ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/OpenCL/OpenCL.xs
(Generate patch)

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.68 by root, Fri May 4 14:56:50 2012 UTC vs.
Revision 1.76 by root, Sat May 5 14:49:52 2012 UTC

28#if PREFER_1_1 28#if PREFER_1_1
29 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS 29 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
30#endif 30#endif
31 31
32#ifdef __APPLE__ 32#ifdef __APPLE__
33 #include <OpenCL/opencl.h> 33 #define CLHDR(name) <OpenCL/name>
34#else 34#else
35 #define CLHDR(name) <CL/name>
36#endif
37
35 #include <CL/opencl.h> 38#include CLHDR(opencl.h)
39
40#ifndef CL_VERSION_1_2
41 #include CLHDR(cl_d3d9.h)
42#endif
43
44#if _WIN32
45 #include CLHDR(cl_d3d10.h)
46 #if CL_VERSION_1_2
47 #include CLHDR<cl_d3d11.h>
48 #endif
49 #include CLHDR<cl_dx9_media_sharing.h.h>
36#endif 50#endif
37 51
38#ifndef CL_VERSION_1_2 52#ifndef CL_VERSION_1_2
39 #undef PREFER_1_1 53 #undef PREFER_1_1
40 #define PREFER_1_1 1 54 #define PREFER_1_1 1
41#endif 55#endif
56
57// make sure all constants we might use are actually defined
58#include "default.h"
42 59
43typedef cl_platform_id OpenCL__Platform; 60typedef cl_platform_id OpenCL__Platform;
44typedef cl_device_id OpenCL__Device; 61typedef cl_device_id OpenCL__Device;
45typedef cl_device_id OpenCL__SubDevice; 62typedef cl_device_id OpenCL__SubDevice;
46typedef cl_context OpenCL__Context; 63typedef cl_context OpenCL__Context;
57typedef cl_kernel OpenCL__Kernel; 74typedef cl_kernel OpenCL__Kernel;
58typedef cl_event OpenCL__Event; 75typedef cl_event OpenCL__Event;
59typedef cl_event OpenCL__UserEvent; 76typedef cl_event OpenCL__UserEvent;
60 77
61typedef struct mapped * OpenCL__Mapped; 78typedef struct mapped * OpenCL__Mapped;
62
63typedef SV *FUTURE;
64 79
65static HV 80static HV
66 *stash_platform, 81 *stash_platform,
67 *stash_device, 82 *stash_device,
68 *stash_subdevice, 83 *stash_subdevice,
153 } 168 }
154 169
155 return buf [idx]; 170 return buf [idx];
156} 171}
157 172
173static const char * ecb_noinline
174cv_get_name (CV *cv)
175{
176 static char fullname [256];
177
178 GV *gv = CvGV (cv); // gv better be nonzero
179
180 HV *stash = GvSTASH (gv);
181 const char *hvname = HvNAME_get (stash); // stash also better be nonzero
182 const char *gvname = GvNAME (gv);
183
184 snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname);
185 return fullname;
186}
187
158/*****************************************************************************/ 188/*****************************************************************************/
159 189
160typedef struct 190typedef struct
161{ 191{
162 IV iv; 192 IV iv;
163 const char *name; 193 const char *name;
164 #define const_iv(name) { (IV)CL_ ## name, # name }, 194 #define const_iv(name) { (IV)CL_ ## name, # name },
165} ivstr; 195} ivstr;
196
197typedef struct
198{
199 NV nv;
200 const char *name;
201 #define const_nv(name) { (NV)CL_ ## name, # name },
202} nvstr;
166 203
167static const char * 204static const char *
168iv2str (IV value, const ivstr *base, int count, const char *fallback) 205iv2str (IV value, const ivstr *base, int count, const char *fallback)
169{ 206{
170 int i; 207 int i;
231#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) 268#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
232 269
233/* cl objects are either \$iv, or [$iv, ...] */ 270/* cl objects are either \$iv, or [$iv, ...] */
234/* they can be upgraded at runtime to the array form */ 271/* they can be upgraded at runtime to the array form */
235static void * ecb_noinline 272static void * ecb_noinline
236SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) 273SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
237{ 274{
238 // sv_derived_from is quite slow :( 275 // sv_derived_from is quite slow :(
239 if (SvROK (sv) && sv_derived_from (sv, pkg)) 276 if (SvROK (sv) && sv_derived_from (sv, pkg))
240 return (void *)SvIV (SvRV (sv)); 277 return (void *)SvIV (SvRV (sv));
241 278
242 croak ("%s: %s is not of type %s", func, svname, pkg); 279 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg);
243} 280}
244 281
245// the "no-inherit" version of the above 282// the "no-inherit" version of the above
246static void * ecb_noinline 283static void * ecb_noinline
247SvCLOBJ_ni (const char *func, const char *svname, SV *sv, HV *stash) 284SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash)
248{ 285{
249 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) 286 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash)
250 return (void *)SvIV (SvRV (sv)); 287 return (void *)SvIV (SvRV (sv));
251 288
252 croak ("%s: %s is not of type %s", func, svname, HvNAME (stash)); 289 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash));
253} 290}
254 291
255/*****************************************************************************/ 292/*****************************************************************************/
256 293
257static cl_context_properties * ecb_noinline 294static cl_context_properties * ecb_noinline
258SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) 295SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
259{ 296{
260 if (!sv || !SvOK (sv)) 297 if (!sv || !SvOK (sv))
261 if (extra) 298 if (extra)
262 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully 299 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
263 else 300 else
269 int i, len = av_len (av) + 1; 306 int i, len = av_len (av) + 1;
270 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); 307 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
271 cl_context_properties *l = p; 308 cl_context_properties *l = p;
272 309
273 if (len & 1) 310 if (len & 1)
274 croak ("%s: %s is not a property list (must contain an even number of elements)", func, svname); 311 croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname);
275 312
276 while (extracount--) 313 while (extracount--)
277 *l++ = *extra++; 314 *l++ = *extra++;
278 315
279 for (i = 0; i < len; i += 2) 316 for (i = 0; i < len; i += 2)
284 321
285 switch (t) 322 switch (t)
286 { 323 {
287 case CL_CONTEXT_PLATFORM: 324 case CL_CONTEXT_PLATFORM:
288 if (SvROK (p_sv)) 325 if (SvROK (p_sv))
289 v = (cl_context_properties)SvCLOBJ (func, svname, p_sv, "OpenCL::Platform"); 326 v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform");
290 break; 327 break;
291 328
292 case CL_GLX_DISPLAY_KHR: 329 case CL_GLX_DISPLAY_KHR:
293 if (!SvOK (p_sv)) 330 if (!SvOK (p_sv))
294 { 331 {
319 *l = 0; 356 *l = 0;
320 357
321 return p; 358 return p;
322 } 359 }
323 360
324 croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); 361 croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname);
362}
363
364// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
365// is a pointer (and only on well-behaved systems).
366static void * ecb_noinline
367object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
368{
369 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
370 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
371 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
372
373 AV *av = (AV *)SvRV (arg);
374 void **list = 0;
375 cl_uint count = av_len (av) + 1;
376
377 if (count)
378 {
379 list = tmpbuf (sizeof (*list) * count);
380 int i;
381 for (i = 0; i < count; ++i)
382 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
383 }
384
385 if (!count && !or_undef)
386 croak ("%s: '%s' must contain at least one %s object",
387 cv_get_name (cv), argname, klass);
388
389 *rcount = count;
390 return (void *)list;
325} 391}
326 392
327/*****************************************************************************/ 393/*****************************************************************************/
328/* callback stuff */ 394/* callback stuff */
329 395
489static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; 555static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
490 556
491static void CL_CALLBACK 557static void CL_CALLBACK
492eq_program_notify (cl_program program, void *user_data) 558eq_program_notify (cl_program program, void *user_data)
493{ 559{
560 clRetainProgram (program);
561
494 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); 562 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
563}
564
565typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data);
566
567static program_callback ecb_noinline
568make_program_callback (SV *notify, void **ruser_data)
569{
570 if (SvOK (notify))
571 {
572 *ruser_data = SvREFCNT_inc (s_get_cv (notify));
573 return eq_program_notify;
574 }
575 else
576 {
577 *ruser_data = 0;
578 return 0;
579 }
495} 580}
496 581
497struct build_args 582struct build_args
498{ 583{
499 cl_program program; 584 cl_program program;
565 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 650 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
566 return res; 651 return res;
567} 652}
568 653
569static cl_event * ecb_noinline 654static cl_event * ecb_noinline
570event_list (SV **items, cl_uint *rcount, cl_event extra) 655event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
571{ 656{
572 cl_uint count = *rcount; 657 cl_uint count = *rcount;
573 658
574 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed 659 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
575 *rcount = count = 0; 660 *rcount = count = 0;
580 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); 665 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
581 int i = 0; 666 int i = 0;
582 667
583 while (count--) 668 while (count--)
584 if (SvOK (items [count])) 669 if (SvOK (items [count]))
585 list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 670 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
586 671
587 if (extra) 672 if (extra)
588 list [i++] = extra; 673 list [i++] = extra;
589 674
590 *rcount = i; 675 *rcount = i;
592 return i ? list : 0; 677 return i ? list : 0;
593} 678}
594 679
595#define EVENT_LIST(skip) \ 680#define EVENT_LIST(skip) \
596 cl_uint event_list_count = items - (skip); \ 681 cl_uint event_list_count = items - (skip); \
597 cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) 682 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
598 683
599#define INFO(class) \ 684#define INFO(class) \
600{ \ 685{ \
601 size_t size; \ 686 size_t size; \
602 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 687 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
684 769
685 mapped->ptr = 0; 770 mapped->ptr = 0;
686} 771}
687 772
688static void 773static void
689mapped_unmap (SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) 774mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count)
690{ 775{
691 cl_event *event_list_ptr = event_list (wait_list, &event_list_count, mapped->event); 776 cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event);
692 cl_event ev; 777 cl_event ev;
693 778
694 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); 779 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev));
695 780
696 clReleaseEvent (mapped->event); 781 clReleaseEvent (mapped->event);
717 eq_signal_arg = (void*)arg; 802 eq_signal_arg = (void*)arg;
718 803
719BOOT: 804BOOT:
720{ 805{
721 HV *stash = gv_stashpv ("OpenCL", 1); 806 HV *stash = gv_stashpv ("OpenCL", 1);
807
722 static const ivstr *civ, const_iv[] = { 808 static const ivstr *civ, const_iv[] = {
723 { sizeof (cl_char ), "SIZEOF_CHAR" }, 809 { sizeof (cl_char ), "SIZEOF_CHAR" },
724 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 810 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
725 { sizeof (cl_short ), "SIZEOF_SHORT" }, 811 { sizeof (cl_short ), "SIZEOF_SHORT" },
726 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 812 { sizeof (cl_ushort), "SIZEOF_USHORT" },
729 { sizeof (cl_long ), "SIZEOF_LONG" }, 815 { sizeof (cl_long ), "SIZEOF_LONG" },
730 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 816 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
731 { sizeof (cl_half ), "SIZEOF_HALF" }, 817 { sizeof (cl_half ), "SIZEOF_HALF" },
732 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 818 { sizeof (cl_float ), "SIZEOF_FLOAT" },
733 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 819 { sizeof (cl_double), "SIZEOF_DOUBLE" },
820 { PREFER_1_1 , "PREFER_1_1" },
734#include "constiv.h" 821#include "constiv.h"
735 }; 822 };
736 823
737 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 824 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
738 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 825 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
826
827 static const nvstr *cnv, const_nv[] = {
828#include "constnv.h"
829 };
830
831 for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--)
832 newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv));
833
834 newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call
739 835
740 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); 836 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
741 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); 837 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
742 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); 838 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD);
743 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); 839 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
798 CONTEXT_NOTIFY_CALLBACK; 894 CONTEXT_NOTIFY_CALLBACK;
799 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); 895 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
800 XPUSH_CLOBJ_CONTEXT; 896 XPUSH_CLOBJ_CONTEXT;
801 897
802void 898void
803context (FUTURE properties, FUTURE devices, FUTURE notify) 899context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
804 PPCODE: 900 PPCODE:
805 /* der Gipfel der Kunst */ 901 cl_uint device_count;
902 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
903
904 CONTEXT_NOTIFY_CALLBACK;
905 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
906 XPUSH_CLOBJ_CONTEXT;
806 907
807void 908void
808wait_for_events (...) 909wait_for_events (...)
809 CODE: 910 CODE:
810 EVENT_LIST (0); 911 EVENT_LIST (0);
861 EXTEND (SP, count); 962 EXTEND (SP, count);
862 for (i = 0; i < count; ++i) 963 for (i = 0; i < count; ++i)
863 PUSH_CLOBJ (stash_device, list [i]); 964 PUSH_CLOBJ (stash_device, list [i]);
864 965
865void 966void
866context (OpenCL::Platform self, SV *properties = 0, SV *devices, SV *notify = &PL_sv_undef) 967context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
867 PPCODE: 968 PPCODE:
868 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
869 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
870
871 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; 969 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
872 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context", "properties", properties, extra, 2); 970 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
873 971
874 AV *av = (AV *)SvRV (devices); 972 cl_uint device_count;
875 cl_uint num_devices = av_len (av) + 1; 973 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
876 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
877
878 int i;
879 for (i = num_devices; i--; )
880 device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
881 974
882 CONTEXT_NOTIFY_CALLBACK; 975 CONTEXT_NOTIFY_CALLBACK;
883 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, num_devices, device_list, pfn_notify, user_data, &res)); 976 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
884 XPUSH_CLOBJ_CONTEXT; 977 XPUSH_CLOBJ_CONTEXT;
885 978
886void 979void
887context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) 980context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
888 PPCODE: 981 PPCODE:
889 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; 982 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
890 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); 983 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
891 984
892 CONTEXT_NOTIFY_CALLBACK; 985 CONTEXT_NOTIFY_CALLBACK;
893 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); 986 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
894 XPUSH_CLOBJ_CONTEXT; 987 XPUSH_CLOBJ_CONTEXT;
895 988
1219 const cl_image_desc desc = { 1312 const cl_image_desc desc = {
1220 type, 1313 type,
1221 width, height, depth, 1314 width, height, depth,
1222 array_size, row_pitch, slice_pitch, 1315 array_size, row_pitch, slice_pitch,
1223 num_mip_level, num_samples, 1316 num_mip_level, num_samples,
1224 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 1317 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1225 }; 1318 };
1226 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); 1319 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1227 HV *stash = stash_image; 1320 HV *stash = stash_image;
1228 switch (type) 1321 switch (type)
1229 { 1322 {
1365 XPUSH_CLOBJ (stash_program, prog); 1458 XPUSH_CLOBJ (stash_program, prog);
1366 1459
1367void 1460void
1368program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) 1461program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1369 PPCODE: 1462 PPCODE:
1370 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 1463 cl_uint device_count;
1371 croak ("OpenCL::Context::program_with_binary: devices must be specified as reference to an array of device objects"); 1464 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1372
1373 devices = SvRV (devices);
1374 1465
1375 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) 1466 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1376 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); 1467 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1377 1468
1378 binaries = SvRV (binaries); 1469 binaries = SvRV (binaries);
1379 1470
1380 if (av_len ((AV *)devices) != av_len ((AV *)binaries)) 1471 if (device_count != av_len ((AV *)binaries) + 1)
1381 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); 1472 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1382 1473
1383 int count = av_len ((AV *)devices) + 1;
1384 cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count);
1385 size_t *length_list = tmpbuf (sizeof (*length_list) * count); 1474 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1386 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * count); 1475 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1387 cl_int *status_list = tmpbuf (sizeof (*status_list) * count); 1476 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1388 1477
1389 int i; 1478 int i;
1390 for (i = 0; i < count; ++i) 1479 for (i = 0; i < device_count; ++i)
1391 { 1480 {
1392 device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_binary", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device");
1393 STRLEN len; 1481 STRLEN len;
1394 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); 1482 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1395 length_list [i] = len; 1483 length_list [i] = len;
1396 } 1484 }
1397 1485
1398 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, count, device_list, length_list, binary_list, GIMME_V == G_ARRAY ? status_list : 0, &res)); 1486 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1487 length_list, binary_list,
1488 GIMME_V == G_ARRAY ? status_list : 0, &res));
1399 1489
1400 EXTEND (SP, 2); 1490 EXTEND (SP, 2);
1401 PUSH_CLOBJ (stash_program, prog); 1491 PUSH_CLOBJ (stash_program, prog);
1402 1492
1403 if (GIMME_V == G_ARRAY) 1493 if (GIMME_V == G_ARRAY)
1404 { 1494 {
1405 AV *av = newAV (); 1495 AV *av = newAV ();
1406 PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); 1496 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1407 1497
1408 for (i = count; i--; ) 1498 for (i = device_count; i--; )
1409 av_store (av, i, newSViv (status_list [i])); 1499 av_store (av, i, newSViv (status_list [i]));
1410 } 1500 }
1411 1501
1412#if CL_VERSION_1_2 1502#if CL_VERSION_1_2
1413 1503
1414void 1504void
1415program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) 1505program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1416 PPCODE: 1506 PPCODE:
1417 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 1507 cl_uint device_count;
1418 croak ("OpenCL::Context::program_with_built_in_kernels: devices must be specified as reference to an array of device objects"); 1508 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1419 1509
1420 devices = SvRV (devices);
1421
1422 int count = av_len ((AV *)devices) + 1;
1423 cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count);
1424
1425 int i;
1426 for (i = 0; i < count; ++i)
1427 device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_built_in_kernels", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device");
1428
1429 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, count, device_list, SvPVbyte_nolen (kernel_names), &res)); 1510 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1511
1512 XPUSH_CLOBJ (stash_program, prog);
1513
1514void
1515link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1516 CODE:
1517 cl_uint device_count = 0;
1518 cl_device_id *device_list = 0;
1519
1520 if (SvOK (devices))
1521 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1522
1523 cl_uint program_count;
1524 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1525
1526 void *user_data;
1527 program_callback pfn_notify = make_program_callback (notify, &user_data);
1528
1529 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1530 program_count, program_list, pfn_notify, user_data, &res));
1430 1531
1431 XPUSH_CLOBJ (stash_program, prog); 1532 XPUSH_CLOBJ (stash_program, prog);
1432 1533
1433#endif 1534#endif
1434 1535
1484void 1585void
1485read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) 1586read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1486 ALIAS: 1587 ALIAS:
1487 enqueue_read_buffer = 0 1588 enqueue_read_buffer = 0
1488 PPCODE: 1589 PPCODE:
1489 cl_event ev = 0;
1490 EVENT_LIST (6); 1590 EVENT_LIST (6);
1491 1591
1492 SvUPGRADE (data, SVt_PV); 1592 SvUPGRADE (data, SVt_PV);
1493 SvGROW (data, len); 1593 SvGROW (data, len);
1494 SvPOK_only (data); 1594 SvPOK_only (data);
1495 SvCUR_set (data, len); 1595 SvCUR_set (data, len);
1596
1597 cl_event ev = 0;
1496 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1598 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1497 1599
1498 if (ev) 1600 if (ev)
1499 XPUSH_CLOBJ (stash_event, ev); 1601 XPUSH_CLOBJ (stash_event, ev);
1500 1602
1501void 1603void
1502write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1604write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1503 ALIAS: 1605 ALIAS:
1504 enqueue_write_buffer = 0 1606 enqueue_write_buffer = 0
1505 PPCODE: 1607 PPCODE:
1506 cl_event ev = 0; 1608 EVENT_LIST (5);
1609
1507 STRLEN len; 1610 STRLEN len;
1508 char *ptr = SvPVbyte (data, len); 1611 char *ptr = SvPVbyte (data, len);
1509 EVENT_LIST (5);
1510 1612
1613 cl_event ev = 0;
1511 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1614 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1512 1615
1513 if (ev) 1616 if (ev)
1514 XPUSH_CLOBJ (stash_event, ev); 1617 XPUSH_CLOBJ (stash_event, ev);
1515 1618
1518void 1621void
1519fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) 1622fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1520 ALIAS: 1623 ALIAS:
1521 enqueue_fill_buffer = 0 1624 enqueue_fill_buffer = 0
1522 PPCODE: 1625 PPCODE:
1523 cl_event ev = 0; 1626 EVENT_LIST (5);
1627
1524 STRLEN len; 1628 STRLEN len;
1525 char *ptr = SvPVbyte (data, len); 1629 char *ptr = SvPVbyte (data, len);
1526 EVENT_LIST (5);
1527 1630
1631 cl_event ev = 0;
1528 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1632 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1529 1633
1530 if (ev) 1634 if (ev)
1531 XPUSH_CLOBJ (stash_event, ev); 1635 XPUSH_CLOBJ (stash_event, ev);
1532 1636
1533void 1637void
1534fill_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, ...) 1638fill_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, ...)
1535 ALIAS: 1639 ALIAS:
1536 enqueue_fill_image = 0 1640 enqueue_fill_image = 0
1537 PPCODE: 1641 PPCODE:
1538 cl_event ev = 0; 1642 EVENT_LIST (12);
1643
1539 const size_t origin [3] = { x, y, z }; 1644 const size_t origin [3] = { x, y, z };
1540 const size_t region [3] = { width, height, depth }; 1645 const size_t region [3] = { width, height, depth };
1541 EVENT_LIST (12);
1542 1646
1543 const cl_float c_f [4] = { r, g, b, a }; 1647 const cl_float c_f [4] = { r, g, b, a };
1544 const cl_uint c_u [4] = { r, g, b, a }; 1648 const cl_uint c_u [4] = { r, g, b, a };
1545 const cl_int c_s [4] = { r, g, b, a }; 1649 const cl_int c_s [4] = { r, g, b, a };
1546 const void *c_fus [3] = { &c_f, &c_u, &c_s }; 1650 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1549 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); 1653 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1550 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); 1654 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1551 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) 1655 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1552 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); 1656 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1553 1657
1658 cl_event ev = 0;
1554 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], 1659 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1555 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1660 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1556 1661
1557 if (ev) 1662 if (ev)
1558 XPUSH_CLOBJ (stash_event, ev); 1663 XPUSH_CLOBJ (stash_event, ev);
1562void 1667void
1563copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1668copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1564 ALIAS: 1669 ALIAS:
1565 enqueue_copy_buffer = 0 1670 enqueue_copy_buffer = 0
1566 PPCODE: 1671 PPCODE:
1672 EVENT_LIST (6);
1673
1567 cl_event ev = 0; 1674 cl_event ev = 0;
1568 EVENT_LIST (6);
1569
1570 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1675 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1571 1676
1572 if (ev) 1677 if (ev)
1573 XPUSH_CLOBJ (stash_event, ev); 1678 XPUSH_CLOBJ (stash_event, ev);
1574 1679
1575void 1680void
1576read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) 1681read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...)
1577 ALIAS: 1682 ALIAS:
1578 enqueue_read_buffer_rect = 0 1683 enqueue_read_buffer_rect = 0
1579 PPCODE: 1684 PPCODE:
1580 cl_event ev = 0; 1685 EVENT_LIST (17);
1686
1581 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1687 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1582 const size_t host_origin[3] = { host_x, host_y, host_z }; 1688 const size_t host_origin[3] = { host_x, host_y, host_z };
1583 const size_t region[3] = { width, height, depth }; 1689 const size_t region[3] = { width, height, depth };
1584 EVENT_LIST (17);
1585 1690
1586 if (!buf_row_pitch) 1691 if (!buf_row_pitch)
1587 buf_row_pitch = region [0]; 1692 buf_row_pitch = region [0];
1588 1693
1589 if (!buf_slice_pitch) 1694 if (!buf_slice_pitch)
1599 1704
1600 SvUPGRADE (data, SVt_PV); 1705 SvUPGRADE (data, SVt_PV);
1601 SvGROW (data, len); 1706 SvGROW (data, len);
1602 SvPOK_only (data); 1707 SvPOK_only (data);
1603 SvCUR_set (data, len); 1708 SvCUR_set (data, len);
1709
1710 cl_event ev = 0;
1604 NEED_SUCCESS (EnqueueReadBufferRect, (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)); 1711 NEED_SUCCESS (EnqueueReadBufferRect, (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));
1605 1712
1606 if (ev) 1713 if (ev)
1607 XPUSH_CLOBJ (stash_event, ev); 1714 XPUSH_CLOBJ (stash_event, ev);
1608 1715
1609void 1716void
1610write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) 1717write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...)
1611 ALIAS: 1718 ALIAS:
1612 enqueue_write_buffer_rect = 0 1719 enqueue_write_buffer_rect = 0
1613 PPCODE: 1720 PPCODE:
1614 cl_event ev = 0; 1721 EVENT_LIST (17);
1722
1615 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1723 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1616 const size_t host_origin[3] = { host_x, host_y, host_z }; 1724 const size_t host_origin[3] = { host_x, host_y, host_z };
1617 const size_t region[3] = { width, height, depth }; 1725 const size_t region[3] = { width, height, depth };
1618 STRLEN len; 1726 STRLEN len;
1619 char *ptr = SvPVbyte (data, len); 1727 char *ptr = SvPVbyte (data, len);
1620 EVENT_LIST (17);
1621 1728
1622 if (!buf_row_pitch) 1729 if (!buf_row_pitch)
1623 buf_row_pitch = region [0]; 1730 buf_row_pitch = region [0];
1624 1731
1625 if (!buf_slice_pitch) 1732 if (!buf_slice_pitch)
1634 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1741 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1635 1742
1636 if (len < min_len) 1743 if (len < min_len)
1637 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1744 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1638 1745
1746 cl_event ev = 0;
1639 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)); 1747 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));
1640 1748
1641 if (ev) 1749 if (ev)
1642 XPUSH_CLOBJ (stash_event, ev); 1750 XPUSH_CLOBJ (stash_event, ev);
1643 1751
1644void 1752void
1645copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) 1753copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...)
1646 ALIAS: 1754 ALIAS:
1647 enqueue_copy_buffer_rect = 0 1755 enqueue_copy_buffer_rect = 0
1648 PPCODE: 1756 PPCODE:
1649 cl_event ev = 0; 1757 EVENT_LIST (16);
1758
1650 const size_t src_origin[3] = { src_x, src_y, src_z }; 1759 const size_t src_origin[3] = { src_x, src_y, src_z };
1651 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1760 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1652 const size_t region[3] = { width, height, depth }; 1761 const size_t region[3] = { width, height, depth };
1653 EVENT_LIST (16);
1654 1762
1763 cl_event ev = 0;
1655 NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1764 NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1656 1765
1657 if (ev) 1766 if (ev)
1658 XPUSH_CLOBJ (stash_event, ev); 1767 XPUSH_CLOBJ (stash_event, ev);
1659 1768
1660void 1769void
1661read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) 1770read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...)
1662 ALIAS: 1771 ALIAS:
1663 enqueue_read_image = 0 1772 enqueue_read_image = 0
1664 PPCODE: 1773 PPCODE:
1665 cl_event ev = 0; 1774 EVENT_LIST (12);
1775
1666 const size_t src_origin[3] = { src_x, src_y, src_z }; 1776 const size_t src_origin[3] = { src_x, src_y, src_z };
1667 const size_t region[3] = { width, height, depth }; 1777 const size_t region[3] = { width, height, depth };
1668 EVENT_LIST (12);
1669 1778
1670 if (!row_pitch) 1779 if (!row_pitch)
1671 row_pitch = img_row_pitch (src); 1780 row_pitch = img_row_pitch (src);
1672 1781
1673 if (depth > 1 && !slice_pitch) 1782 if (depth > 1 && !slice_pitch)
1677 1786
1678 SvUPGRADE (data, SVt_PV); 1787 SvUPGRADE (data, SVt_PV);
1679 SvGROW (data, len); 1788 SvGROW (data, len);
1680 SvPOK_only (data); 1789 SvPOK_only (data);
1681 SvCUR_set (data, len); 1790 SvCUR_set (data, len);
1791
1792 cl_event ev = 0;
1682 NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1793 NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1683 1794
1684 if (ev) 1795 if (ev)
1685 XPUSH_CLOBJ (stash_event, ev); 1796 XPUSH_CLOBJ (stash_event, ev);
1686 1797
1687void 1798void
1688write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) 1799write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...)
1689 ALIAS: 1800 ALIAS:
1690 enqueue_write_image = 0 1801 enqueue_write_image = 0
1691 PPCODE: 1802 PPCODE:
1692 cl_event ev = 0; 1803 EVENT_LIST (12);
1804
1693 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1805 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1694 const size_t region[3] = { width, height, depth }; 1806 const size_t region[3] = { width, height, depth };
1695 STRLEN len; 1807 STRLEN len;
1696 char *ptr = SvPVbyte (data, len); 1808 char *ptr = SvPVbyte (data, len);
1697 EVENT_LIST (12);
1698 1809
1699 if (!row_pitch) 1810 if (!row_pitch)
1700 row_pitch = img_row_pitch (dst); 1811 row_pitch = img_row_pitch (dst);
1701 1812
1702 if (depth > 1 && !slice_pitch) 1813 if (depth > 1 && !slice_pitch)
1705 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1816 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1706 1817
1707 if (len < min_len) 1818 if (len < min_len)
1708 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1819 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1709 1820
1821 cl_event ev = 0;
1710 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)); 1822 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));
1711 1823
1712 if (ev) 1824 if (ev)
1713 XPUSH_CLOBJ (stash_event, ev); 1825 XPUSH_CLOBJ (stash_event, ev);
1714 1826
1715void 1827void
1716copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) 1828copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...)
1717 ALIAS: 1829 ALIAS:
1718 enqueue_copy_image = 0 1830 enqueue_copy_image = 0
1719 PPCODE: 1831 PPCODE:
1720 cl_event ev = 0; 1832 EVENT_LIST (12);
1833
1721 const size_t src_origin[3] = { src_x, src_y, src_z }; 1834 const size_t src_origin[3] = { src_x, src_y, src_z };
1722 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1835 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1723 const size_t region[3] = { width, height, depth }; 1836 const size_t region[3] = { width, height, depth };
1724 EVENT_LIST (12);
1725 1837
1838 cl_event ev = 0;
1726 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1839 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1727 1840
1728 if (ev) 1841 if (ev)
1729 XPUSH_CLOBJ (stash_event, ev); 1842 XPUSH_CLOBJ (stash_event, ev);
1730 1843
1731void 1844void
1732copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) 1845copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...)
1733 ALIAS: 1846 ALIAS:
1734 enqueue_copy_image_to_buffer = 0 1847 enqueue_copy_image_to_buffer = 0
1735 PPCODE: 1848 PPCODE:
1736 cl_event ev = 0; 1849 EVENT_LIST (10);
1850
1737 const size_t src_origin[3] = { src_x, src_y, src_z }; 1851 const size_t src_origin[3] = { src_x, src_y, src_z };
1738 const size_t region [3] = { width, height, depth }; 1852 const size_t region [3] = { width, height, depth };
1739 EVENT_LIST (10);
1740 1853
1854 cl_event ev = 0;
1741 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1855 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1742 1856
1743 if (ev) 1857 if (ev)
1744 XPUSH_CLOBJ (stash_event, ev); 1858 XPUSH_CLOBJ (stash_event, ev);
1745 1859
1746void 1860void
1747copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) 1861copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...)
1748 ALIAS: 1862 ALIAS:
1749 enqueue_copy_buffer_to_image = 0 1863 enqueue_copy_buffer_to_image = 0
1750 PPCODE: 1864 PPCODE:
1751 cl_event ev = 0; 1865 EVENT_LIST (10);
1866
1752 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1867 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1753 const size_t region [3] = { width, height, depth }; 1868 const size_t region [3] = { width, height, depth };
1754 EVENT_LIST (10);
1755 1869
1870 cl_event ev = 0;
1756 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1871 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1757 1872
1758 if (ev) 1873 if (ev)
1759 XPUSH_CLOBJ (stash_event, ev); 1874 XPUSH_CLOBJ (stash_event, ev);
1760 1875
1761void 1876void
1762map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...) 1877map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...)
1763 ALIAS: 1878 ALIAS:
1764 enqueue_map_buffer = 0 1879 enqueue_map_buffer = 0
1765 PPCODE: 1880 PPCODE:
1766 cl_event ev;
1767 EVENT_LIST (6); 1881 EVENT_LIST (6);
1882
1768 size_t cb = SvIV (cb_); 1883 size_t cb = SvIV (cb_);
1769 1884
1770 if (!SvOK (cb_)) 1885 if (!SvOK (cb_))
1771 { 1886 {
1772 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); 1887 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1773 cb -= offset; 1888 cb -= offset;
1774 } 1889 }
1775 1890
1891 cl_event ev;
1776 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); 1892 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1777 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); 1893 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
1778 1894
1779void 1895void
1780map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...) 1896map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...)
1821 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); 1937 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1822 1938
1823void 1939void
1824unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) 1940unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1825 PPCODE: 1941 PPCODE:
1826 mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); 1942 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1827 if (GIMME_V != G_VOID) 1943 if (GIMME_V != G_VOID)
1828 { 1944 {
1829 clRetainEvent (mapped->event); 1945 clRetainEvent (mapped->event);
1830 XPUSH_CLOBJ (stash_event, mapped->event); 1946 XPUSH_CLOBJ (stash_event, mapped->event);
1831 } 1947 }
1833void 1949void
1834task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) 1950task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1835 ALIAS: 1951 ALIAS:
1836 enqueue_task = 0 1952 enqueue_task = 0
1837 PPCODE: 1953 PPCODE:
1954 EVENT_LIST (2);
1955
1838 cl_event ev = 0; 1956 cl_event ev = 0;
1839 EVENT_LIST (2);
1840
1841 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1957 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1842 1958
1843 if (ev) 1959 if (ev)
1844 XPUSH_CLOBJ (stash_event, ev); 1960 XPUSH_CLOBJ (stash_event, ev);
1845 1961
1846void 1962void
1847nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) 1963nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1848 ALIAS: 1964 ALIAS:
1849 enqueue_nd_range_kernel = 0 1965 enqueue_nd_range_kernel = 0
1850 PPCODE: 1966 PPCODE:
1851 cl_event ev = 0; 1967 EVENT_LIST (5);
1968
1852 size_t *gwo = 0, *gws, *lws = 0; 1969 size_t *gwo = 0, *gws, *lws = 0;
1853 int gws_len; 1970 int gws_len;
1854 size_t *lists; 1971 size_t *lists;
1855 int i; 1972 int i;
1856 EVENT_LIST (5);
1857 1973
1858 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1974 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
1859 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1975 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
1860 1976
1861 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1977 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
1900 if (!lws [i]) 2016 if (!lws [i])
1901 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); 2017 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
1902 } 2018 }
1903 } 2019 }
1904 2020
2021 cl_event ev = 0;
1905 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2022 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1906 2023
1907 if (ev) 2024 if (ev)
1908 XPUSH_CLOBJ (stash_event, ev); 2025 XPUSH_CLOBJ (stash_event, ev);
1909 2026
1914 ALIAS: 2031 ALIAS:
1915 enqueue_migrate_mem_objects = 0 2032 enqueue_migrate_mem_objects = 0
1916 PPCODE: 2033 PPCODE:
1917 EVENT_LIST (3); 2034 EVENT_LIST (3);
1918 2035
1919 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 2036 cl_uint obj_count;
1920 croak ("OpenCL::Queue::migrate_mem_objects: objects must be an array reference with OpenCL::Memory objects"); 2037 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
1921
1922 objects = SvRV (objects);
1923
1924 cl_uint object_count = av_len ((AV *)objects) + 1;
1925 cl_mem *object_list = tmpbuf (sizeof (*object_list) * object_count);
1926
1927 int i;
1928 for (i = object_count; i--; )
1929 object_list [i] = SvCLOBJ ("OpenCL::Queue::migrate_mem_objects", "objects", *av_fetch ((AV *)objects, i, 0), "OpenCL::Memory");
1930 2038
1931 cl_event ev = 0; 2039 cl_event ev = 0;
1932 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, object_count, object_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2040 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1933 2041
1934 if (ev) 2042 if (ev)
1935 XPUSH_CLOBJ (stash_event, ev); 2043 XPUSH_CLOBJ (stash_event, ev);
1936 2044
1937#endif 2045#endif
1943 ALIAS: 2051 ALIAS:
1944 release_gl_objects = 1 2052 release_gl_objects = 1
1945 enqueue_acquire_gl_objects = 0 2053 enqueue_acquire_gl_objects = 0
1946 enqueue_release_gl_objects = 1 2054 enqueue_release_gl_objects = 1
1947 PPCODE: 2055 PPCODE:
1948 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 2056 EVENT_LIST (2);
1949 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); 2057
2058 cl_uint obj_count;
2059 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
1950 2060
1951 cl_event ev = 0; 2061 cl_event ev = 0;
1952 EVENT_LIST (2);
1953 AV *av = (AV *)SvRV (objects);
1954 cl_uint num_objects = av_len (av) + 1;
1955 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1956
1957 int i;
1958 for (i = num_objects; i--; )
1959 object_list [i] = SvCLOBJ ("OpenCL::Queue::acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1960 2062
1961 if (ix) 2063 if (ix)
1962 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2064 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1963 else 2065 else
1964 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2066 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1965 2067
1966 if (ev) 2068 if (ev)
1967 XPUSH_CLOBJ (stash_event, ev); 2069 XPUSH_CLOBJ (stash_event, ev);
1968 2070
1969#endif 2071#endif
1983void 2085void
1984marker (OpenCL::Queue self, ...) 2086marker (OpenCL::Queue self, ...)
1985 ALIAS: 2087 ALIAS:
1986 enqueue_marker = 0 2088 enqueue_marker = 0
1987 PPCODE: 2089 PPCODE:
2090 EVENT_LIST (1);
1988 cl_event ev = 0; 2091 cl_event ev = 0;
1989 EVENT_LIST (1);
1990#if PREFER_1_1 2092#if PREFER_1_1
1991 if (!event_list_count) 2093 if (!event_list_count)
1992 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); 2094 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1993 else 2095 else
1994#if CL_VERSION_1_2 2096#if CL_VERSION_1_2
2008void 2110void
2009barrier (OpenCL::Queue self, ...) 2111barrier (OpenCL::Queue self, ...)
2010 ALIAS: 2112 ALIAS:
2011 enqueue_barrier = 0 2113 enqueue_barrier = 0
2012 PPCODE: 2114 PPCODE:
2115 EVENT_LIST (1);
2013 cl_event ev = 0; 2116 cl_event ev = 0;
2014 EVENT_LIST (1);
2015#if PREFER_1_1 2117#if PREFER_1_1
2016 if (!event_list_count && GIMME_V == G_VOID) 2118 if (!event_list_count && GIMME_V == G_VOID)
2017 NEED_SUCCESS (EnqueueBarrier, (self)); 2119 NEED_SUCCESS (EnqueueBarrier, (self));
2018 else 2120 else
2019#if CL_VERSION_1_2 2121#if CL_VERSION_1_2
2347void 2449void
2348build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) 2450build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2349 ALIAS: 2451 ALIAS:
2350 build_async = 1 2452 build_async = 1
2351 CODE: 2453 CODE:
2352 void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0;
2353 void *user_data = 0;
2354 cl_uint num_devices = 0; 2454 cl_uint device_count = 0;
2355 cl_device_id *device_list = 0; 2455 cl_device_id *device_list = 0;
2356 2456
2357 if (SvOK (devices)) 2457 if (SvOK (devices))
2358 { 2458 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2459
2460 void *user_data;
2461 program_callback pfn_notify = make_program_callback (notify, &user_data);
2462
2463 if (ix)
2464 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2465 else
2466 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2467
2468#if CL_VERSION_1_2
2469
2470void
2471compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
2472 CODE:
2473 cl_uint device_count = 0;
2474 cl_device_id *device_list = 0;
2475
2476 if (SvOK (devices))
2477 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2478
2479 cl_uint header_count = 0;
2480 cl_program *header_list = 0;
2481 const char **header_name = 0;
2482
2483 if (SvOK (headers))
2484 {
2359 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 2485 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2360 croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); 2486 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2361 2487
2362 AV *av = (AV *)SvRV (devices); 2488 HV *hv = (HV *)SvRV (devices);
2363 num_devices = av_len (av) + 1;
2364 2489
2365 if (num_devices) 2490 header_count = hv_iterinit (hv);
2491 header_list = tmpbuf (sizeof (*header_list) * header_count);
2492 header_name = tmpbuf (sizeof (*header_name) * header_count);
2493
2494 HE *he;
2495 int i = 0;
2496 while (he = hv_iternext (hv))
2366 { 2497 {
2367 device_list = tmpbuf (sizeof (*device_list) * num_devices); 2498 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2499 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2368 int count; 2500 ++i;
2369 for (count = 0; count < num_devices; ++count)
2370 device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device");
2371 } 2501 }
2372 } 2502 }
2373 2503
2374 if (SvOK (notify)) 2504 void *user_data;
2375 { 2505 program_callback pfn_notify = make_program_callback (notify, &user_data);
2376 NEED_SUCCESS (RetainProgram, (self));
2377 pfn_notify = eq_program_notify;
2378 user_data = SvREFCNT_inc (s_get_cv (notify));
2379 }
2380 2506
2381 if (ix)
2382 build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data);
2383 else
2384 NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); 2507 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2508 header_count, header_list, header_name, pfn_notify, user_data));
2509
2510#endif
2385 2511
2386void 2512void
2387build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) 2513build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
2388 PPCODE: 2514 PPCODE:
2389 size_t size; 2515 size_t size;
2417 char *value = tmpbuf (size); 2543 char *value = tmpbuf (size);
2418 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); 2544 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2419 EXTEND (SP, 1); 2545 EXTEND (SP, 1);
2420 const int i = 0; 2546 const int i = 0;
2421 PUSHs (sv_2mortal (newSVpv (value, 0))); 2547 PUSHs (sv_2mortal (newSVpv (value, 0)));
2548
2549void
2550binary_type (OpenCL::Program self, OpenCL::Device device)
2551 PPCODE:
2552 cl_program_binary_type value [1];
2553 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2554 EXTEND (SP, 1);
2555 const int i = 0;
2556 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
2422 2557
2423#END:program_build 2558#END:program_build
2424 2559
2425void 2560void
2426kernel (OpenCL::Program program, SV *function) 2561kernel (OpenCL::Program program, SV *function)
2593 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; 2728 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2594 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; 2729 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2595 2730
2596 case 'z': nullarg = 1; size = SvIV (sv); break; 2731 case 'z': nullarg = 1; size = SvIV (sv); break;
2597 2732
2598 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; 2733 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2599 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; 2734 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2600 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; 2735 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2601 2736
2602 default: 2737 default:
2603 croak ("OpenCL::Kernel::setf format character '%c' not supported", type); 2738 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2604 } 2739 }
2605 2740
2676 CODE: 2811 CODE:
2677 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); 2812 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2678 2813
2679void 2814void
2680set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) 2815set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2681 ALIAS:
2682 set_image2d = 0
2683 set_image3d = 0
2684 CODE: 2816 CODE:
2685 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); 2817 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2686 2818
2687void 2819void
2688set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) 2820set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2984 Safefree (mapped); 3116 Safefree (mapped);
2985 3117
2986void 3118void
2987unmap (OpenCL::Mapped self, ...) 3119unmap (OpenCL::Mapped self, ...)
2988 CODE: 3120 CODE:
2989 mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); 3121 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
2990 3122
2991bool 3123bool
2992mapped (OpenCL::Mapped self) 3124mapped (OpenCL::Mapped self)
2993 CODE: 3125 CODE:
2994 RETVAL = !!self->ptr; 3126 RETVAL = !!self->ptr;

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines