… | |
… | |
2 | #include "perl.h" |
2 | #include "perl.h" |
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 |
|
|
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_2_APIS /* just guessing, you stupid idiots */ |
|
|
17 | |
|
|
18 | #ifndef PREFER_1_1 |
|
|
19 | #define PREFER_1_1 1 |
|
|
20 | #endif |
|
|
21 | |
|
|
22 | #if PREFER_1_1 |
|
|
23 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
7 | #endif |
24 | #endif |
8 | |
25 | |
9 | #ifdef __APPLE__ |
26 | #ifdef __APPLE__ |
10 | #include <OpenCL/opencl.h> |
27 | #include <OpenCL/opencl.h> |
11 | #else |
28 | #else |
12 | #include <CL/opencl.h> |
29 | #include <CL/opencl.h> |
|
|
30 | #endif |
|
|
31 | |
|
|
32 | #ifndef CL_VERSION_1_2 |
|
|
33 | #undef PREFER_1_1 |
|
|
34 | #define PREFER_1_1 1 |
13 | #endif |
35 | #endif |
14 | |
36 | |
15 | typedef cl_platform_id OpenCL__Platform; |
37 | typedef cl_platform_id OpenCL__Platform; |
16 | typedef cl_device_id OpenCL__Device; |
38 | typedef cl_device_id OpenCL__Device; |
17 | typedef cl_context OpenCL__Context; |
39 | typedef cl_context OpenCL__Context; |
… | |
… | |
89 | enum { buffers = 3 }; |
111 | enum { buffers = 3 }; |
90 | static int idx; |
112 | static int idx; |
91 | static void *buf [buffers]; |
113 | static void *buf [buffers]; |
92 | static size_t len [buffers]; |
114 | static size_t len [buffers]; |
93 | |
115 | |
94 | idx = ++idx % buffers; |
116 | idx = (idx + 1) % buffers; |
95 | |
117 | |
96 | if (len [idx] < size) |
118 | if (len [idx] < size) |
97 | { |
119 | { |
98 | free (buf [idx]); |
120 | free (buf [idx]); |
99 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
121 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
257 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
279 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
258 | return res; |
280 | return res; |
259 | } |
281 | } |
260 | |
282 | |
261 | static cl_event * |
283 | static cl_event * |
262 | event_list (SV **items, int count) |
284 | event_list (SV **items, cl_uint *rcount) |
263 | { |
285 | { |
|
|
286 | cl_uint count = *rcount; |
|
|
287 | |
264 | if (!count) |
288 | if (!count) |
265 | return 0; |
289 | return 0; |
266 | |
290 | |
267 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
291 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
|
|
292 | int i = 0; |
268 | |
293 | |
269 | while (count--) |
294 | do |
|
|
295 | { |
|
|
296 | --count; |
|
|
297 | if (SvOK (items [count])) |
270 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
298 | list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
|
|
299 | } |
|
|
300 | while (count); |
271 | |
301 | |
|
|
302 | *rcount = i; |
|
|
303 | |
272 | return list; |
304 | return i ? list : 0; |
273 | } |
305 | } |
274 | |
306 | |
275 | #define EVENT_LIST(items,count) \ |
307 | #define EVENT_LIST(items,count) \ |
276 | cl_uint event_list_count = (count); \ |
308 | cl_uint event_list_count = (count); \ |
277 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
309 | cl_event *event_list_ptr = event_list (&ST (items), &event_list_count) |
278 | |
310 | |
279 | #define INFO(class) \ |
311 | #define INFO(class) \ |
280 | { \ |
312 | { \ |
281 | size_t size; \ |
313 | size_t size; \ |
282 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
314 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
314 | } |
346 | } |
315 | |
347 | |
316 | cl_int |
348 | cl_int |
317 | errno () |
349 | errno () |
318 | CODE: |
350 | CODE: |
319 | errno = res; |
351 | RETVAL = res; |
|
|
352 | OUTPUT: |
|
|
353 | RETVAL |
320 | |
354 | |
321 | const char * |
355 | const char * |
322 | err2str (cl_int err) |
356 | err2str (cl_int err) |
323 | |
357 | |
324 | const char * |
358 | const char * |
… | |
… | |
343 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
377 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
344 | PPCODE: |
378 | PPCODE: |
345 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
379 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
346 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
380 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
347 | |
381 | |
|
|
382 | #if 0 |
|
|
383 | |
348 | void |
384 | void |
349 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
385 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
350 | PPCODE: |
386 | PPCODE: |
351 | /* der Gipfel der Kunst */ |
387 | /* der Gipfel der Kunst */ |
|
|
388 | |
|
|
389 | #endif |
352 | |
390 | |
353 | void |
391 | void |
354 | wait_for_events (...) |
392 | wait_for_events (...) |
355 | CODE: |
393 | CODE: |
356 | EVENT_LIST (0, items); |
394 | EVENT_LIST (0, items); |
… | |
… | |
362 | |
400 | |
363 | void |
401 | void |
364 | info (OpenCL::Platform self, cl_platform_info name) |
402 | info (OpenCL::Platform self, cl_platform_info name) |
365 | PPCODE: |
403 | PPCODE: |
366 | INFO (Platform) |
404 | INFO (Platform) |
|
|
405 | |
|
|
406 | void |
|
|
407 | unload_compiler (OpenCL::Platform self) |
|
|
408 | CODE: |
|
|
409 | #if CL_VERSION_1_2 |
|
|
410 | clUnloadPlatformCompiler (self); |
|
|
411 | #endif |
367 | |
412 | |
368 | #BEGIN:platform |
413 | #BEGIN:platform |
369 | |
414 | |
370 | void |
415 | void |
371 | profile (OpenCL::Platform self) |
416 | profile (OpenCL::Platform self) |
… | |
… | |
471 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
516 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
472 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
517 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
473 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
518 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
474 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
519 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
475 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
520 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
476 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
521 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
477 | PPCODE: |
522 | PPCODE: |
478 | cl_uint value [1]; |
523 | cl_uint value [1]; |
479 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
524 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
480 | EXTEND (SP, 1); |
525 | EXTEND (SP, 1); |
481 | const int i = 0; |
526 | const int i = 0; |
… | |
… | |
698 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
743 | 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?"); |
744 | 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)); |
745 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
701 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
746 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
702 | |
747 | |
|
|
748 | #if CL_VERSION_1_2 |
|
|
749 | |
703 | void |
750 | 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) |
751 | 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: |
752 | PPCODE: |
706 | STRLEN len; |
753 | STRLEN len; |
707 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
754 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
708 | const cl_image_format format = { channel_order, channel_type }; |
755 | const cl_image_format format = { channel_order, channel_type }; |
|
|
756 | const cl_image_desc desc = { |
|
|
757 | type, |
|
|
758 | width, height, depth, |
|
|
759 | array_size, row_pitch, slice_pitch, |
|
|
760 | num_mip_level, num_samples, |
|
|
761 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
762 | }; |
709 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
763 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
764 | char *klass = "OpenCL::Image"; |
|
|
765 | switch (type) |
|
|
766 | { |
|
|
767 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: klass = "OpenCL::Image1DBuffer"; break; |
|
|
768 | case CL_MEM_OBJECT_IMAGE1D: klass = "OpenCL::Image1D"; break; |
|
|
769 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
770 | case CL_MEM_OBJECT_IMAGE2D: klass = "OpenCL::Image2D"; break; |
|
|
771 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
772 | case CL_MEM_OBJECT_IMAGE3D: klass = "OpenCL::Image3D"; break; |
|
|
773 | } |
710 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
774 | XPUSH_NEW_OBJ (klass, mem); |
711 | |
775 | |
|
|
776 | #endif |
|
|
777 | |
712 | void |
778 | 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) |
779 | 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: |
780 | PPCODE: |
715 | STRLEN len; |
781 | STRLEN len; |
716 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
782 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
717 | const cl_image_format format = { channel_order, channel_type }; |
783 | const cl_image_format format = { channel_order, channel_type }; |
|
|
784 | #if PREFER_1_1 |
|
|
785 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
786 | #else |
|
|
787 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
788 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
789 | #endif |
|
|
790 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
791 | |
|
|
792 | void |
|
|
793 | 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) |
|
|
794 | PPCODE: |
|
|
795 | STRLEN len; |
|
|
796 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
797 | const cl_image_format format = { channel_order, channel_type }; |
|
|
798 | #if PREFER_1_1 |
718 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
799 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
|
|
800 | #else |
|
|
801 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
802 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
803 | #endif |
719 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
804 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
720 | |
805 | |
721 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
806 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
722 | |
807 | |
723 | void |
808 | void |
… | |
… | |
725 | PPCODE: |
810 | PPCODE: |
726 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
811 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
727 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
812 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
728 | |
813 | |
729 | void |
814 | void |
730 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
731 | PPCODE: |
|
|
732 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
733 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
734 | |
|
|
735 | void |
|
|
736 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
737 | PPCODE: |
|
|
738 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
739 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
|
|
740 | |
|
|
741 | void |
|
|
742 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
815 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
743 | PPCODE: |
816 | PPCODE: |
744 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
817 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
745 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
818 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
819 | |
|
|
820 | #if CL_VERSION_1_2 |
|
|
821 | |
|
|
822 | void |
|
|
823 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
824 | ALIAS: |
|
|
825 | PPCODE: |
|
|
826 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
827 | cl_gl_object_type type; |
|
|
828 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
829 | char *klass = "OpenCL::Memory"; |
|
|
830 | switch (type) |
|
|
831 | { |
|
|
832 | case CL_GL_OBJECT_TEXTURE_BUFFER: klass = "OpenCL::Image1DBuffer"; break; |
|
|
833 | case CL_GL_OBJECT_TEXTURE1D: klass = "OpenCL::Image1D"; break; |
|
|
834 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
835 | case CL_GL_OBJECT_TEXTURE2D: klass = "OpenCL::Image2D"; break; |
|
|
836 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: klass = "OpenCL::Image2DArray"; break; |
|
|
837 | case CL_GL_OBJECT_TEXTURE3D: klass = "OpenCL::Image3D"; break; |
|
|
838 | } |
|
|
839 | XPUSH_NEW_OBJ (klass, mem); |
|
|
840 | |
|
|
841 | #endif |
|
|
842 | |
|
|
843 | void |
|
|
844 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
845 | PPCODE: |
|
|
846 | #if PREFER_1_1 |
|
|
847 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
848 | #else |
|
|
849 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
850 | #endif |
|
|
851 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
852 | |
|
|
853 | void |
|
|
854 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
855 | PPCODE: |
|
|
856 | #if PREFER_1_1 |
|
|
857 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
858 | #else |
|
|
859 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
860 | #endif |
|
|
861 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
746 | |
862 | |
747 | #endif |
863 | #endif |
748 | |
864 | |
749 | void |
865 | void |
750 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
866 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
855 | cl_event ev = 0; |
971 | cl_event ev = 0; |
856 | STRLEN len; |
972 | STRLEN len; |
857 | char *ptr = SvPVbyte (data, len); |
973 | char *ptr = SvPVbyte (data, len); |
858 | EVENT_LIST (5, items - 5); |
974 | EVENT_LIST (5, items - 5); |
859 | |
975 | |
860 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
976 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
861 | |
977 | |
862 | if (ev) |
978 | if (ev) |
863 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
979 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
980 | |
|
|
981 | #if CL_VERSION_1_2 |
|
|
982 | |
|
|
983 | void |
|
|
984 | enqueue_fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
985 | PPCODE: |
|
|
986 | cl_event ev = 0; |
|
|
987 | STRLEN len; |
|
|
988 | char *ptr = SvPVbyte (data, len); |
|
|
989 | EVENT_LIST (5, items - 5); |
|
|
990 | |
|
|
991 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
992 | |
|
|
993 | if (ev) |
|
|
994 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
995 | |
|
|
996 | void |
|
|
997 | enqueue_fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...) |
|
|
998 | PPCODE: |
|
|
999 | cl_event ev = 0; |
|
|
1000 | STRLEN len; |
|
|
1001 | const size_t origin [3] = { x, y, z }; |
|
|
1002 | const size_t region [3] = { width, height, depth }; |
|
|
1003 | EVENT_LIST (12, items - 12); |
|
|
1004 | |
|
|
1005 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1006 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1007 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1008 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1009 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1010 | cl_image_format format; |
|
|
1011 | NEED_SUCCESS (clGetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1012 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1013 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1014 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1015 | |
|
|
1016 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type]], |
|
|
1017 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1018 | |
|
|
1019 | if (ev) |
|
|
1020 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1021 | |
|
|
1022 | #endif |
864 | |
1023 | |
865 | void |
1024 | void |
866 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1025 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
867 | PPCODE: |
1026 | PPCODE: |
868 | cl_event ev = 0; |
1027 | cl_event ev = 0; |
… | |
… | |
931 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1090 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
932 | |
1091 | |
933 | if (len < min_len) |
1092 | if (len < min_len) |
934 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1093 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
935 | |
1094 | |
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)); |
1095 | 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 | |
1096 | |
938 | if (ev) |
1097 | if (ev) |
939 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1098 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
940 | |
1099 | |
941 | void |
1100 | void |
… | |
… | |
996 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1155 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
997 | |
1156 | |
998 | if (len < min_len) |
1157 | if (len < min_len) |
999 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1158 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1000 | |
1159 | |
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)); |
1160 | 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 | |
1161 | |
1003 | if (ev) |
1162 | if (ev) |
1004 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1163 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1005 | |
1164 | |
1006 | void |
1165 | void |
… | |
… | |
1088 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1247 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1089 | } |
1248 | } |
1090 | |
1249 | |
1091 | if (SvOK (local_work_size)) |
1250 | if (SvOK (local_work_size)) |
1092 | { |
1251 | { |
1093 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1252 | 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"); |
1253 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1095 | |
1254 | |
1096 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1255 | 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"); |
1256 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1098 | |
1257 | |
… | |
… | |
1110 | |
1269 | |
1111 | void |
1270 | void |
1112 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1271 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1113 | ALIAS: |
1272 | ALIAS: |
1114 | enqueue_release_gl_objects = 1 |
1273 | enqueue_release_gl_objects = 1 |
1115 | CODE: |
1274 | PPCODE: |
1116 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
1275 | 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"); |
1276 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
1118 | |
1277 | |
1119 | cl_event ev = 0; |
1278 | cl_event ev = 0; |
1120 | EVENT_LIST (2, items - 2); |
1279 | EVENT_LIST (2, items - 2); |
… | |
… | |
1135 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1294 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1136 | |
1295 | |
1137 | #endif |
1296 | #endif |
1138 | |
1297 | |
1139 | void |
1298 | void |
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 | |
|
|
1146 | void |
|
|
1147 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1299 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1148 | CODE: |
1300 | CODE: |
1149 | EVENT_LIST (1, items - 1); |
1301 | EVENT_LIST (1, items - 1); |
|
|
1302 | #if PREFER_1_1 |
1150 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1303 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1304 | #else |
|
|
1305 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
1306 | #endif |
1151 | |
1307 | |
1152 | void |
1308 | void |
|
|
1309 | enqueue_marker (OpenCL::Queue self, ...) |
|
|
1310 | PPCODE: |
|
|
1311 | cl_event ev = 0; |
|
|
1312 | EVENT_LIST (1, items - 1); |
|
|
1313 | #if PREFER_1_1 |
|
|
1314 | if (!event_list_count) |
|
|
1315 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1316 | else |
|
|
1317 | #if CL_VERSION_1_2 |
|
|
1318 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1319 | #else |
|
|
1320 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); // also a barrier |
|
|
1321 | #endif |
|
|
1322 | #else |
|
|
1323 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1324 | #endif |
|
|
1325 | if (ev) |
|
|
1326 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1327 | |
|
|
1328 | void |
1153 | enqueue_barrier (OpenCL::Queue self) |
1329 | enqueue_barrier (OpenCL::Queue self, ...) |
1154 | CODE: |
1330 | PPCODE: |
|
|
1331 | cl_event ev = 0; |
|
|
1332 | EVENT_LIST (1, items - 1); |
|
|
1333 | #if PREFER_1_1 |
|
|
1334 | if (!event_list_count && GIMME_V == G_VOID) |
1155 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1335 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1336 | else |
|
|
1337 | #if CL_VERSION_1_2 |
|
|
1338 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1339 | #else |
|
|
1340 | { |
|
|
1341 | if (event_list_count) |
|
|
1342 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1343 | |
|
|
1344 | if (GIMME_V != G_VOID) |
|
|
1345 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1346 | } |
|
|
1347 | #endif |
|
|
1348 | #else |
|
|
1349 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1350 | #endif |
|
|
1351 | if (ev) |
|
|
1352 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1156 | |
1353 | |
1157 | void |
1354 | void |
1158 | flush (OpenCL::Queue self) |
1355 | flush (OpenCL::Queue self) |
1159 | CODE: |
1356 | CODE: |
1160 | NEED_SUCCESS (Flush, (self)); |
1357 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1505 | PPCODE: |
1702 | PPCODE: |
1506 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1703 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1507 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
1704 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
1508 | |
1705 | |
1509 | void |
1706 | void |
|
|
1707 | kernels_in_program (OpenCL::Program program) |
|
|
1708 | PPCODE: |
|
|
1709 | cl_uint num_kernels; |
|
|
1710 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
1711 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
1712 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
1713 | |
|
|
1714 | int i; |
|
|
1715 | EXTEND (SP, num_kernels); |
|
|
1716 | for (i = 0; i < num_kernels; ++i) |
|
|
1717 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Kernel", kernels [i])); |
|
|
1718 | |
|
|
1719 | void |
1510 | info (OpenCL::Program self, cl_program_info name) |
1720 | info (OpenCL::Program self, cl_program_info name) |
1511 | PPCODE: |
1721 | PPCODE: |
1512 | INFO (Program) |
1722 | INFO (Program) |
1513 | |
1723 | |
1514 | void |
1724 | void |
… | |
… | |
1530 | { |
1740 | { |
1531 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1741 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1532 | SvUPGRADE (sv, SVt_PV); |
1742 | SvUPGRADE (sv, SVt_PV); |
1533 | SvPOK_only (sv); |
1743 | SvPOK_only (sv); |
1534 | SvCUR_set (sv, sizes [i]); |
1744 | SvCUR_set (sv, sizes [i]); |
1535 | ptrs [i] = SvPVX (sv); |
1745 | ptrs [i] = (void *)SvPVX (sv); |
1536 | PUSHs (sv); |
1746 | PUSHs (sv); |
1537 | } |
1747 | } |
1538 | |
1748 | |
1539 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1749 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1540 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
1750 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |