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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.32 by root, Thu Apr 19 19:36:24 2012 UTC vs.
Revision 1.38 by root, Sat Apr 21 19:17:09 2012 UTC

… …
3#include "XSUB.h" 3#include "XSUB.h"
4 4
5#ifdef I_DLFCN 5#ifdef I_DLFCN
6 #include <dlfcn.h> 6 #include <dlfcn.h>
7#endif 7#endif
8
9// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
10// but then fail to define the api functions unless you ALSO define
11// this. This breaks 100% of the opencl 1.1 apps, for what reason?
12// after all, the functions are deprecated, not removed.
13// in addition, you cannot test for this in any future-proof way.
14// each time a new opencl version comes out, you need to make a new
15// release.
16#define CL_USE_DEPRECATED_OPENCL_1_1_APIS
17#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
8 18
9#ifdef __APPLE__ 19#ifdef __APPLE__
10 #include <OpenCL/opencl.h> 20 #include <OpenCL/opencl.h>
11#else 21#else
12 #include <CL/opencl.h> 22 #include <CL/opencl.h>
23#endif
24
25#if 0 // testing
26#undef CL_USE_DEPRECATED_OPENCL_1_1_APIS
27#undef CL_VERSION_1_2
13#endif 28#endif
14 29
15typedef cl_platform_id OpenCL__Platform; 30typedef cl_platform_id OpenCL__Platform;
16typedef cl_device_id OpenCL__Device; 31typedef cl_device_id OpenCL__Device;
17typedef cl_context OpenCL__Context; 32typedef cl_context OpenCL__Context;
… …
89 enum { buffers = 3 }; 104 enum { buffers = 3 };
90 static int idx; 105 static int idx;
91 static void *buf [buffers]; 106 static void *buf [buffers];
92 static size_t len [buffers]; 107 static size_t len [buffers];
93 108
94 idx = ++idx % buffers; 109 idx = (idx + 1) % buffers;
95 110
96 if (len [idx] < size) 111 if (len [idx] < size)
97 { 112 {
98 free (buf [idx]); 113 free (buf [idx]);
99 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 114 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
… …
257 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 272 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
258 return res; 273 return res;
259} 274}
260 275
261static cl_event * 276static cl_event *
262event_list (SV **items, int count) 277event_list (SV **items, cl_uint *rcount)
263{ 278{
279 cl_uint count = *rcount;
280
264 if (!count) 281 if (!count)
265 return 0; 282 return 0;
266 283
267 cl_event *list = tmpbuf (sizeof (cl_event) * count); 284 cl_event *list = tmpbuf (sizeof (cl_event) * count);
285 int i = 0;
268 286
269 while (count--) 287 do
288 {
289 --count;
290 if (SvOK (items [count]))
270 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 291 list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
292 }
293 while (count);
271 294
295 *rcount = i;
296
272 return list; 297 return i ? list : 0;
273} 298}
274 299
275#define EVENT_LIST(items,count) \ 300#define EVENT_LIST(items,count) \
276 cl_uint event_list_count = (count); \ 301 cl_uint event_list_count = (count); \
277 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 302 cl_event *event_list_ptr = event_list (&ST (items), &event_list_count)
278 303
279#define INFO(class) \ 304#define INFO(class) \
280{ \ 305{ \
281 size_t size; \ 306 size_t size; \
282 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 307 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
… …
314} 339}
315 340
316cl_int 341cl_int
317errno () 342errno ()
318 CODE: 343 CODE:
319 errno = res; 344 RETVAL = res;
345 OUTPUT:
346 RETVAL
320 347
321const char * 348const char *
322err2str (cl_int err) 349err2str (cl_int err)
323 350
324const char * 351const char *
… …
343context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 370context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
344 PPCODE: 371 PPCODE:
345 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); 372 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
346 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 373 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
347 374
375#if 0
376
348void 377void
349context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) 378context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
350 PPCODE: 379 PPCODE:
351 /* der Gipfel der Kunst */ 380 /* der Gipfel der Kunst */
381
382#endif
352 383
353void 384void
354wait_for_events (...) 385wait_for_events (...)
355 CODE: 386 CODE:
356 EVENT_LIST (0, items); 387 EVENT_LIST (0, items);
… …
698 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 729 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
699 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); 730 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
700 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); 731 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
701 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 732 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
702 733
734#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
735
703void 736void
704image2d (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) 737image2d (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)
705 PPCODE: 738 PPCODE:
706 STRLEN len; 739 STRLEN len;
707 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 740 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
… …
716 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 749 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
717 const cl_image_format format = { channel_order, channel_type }; 750 const cl_image_format format = { channel_order, channel_type };
718 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 751 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
719 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 752 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
720 753
754#endif
755
721#if cl_apple_gl_sharing || cl_khr_gl_sharing 756#if cl_apple_gl_sharing || cl_khr_gl_sharing
722 757
723void 758void
724gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) 759gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
725 PPCODE: 760 PPCODE:
726 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); 761 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
727 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 762 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
728 763
764#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
765
729void 766void
730gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) 767gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
731 PPCODE: 768 PPCODE:
732 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); 769 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
733 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 770 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
… …
735void 772void
736gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) 773gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
737 PPCODE: 774 PPCODE:
738 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); 775 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
739 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 776 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
777
778#endif
740 779
741void 780void
742gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) 781gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
743 PPCODE: 782 PPCODE:
744 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); 783 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
… …
855 cl_event ev = 0; 894 cl_event ev = 0;
856 STRLEN len; 895 STRLEN len;
857 char *ptr = SvPVbyte (data, len); 896 char *ptr = SvPVbyte (data, len);
858 EVENT_LIST (5, items - 5); 897 EVENT_LIST (5, items - 5);
859 898
860 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 899 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
861 900
862 if (ev) 901 if (ev)
863 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 902 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
864 903
865void 904void
… …
931 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 970 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
932 971
933 if (len < min_len) 972 if (len < min_len)
934 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 973 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
935 974
936 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)); 975 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));
937 976
938 if (ev) 977 if (ev)
939 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 978 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
940 979
941void 980void
… …
996 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1035 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
997 1036
998 if (len < min_len) 1037 if (len < min_len)
999 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1038 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1000 1039
1001 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)); 1040 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));
1002 1041
1003 if (ev) 1042 if (ev)
1004 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1043 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1005 1044
1006void 1045void
… …
1088 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1127 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
1089 } 1128 }
1090 1129
1091 if (SvOK (local_work_size)) 1130 if (SvOK (local_work_size))
1092 { 1131 {
1093 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1132 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
1094 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); 1133 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference");
1095 1134
1096 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1135 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
1097 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); 1136 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
1098 1137
… …
1110 1149
1111void 1150void
1112enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) 1151enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1113 ALIAS: 1152 ALIAS:
1114 enqueue_release_gl_objects = 1 1153 enqueue_release_gl_objects = 1
1115 CODE: 1154 PPCODE:
1116 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 1155 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1117 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); 1156 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1118 1157
1119 cl_event ev = 0; 1158 cl_event ev = 0;
1120 EVENT_LIST (2, items - 2); 1159 EVENT_LIST (2, items - 2);
… …
1134 if (ev) 1173 if (ev)
1135 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1174 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1136 1175
1137#endif 1176#endif
1138 1177
1139void 1178#if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS
1140enqueue_marker (OpenCL::Queue self)
1141 PPCODE:
1142 cl_event ev;
1143 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1144 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1145 1179
1146void 1180void
1147enqueue_wait_for_events (OpenCL::Queue self, ...) 1181enqueue_wait_for_events (OpenCL::Queue self, ...)
1148 CODE: 1182 CODE:
1149 EVENT_LIST (1, items - 1); 1183 EVENT_LIST (1, items - 1);
1150 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 1184 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
1151 1185
1186#endif
1187
1152void 1188void
1189enqueue_marker (OpenCL::Queue self, ...)
1190 PPCODE:
1191 cl_event ev = 0;
1192 EVENT_LIST (1, items - 1);
1193#if CL_VERSION_1_2
1194 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1195#else
1196 if (event_list_count)
1197 croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1198 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1199#endif
1200 if (ev)
1201 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1202
1203void
1153enqueue_barrier (OpenCL::Queue self) 1204enqueue_barrier (OpenCL::Queue self, ...)
1154 CODE: 1205 PPCODE:
1206 cl_event ev = 0;
1207 EVENT_LIST (1, items - 1);
1208#if CL_VERSION_1_2
1209 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev));
1210#else
1211 if (event_list_count)
1212 croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2");
1213 if (GIMME_V != G_VOID)
1214 croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2");
1155 NEED_SUCCESS (EnqueueBarrier, (self)); 1215 NEED_SUCCESS (EnqueueBarrier, (self));
1216#endif
1217 if (ev)
1218 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1156 1219
1157void 1220void
1158flush (OpenCL::Queue self) 1221flush (OpenCL::Queue self)
1159 CODE: 1222 CODE:
1160 NEED_SUCCESS (Flush, (self)); 1223 NEED_SUCCESS (Flush, (self));
… …
1530 { 1593 {
1531 SV *sv = sv_2mortal (newSV (sizes [i])); 1594 SV *sv = sv_2mortal (newSV (sizes [i]));
1532 SvUPGRADE (sv, SVt_PV); 1595 SvUPGRADE (sv, SVt_PV);
1533 SvPOK_only (sv); 1596 SvPOK_only (sv);
1534 SvCUR_set (sv, sizes [i]); 1597 SvCUR_set (sv, sizes [i]);
1535 ptrs [i] = SvPVX (sv); 1598 ptrs [i] = (void *)SvPVX (sv);
1536 PUSHs (sv); 1599 PUSHs (sv);
1537 } 1600 }
1538 1601
1539 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 1602 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1540 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 1603 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
… …
1690set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1753set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1691 CODE: 1754 CODE:
1692 clSetKernelArg (self, idx, sizeof (value), &value); 1755 clSetKernelArg (self, idx, sizeof (value), &value);
1693 1756
1694void 1757void
1758set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1759 CODE:
1760 clSetKernelArg (self, idx, size, 0);
1761
1762void
1695set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1763set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1696 CODE: 1764 CODE:
1697 clSetKernelArg (self, idx, sizeof (value), &value); 1765 clSetKernelArg (self, idx, sizeof (value), &value);
1698 1766
1699void 1767void

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines