIST (5, items - 5);
792 968
793 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));
794 970
795 if (ev) 971 if (ev)
796 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 972 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
797 973
798void 974void
864 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];
865 1041
866 if (len < min_len) 1042 if (len < min_len)
867 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1043 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
868 1044
869 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));
870 1046
871 if (ev) 1047 if (ev)
872 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1048 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
873 1049
874void 1050void
929 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;
930 1106
931 if (len < min_len) 1107 if (len < min_len)
932 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1108 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
933 1109
934 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));
935 1111
936 if (ev) 1112 if (ev)
937 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1113 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
938 1114
939void 1115void
1021 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1197 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
1022 } 1198 }
1023 1199
1024 if (SvOK (local_work_size)) 1200 if (SvOK (local_work_size))
1025 { 1201 {
1026 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)
1027 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");
1028 1204
1029 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1205 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
1030 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");
1031 1207
1043 1219
1044void 1220void
1045enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) 1221enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1046 ALIAS: 1222 ALIAS:
1047 enqueue_release_gl_objects = 1 1223 enqueue_release_gl_objects = 1
1048 CODE: 1224 PPCODE:
1049 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 1225 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1050 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); 1226 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1051 1227
1052 cl_event ev = 0; 1228 cl_event ev = 0;
1053 EVENT_LIST (2, items - 2); 1229 EVENT_LIST (2, items - 2);
1067 if (ev) 1243 if (ev)
1068 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1244 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1069 1245
1070#endif 1246#endif
1071 1247
1072void 1248#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
1073enqueue_marker (OpenCL::Queue self)
1074 PPCODE:
1075 cl_event ev;
1076 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1077 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1078 1249
1079void 1250void
1080enqueue_wait_for_events (OpenCL::Queue self, ...) 1251enqueue_wait_for_events (OpenCL::Queue self, ...)
1081 CODE: 1252 CODE:
1082 EVENT_LIST (1, items - 1); 1253 EVENT_LIST (1, items - 1);
1083 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 1254 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
1084 1255
1256#endif
1257
1085void 1258void
1259enqueue_marker (OpenCL::Queue self, ...)
1260 PPCODE:
1261 cl_event ev = 0;
1262 EVENT_LIST (1, items - 1);
1263#if CL_VERSION_1_2
1264 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1265#else
1266 if (event_list_count)
1267 croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1268 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1269#endif
1270 if (ev)
1271 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1272
1273void
1086enqueue_barrier (OpenCL::Queue self) 1274enqueue_barrier (OpenCL::Queue self, ...)
1087 CODE: 1275 PPCODE:
1276 cl_event ev = 0;
1277 EVENT_LIST (1, items - 1);
1278#if CL_VERSION_1_2
1279 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev));
1280#else
1281 if (event_list_count)
1282 croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1283 if (GIMME_V != G_VOID)
1284 croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2");
1088 NEED_SUCCESS (EnqueueBarrier, (self)); 1285 NEED_SUCCESS (EnqueueBarrier, (self));
1286#endif
1287 if (ev)
1288 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1089 1289
1090void 1290void
1091flush (OpenCL::Queue self) 1291flush (OpenCL::Queue self)
1092 CODE: 1292 CODE:
1093 NEED_SUCCESS (Flush, (self)); 1293 NEED_SUCCESS (Flush, (self));
1243void 1443void
1244gl_object_info (OpenCL::Memory self) 1444gl_object_info (OpenCL::Memory self)
1245 PPCODE: 1445 PPCODE:
1246 cl_gl_object_type type; 1446 cl_gl_object_type type;
1247 cl_GLuint name; 1447 cl_GLuint name;
1248 NEED_SUCCESS (clGetGLObjectInfo, (self, &type, &name)); 1448 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1249 EXTEND (SP, 2); 1449 EXTEND (SP, 2);
1250 PUSHs (sv_2mortal (newSVuv (type))); 1450 PUSHs (sv_2mortal (newSVuv (type)));
1251 PUSHs (sv_2mortal (newSVuv (name))); 1451 PUSHs (sv_2mortal (newSVuv (name)));
1252 1452
1253#endif 1453#endif
1298 1498
1299void 1499void
1300target (OpenCL::Image self) 1500target (OpenCL::Image self)
1301 PPCODE: 1501 PPCODE:
1302 cl_GLenum value [1]; 1502 cl_GLenum value [1];
1303 NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); 1503 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1304 EXTEND (SP, 1); 1504 EXTEND (SP, 1);
1305 const int i = 0; 1505 const int i = 0;
1306 PUSHs (sv_2mortal (newSVuv (value [i]))); 1506 PUSHs (sv_2mortal (newSVuv (value [i])));
1307 1507
1308void 1508void
1309gl_mipmap_level (OpenCL::Image self) 1509gl_mipmap_level (OpenCL::Image self)
1310 PPCODE: 1510 PPCODE:
1311 cl_GLint value [1]; 1511 cl_GLint value [1];
1312 NEED_SUCCESS (GetGlTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); 1512 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1313 EXTEND (SP, 1); 1513 EXTEND (SP, 1);
1314 const int i = 0; 1514 const int i = 0;
1315 PUSHs (sv_2mortal (newSViv (value [i]))); 1515 PUSHs (sv_2mortal (newSViv (value [i])));
1316 1516
1317#END:gl_texture 1517#END:gl_texture
1463 { 1663 {
1464 SV *sv = sv_2mortal (newSV (sizes [i])); 1664 SV *sv = sv_2mortal (newSV (sizes [i]));
1465 SvUPGRADE (sv, SVt_PV); 1665 SvUPGRADE (sv, SVt_PV);
1466 SvPOK_only (sv); 1666 SvPOK_only (sv);
1467 SvCUR_set (sv, sizes [i]); 1667 SvCUR_set (sv, sizes [i]);
1468 ptrs [i] = SvPVX (sv); 1668 ptrs [i] = (void *)SvPVX (sv);
1469 PUSHs (sv); 1669 PUSHs (sv);
1470 } 1670 }
1471 1671
1472 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));
1473 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1673 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1623set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1823set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1624 CODE: 1824 CODE:
1625 clSetKernelArg (self, idx, sizeof (value), &value); 1825 clSetKernelArg (self, idx, sizeof (value), &value);
1626 1826
1627void 1827void
1828set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1829 CODE:
1830 clSetKernelArg (self, idx, size, 0);
1831
1832void
1628set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1833set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1629 CODE: 1834 CODE:
1630 clSetKernelArg (self, idx, sizeof (value), &value); 1835 clSetKernelArg (self, idx, sizeof (value), &value);
1631 1836
1632void 1837void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines