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> |
… | |
… | |
31 | |
35 | |
32 | typedef SV *FUTURE; |
36 | typedef SV *FUTURE; |
33 | |
37 | |
34 | /*****************************************************************************/ |
38 | /*****************************************************************************/ |
35 | |
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. |
|
|
43 | static void * |
|
|
44 | glsym (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 | |
36 | /* up to two temporary buffers */ |
85 | /* up to two temporary buffers */ |
37 | static void * |
86 | static void * |
38 | tmpbuf (size_t size) |
87 | tmpbuf (size_t size) |
39 | { |
88 | { |
40 | enum { buffers = 3 }; |
89 | enum { buffers = 3 }; |
… | |
… | |
144 | return 0; |
193 | return 0; |
145 | |
194 | |
146 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
195 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
147 | { |
196 | { |
148 | AV *av = (AV *)SvRV (sv); |
197 | AV *av = (AV *)SvRV (sv); |
149 | int i, len = av_len (av); |
198 | int i, len = av_len (av) + 1; |
150 | 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)); |
151 | cl_context_properties *l = p; |
200 | cl_context_properties *l = p; |
152 | |
201 | |
153 | if (len & 1) |
202 | if (len & 1) |
154 | 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); |
… | |
… | |
156 | while (extracount--) |
205 | while (extracount--) |
157 | *l++ = *extra++; |
206 | *l++ = *extra++; |
158 | |
207 | |
159 | for (i = 0; i < len; i += 2) |
208 | for (i = 0; i < len; i += 2) |
160 | { |
209 | { |
161 | cl_context_properties t = SvIV (*av_fetch (av, i, 0)); |
210 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
162 | 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 |
163 | |
213 | |
164 | switch (t) |
214 | switch (t) |
165 | { |
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 | |
166 | default: |
234 | default: |
167 | /* unknown property, treat as int */ |
235 | /* unknown property, treat as int */ |
168 | v = SvIV (*av_fetch (av, i, 0)); |
|
|
169 | break; |
236 | break; |
170 | } |
237 | } |
171 | |
238 | |
172 | *l++ = t; |
239 | *l++ = t; |
173 | *l++ = v; |
240 | *l++ = v; |
… | |
… | |
190 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
257 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
191 | return res; |
258 | return res; |
192 | } |
259 | } |
193 | |
260 | |
194 | static cl_event * |
261 | static cl_event * |
195 | event_list (SV **items, int count) |
262 | event_list (SV **items, cl_uint *rcount) |
196 | { |
263 | { |
|
|
264 | cl_uint count = *rcount; |
|
|
265 | |
197 | if (!count) |
266 | if (!count) |
198 | return 0; |
267 | return 0; |
199 | |
268 | |
200 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
269 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
|
|
270 | int i = 0; |
201 | |
271 | |
202 | while (count--) |
272 | do |
|
|
273 | { |
|
|
274 | --count; |
|
|
275 | if (SvOK (items [count])) |
203 | 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); |
204 | |
279 | |
|
|
280 | *rcount = i; |
|
|
281 | |
205 | return list; |
282 | return i ? list : 0; |
206 | } |
283 | } |
207 | |
284 | |
208 | #define EVENT_LIST(items,count) \ |
285 | #define EVENT_LIST(items,count) \ |
209 | cl_uint event_list_count = (count); \ |
286 | cl_uint event_list_count = (count); \ |
210 | 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) |
211 | |
288 | |
212 | #define INFO(class) \ |
289 | #define INFO(class) \ |
213 | { \ |
290 | { \ |
214 | size_t size; \ |
291 | size_t size; \ |
215 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
292 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
336 | |
413 | |
337 | void |
414 | void |
338 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
415 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
339 | PPCODE: |
416 | PPCODE: |
340 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
417 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
341 | 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"); |
342 | |
419 | |
343 | AV *av = (AV *)SvRV (devices); |
420 | AV *av = (AV *)SvRV (devices); |
344 | cl_uint num_devices = av_len (av) + 1; |
421 | cl_uint num_devices = av_len (av) + 1; |
345 | 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); |
346 | int i; |
423 | int i; |
… | |
… | |
616 | |
693 | |
617 | void |
694 | void |
618 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
695 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
619 | PPCODE: |
696 | PPCODE: |
620 | 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)) |
621 | 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?"); |
622 | |
699 | |
623 | 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)); |
624 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
701 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
625 | |
702 | |
626 | void |
703 | void |
627 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
704 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
628 | PPCODE: |
705 | PPCODE: |
629 | STRLEN len; |
706 | STRLEN len; |
630 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
707 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
631 | 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))) |
632 | 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?"); |
633 | 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)); |
634 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
711 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
635 | |
712 | |
636 | void |
713 | void |
637 | image2d (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) |
714 | image2d (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) |
… | |
… | |
672 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
749 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
673 | |
750 | |
674 | void |
751 | void |
675 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
752 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
676 | PPCODE: |
753 | PPCODE: |
677 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderBuffer, (self, flags, renderbuffer, &res)); |
754 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
678 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
755 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
679 | |
756 | |
680 | #endif |
757 | #endif |
681 | |
758 | |
682 | void |
759 | void |
… | |
… | |
788 | cl_event ev = 0; |
865 | cl_event ev = 0; |
789 | STRLEN len; |
866 | STRLEN len; |
790 | char *ptr = SvPVbyte (data, len); |
867 | char *ptr = SvPVbyte (data, len); |
791 | EVENT_LIST (5, items - 5); |
868 | EVENT_LIST (5, items - 5); |
792 | |
869 | |
793 | 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)); |
794 | |
871 | |
795 | if (ev) |
872 | if (ev) |
796 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
873 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
797 | |
874 | |
798 | void |
875 | void |
… | |
… | |
1037 | 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)); |
1038 | |
1115 | |
1039 | if (ev) |
1116 | if (ev) |
1040 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1117 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1041 | |
1118 | |
1042 | #if 0 |
|
|
1043 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1119 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1044 | |
1120 | |
1045 | void |
1121 | void |
1046 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1122 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
1123 | ALIAS: |
|
|
1124 | enqueue_release_gl_objects = 1 |
1047 | CODE: |
1125 | CODE: |
|
|
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 | |
1048 | cl_event ev = 0; |
1129 | cl_event ev = 0; |
1049 | EVENT_LIST (2, items - 2); |
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; |
1050 | |
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 |
1051 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, ..., event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1142 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1052 | |
1143 | |
1053 | if (ev) |
1144 | if (ev) |
1054 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1145 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1055 | |
1146 | |
1056 | void |
|
|
1057 | enqueue_release_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
1058 | CODE: |
|
|
1059 | cl_event ev = 0; |
|
|
1060 | EVENT_LIST (2, items - 2); |
|
|
1061 | |
|
|
1062 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, ..., event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1063 | |
|
|
1064 | if (ev) |
|
|
1065 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1066 | |
|
|
1067 | #endif |
|
|
1068 | #endif |
1147 | #endif |
1069 | |
1148 | |
1070 | void |
1149 | void |
1071 | enqueue_marker (OpenCL::Queue self) |
1150 | enqueue_marker (OpenCL::Queue self) |
1072 | PPCODE: |
1151 | PPCODE: |
… | |
… | |
1233 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1312 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1234 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
1313 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
1235 | } |
1314 | } |
1236 | |
1315 | |
1237 | #END:mem |
1316 | #END:mem |
|
|
1317 | |
|
|
1318 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1319 | |
|
|
1320 | void |
|
|
1321 | gl_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 |
1238 | |
1331 | |
1239 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
1332 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
1240 | |
1333 | |
1241 | void |
1334 | void |
1242 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
1335 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
… | |
… | |
1273 | EXTEND (SP, 1); |
1366 | EXTEND (SP, 1); |
1274 | const int i = 0; |
1367 | const int i = 0; |
1275 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1368 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1276 | |
1369 | |
1277 | #END:image |
1370 | #END:image |
|
|
1371 | |
|
|
1372 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1373 | |
|
|
1374 | #BEGIN:gl_texture |
|
|
1375 | |
|
|
1376 | void |
|
|
1377 | target (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 | |
|
|
1385 | void |
|
|
1386 | gl_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 |
1278 | |
1397 | |
1279 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
1398 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
1280 | |
1399 | |
1281 | void |
1400 | void |
1282 | DESTROY (OpenCL::Sampler self) |
1401 | DESTROY (OpenCL::Sampler self) |
… | |
… | |
1581 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
1700 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
1582 | CODE: |
1701 | CODE: |
1583 | clSetKernelArg (self, idx, sizeof (value), &value); |
1702 | clSetKernelArg (self, idx, sizeof (value), &value); |
1584 | |
1703 | |
1585 | void |
1704 | void |
|
|
1705 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
|
|
1706 | CODE: |
|
|
1707 | clSetKernelArg (self, idx, size, 0); |
|
|
1708 | |
|
|
1709 | void |
1586 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1710 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1587 | CODE: |
1711 | CODE: |
1588 | clSetKernelArg (self, idx, sizeof (value), &value); |
1712 | clSetKernelArg (self, idx, sizeof (value), &value); |
1589 | |
1713 | |
1590 | void |
1714 | void |