… | |
… | |
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 | |
15 | typedef cl_platform_id OpenCL__Platform; |
30 | typedef cl_platform_id OpenCL__Platform; |
16 | typedef cl_device_id OpenCL__Device; |
31 | typedef cl_device_id OpenCL__Device; |
17 | typedef cl_context OpenCL__Context; |
32 | typedef 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; |
… | |
… | |
324 | } |
339 | } |
325 | |
340 | |
326 | cl_int |
341 | cl_int |
327 | errno () |
342 | errno () |
328 | CODE: |
343 | CODE: |
329 | errno = res; |
344 | RETVAL = res; |
|
|
345 | OUTPUT: |
|
|
346 | RETVAL |
330 | |
347 | |
331 | const char * |
348 | const char * |
332 | err2str (cl_int err) |
349 | err2str (cl_int err) |
333 | |
350 | |
334 | const char * |
351 | const char * |
… | |
… | |
353 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
370 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
354 | PPCODE: |
371 | PPCODE: |
355 | 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)); |
356 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
373 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
357 | |
374 | |
|
|
375 | #if 0 |
|
|
376 | |
358 | void |
377 | void |
359 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
378 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
360 | PPCODE: |
379 | PPCODE: |
361 | /* der Gipfel der Kunst */ |
380 | /* der Gipfel der Kunst */ |
|
|
381 | |
|
|
382 | #endif |
362 | |
383 | |
363 | void |
384 | void |
364 | wait_for_events (...) |
385 | wait_for_events (...) |
365 | CODE: |
386 | CODE: |
366 | EVENT_LIST (0, items); |
387 | EVENT_LIST (0, items); |
… | |
… | |
708 | 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))) |
709 | 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?"); |
710 | 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)); |
711 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
732 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
712 | |
733 | |
|
|
734 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
735 | |
713 | void |
736 | void |
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) |
737 | 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) |
715 | PPCODE: |
738 | PPCODE: |
716 | STRLEN len; |
739 | STRLEN len; |
717 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
740 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
… | |
… | |
726 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
749 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
727 | const cl_image_format format = { channel_order, channel_type }; |
750 | const cl_image_format format = { channel_order, channel_type }; |
728 | 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)); |
729 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
752 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
730 | |
753 | |
|
|
754 | #endif |
|
|
755 | |
731 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
756 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
732 | |
757 | |
733 | void |
758 | void |
734 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
759 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
735 | PPCODE: |
760 | PPCODE: |
736 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
761 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
737 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
762 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
738 | |
763 | |
739 | void |
764 | void |
|
|
765 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
766 | PPCODE: |
|
|
767 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
768 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
769 | |
|
|
770 | #if CL_VERSION_1_2 |
|
|
771 | |
|
|
772 | void |
|
|
773 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
774 | PPCODE: |
|
|
775 | char *klass = "OpenCL::Memory"; |
|
|
776 | cl_gl_object_type t; |
|
|
777 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
778 | NEED_SUCCESS (GetGLObjectInfo, (mem, &t, 0)); |
|
|
779 | switch (t) |
|
|
780 | { |
|
|
781 | case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break; |
|
|
782 | case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break; |
|
|
783 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
784 | case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break; |
|
|
785 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
786 | case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break; |
|
|
787 | } |
|
|
788 | XPUSH_NEW_OBJ (klass, mem); |
|
|
789 | |
|
|
790 | #endif |
|
|
791 | |
|
|
792 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
793 | |
|
|
794 | void |
740 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
795 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
741 | PPCODE: |
796 | PPCODE: |
742 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
797 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
743 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
798 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
744 | |
799 | |
… | |
… | |
746 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
801 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
747 | PPCODE: |
802 | PPCODE: |
748 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
803 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
749 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
804 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
750 | |
805 | |
751 | void |
806 | #endif |
752 | gl_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 | |
807 | |
757 | #endif |
808 | #endif |
758 | |
809 | |
759 | void |
810 | void |
760 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
811 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
941 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
992 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
942 | |
993 | |
943 | if (len < min_len) |
994 | if (len < min_len) |
944 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
995 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
945 | |
996 | |
946 | 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)); |
997 | 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)); |
947 | |
998 | |
948 | if (ev) |
999 | if (ev) |
949 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1000 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
950 | |
1001 | |
951 | void |
1002 | void |
… | |
… | |
1006 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1057 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1007 | |
1058 | |
1008 | if (len < min_len) |
1059 | if (len < min_len) |
1009 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1060 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1010 | |
1061 | |
1011 | 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)); |
1062 | 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)); |
1012 | |
1063 | |
1013 | if (ev) |
1064 | if (ev) |
1014 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1065 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1015 | |
1066 | |
1016 | void |
1067 | void |
… | |
… | |
1098 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1149 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1099 | } |
1150 | } |
1100 | |
1151 | |
1101 | if (SvOK (local_work_size)) |
1152 | if (SvOK (local_work_size)) |
1102 | { |
1153 | { |
1103 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1154 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1104 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1155 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1105 | |
1156 | |
1106 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1157 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1107 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1158 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1108 | |
1159 | |
… | |
… | |
1120 | |
1171 | |
1121 | void |
1172 | void |
1122 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1173 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1123 | ALIAS: |
1174 | ALIAS: |
1124 | enqueue_release_gl_objects = 1 |
1175 | enqueue_release_gl_objects = 1 |
1125 | CODE: |
1176 | PPCODE: |
1126 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
1177 | 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"); |
1178 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
1128 | |
1179 | |
1129 | cl_event ev = 0; |
1180 | cl_event ev = 0; |
1130 | EVENT_LIST (2, items - 2); |
1181 | EVENT_LIST (2, items - 2); |
… | |
… | |
1144 | if (ev) |
1195 | if (ev) |
1145 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1196 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1146 | |
1197 | |
1147 | #endif |
1198 | #endif |
1148 | |
1199 | |
1149 | void |
1200 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
1150 | enqueue_marker (OpenCL::Queue self) |
|
|
1151 | PPCODE: |
|
|
1152 | cl_event ev; |
|
|
1153 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1154 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1155 | |
1201 | |
1156 | void |
1202 | void |
1157 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1203 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1158 | CODE: |
1204 | CODE: |
1159 | EVENT_LIST (1, items - 1); |
1205 | EVENT_LIST (1, items - 1); |
1160 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1206 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1161 | |
1207 | |
|
|
1208 | #endif |
|
|
1209 | |
1162 | void |
1210 | void |
|
|
1211 | enqueue_marker (OpenCL::Queue self, ...) |
|
|
1212 | PPCODE: |
|
|
1213 | cl_event ev = 0; |
|
|
1214 | EVENT_LIST (1, items - 1); |
|
|
1215 | #if CL_VERSION_1_2 |
|
|
1216 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1217 | #else |
|
|
1218 | if (event_list_count) |
|
|
1219 | croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2"); |
|
|
1220 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1221 | #endif |
|
|
1222 | if (ev) |
|
|
1223 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1224 | |
|
|
1225 | void |
1163 | enqueue_barrier (OpenCL::Queue self) |
1226 | enqueue_barrier (OpenCL::Queue self, ...) |
1164 | CODE: |
1227 | PPCODE: |
|
|
1228 | cl_event ev = 0; |
|
|
1229 | EVENT_LIST (1, items - 1); |
|
|
1230 | #if CL_VERSION_1_2 |
|
|
1231 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev)); |
|
|
1232 | #else |
|
|
1233 | if (event_list_count) |
|
|
1234 | croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2"); |
|
|
1235 | if (GIMME_V != G_VOID) |
|
|
1236 | croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2"); |
1165 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1237 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1238 | #endif |
|
|
1239 | if (ev) |
|
|
1240 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1166 | |
1241 | |
1167 | void |
1242 | void |
1168 | flush (OpenCL::Queue self) |
1243 | flush (OpenCL::Queue self) |
1169 | CODE: |
1244 | CODE: |
1170 | NEED_SUCCESS (Flush, (self)); |
1245 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1540 | { |
1615 | { |
1541 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1616 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1542 | SvUPGRADE (sv, SVt_PV); |
1617 | SvUPGRADE (sv, SVt_PV); |
1543 | SvPOK_only (sv); |
1618 | SvPOK_only (sv); |
1544 | SvCUR_set (sv, sizes [i]); |
1619 | SvCUR_set (sv, sizes [i]); |
1545 | ptrs [i] = SvPVX (sv); |
1620 | ptrs [i] = (void *)SvPVX (sv); |
1546 | PUSHs (sv); |
1621 | PUSHs (sv); |
1547 | } |
1622 | } |
1548 | |
1623 | |
1549 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1624 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1550 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
1625 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |