… | |
… | |
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> |
… | |
… | |
89 | enum { buffers = 3 }; |
99 | enum { buffers = 3 }; |
90 | static int idx; |
100 | static int idx; |
91 | static void *buf [buffers]; |
101 | static void *buf [buffers]; |
92 | static size_t len [buffers]; |
102 | static size_t len [buffers]; |
93 | |
103 | |
94 | idx = ++idx % buffers; |
104 | idx = (idx + 1) % buffers; |
95 | |
105 | |
96 | if (len [idx] < size) |
106 | if (len [idx] < size) |
97 | { |
107 | { |
98 | free (buf [idx]); |
108 | free (buf [idx]); |
99 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
109 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
324 | } |
334 | } |
325 | |
335 | |
326 | cl_int |
336 | cl_int |
327 | errno () |
337 | errno () |
328 | CODE: |
338 | CODE: |
329 | errno = res; |
339 | RETVAL = res; |
|
|
340 | OUTPUT: |
|
|
341 | RETVAL |
330 | |
342 | |
331 | const char * |
343 | const char * |
332 | err2str (cl_int err) |
344 | err2str (cl_int err) |
333 | |
345 | |
334 | const char * |
346 | const char * |
… | |
… | |
353 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
365 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
354 | PPCODE: |
366 | PPCODE: |
355 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
367 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
356 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
368 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
357 | |
369 | |
|
|
370 | #if 0 |
|
|
371 | |
358 | void |
372 | void |
359 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
373 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
360 | PPCODE: |
374 | PPCODE: |
361 | /* der Gipfel der Kunst */ |
375 | /* der Gipfel der Kunst */ |
|
|
376 | |
|
|
377 | #endif |
362 | |
378 | |
363 | void |
379 | void |
364 | wait_for_events (...) |
380 | wait_for_events (...) |
365 | CODE: |
381 | CODE: |
366 | EVENT_LIST (0, items); |
382 | EVENT_LIST (0, items); |
… | |
… | |
708 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
724 | 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?"); |
725 | 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)); |
726 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
711 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
727 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
712 | |
728 | |
|
|
729 | #if !defined CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
730 | |
713 | void |
731 | 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) |
732 | 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: |
733 | PPCODE: |
716 | STRLEN len; |
734 | STRLEN len; |
717 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
735 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
… | |
… | |
726 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
744 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
727 | const cl_image_format format = { channel_order, channel_type }; |
745 | 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)); |
746 | 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); |
747 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
730 | |
748 | |
|
|
749 | #endif |
|
|
750 | |
731 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
751 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
732 | |
752 | |
733 | void |
753 | void |
734 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
754 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
735 | PPCODE: |
755 | PPCODE: |
736 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
756 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
737 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
757 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
738 | |
758 | |
|
|
759 | #if !defined CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
760 | |
739 | void |
761 | void |
740 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
762 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
741 | PPCODE: |
763 | PPCODE: |
742 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
764 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
743 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
765 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
… | |
… | |
745 | void |
767 | void |
746 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
768 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
747 | PPCODE: |
769 | PPCODE: |
748 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
770 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
749 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
771 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
|
|
772 | |
|
|
773 | #endif |
750 | |
774 | |
751 | void |
775 | void |
752 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
776 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
753 | PPCODE: |
777 | PPCODE: |
754 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
778 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
… | |
… | |
941 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
965 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
942 | |
966 | |
943 | if (len < min_len) |
967 | if (len < min_len) |
944 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
968 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
945 | |
969 | |
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)); |
970 | 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 | |
971 | |
948 | if (ev) |
972 | if (ev) |
949 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
973 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
950 | |
974 | |
951 | void |
975 | void |
… | |
… | |
1006 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1030 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1007 | |
1031 | |
1008 | if (len < min_len) |
1032 | if (len < min_len) |
1009 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1033 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1010 | |
1034 | |
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)); |
1035 | 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 | |
1036 | |
1013 | if (ev) |
1037 | if (ev) |
1014 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1038 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1015 | |
1039 | |
1016 | void |
1040 | void |
… | |
… | |
1098 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1122 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1099 | } |
1123 | } |
1100 | |
1124 | |
1101 | if (SvOK (local_work_size)) |
1125 | if (SvOK (local_work_size)) |
1102 | { |
1126 | { |
1103 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1127 | 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"); |
1128 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1105 | |
1129 | |
1106 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1130 | 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"); |
1131 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1108 | |
1132 | |
… | |
… | |
1144 | if (ev) |
1168 | if (ev) |
1145 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1169 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1146 | |
1170 | |
1147 | #endif |
1171 | #endif |
1148 | |
1172 | |
|
|
1173 | #if !defined CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
1174 | |
1149 | void |
1175 | void |
1150 | enqueue_marker (OpenCL::Queue self) |
1176 | enqueue_marker (OpenCL::Queue self) |
1151 | PPCODE: |
1177 | PPCODE: |
1152 | cl_event ev; |
1178 | cl_event ev; |
1153 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
1179 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
… | |
… | |
1161 | |
1187 | |
1162 | void |
1188 | void |
1163 | enqueue_barrier (OpenCL::Queue self) |
1189 | enqueue_barrier (OpenCL::Queue self) |
1164 | CODE: |
1190 | CODE: |
1165 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1191 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1192 | |
|
|
1193 | #endif |
1166 | |
1194 | |
1167 | void |
1195 | void |
1168 | flush (OpenCL::Queue self) |
1196 | flush (OpenCL::Queue self) |
1169 | CODE: |
1197 | CODE: |
1170 | NEED_SUCCESS (Flush, (self)); |
1198 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1540 | { |
1568 | { |
1541 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1569 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1542 | SvUPGRADE (sv, SVt_PV); |
1570 | SvUPGRADE (sv, SVt_PV); |
1543 | SvPOK_only (sv); |
1571 | SvPOK_only (sv); |
1544 | SvCUR_set (sv, sizes [i]); |
1572 | SvCUR_set (sv, sizes [i]); |
1545 | ptrs [i] = SvPVX (sv); |
1573 | ptrs [i] = (void *)SvPVX (sv); |
1546 | PUSHs (sv); |
1574 | PUSHs (sv); |
1547 | } |
1575 | } |
1548 | |
1576 | |
1549 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1577 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1550 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
1578 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |