[ViewVC] Diff of: cvs/OpenCL/OpenCL.xs

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.50 by root, Tue Apr 24 14:57:06 2012 UTC vs.
Revision 1.51 by root, Tue Apr 24 22:45:38 2012 UTC

 #include "EXTERN.h"
 #include "perl.h"
 #include "XSUB.h"
+#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
+#include "xthread.h"
+#include "schmorp.h"
 #ifdef I_DLFCN
   #include <dlfcn.h>
 #endif
   if (res) \
     FAIL (name);
 /*****************************************************************************/
-#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr)
-#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr))
-static void *
-SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg)
-{
-  if (SvROK (sv) && sv_derived_from (sv, pkg))
-    return (void *)SvIV (SvRV (sv));
-   croak ("%s: %s is not of type %s", func, svname, pkg);
-}
-/*****************************************************************************/
 static cl_context_properties *
 SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
 {
   if (!sv || !SvOK (sv))
     if (extra)
    croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname);
 }
 /*****************************************************************************/
+#define NEW_CLOBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr)
+#define PUSH_CLOBJ(class,ptr)  PUSHs  (NEW_CLOBJ (class, ptr))
+#define XPUSH_CLOBJ(class,ptr) XPUSHs (NEW_CLOBJ (class, ptr))
+/* cl objects are either \$iv, or [$iv, ...] */
+/* they can be upgraded at runtime to the array form */
+static void *
+SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg)
+{
+  if (SvROK (sv) && sv_derived_from (sv, pkg))
+    {
+      SV *rv = SvRV (sv);
+      if (SvTYPE (rv) == SVt_PVAV)
+        rv = AvARRAY (rv)[0];
+      return (void *)SvIV (SvRV (sv));
+    }
+   croak ("%s: %s is not of type %s", func, svname, pkg);
+}
+static void
+CLOBJ_push (SV *self, SV *data)
+{
+  SV *rv = SvRV (self);
+  if (SvTYPE (rv) != SVt_PVAV)
+    {
+      AV *av = newAV ();
+      av_push (av, rv);
+      rv = (SV *)av;
+      SvRV_set (self, rv);
+    }
+  av_push ((AV *)rv, data);
+}
+static SV *
+sv_struct (STRLEN size)
+{
+  SV *sv = newSV (size);
+  SvPOK_only (sv);
+  return sv;
+}
+static void *
+CLOBJ_push_struct (SV *self, STRLEN size)
+{
+  SV *sv = sv_struct (size);
+  CLOBJ_push (self, sv);
+  return SvPVX (sv);
+}
+/*****************************************************************************/
+/* callback stuff */
+/* default context callback, log to stderr */
+static void CL_CALLBACK
+context_default_notify (const char *msg, const void *info, size_t cb, void *data)
+{
+  fprintf (stderr, "OpenCL Context Notify: %s\n", msg);
+}
+typedef struct
+{
+  int free_cb;
+  void (*push)(void *data1, void *data2, void *data3);
+} eq_vtbl;
+typedef struct eq_item
+{
+  struct eq_item *next;
+  eq_vtbl *vtbl;
+  SV *cb;
+  void *data1, *data2, *data3;
+} eq_item;
+static void (*eq_signal_func)(void *signal_arg, int value);
+static void *eq_signal_arg;
+static xmutex_t eq_lock = X_MUTEX_INIT;
+static eq_item *eq_head, *eq_tail;
+static void
+eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
+{
+  eq_item *item = malloc (sizeof (eq_item));
+  printf ("enq(%p,%p,%p,%p,%p)\n", vtbl, cb, data1, data2, data3);//D
+  item->next  = 0;
+  item->vtbl  = vtbl;
+  item->cb    = cb;
+  item->data1 = data1;
+  item->data2 = data2;
+  item->data3 = data3;
+  X_LOCK (eq_lock);
+  *(eq_head ? &eq_tail->next : &eq_head) = item;
+  eq_tail = item;
+  X_UNLOCK (eq_lock);
+  eq_signal_func (eq_signal_arg, 0);
+}
+static eq_item *
+eq_dec (void)
+{
+  eq_item *res;
+  X_LOCK (eq_lock);
+  res = eq_head;
+  if (res)
+    eq_head = res->next;
+  X_UNLOCK (eq_lock);
+  return res;
+}
+#if 0
+static void
+mem_free (pTHX_ void *p)
+{
+  free (p);
+}
+//SAVEDESTRUCTOR_X (mem_free, item);
+#endif
+static void
+eq_poll (void)
+{
+  eq_item *item;
+  while ((item = eq_dec ()))
+    {
+      ENTER;
+      SAVETMPS;
+      dSP;
+      PUSHMARK (SP);
+      EXTEND (SP, 2);
+      if (item->vtbl->free_cb)
+        sv_2mortal (item->cb);
+      PUTBACK;
+      item->vtbl->push (item->data1, item->data2, item->data3);
+      SV *cb = item->cb;
+      free (item);
+      call_sv (cb, G_DISCARD | G_VOID);
+      FREETMPS;
+      LEAVE;
+    }
+}
+static void
+eq_poll_interrupt (pTHX_ void *c_arg, int value)
+{
+  eq_poll ();
+}
+/* context notify */
+static void
+eq_context_push (void *data1, void *data2, void *data3)
+{
+  dSP;
+  PUSHs (sv_2mortal (newSVpv  (data1, 0)));
+  PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
+  PUTBACK;
+}
+static eq_vtbl eq_context_vtbl = { 0, eq_context_push };
+/* build/compile/link notify */
+static void
+eq_program_push (void *data1, void *data2, void *data3)
+{
+  dSP;
+  PUSH_CLOBJ ("OpenCL::Program", data1);
+  PUTBACK;
+}
+static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
+static void CL_CALLBACK
+eq_program_notify (cl_program program, void *user_data)
+{
+  eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
+}
+struct build_args
+{
+  cl_program program;
+  char *options;
+  void *user_data;
+  cl_uint num_devices;
+};
+X_THREAD_PROC (build_program_thread)
+{
+  struct build_args *arg = thr_arg;
+  clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0);
+  if (arg->user_data)
+    eq_program_notify (arg->program, arg->user_data);
+  else
+    clReleaseProgram (arg->program);
+  free (arg->options);
+  free (arg);
+}
+static void
+build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
+{
+  struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices);
+  arg->program     = program;
+  arg->options     = strdup (options);
+  arg->user_data   = user_data;
+  arg->num_devices = num_devices;
+  memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices);
+  xthread_t id;
+  thread_create (&id, build_program_thread, arg);
+}
+/* event objects */
+static void
+eq_event_push (void *data1, void *data2, void *data3)
+{
+  dSP;
+  PUSH_CLOBJ ("OpenCL::Event", data1);
+  PUSHs (sv_2mortal (newSViv ((IV)data2)));
+  PUTBACK;
+}
+static eq_vtbl eq_event_vtbl = { 1, eq_event_push };
+static void CL_CALLBACK
+eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data)
+{
+  clRetainEvent (event);
+  eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)event_command_exec_status, 0);
+}
+/*****************************************************************************/
 static size_t
 img_row_pitch (cl_mem img)
 {
   size_t res;
   clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
   do
     {
       --count;
       if (SvOK (items [count]))
-        list [i++] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
+        list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
     }
   while (count);
   *rcount = i;
 }
 MODULE = OpenCL		PACKAGE = OpenCL
 PROTOTYPES: ENABLE
+void
+poll ()
+	CODE:
+        eq_poll ();
+void
+_eq_initialise (IV func, IV arg)
+	CODE:
+        eq_signal_func = (void (*)(void *, int))func;
+        eq_signal_arg  = (void*)arg;
 BOOT:
 {
   HV *stash = gv_stashpv ("OpenCL", 1);
   static const ivstr *civ, const_iv[] = {
     { sizeof (cl_half  ), "SIZEOF_HALF"   },
     { sizeof (cl_float ), "SIZEOF_FLOAT"  },
     { sizeof (cl_double), "SIZEOF_DOUBLE" },
 #include "constiv.h"
   };
   for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
     newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
+  sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt);
 }
 cl_int
 errno ()
 	CODE:
         list = tmpbuf (sizeof (*list) * count);
 	NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
         EXTEND (SP, count);
         for (i = 0; i < count; ++i)
-          PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i]));
+          PUSH_CLOBJ ("OpenCL::Platform", list [i]);
 void
 context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
 	PPCODE:
         NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
+        XPUSH_CLOBJ ("OpenCL::Context", ctx);
 #if 0
 void
 context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0)
         EXTEND (SP, count);
         for (i = 0; i < count; ++i)
           PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i]));
 void
-context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0)
+context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = 0)
 	PPCODE:
         if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
           croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
         AV *av = (AV *)SvRV (devices);
         cl_uint num_devices = av_len (av) + 1;
         cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
         int i;
         for (i = num_devices; i--; )
-          device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
+          device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
+        void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify;
+        void *user_data = 0;
-	NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res));
+	NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
+        XPUSH_CLOBJ ("OpenCL::Context", ctx);
 void
 context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0)
 	PPCODE:
 	cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
         cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2);
         NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
+        XPUSH_CLOBJ ("OpenCL::Context", ctx);
 MODULE = OpenCL		PACKAGE = OpenCL::Device
 void
 info (OpenCL::Device self, cl_device_info name)
  cl_platform_id value [1];
  NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i]));
+   PUSH_CLOBJ ("OpenCL::Platform", value [i]);
  }
 void
 name (OpenCL::Device self)
  ALIAS:
  cl_device_id value [1];
  NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i]));
+   PUSH_CLOBJ ("OpenCL::Device", value [i]);
  }
 void
 partition_types_ext (OpenCL::Device self)
  ALIAS:
 void
 queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
 	PPCODE:
 	NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Queue", queue);
+        XPUSH_CLOBJ ("OpenCL::Queue", queue);
 void
 user_event (OpenCL::Context self)
 	PPCODE:
 	NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
-        XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev);
+        XPUSH_CLOBJ ("OpenCL::UserEvent", ev);
 void
 buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
 	PPCODE:
         if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
           croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
         NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
-        XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
+        XPUSH_CLOBJ ("OpenCL::BufferObj", mem);
 void
 buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
 	PPCODE:
 	STRLEN len;
         char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
         if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
           croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
         NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
-        XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
+        XPUSH_CLOBJ ("OpenCL::BufferObj", mem);
 #if CL_VERSION_1_2
 void
 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)
         const cl_image_desc desc = {
           type,
           width, height, depth,
           array_size, row_pitch, slice_pitch,
           num_mip_level, num_samples,
-	  type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvPTROBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
+	  type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
         };
 	NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
         char *klass = "OpenCL::Image";
         switch (type)
           {
             case CL_MEM_OBJECT_IMAGE1D_ARRAY:	klass = "OpenCL::Image2DArray";  break;
             case CL_MEM_OBJECT_IMAGE2D:		klass = "OpenCL::Image2D";       break;
             case CL_MEM_OBJECT_IMAGE2D_ARRAY:	klass = "OpenCL::Image2DArray";  break;
             case CL_MEM_OBJECT_IMAGE3D:		klass = "OpenCL::Image3D";       break;
           }
-        XPUSH_NEW_OBJ (klass, mem);
+        XPUSH_CLOBJ (klass, mem);
 #endif
 void
 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)
 	NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
 #else
         const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
 	NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
 #endif
-        XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
+        XPUSH_CLOBJ ("OpenCL::Image2D", mem);
 void
 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)
 	PPCODE:
 	STRLEN len;
 	NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
 #else
         const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
 	NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
 #endif
-        XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
+        XPUSH_CLOBJ ("OpenCL::Image3D", mem);
 #if cl_apple_gl_sharing || cl_khr_gl_sharing
 void
 gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
 	PPCODE:
         NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
-        XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem);
+        XPUSH_CLOBJ ("OpenCL::BufferObj", mem);
 void
 gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
 	PPCODE:
 	NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
+        XPUSH_CLOBJ ("OpenCL::Image2D", mem);
 #if CL_VERSION_1_2
 void
 gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
             case CL_GL_OBJECT_TEXTURE1D_ARRAY:	klass = "OpenCL::Image2DArray";  break;
             case CL_GL_OBJECT_TEXTURE2D:	klass = "OpenCL::Image2D";       break;
             case CL_GL_OBJECT_TEXTURE2D_ARRAY:	klass = "OpenCL::Image2DArray";  break;
             case CL_GL_OBJECT_TEXTURE3D:	klass = "OpenCL::Image3D";       break;
           }
-        XPUSH_NEW_OBJ (klass, mem);
+        XPUSH_CLOBJ (klass, mem);
 #endif
 void
 gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
 #if PREFER_1_1
 	NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
 #else
 	NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture  , (self, flags, target, miplevel, texture, &res));
 #endif
-        XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
+        XPUSH_CLOBJ ("OpenCL::Image2D", mem);
 void
 gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
 	PPCODE:
 #if PREFER_1_1
 	NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
 #else
 	NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture  , (self, flags, target, miplevel, texture, &res));
 #endif
-        XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
+        XPUSH_CLOBJ ("OpenCL::Image3D", mem);
 #endif
 void
 supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
 void
 sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
 	PPCODE:
 	NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler);
+        XPUSH_CLOBJ ("OpenCL::Sampler", sampler);
 void
 program_with_source (OpenCL::Context self, SV *program)
 	PPCODE:
 	STRLEN len;
         size_t len2;
         const char *ptr = SvPVbyte (program, len);
         len2 = len;
 	NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Program", prog);
+        XPUSH_CLOBJ ("OpenCL::Program", prog);
 #BEGIN:context
 void
 reference_count (OpenCL::Context self)
  NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value,     0));
  int i, n = size / sizeof (*value);
  EXTEND (SP, n);
  for (i = 0; i < n; ++i)
  {
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i]));
+   PUSH_CLOBJ ("OpenCL::Device", value [i]);
  }
 void
 properties (OpenCL::Context self)
  PPCODE:
         SvPOK_only (data);
         SvCUR_set (data, len);
         NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (5, items - 5);
         NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 #if CL_VERSION_1_2
 void
 enqueue_fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
         EVENT_LIST (5, items - 5);
         NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 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, ...)
 	PPCODE:
 	cl_event ev = 0;
         NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type]],
                                          origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 #endif
 void
 enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
         EVENT_LIST (6, items - 6);
         NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_read_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         SvPOK_only (data);
         SvCUR_set (data, len);
         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));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_write_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, ...)
 	PPCODE:
 	cl_event ev = 0;
           croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
         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));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_copy_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (16, items - 16);
         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));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_read_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         SvPOK_only (data);
         SvCUR_set (data, len);
         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));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_write_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, ...)
 	PPCODE:
 	cl_event ev = 0;
           croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
         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));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_copy_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (12, items - 12);
         NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_copy_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (10, items - 10);
         NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_copy_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, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (10, items - 10);
         NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
 	PPCODE:
 	cl_event ev = 0;
         EVENT_LIST (2, items - 2);
         NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
 	PPCODE:
 	cl_event ev = 0;
           }
         NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 #if cl_apple_gl_sharing || cl_khr_gl_sharing
 void
 enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
         cl_uint num_objects = av_len (av) + 1;
         cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
         int i;
         for (i = num_objects; i--; )
-          object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
+          object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
         if (ix)
           NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         else
           NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 #endif
 void
 enqueue_wait_for_events (OpenCL::Queue self, ...)
 #endif
 #else
         NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
 #endif
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 enqueue_barrier (OpenCL::Queue self, ...)
 	PPCODE:
 	cl_event ev = 0;
 #endif
 #else
         NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
 #endif
         if (ev)
-          XPUSH_NEW_OBJ ("OpenCL::Event", ev);
+          XPUSH_CLOBJ ("OpenCL::Event", ev);
 void
 flush (OpenCL::Queue self)
 	CODE:
         NEED_SUCCESS (Flush, (self));
  NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 void
 device (OpenCL::Queue self)
  PPCODE:
  cl_device_id value [1];
  NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i]));
+   PUSH_CLOBJ ("OpenCL::Device", value [i]);
  }
 void
 reference_count (OpenCL::Queue self)
  PPCODE:
  NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 void
 associated_memobject (OpenCL::Memory self)
  PPCODE:
  NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainMemObject, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i]));
+   PUSH_CLOBJ ("OpenCL::Memory", value [i]);
  }
 #END:mem
 #if cl_apple_gl_sharing || cl_khr_gl_sharing
           croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
         cl_buffer_region crdata = { origin, size };
         NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
-        XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
+        XPUSH_CLOBJ ("OpenCL::Buffer", mem);
 MODULE = OpenCL		PACKAGE = OpenCL::Image
 void
 image_info (OpenCL::Image self, cl_image_info name)
  NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 void
 normalized_coords (OpenCL::Sampler self)
  PPCODE:
 DESTROY (OpenCL::Program self)
 	CODE:
         clReleaseProgram (self);
 void
-build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef)
+build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
+	ALIAS:
+        build_async = 1
 	CODE:
+        void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0;
+        void *user_data = 0;
+        cl_uint num_devices = 0;
+        cl_device_id *device_list = 0;
+        if (SvOK (devices))
+	  {
+            if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
+              croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects.");
+            AV *av = (AV *)SvRV (devices);
+            num_devices = av_len (av) + 1;
+            if (num_devices)
+              {
+                device_list = tmpbuf (sizeof (*device_list) * num_devices);
+                int count;
+                for (count = 0; count < num_devices; ++count)
+                  device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device");
+              }
+          }
+        if (SvOK (notify))
+          {
+            NEED_SUCCESS (RetainProgram, (self));
+            pfn_notify = eq_program_notify;
+            user_data = SvREFCNT_inc (s_get_cv (notify));
+          }
+	if (ix)
+          build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data);
+        else
-        NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0));
+          NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
 void
 build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
 	PPCODE:
 	size_t size;
 void
 kernel (OpenCL::Program program, SV *function)
 	PPCODE:
 	NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
-        XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel);
+        XPUSH_CLOBJ ("OpenCL::Kernel", kernel);
 void
 kernels_in_program (OpenCL::Program program)
 	PPCODE:
         cl_uint num_kernels;
 	NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
         int i;
         EXTEND (SP, num_kernels);
         for (i = 0; i < num_kernels; ++i)
-          PUSHs (NEW_MORTAL_OBJ ("OpenCL::Kernel", kernels [i]));
+          PUSH_CLOBJ ("OpenCL::Kernel", kernels [i]);
 void
 info (OpenCL::Program self, cl_program_info name)
 	PPCODE:
         INFO (Program)
  NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 void
 devices (OpenCL::Program self)
  PPCODE:
  NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value,     0));
  int i, n = size / sizeof (*value);
  EXTEND (SP, n);
  for (i = 0; i < n; ++i)
  {
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i]));
+   PUSH_CLOBJ ("OpenCL::Device", value [i]);
  }
 void
 source (OpenCL::Program self)
  PPCODE:
  NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 void
 program (OpenCL::Kernel self)
  PPCODE:
  NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainProgram, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i]));
+   PUSH_CLOBJ ("OpenCL::Program", value [i]);
  }
 #END:kernel
 void
 wait (OpenCL::Event self)
 	CODE:
 	clWaitForEvents (1, &self);
 void
+cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
+	CODE:
+        clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
+void
 info (OpenCL::Event self, cl_event_info name)
 	PPCODE:
         INFO (Event)
 #BEGIN:event
  NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainCommandQueue, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i]));
+   PUSH_CLOBJ ("OpenCL::Queue", value [i]);
  }
 void
 command_type (OpenCL::Event self)
  PPCODE:
  NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
  EXTEND (SP, 1);
  const int i = 0;
  {
    NEED_SUCCESS (RetainContext, (value [i]));
-   PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i]));
+   PUSH_CLOBJ ("OpenCL::Context", value [i]);
  }
 #END:event
 void

Diff Legend

-–
+Removed lines
-+
+Added lines
-<
+Changed lines
->
+Changed lines

Comparing OpenCL/OpenCL.xs (file contents): Revision 1.50 by root, Tue Apr 24 14:57:06 2012 UTC vs. Revision 1.51 by root, Tue Apr 24 22:45:38 2012 UTC

Diff Legend

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.50 by root, Tue Apr 24 14:57:06 2012 UTC vs.
Revision 1.51 by root, Tue Apr 24 22:45:38 2012 UTC