… | |
… | |
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; |
… | |
… | |
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 | |
261 | static cl_event * |
276 | static cl_event * |
262 | event_list (SV **items, int count) |
277 | event_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 | |
316 | cl_int |
341 | cl_int |
317 | errno () |
342 | errno () |
318 | CODE: |
343 | CODE: |
319 | errno = res; |
344 | RETVAL = res; |
|
|
345 | OUTPUT: |
|
|
346 | RETVAL |
320 | |
347 | |
321 | const char * |
348 | const char * |
322 | err2str (cl_int err) |
349 | err2str (cl_int err) |
323 | |
350 | |
324 | const char * |
351 | const char * |
… | |
… | |
343 | 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) |
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 | |
348 | void |
377 | void |
349 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
378 | context (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 | |
353 | void |
384 | void |
354 | wait_for_events (...) |
385 | wait_for_events (...) |
355 | CODE: |
386 | CODE: |
356 | EVENT_LIST (0, items); |
387 | EVENT_LIST (0, items); |
… | |
… | |
471 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
502 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
472 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
503 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
473 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
504 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
474 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
505 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
475 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
506 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
476 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
507 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
477 | PPCODE: |
508 | PPCODE: |
478 | cl_uint value [1]; |
509 | cl_uint value [1]; |
479 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
510 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
480 | EXTEND (SP, 1); |
511 | EXTEND (SP, 1); |
481 | const int i = 0; |
512 | const int i = 0; |
… | |
… | |
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 |
|
|
735 | |
703 | void |
736 | void |
704 | 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 | image (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, cl_mem_object_type type, size_t width, size_t height, size_t depth, size_t array_size = 0, size_t row_pitch = 0, size_t slice_pitch = 0, cl_uint num_mip_level = 0, cl_uint num_samples = 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; |
708 | const cl_image_format format = { channel_order, channel_type }; |
741 | const cl_image_format format = { channel_order, channel_type }; |
|
|
742 | const cl_image_desc desc = { |
|
|
743 | type, |
|
|
744 | width, height, depth, |
|
|
745 | array_size, row_pitch, slice_pitch, |
|
|
746 | num_mip_level, num_samples, |
|
|
747 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
748 | }; |
709 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
749 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
750 | char *klass = "OpenCL::Image"; |
|
|
751 | switch (type) |
|
|
752 | { |
|
|
753 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: klass = "OpenCL::Image1DBuffer"; break; |
|
|
754 | case CL_MEM_OBJECT_IMAGE1D: klass = "OpenCL::Image1D"; break; |
|
|
755 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
756 | case CL_MEM_OBJECT_IMAGE2D: klass = "OpenCL::Image2D"; break; |
|
|
757 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
758 | case CL_MEM_OBJECT_IMAGE3D: klass = "OpenCL::Image3D"; break; |
|
|
759 | } |
710 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
760 | XPUSH_NEW_OBJ (klass, mem); |
711 | |
761 | |
|
|
762 | #endif |
|
|
763 | |
|
|
764 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
765 | |
712 | void |
766 | void |
713 | image3d (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 depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef) |
767 | 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 | PPCODE: |
768 | PPCODE: |
715 | STRLEN len; |
769 | STRLEN len; |
716 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
770 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
717 | const cl_image_format format = { channel_order, channel_type }; |
771 | const cl_image_format format = { channel_order, channel_type }; |
|
|
772 | #if CL_VERSION_1_2 |
|
|
773 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
774 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
775 | #else |
|
|
776 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
777 | #endif |
|
|
778 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
779 | |
|
|
780 | void |
|
|
781 | image3d (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 depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef) |
|
|
782 | PPCODE: |
|
|
783 | STRLEN len; |
|
|
784 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
785 | const cl_image_format format = { channel_order, channel_type }; |
|
|
786 | #if CL_VERSION_1_2 |
|
|
787 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
788 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
789 | #else |
718 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
790 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
|
|
791 | #endif |
719 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
792 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
|
|
793 | |
|
|
794 | #endif |
720 | |
795 | |
721 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
796 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
722 | |
797 | |
723 | void |
798 | void |
724 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
799 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
725 | PPCODE: |
800 | PPCODE: |
726 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
801 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
727 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
802 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
728 | |
803 | |
729 | void |
804 | void |
|
|
805 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
806 | PPCODE: |
|
|
807 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
808 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
809 | |
|
|
810 | #if CL_VERSION_1_2 |
|
|
811 | |
|
|
812 | void |
|
|
813 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
814 | PPCODE: |
|
|
815 | char *klass = "OpenCL::Memory"; |
|
|
816 | cl_gl_object_type type; |
|
|
817 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
818 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); |
|
|
819 | switch (type) |
|
|
820 | { |
|
|
821 | case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break; |
|
|
822 | case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break; |
|
|
823 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
824 | case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break; |
|
|
825 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
826 | case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break; |
|
|
827 | } |
|
|
828 | XPUSH_NEW_OBJ (klass, mem); |
|
|
829 | |
|
|
830 | #endif |
|
|
831 | |
|
|
832 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
833 | |
|
|
834 | void |
730 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
835 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
731 | PPCODE: |
836 | PPCODE: |
732 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
837 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
733 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
838 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
734 | |
839 | |
… | |
… | |
736 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
841 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
737 | PPCODE: |
842 | PPCODE: |
738 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
843 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
739 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
844 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
740 | |
845 | |
741 | void |
846 | #endif |
742 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
743 | PPCODE: |
|
|
744 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
745 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
746 | |
847 | |
747 | #endif |
848 | #endif |
748 | |
849 | |
749 | void |
850 | void |
750 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
851 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
931 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1032 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
932 | |
1033 | |
933 | if (len < min_len) |
1034 | if (len < min_len) |
934 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1035 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
935 | |
1036 | |
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)); |
1037 | 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 | |
1038 | |
938 | if (ev) |
1039 | if (ev) |
939 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1040 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
940 | |
1041 | |
941 | void |
1042 | void |
… | |
… | |
996 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1097 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
997 | |
1098 | |
998 | if (len < min_len) |
1099 | if (len < min_len) |
999 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1100 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1000 | |
1101 | |
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)); |
1102 | 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 | |
1103 | |
1003 | if (ev) |
1104 | if (ev) |
1004 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1105 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1005 | |
1106 | |
1006 | void |
1107 | void |
… | |
… | |
1088 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1189 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1089 | } |
1190 | } |
1090 | |
1191 | |
1091 | if (SvOK (local_work_size)) |
1192 | if (SvOK (local_work_size)) |
1092 | { |
1193 | { |
1093 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1194 | 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"); |
1195 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1095 | |
1196 | |
1096 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1197 | 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"); |
1198 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1098 | |
1199 | |
… | |
… | |
1110 | |
1211 | |
1111 | void |
1212 | void |
1112 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1213 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1113 | ALIAS: |
1214 | ALIAS: |
1114 | enqueue_release_gl_objects = 1 |
1215 | enqueue_release_gl_objects = 1 |
1115 | CODE: |
1216 | PPCODE: |
1116 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
1217 | 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"); |
1218 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
1118 | |
1219 | |
1119 | cl_event ev = 0; |
1220 | cl_event ev = 0; |
1120 | EVENT_LIST (2, items - 2); |
1221 | EVENT_LIST (2, items - 2); |
… | |
… | |
1134 | if (ev) |
1235 | if (ev) |
1135 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1236 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1136 | |
1237 | |
1137 | #endif |
1238 | #endif |
1138 | |
1239 | |
1139 | void |
1240 | #if !CL_VERSION_1_2 || defined CL_USE_DEPRECATED_OPENCL_1_1_APIS |
1140 | enqueue_marker (OpenCL::Queue self) |
|
|
1141 | PPCODE: |
|
|
1142 | cl_event ev; |
|
|
1143 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1144 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1145 | |
1241 | |
1146 | void |
1242 | void |
1147 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1243 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1148 | CODE: |
1244 | CODE: |
1149 | EVENT_LIST (1, items - 1); |
1245 | EVENT_LIST (1, items - 1); |
1150 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1246 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1151 | |
1247 | |
|
|
1248 | #endif |
|
|
1249 | |
1152 | void |
1250 | void |
|
|
1251 | enqueue_marker (OpenCL::Queue self, ...) |
|
|
1252 | PPCODE: |
|
|
1253 | cl_event ev = 0; |
|
|
1254 | EVENT_LIST (1, items - 1); |
|
|
1255 | #if CL_VERSION_1_2 |
|
|
1256 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1257 | #else |
|
|
1258 | if (event_list_count) |
|
|
1259 | croak ("OpenCL::Queue->enqueue_marker does not support a wait list in OpenCL 1.1 - upgrade to 1.2"); |
|
|
1260 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1261 | #endif |
|
|
1262 | if (ev) |
|
|
1263 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1264 | |
|
|
1265 | void |
1153 | enqueue_barrier (OpenCL::Queue self) |
1266 | enqueue_barrier (OpenCL::Queue self, ...) |
1154 | CODE: |
1267 | PPCODE: |
|
|
1268 | cl_event ev = 0; |
|
|
1269 | EVENT_LIST (1, items - 1); |
|
|
1270 | #if CL_VERSION_1_2 |
|
|
1271 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, &ev)); |
|
|
1272 | #else |
|
|
1273 | if (event_list_count) |
|
|
1274 | croak ("OpenCL::Queue->enqueue_barrier does not support a wait list in OpenCL 1.1 - upgrade to 1.2"); |
|
|
1275 | if (GIMME_V != G_VOID) |
|
|
1276 | croak ("OpenCL::Queue->enqueue_barrier does not return an event object in OpenCL 1.1 - upgrade to 1.2"); |
1155 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1277 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1278 | #endif |
|
|
1279 | if (ev) |
|
|
1280 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1156 | |
1281 | |
1157 | void |
1282 | void |
1158 | flush (OpenCL::Queue self) |
1283 | flush (OpenCL::Queue self) |
1159 | CODE: |
1284 | CODE: |
1160 | NEED_SUCCESS (Flush, (self)); |
1285 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1530 | { |
1655 | { |
1531 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1656 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1532 | SvUPGRADE (sv, SVt_PV); |
1657 | SvUPGRADE (sv, SVt_PV); |
1533 | SvPOK_only (sv); |
1658 | SvPOK_only (sv); |
1534 | SvCUR_set (sv, sizes [i]); |
1659 | SvCUR_set (sv, sizes [i]); |
1535 | ptrs [i] = SvPVX (sv); |
1660 | ptrs [i] = (void *)SvPVX (sv); |
1536 | PUSHs (sv); |
1661 | PUSHs (sv); |
1537 | } |
1662 | } |
1538 | |
1663 | |
1539 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1664 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1540 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
1665 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |