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 1.36 by root, Sat Apr 21 17:56:21 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
4 8
5#ifdef __APPLE__ 9#ifdef __APPLE__
6 #include <OpenCL/opencl.h> 10 #include <OpenCL/opencl.h>
7#else 11#else
8 #include <CL/opencl.h> 12 #include <CL/opencl.h>
9#endif
10#if HAVE_OPENGL
11 #ifdef __APPLE__
12 #include <OpenGL/gl.h>
13 #include <OpenCL/cl_gl.h>
14 #else
15 #include <GL/gl.h>
16 #include <CL/cl_gl.h>
17 #endif
18#endif 13#endif
19 14
20typedef cl_platform_id OpenCL__Platform; 15typedef cl_platform_id OpenCL__Platform;
21typedef cl_device_id OpenCL__Device; 16typedef cl_device_id OpenCL__Device;
22typedef cl_context OpenCL__Context; 17typedef cl_context OpenCL__Context;
40 35
41typedef SV *FUTURE; 36typedef SV *FUTURE;
42 37
43/*****************************************************************************/ 38/*****************************************************************************/
44 39
40// name must include a leading underscore
41// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
42// for perl. doh.
43static void *
44glsym (const char *name)
45{
46 void *fun = 0;
47
48 #if defined I_DLFCN && defined RTLD_DEFAULT
49 fun = dlsym (RTLD_DEFAULT, name + 1);
50 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
51
52 if (!fun)
53 {
54 static void *libgl;
55 static const char *glso[] = {
56 "libGL.so.1",
57 "libGL.so.3",
58 "libGL.so.4.0",
59 "libGL.so",
60 "/usr/lib/libGL.so",
61 "/usr/X11R6/lib/libGL.1.dylib"
62 };
63 int i;
64
65 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
66 {
67 libgl = dlopen (glso [i], RTLD_LAZY);
68 if (libgl)
69 break;
70 }
71
72 if (libgl)
73 {
74 fun = dlsym (libgl, name + 1);
75 if (!fun) fun = dlsym (libgl, name);
76 }
77 }
78 #endif
79
80 return fun;
81}
82
83/*****************************************************************************/
84
45/* up to two temporary buffers */ 85/* up to two temporary buffers */
46static void * 86static void *
47tmpbuf (size_t size) 87tmpbuf (size_t size)
48{ 88{
49 enum { buffers = 3 }; 89 enum { buffers = 3 };
153 return 0; 193 return 0;
154 194
155 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) 195 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
156 { 196 {
157 AV *av = (AV *)SvRV (sv); 197 AV *av = (AV *)SvRV (sv);
158 int i, len = av_len (av); 198 int i, len = av_len (av) + 1;
159 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); 199 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
160 cl_context_properties *l = p; 200 cl_context_properties *l = p;
161 201
162 if (len & 1) 202 if (len & 1)
163 croak ("%s: %s is not a property list (must be even number of elements)", func, svname); 203 croak ("%s: %s is not a property list (must be even number of elements)", func, svname);
165 while (extracount--) 205 while (extracount--)
166 *l++ = *extra++; 206 *l++ = *extra++;
167 207
168 for (i = 0; i < len; i += 2) 208 for (i = 0; i < len; i += 2)
169 { 209 {
170 cl_context_properties t = SvIV (*av_fetch (av, i, 0)); 210 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
171 cl_context_properties v; 211 SV *p_sv = *av_fetch (av, i + 1, 0);
212 cl_context_properties v = SvIV (p_sv); // code below can override
172 213
173 switch (t) 214 switch (t)
174 { 215 {
216 case CL_GLX_DISPLAY_KHR:
217 if (!SvOK (p_sv))
218 {
219 void *func = glsym ("_glXGetCurrentDisplay");
220 if (func)
221 v = (cl_context_properties)((void *(*)(void))func)();
222 }
223 break;
224
225 case CL_GL_CONTEXT_KHR:
226 if (!SvOK (p_sv))
227 {
228 void *func = glsym ("_glXGetCurrentContext");
229 if (func)
230 v = (cl_context_properties)((void *(*)(void))func)();
231 }
232 break;
233
175 default: 234 default:
176 /* unknown property, treat as int */ 235 /* unknown property, treat as int */
177 v = SvIV (*av_fetch (av, i, 0));
178 break; 236 break;
179 } 237 }
180 238
181 *l++ = t; 239 *l++ = t;
182 *l++ = v; 240 *l++ = v;
199 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 257 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
200 return res; 258 return res;
201} 259}
202 260
203static cl_event * 261static cl_event *
204event_list (SV **items, int count) 262event_list (SV **items, cl_uint *rcount)
205{ 263{
264 cl_uint count = *rcount;
265
206 if (!count) 266 if (!count)
207 return 0; 267 return 0;
208 268
209 cl_event *list = tmpbuf (sizeof (cl_event) * count); 269 cl_event *list = tmpbuf (sizeof (cl_event) * count);
270 int i = 0;
210 271
211 while (count--) 272 do
273 {
274 --count;
275 if (SvOK (items [count]))
212 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 276 list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
277 }
278 while (count);
213 279
280 *rcount = i;
281
214 return list; 282 return i ? list : 0;
215} 283}
216 284
217#define EVENT_LIST(items,count) \ 285#define EVENT_LIST(items,count) \
218 cl_uint event_list_count = (count); \ 286 cl_uint event_list_count = (count); \
219 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 287 cl_event *event_list_ptr = event_list (&ST (items), &event_list_count)
220 288
221#define INFO(class) \ 289#define INFO(class) \
222{ \ 290{ \
223 size_t size; \ 291 size_t size; \
224 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 292 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
247 { sizeof (cl_long ), "SIZEOF_LONG" }, 315 { sizeof (cl_long ), "SIZEOF_LONG" },
248 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 316 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
249 { sizeof (cl_half ), "SIZEOF_HALF" }, 317 { sizeof (cl_half ), "SIZEOF_HALF" },
250 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 318 { sizeof (cl_float ), "SIZEOF_FLOAT" },
251 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 319 { sizeof (cl_double), "SIZEOF_DOUBLE" },
252 { HAVE_OPENGL , "HAVE_OPENGL" },
253#include "constiv.h" 320#include "constiv.h"
254 }; 321 };
255 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 322 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)); 323 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
257} 324}
346 413
347void 414void
348context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) 415context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0)
349 PPCODE: 416 PPCODE:
350 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 417 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
351 croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); 418 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
352 419
353 AV *av = (AV *)SvRV (devices); 420 AV *av = (AV *)SvRV (devices);
354 cl_uint num_devices = av_len (av) + 1; 421 cl_uint num_devices = av_len (av) + 1;
355 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 422 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
356 int i; 423 int i;
626 693
627void 694void
628buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) 695buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
629 PPCODE: 696 PPCODE:
630 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 697 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?"); 698 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
632 699
633 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); 700 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
634 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 701 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
635 702
636void 703void
637buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) 704buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
638 PPCODE: 705 PPCODE:
639 STRLEN len; 706 STRLEN len;
640 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 707 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
641 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 708 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?"); 709 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)); 710 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
644 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 711 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
645 712
646void 713void
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) 714image2d (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)
658 STRLEN len; 725 STRLEN len;
659 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 726 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
660 const cl_image_format format = { channel_order, channel_type }; 727 const cl_image_format format = { channel_order, channel_type };
661 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 728 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
662 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 729 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
730
731#if cl_apple_gl_sharing || cl_khr_gl_sharing
732
733void
734gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
735 PPCODE:
736 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
737 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
738
739void
740gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
741 PPCODE:
742 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
743 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
744
745void
746gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
747 PPCODE:
748 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
749 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
750
751void
752gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
753 PPCODE:
754 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
755 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
756
757#endif
663 758
664void 759void
665supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) 760supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
666 PPCODE: 761 PPCODE:
667{ 762{
770 cl_event ev = 0; 865 cl_event ev = 0;
771 STRLEN len; 866 STRLEN len;
772 char *ptr = SvPVbyte (data, len); 867 char *ptr = SvPVbyte (data, len);
773 EVENT_LIST (5, items - 5); 868 EVENT_LIST (5, items - 5);
774 869
775 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 870 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
776 871
777 if (ev) 872 if (ev)
778 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 873 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
779 874
780void 875void
1019 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1114 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1020 1115
1021 if (ev) 1116 if (ev)
1022 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1117 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1023 1118
1119#if cl_apple_gl_sharing || cl_khr_gl_sharing
1120
1121void
1122enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1123 ALIAS:
1124 enqueue_release_gl_objects = 1
1125 PPCODE:
1126 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1127 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1128
1129 cl_event ev = 0;
1130 EVENT_LIST (2, items - 2);
1131 AV *av = (AV *)SvRV (objects);
1132 cl_uint num_objects = av_len (av) + 1;
1133 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1134 int i;
1135
1136 for (i = num_objects; i--; )
1137 object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1138
1139 if (ix)
1140 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1141 else
1142 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1143
1144 if (ev)
1145 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
1146
1147#endif
1148
1024void 1149void
1025enqueue_marker (OpenCL::Queue self) 1150enqueue_marker (OpenCL::Queue self)
1026 PPCODE: 1151 PPCODE:
1027 cl_event ev; 1152 cl_event ev;
1028 NEED_SUCCESS (EnqueueMarker, (self, &ev)); 1153 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1187 NEED_SUCCESS (RetainMemObject, (value [i])); 1312 NEED_SUCCESS (RetainMemObject, (value [i]));
1188 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); 1313 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i]));
1189 } 1314 }
1190 1315
1191#END:mem 1316#END:mem
1317
1318#if cl_apple_gl_sharing || cl_khr_gl_sharing
1319
1320void
1321gl_object_info (OpenCL::Memory self)
1322 PPCODE:
1323 cl_gl_object_type type;
1324 cl_GLuint name;
1325 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1326 EXTEND (SP, 2);
1327 PUSHs (sv_2mortal (newSVuv (type)));
1328 PUSHs (sv_2mortal (newSVuv (name)));
1329
1330#endif
1192 1331
1193MODULE = OpenCL PACKAGE = OpenCL::BufferObj 1332MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1194 1333
1195void 1334void
1196sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) 1335sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1227 EXTEND (SP, 1); 1366 EXTEND (SP, 1);
1228 const int i = 0; 1367 const int i = 0;
1229 PUSHs (sv_2mortal (newSVuv (value [i]))); 1368 PUSHs (sv_2mortal (newSVuv (value [i])));
1230 1369
1231#END:image 1370#END:image
1371
1372#if cl_apple_gl_sharing || cl_khr_gl_sharing
1373
1374#BEGIN:gl_texture
1375
1376void
1377target (OpenCL::Image self)
1378 PPCODE:
1379 cl_GLenum value [1];
1380 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
1381 EXTEND (SP, 1);
1382 const int i = 0;
1383 PUSHs (sv_2mortal (newSVuv (value [i])));
1384
1385void
1386gl_mipmap_level (OpenCL::Image self)
1387 PPCODE:
1388 cl_GLint value [1];
1389 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
1390 EXTEND (SP, 1);
1391 const int i = 0;
1392 PUSHs (sv_2mortal (newSViv (value [i])));
1393
1394#END:gl_texture
1395
1396#endif
1232 1397
1233MODULE = OpenCL PACKAGE = OpenCL::Sampler 1398MODULE = OpenCL PACKAGE = OpenCL::Sampler
1234 1399
1235void 1400void
1236DESTROY (OpenCL::Sampler self) 1401DESTROY (OpenCL::Sampler self)
1535set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 1700set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1536 CODE: 1701 CODE:
1537 clSetKernelArg (self, idx, sizeof (value), &value); 1702 clSetKernelArg (self, idx, sizeof (value), &value);
1538 1703
1539void 1704void
1705set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1706 CODE:
1707 clSetKernelArg (self, idx, size, 0);
1708
1709void
1540set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 1710set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1541 CODE: 1711 CODE:
1542 clSetKernelArg (self, idx, sizeof (value), &value); 1712 clSetKernelArg (self, idx, sizeof (value), &value);
1543 1713
1544void 1714void

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines