OSD support, optinal queue, optional use of sleep(), and benchmark mode

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@10699 b3059339-0415-0410-9bf9-f77b7e298cf2

1 files changed, 709 insertions, 139 deletions

diff --git a/libvo/vo_xvmc.c b/libvo/vo_xvmc.c
index 4742572095..939f263fd5 100644
--- a/libvo/vo_xvmc.c
+++ b/libvo/vo_xvmc.c
@@ -1,20 +1,29 @@
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#include <unistd.h>
+
+#include "config.h"
+#include "mp_msg.h"
+#include "video_out.h"
+#include "video_out_internal.h"
+#include "fastmemcpy.h"
+#include "osdep/timer.h"
 
 #include <X11/Xlib.h>
 #include <X11/Xutil.h>
 #include <X11/Xatom.h>
+
+#ifdef HAVE_SHM
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <X11/extensions/XShm.h>
+#endif
+
 #include <X11/extensions/Xv.h>
 #include <X11/extensions/Xvlib.h>
 #include <X11/extensions/XvMClib.h>
 
-#include "config.h"
-#include "mp_msg.h"
-#include "video_out.h"
-#include "video_out_internal.h"
-#include "fastmemcpy.h"
-
 #include "x11_common.h"
 #include "xvmc_render.h"
 
@@ -25,12 +34,12 @@
 #include "../Gui/interface.h"
 #endif
 
+//no chanse xinerama to be suported in near future
 #undef HAVE_XINERAMA
 
 #undef NDEBUG 
 #include <assert.h>
 
-//no chanse xinerama to be suported in near future
 
 #define UNUSED(x) ((void)(x))
 
@@ -38,17 +47,33 @@
 extern int vo_directrendering;
 extern int vo_verbose;
 
-static void xvmc_free(void);
+static int benchmark;
+static int busy_wait;
+static int use_queue;
 
 static int image_width,image_height;
 static uint32_t  drwX,drwY;
 
 static XvPortID xv_port;
 
-//0-auto;1-backgound always keycolor;2-autopaint(by X);3-manual fill 
+#define AUTO_COLORKEY       0
+#define BACKGROUND_COLORKEY 1
+#define AUTOPAINT_COLORKEY  2
+#define MANUALFILL_COLORKEY 3
 static int keycolor_handling;
 static unsigned long keycolor;
 
+#define NO_SUBPICTURE      0
+#define OVERLAY_SUBPICTURE 1
+#define BLEND_SUBPICTURE   2
+#define BACKEND_SUBPICTURE 3
+
+static int subpicture_mode;
+static int subpicture_alloc;
+static XvMCSubpicture subpicture;
+static XvImageFormatValues subpicture_info;
+static int subpicture_clear_color;//transparent color for the subpicture or color key for overlay
+
 static XvMCSurfaceInfo surface_info;
 static XvMCContext ctx;
 static XvMCBlockArray data_blocks;
@@ -62,6 +87,36 @@ static xvmc_render_state_t * surface_render;
 static xvmc_render_state_t * p_render_surface_to_show=NULL;
 static xvmc_render_state_t * p_render_surface_visible=NULL;
 
+//display queue, kinda render ahead
+static xvmc_render_state_t * show_queue[MAX_SURFACES];
+static int free_element;
+
+
+static void (*draw_osd_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void (*clear_osd_fnc)(int x0,int y0, int w,int h);
+static void (*init_osd_fnc)(void);
+
+static void   draw_osd_AI44(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void   draw_osd_IA44(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void   clear_osd_subpic(int x0,int y0, int w,int h);
+static void   init_osd_yuv_pal(void);
+
+
+static const struct{
+   int id;//id as xvimages or as mplayer RGB|{8,15,16,24,32}
+   void (* init_func_ptr)();
+   void (* draw_func_ptr)();
+   void (* clear_func_ptr)();
+   } osd_render[]={
+                     {0x34344149,init_osd_yuv_pal,draw_osd_AI44,clear_osd_subpic},
+                     {0x34344941,init_osd_yuv_pal,draw_osd_IA44,clear_osd_subpic},
+                     {0,NULL,NULL,NULL}
+                  };
+
+static void xvmc_free(void);
+static int count_free_surfaces();
+static xvmc_render_state_t * find_free_surface();
+
 static vo_info_t info = {
   "XVideo Motion Compensation",
   "xvmc",
@@ -71,6 +126,76 @@ static vo_info_t info = {
 
 LIBVO_EXTERN(xvmc);
 
+//shm stuff from vo_xv
+#ifdef HAVE_SHM
+/* since it doesn't seem to be defined on some platforms */
+int XShmGetEventBase(Display*);
+static XShmSegmentInfo Shminfo;
+static int Shmem_Flag;
+#endif
+XvImage * xvimage;
+
+
+static void allocate_xvimage(int xvimage_width,int xvimage_height,int xv_format)
+{
+ /*
+  * allocate XvImages.  FIXME: no error checking, without
+  * mit-shm this will bomb... trzing to fix ::atmos
+  */
+#ifdef HAVE_SHM
+   if ( mLocalDisplay && XShmQueryExtension( mDisplay ) ) Shmem_Flag = 1;
+   else
+   {
+      Shmem_Flag = 0;
+      mp_msg(MSGT_VO,MSGL_INFO, "Shared memory not supported\nReverting to normal Xv\n" );
+   }
+   if ( Shmem_Flag )
+   {
+      xvimage = (XvImage *) XvShmCreateImage(mDisplay, xv_port, xv_format, 
+                             NULL, xvimage_width, xvimage_height, &Shminfo);
+
+      Shminfo.shmid    = shmget(IPC_PRIVATE, xvimage->data_size, IPC_CREAT | 0777);
+      Shminfo.shmaddr  = (char *) shmat(Shminfo.shmid, 0, 0);
+      Shminfo.readOnly = False;
+
+      xvimage->data = Shminfo.shmaddr;
+      XShmAttach(mDisplay, &Shminfo);
+      XSync(mDisplay, False);
+      shmctl(Shminfo.shmid, IPC_RMID, 0);
+   }
+   else
+#endif
+   {
+      xvimage = (XvImage *) XvCreateImage(mDisplay, xv_port, xv_format, NULL, xvimage_width, xvimage_height);
+      xvimage->data = malloc(xvimage->data_size);
+      XSync(mDisplay,False);
+   }
+// memset(xvimage->data,128,xvimage->data_size);
+   return;
+}
+
+static void deallocate_xvimage()
+{
+#ifdef HAVE_SHM
+   if ( Shmem_Flag )
+   {
+      XShmDetach( mDisplay,&Shminfo );
+      shmdt( Shminfo.shmaddr );
+   }
+   else
+#endif
+   {
+      free(xvimage->data);
+   }
+   XFree(xvimage);
+
+   XFlush( mDisplay );
+   XSync(mDisplay, False);
+   return;
+}
+//end of vo_xv shm/xvimage code
+
+
 static void  init_keycolor(){
 Atom xv_atom;
 XvAttribute * attributes;
@@ -80,10 +205,10 @@ int attrib_count,i;
 
    keycolor=2110;
 
-   if(keycolor_handling == 0){
+   if(keycolor_handling == AUTO_COLORKEY){
    //XV_AUTOPING_COLORKEY doesn't work for XvMC yet(NVidia 43.63)
       attributes = XvQueryPortAttributes(mDisplay, xv_port, &attrib_count);
-      if(attributes!=NULL) 
+      if(attributes!=NULL)
          for (i = 0; i < attrib_count; i++)
             if (!strcmp(attributes[i].name, "XV_AUTOPAINT_COLORKEY"))
             {
@@ -92,7 +217,7 @@ int attrib_count,i;
                {
                   rez=XvSetPortAttribute(mDisplay, xv_port, xv_atom, 1);
                   if(rez == Success) 
-                     keycolor_handling = 2;//this is the way vo_xv works
+                     keycolor_handling = AUTOPAINT_COLORKEY;
                }
                break;
             }
@@ -103,9 +228,9 @@ int attrib_count,i;
    if(xv_atom == None) return;
    rez=XvGetPortAttribute(mDisplay,xv_port, xv_atom, &colorkey);
    if(rez == Success){
-      keycolor=colorkey;
-      if(keycolor_handling == 0){
-         keycolor_handling = 3;
+      keycolor = colorkey;
+      if(keycolor_handling == AUTO_COLORKEY){
+         keycolor_handling = MANUALFILL_COLORKEY;
       }
    }
 }
@@ -113,20 +238,21 @@ int attrib_count,i;
 //from vo_xmga
 static void mDrawColorKey(uint32_t x,uint32_t  y, uint32_t w, uint32_t h)
 {
-   if( (keycolor_handling != 2) && (keycolor_handling != 3) ) 
-      return ;//unknown method
+   if( (keycolor_handling != AUTOPAINT_COLORKEY) && 
+       (keycolor_handling != MANUALFILL_COLORKEY) ) 
+      return;
 
    XSetBackground( mDisplay,vo_gc,0 );
    XClearWindow( mDisplay,vo_window );
- 
-   if(keycolor_handling == 3){
+
+   if(keycolor_handling == MANUALFILL_COLORKEY){
       XSetForeground( mDisplay,vo_gc,keycolor );
       XFillRectangle( mDisplay,vo_window,vo_gc,x,y,w,h);
    }
    XFlush( mDisplay );
 }
 
-// now it is ugly, but i need it working
+
 static int xvmc_check_surface_format(uint32_t format, XvMCSurfaceInfo * surf_info){
    if ( format == IMGFMT_XVMC_IDCT_MPEG2 ){ 
       if( surf_info->mc_type != (XVMC_IDCT|XVMC_MPEG_2) ) return -1;
@@ -141,16 +267,72 @@ static int xvmc_check_surface_format(uint32_t format, XvMCSurfaceInfo * surf_inf
 return -1;//fail
 }
 
+//print all info needed to add new format
+static void print_xvimage_format_values(XvImageFormatValues *xifv){
+int i;
+   printf("Format_ID = 0x%X\n",xifv->id);
+   
+   printf("  type = ");
+   if(xifv->type == XvRGB) printf("RGB\n");
+   else if(xifv->type == XvYUV) printf("YUV\n");
+   else printf("Unknown\n");
+
+   printf("  byte_order = ");
+   if(xifv->byte_order == LSBFirst) printf("LSB First\n");
+   else if(xifv->type == MSBFirst) printf("MSB First\n");
+   else printf("Unknown\n");//yes Linux support other types too
+
+   printf("  guid = ");
+   for(i=0;i<16;i++)
+      printf("%02X ",(unsigned char)xifv->guid[i]);
+   printf("\n");
+
+   printf("  bits_per_pixel = %d\n",xifv->bits_per_pixel);
+
+   printf("  format = ");
+   if(xifv->format == XvPacked) printf("XvPacked\n");
+   else if(xifv->format == XvPlanar) printf("XvPlanar\n");
+   else printf("Unknown\n");
+
+   printf("  num_planes = %d\n",xifv->num_planes);
+
+   if(xifv->type == XvRGB){
+      printf("  red_mask = %0X\n",  xifv->red_mask);
+      printf("  green_mask = %0X\n",xifv->green_mask);
+      printf("  blue_mask = %0X\n", xifv->blue_mask);
+   }
+   if(xifv->type == XvYUV){
+      printf("  y_sample_bits = %d\n  u_sample_bits = %d\n  v_sample_bits = %d\n",
+             xifv->y_sample_bits,xifv->u_sample_bits,xifv->v_sample_bits);
+      printf("  horz_y_period = %d\n  horz_u_period = %d\n  horz_v_period = %d\n",
+            xifv->horz_y_period,xifv->horz_u_period,xifv->horz_v_period);
+      printf("  vert_y_period = %d\n  vert_u_period = %d\n  vert_v_period = %d\n",
+            xifv->vert_y_period,xifv->vert_u_period,xifv->vert_v_period);
+
+      printf("  component_order = ");
+      for(i=0;i<32;i++)
+         if(xifv->component_order[i]>=32) 
+            printf("%c",xifv->component_order[i]);
+      printf("\n");
+
+      printf("  scanline = ");
+      if(xifv->scanline_order == XvTopToBottom) printf("XvTopToBottom\n");
+      else if(xifv->scanline_order == XvBottomToTop) printf("XvBottomToTop\n");
+      else printf("Unknown\n");
+   }
+   printf("\n");
+}
+
 // WARNING This function may changes xv_port and surface_info!
-static int xvmc_find_surface_by_format(int format,int width,int height, 
-                       XvMCSurfaceInfo * surf_info,int query){
+static int xvmc_find_surface_by_format(int format,int width,int height,
+                                  XvMCSurfaceInfo * surf_info,int query){
 int rez;
 XvAdaptorInfo * ai;
 int num_adaptors,i;
 unsigned long p;
 int s,mc_surf_num;
 XvMCSurfaceInfo * mc_surf_list;
-   
+
    rez = XvQueryAdaptors(mDisplay,DefaultRootWindow(mDisplay),&num_adaptors,&ai);
    if( rez != Success ) return -1;
    if( verbose > 2 ) printf("vo_xvmc: Querying %d adaptors\n",num_adaptors);
@@ -171,13 +353,13 @@ XvMCSurfaceInfo * mc_surf_list;
 //we have XvMC list!
          for(s=0; s<mc_surf_num; s++)
          {
-	    if( width > mc_surf_list[s].max_width ) continue;
-	    if( height > mc_surf_list[s].max_height ) continue;
+            if( width > mc_surf_list[s].max_width ) continue;
+            if( height > mc_surf_list[s].max_height ) continue;
             if( xvmc_check_surface_format(format,&mc_surf_list[s])<0 ) continue;
 //we have match!
-	    
-	    if(!query){
-	       rez = XvGrabPort(mDisplay,p,CurrentTime);
+
+            if(!query){
+               rez = XvGrabPort(mDisplay,p,CurrentTime);
 	       if(rez != Success){
 	          if (verbose > 2) printf("vo_xvmc: Fail to grab port %ld\n",p);
 	          continue;
@@ -190,7 +372,7 @@ XvMCSurfaceInfo * mc_surf_list;
 	 XFree(mc_surf_list);//if mc_surf_num==0 is list==NULL ?
       }//for ports
    }//for adaptors
-   
+
    if(!query) printf("vo_xvmc: Could not find free matching surface. Sorry.\n");
    return 0;
 
@@ -210,7 +392,7 @@ xvmc_render_state_t * rndr;
    assert(mpi!=NULL);
    assert(mpi->flags &MP_IMGFLAG_DIRECT);
 //   assert(mpi->flags &MP_IMGFLAGS_DRAWBACK);
-   
+
    rndr = (xvmc_render_state_t*)mpi->priv;//there is copy in plane[2]
    assert( rndr != NULL );
    assert( rndr->magic == MP_XVMC_RENDER_MAGIC );
@@ -227,10 +409,10 @@ int xv_version,xv_release,xv_request_base,xv_event_base,xv_error_base;
 int mc_eventBase,mc_errorBase;
 int mc_ver,mc_rev;
 
-//Obtain display handler
+   //Obtain display handler
    if (!vo_init()) return -1;//vo_xv
 
-  //XvMC is subdivision of XVideo
+   //XvMC is subdivision of XVideo
    if (Success != XvQueryExtension(mDisplay,&xv_version,&xv_release,&xv_request_base,
 			 &xv_event_base,&xv_error_base) ){
       mp_msg(MSGT_VO,MSGL_ERR,"Sorry, Xv(MC) not supported by this X11 version/driver\n");
@@ -238,13 +420,13 @@ int mc_ver,mc_rev;
       return -1;
    }
    printf("vo_xvmc: X-Video extension %d.%d\n",xv_version,xv_release);
-   
+
    if( True != XvMCQueryExtension(mDisplay,&mc_eventBase,&mc_errorBase) ){	
       printf("vo_xvmc: No X-Video MotionCompensation Extension on %s\n",
 	      XDisplayName(NULL));
       return -1;
    }
-    
+
    if(Success == XvMCQueryVersion(mDisplay, &mc_ver, &mc_rev) ){
       printf("vo_xvmc: X-Video MotionCompensation Extension version %i.%i\n",
 		mc_ver,mc_rev);
@@ -253,11 +435,44 @@ int mc_ver,mc_rev;
       printf("vo_xvmc: Error querying version info!\n");
       return -1;
    }
+   surface_render = NULL;
    xv_port = 0;
    number_of_surfaces = 0;
-   keycolor_handling = 3;//!!fixme
-   surface_render=NULL;
+   keycolor_handling = MANUALFILL_COLORKEY;//fixme
+   subpicture_alloc = 0;
    
+   benchmark = 0; //disable PutImageto allow faster display than screen refresh
+   busy_wait = 1;
+   use_queue = 0;
+   if(arg)
+      while(*arg){
+         if(strncmp(arg,"benchmark",9) == 0){
+            arg+=9;
+            if(*arg == ':') arg++;
+            benchmark = 1;//disable PutImageto allow faster display than screen refresh
+            continue;
+         }
+         if(strncmp(arg,"wait",4) == 0){
+            arg+=4;
+            if(*arg == ':') arg++;
+            busy_wait = 1;
+            continue;
+         }
+         if(strncmp(arg,"sleep",5) == 0){
+            arg+=5;
+            if(*arg == ':') arg++;
+            busy_wait = 0;
+            continue;
+         }
+         if(strncmp(arg,"queue",5) == 0){
+            arg+=5;
+            if(*arg == ':') arg++;
+            use_queue = 1;
+            continue;
+         }
+         break;
+      }
+
    return 0;
 }
 
@@ -297,7 +512,7 @@ static uint32_t vm_height;
    numblocks=((width+15)/16)*((height+15)/16);
 // Find Supported Surface Type
    mode_id = xvmc_find_surface_by_format(format,width,height,&surface_info,0);//false=1 to grab port, not query
-   
+
    rez = XvMCCreateContext(mDisplay, xv_port,mode_id,width,height,XVMC_DIRECT,&ctx);
    if( rez != Success ) return -1;
    if( ctx.flags & XVMC_DIRECT ){
@@ -306,10 +521,11 @@ static uint32_t vm_height;
       printf("vo_xvmc: Allocated Indirect Context!\n");
    }
 
+
    blocks_per_macroblock = 6;
    if(surface_info.chroma_format == XVMC_CHROMA_FORMAT_422)
       blocks_per_macroblock = 8;
-   if(surface_info.chroma_format ==  XVMC_CHROMA_FORMAT_444)
+   if(surface_info.chroma_format == XVMC_CHROMA_FORMAT_444)
       blocks_per_macroblock = 12;
 
    rez = XvMCCreateBlocks(mDisplay,&ctx,numblocks*blocks_per_macroblock,&data_blocks);
@@ -329,7 +545,7 @@ static uint32_t vm_height;
 
    if(surface_render==NULL)
       surface_render=malloc(MAX_SURFACES*sizeof(xvmc_render_state_t));//easy mem debug
-   
+
    for(i=0; i<MAX_SURFACES; i++){
       rez=XvMCCreateSurface(mDisplay,&ctx,&surface_array[i]);
       if( rez != Success )
@@ -362,6 +578,71 @@ static uint32_t vm_height;
            (surface_info.mc_type & XVMC_IDCT) == XVMC_IDCT,
 	   (surface_info.flags & XVMC_INTRA_UNSIGNED) == XVMC_INTRA_UNSIGNED);
 
+// Find way to display OSD & subtitle
+   printf("vo_xvmc: looking for OSD support\n");
+   subpicture_mode = NO_SUBPICTURE;
+   if(surface_info.flags & XVMC_OVERLAID_SURFACE)
+      subpicture_mode = OVERLAY_SUBPICTURE;
+
+   if(surface_info.subpicture_max_width  != 0 && 
+      surface_info.subpicture_max_height != 0  ){
+      int s,k,num_subpic;
+
+      XvImageFormatValues * xvfmv;
+      xvfmv = XvMCListSubpictureTypes(mDisplay, xv_port,
+                      surface_info.surface_type_id, &num_subpic);
+
+      if(num_subpic != 0 && xvfmv != NULL){
+         if(verbose > 3){//Print All subpicture types for debug
+            for(s=0;s<num_subpic;s++)
+               printf("    Subpicture id 0x%08X\n",xvfmv[s].id);
+         }
+
+         if(verbose > 0)
+         for(s=0;s<num_subpic;s++){
+            for(k=0;osd_render[k].draw_func_ptr!=NULL;k++)
+
+            if(xvfmv[s].id == osd_render[k].id)
+            {  
+               init_osd_fnc  = osd_render[k].init_func_ptr;
+               draw_osd_fnc  = osd_render[k].draw_func_ptr;
+               clear_osd_fnc = osd_render[k].clear_func_ptr;
+
+               subpicture_mode = BLEND_SUBPICTURE;
+               subpicture_info = xvfmv[s];
+               print_xvimage_format_values(&subpicture_info);
+               goto found_subpic;
+            }
+         }
+found_subpic:
+         XFree(xvfmv); 
+      }
+      //Blend2 supicture is always possible, blend1 only at backend
+      if( (subpicture_mode == BLEND_SUBPICTURE) &&
+          (surface_info.flags & XVMC_BACKEND_SUBPICTURE) )
+      {
+         subpicture_mode = BACKEND_SUBPICTURE;
+      }
+
+   }
+
+   switch(subpicture_mode){
+      case NO_SUBPICTURE:
+         printf("vo_xvmc: No OSD support for this mode\n");
+         break;
+      case OVERLAY_SUBPICTURE:
+         printf("vo_xvmc: OSD support via color key tricks\n");
+         printf("vo_xvmc: not yet implemented:(\n");
+         break;
+      case BLEND_SUBPICTURE:
+         printf("vo_xvmc: OSD support by additional frontend rendering\n");
+         break;
+      case BACKEND_SUBPICTURE:
+         printf("vo_xvmc: OSD support by beckend rendering (fast)\n");
+         printf("vo_xvmc: Pleace send feedback to configrm that it work,otherwise send bugreport!\n");
+         break;
+   }
+
    init_keycolor();// take keycolor value and choose method for handling it
 
 //taken from vo_xv
@@ -372,13 +653,13 @@ static uint32_t vm_height;
 
    image_height = height;
    image_width = width;
- 
+
    vo_mouse_autohide = 1;
 
    vo_dx=( vo_screenwidth - d_width ) / 2; vo_dy=( vo_screenheight - d_height ) / 2;
    geometry(&vo_dx, &vo_dy, &d_width, &d_height, vo_screenwidth, vo_screenheight);
    vo_dwidth=d_width; vo_dheight=d_height;
-     
+
 #ifdef HAVE_XF86VM
    if( flags&0x02 ) vm = 1;
 #endif
@@ -435,7 +716,7 @@ static uint32_t vm_height;
    XMatchVisualInfo(mDisplay, mScreen, depth, TrueColor, &vinfo);
 
    xswa.background_pixel = 0;
-   if (keycolor_handling == 1)
+   if (keycolor_handling == BACKGROUND_COLORKEY)
       xswa.background_pixel = keycolor;// 2110;
    xswa.border_pixel     = 0;
    xswamask = CWBackPixel | CWBorderPixel;
@@ -477,9 +758,9 @@ static uint32_t vm_height;
 	 if ( !vo_fs ) XMoveResizeWindow( mDisplay,vo_window,hint.x,hint.y,hint.width,hint.height );
 	 if ( flags&1 && !vo_fs ) vo_x11_fullscreen(); // handle -fs on non-first file
       }
-    
+
 //    vo_x11_sizehint( hint.x, hint.y, hint.width, hint.height,0 );   
-    
+
       if ( vo_gc != None ) XFreeGC( mDisplay,vo_gc );
       vo_gc = XCreateGC(mDisplay, vo_window, GCForeground, &xgcv);
       XFlush(mDisplay);
@@ -514,74 +795,326 @@ static uint32_t vm_height;
 
    saver_off(mDisplay);  // turning off screen saver
 //end vo_xv
-   
+
    /* store image dimesions for displaying */
    p_render_surface_visible = NULL;
    p_render_surface_to_show = NULL;
- 
+
+   free_element = 0;
+
    vo_directrendering = 1;//ugly hack, coz xvmc works only with direct rendering
    return 0;		
 }
 
 static uint32_t draw_frame(uint8_t *srcp[]){
-assert(0 && srcp==NULL);//silense unused srcp warning
+   UNUSED(srcp);
+   assert(0);
+}
+
+static void init_osd_yuv_pal(){
+   char * palette;
+   int rez;
+   int i,j;
+   int snum,seb;
+   int Y,U,V;
+
+   subpicture_clear_color = 0;
+
+   if(subpicture.num_palette_entries > 0){
+
+      snum = subpicture.num_palette_entries;
+      seb = subpicture.entry_bytes;
+      palette = (char*)malloc(snum*seb);//check fail
+      if(palette == NULL) return;
+      for(i=0; i<snum; i++){
+         // 0-black max-white the other are gradients
+         Y = i*(1 << subpicture_info.y_sample_bits)/snum;//snum=2;->(0),(1*(1<<1)/2)
+         U = 1 << (subpicture_info.u_sample_bits - 1);
+         V = 1 << (subpicture_info.v_sample_bits - 1);
+         for(j=0; j<seb; j++)
+            switch(subpicture.component_order[j]){
+               case 'U': palette[i*seb+j] = U; break;
+               case 'V': palette[i*seb+j] = V; break;
+               case 'Y': 
+               default:
+                         palette[i*seb+j] = Y; break;
+         }
+      }
+      rez = XvMCSetSubpicturePalette(mDisplay, &subpicture, palette);
+      if(rez!=Success){
+         printf("vo_xvmc: set pallete fail\n");
+      }
+   }
+}
+
+static void clear_osd_subpic(int x0, int y0, int w, int h){
+int rez;
+   rez=XvMCClearSubpicture(mDisplay, &subpicture,
+                       x0, y0, w,h,
+                       subpicture_clear_color);
+   if(rez != Success)
+      printf("vo_xvmc: XvMCClearSubpicture failed!\n");
+}
+
+static void OSD_init(){
+unsigned short osd_height, osd_width;
+int rez;
+
+   if(subpicture_alloc){
+      if(verbose>3)
+         printf("vo_xvmc: destroying subpicture\n");
+      XvMCDestroySubpicture(mDisplay,&subpicture);
+      deallocate_xvimage();
+      subpicture_alloc = 0;
+   }
+
+/*   if(surface_info.flags & XVMC_SUBPICTURE_INDEPENDENT_SCALING){
+      osd_width = vo_dwidth;
+      osd_height = vo_dheight;
+   }else*/
+   {
+      osd_width = image_width;
+      osd_height = image_height;
+   }
+
+   if(osd_width > surface_info.subpicture_max_width)
+      osd_width = surface_info.subpicture_max_width;
+   if(osd_height > surface_info.subpicture_max_height)
+      osd_height = surface_info.subpicture_max_height;
+   if(osd_width == 0 || osd_height == 0) 
+      return;//if called before window size is known
+
+   if(verbose > 3)
+      printf("vo_xvmc: creating subpicture (%d,%d) format %X\n",
+              osd_width,osd_height,subpicture_info.id);
+
+   rez = XvMCCreateSubpicture(mDisplay,&ctx,&subpicture,
+                           osd_width,osd_height,subpicture_info.id);
+   if(rez != Success){
+      subpicture_mode = NO_SUBPICTURE;
+      printf("vo_xvmc: Create Subpicture failed, OSD disabled\n");
+      return;
+   }
+   if(verbose > 3){
+   int i;
+      printf("vo_xvmc: Created Subpicture:\n");
+      printf("         xvimage_id=0x%X\n",subpicture.xvimage_id);
+      printf("         width=%d\n",subpicture.width);
+      printf("         height=%d\n",subpicture.height);
+      printf("         num_palette_entries=0x%X\n",subpicture.num_palette_entries);
+      printf("         entry_bytes=0x%X\n",subpicture.entry_bytes);
+
+      printf("         component_order=\"");
+      for(i=0; i<4; i++)
+         if(subpicture.component_order[i] >= 32)
+            printf("%c", subpicture.component_order[i]);
+      printf("\"\n");
+   }
+   
+   //call init for the surface type
+   init_osd_fnc();//init palete,clear color etc ...
+   if(verbose > 3)
+      printf("vo_xvmc: clearing subpicture\n");
+   clear_osd_fnc(0, 0, subpicture.width, subpicture.height);
+
+   allocate_xvimage(subpicture.width, subpicture.height, subpicture_info.id);
+   subpicture_alloc = 1;
+}
+
+static void draw_osd_IA44(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+int ox,oy;
+int rez;
+
+   if(verbose > 3)
+      printf("vo_xvmc:composite AI44 subpicture (%d,%d - %d,%d)\n",x0,y0,w,h);
+
+   for(ox=0; ox<w; ox++){
+      for(oy=0; oy<h; oy++){
+         xvimage->data[oy*xvimage->width+ox] = (src[oy*stride+ox]>>4) | ((0-srca[oy*stride+ox])&0xf0);
+      }
+   }
+   rez = XvMCCompositeSubpicture(mDisplay, &subpicture, xvimage, 0, 0,
+                           w,h,x0,y0);
+   if(rez != Success){
+      printf("vo_xvmc: composite subpicture failed\n");
+      assert(0);
+   }
+}
+
+static void draw_osd_AI44(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+int ox,oy;
+int rez;
+   if( verbose > 3)
+      printf("vo_xvmc:composite AI44 subpicture (%d,%d - %d,%d)\n",x0,y0,w,h);
+
+   for(ox=0; ox<w; ox++){
+      for(oy=0; oy<h; oy++){
+         xvimage->data[oy*xvimage->width+ox] = (src[oy*stride+ox]&0xf0) | (((0-srca[oy*stride+ox])>>4)&0xf);
+      }
+   }
+   rez = XvMCCompositeSubpicture(mDisplay, &subpicture, xvimage, 0, 0,
+                           w,h,x0,y0);
+   if(rez != Success){
+      printf("vo_xvmc: composite subpicture failed\n");
+      assert(0);
+   }
 }
 
 static void draw_osd(void){
+xvmc_render_state_t * osd_rndr;
+int osd_has_changed;
+int have_osd_to_draw;
+int rez;
+
+   if(verbose > 3)
+      printf("vo_xvmc: draw_osd ,OSD_mode=%d, surface_to_show=%p\n",
+             subpicture_mode,p_render_surface_to_show);
+
+   if(subpicture_mode == BLEND_SUBPICTURE || 
+      subpicture_mode == BACKEND_SUBPICTURE ){
+
+      if(!subpicture_alloc) //allocate subpicture when dimensions are known
+         OSD_init();
+      if(!subpicture_alloc) 
+         return;//dimensions still unknown.
+
+      osd_has_changed = vo_update_osd(subpicture.width, subpicture.height);
+      have_osd_to_draw = vo_osd_check_range_update(0, 0, subpicture.width, 
+                                                         subpicture.height); 
+
+      if(!have_osd_to_draw)
+         return;//nothing to draw,no subpic, no blend
+
+      if(osd_has_changed){
+         //vo_remove_text(subpicture.width, subpicture.height,clear_osd_fnc)
+         clear_osd_fnc(0,0,subpicture.width,subpicture.height);
+         vo_draw_text(subpicture.width, subpicture.height, draw_osd_fnc);
+      }
+      XvMCSyncSubpicture(mDisplay,&subpicture);//todo usleeep wait!
+
+      if(subpicture_mode == BLEND_SUBPICTURE){
+         osd_rndr = find_free_surface();
+         if(osd_rndr == NULL) 
+            return;// no free surface to draw OSD in
+
+         rez = XvMCBlendSubpicture2(mDisplay,
+                       p_render_surface_to_show->p_surface, osd_rndr->p_surface,
+                       &subpicture,
+                       0, 0, subpicture.width, subpicture.height,
+                       0, 0, image_width, image_height);
+         if(rez!=Success){
+            printf("vo_xvmc: BlendSubpicture failed rez=%d\n",rez);
+            assert(0);
+            return;
+         }
+//       XvMCFlushSurface(mDisplay,osd_rndr->p_surface);//fixme- should I?
+
+         //When replaceing the surface with osd one, save the flags too!
+         osd_rndr->picture_structure = p_render_surface_to_show->picture_structure;
+         osd_rndr->display_flags = p_render_surface_to_show->display_flags;
+//add more if needed    osd_rndr-> = p_render_surface_to_show->;
+
+         p_render_surface_to_show->state &= ~MP_XVMC_STATE_DISPLAY_PENDING;
+         p_render_surface_to_show->state |= MP_XVMC_STATE_OSD_SOURCE;
+         p_render_surface_to_show->p_osd_target_surface_render = osd_rndr;
+
+         p_render_surface_to_show = osd_rndr;
+         p_render_surface_to_show->state = MP_XVMC_STATE_DISPLAY_PENDING;
+
+         if(verbose > 3)
+            printf("vo_xvmc:draw_osd: surface_to_show changed to %p\n",osd_rndr);
+      }//endof if(BLEND)
+      if(subpicture_mode == BACKEND_SUBPICTURE){
+         rez = XvMCBlendSubpicture(mDisplay,
+                       p_render_surface_to_show->p_surface,
+                       &subpicture,
+                       0, 0, subpicture.width, subpicture.height,
+                       0, 0, image_width, image_height);
+
+      }
+   
+   }//if(BLEND||BACKEND)
 }
 
 static void xvmc_sync_surface(XvMCSurface * srf){
 int status,rez;
    rez = XvMCGetSurfaceStatus(mDisplay,srf,&status);
    assert(rez==Success);
-   if( status & XVMC_RENDERING )
-      XvMCSyncSurface(mDisplay, srf);
-/*
-   rez = XvMCFlushSurface(mDisplay, srf);
-   assert(rez==Success);
-
-   do {
-   usleep(1);
-   printf("waiting...\n");
-   XvMCGetSurfaceStatus(mDisplay,srf,&status);
-   } while (status & XVMC_RENDERING);       
-*/
+   if((status & XVMC_RENDERING) == 0)
+      return;//surface is already complete
+   if(!busy_wait){
+      rez = XvMCFlushSurface(mDisplay, srf);
+      assert(rez==Success);
+
+      do{
+         usec_sleep(1000);//1ms (may be 20ms on linux)
+         XvMCGetSurfaceStatus(mDisplay,srf,&status);
+      } while (status & XVMC_RENDERING);
+      return;//done
+   }       
+
+   XvMCSyncSurface(mDisplay, srf);
 }
 
 static void flip_page(void){
 int rez;
 int clipX,clipY,clipW,clipH;
+int i,cfs;
 
    clipX = drwX-(vo_panscan_x>>1);
    clipY = drwY-(vo_panscan_y>>1); 
    clipW = vo_dwidth+vo_panscan_x;
-   clipH = vo_dheight+vo_panscan_y;//
-   
+   clipH = vo_dheight+vo_panscan_y;
+
    if( verbose > 3 ) 
       printf("vo_xvmc: flip_page  show(rndr=%p)\n\n",p_render_surface_to_show);
 
    if(p_render_surface_to_show == NULL) return;
    assert( p_render_surface_to_show->magic == MP_XVMC_RENDER_MAGIC );
 //fixme   assert( p_render_surface_to_show != p_render_surface_visible);
-   
+
+   if(use_queue){
+      // fill the queue until only n free surfaces remain
+      // after that start displaying
+      cfs = count_free_surfaces();
+      show_queue[free_element++] = p_render_surface_to_show;
+      if(cfs > 3){//well have 3 free surfaces after add queue
+         if(free_element > 1)//a little voodoo magic
+            xvmc_sync_surface(show_queue[0]->p_surface);
+         return; 
+      }
+      p_render_surface_to_show=show_queue[0];
+      if(verbose > 4)
+         printf("vo_xvmc: flip_queue free_element=%d\n",free_element);
+      free_element--;
+      for(i=0; i<free_element; i++){
+         show_queue[i] = show_queue[i+1];
+      }
+      show_queue[free_element] = NULL;
+   }
+
 // make sure the rendering is done
    xvmc_sync_surface(p_render_surface_to_show->p_surface);
 
 //the visible surface won't be displayed anymore, mark it as free
-   if( p_render_surface_visible!=NULL ) 
+   if(p_render_surface_visible != NULL)
       p_render_surface_visible->state &= ~MP_XVMC_STATE_DISPLAY_PENDING;
 
 //!!fixme   assert(p_render_surface_to_show->state & MP_XVMC_STATE_DISPLAY_PENDING);
 
-// show it
-// if(benchmark)
-   rez=XvMCPutSurface(mDisplay, p_render_surface_to_show->p_surface, vo_window,
-                      0, 0, image_width, image_height,
-                      clipX, clipY, clipW, clipH,
-                      3);//p_render_surface_to_show->display_flags);
+   //show it, displaying is always vsynced, so skip it for benchmark
+   if(!benchmark){
+      rez = XvMCPutSurface(mDisplay, p_render_surface_to_show->p_surface, 
+                         vo_window,
+                         0, 0, image_width, image_height,
+                         clipX, clipY, clipW, clipH,
+                         3);//p_render_surface_to_show->display_flags);
+      if(rez != Success){
+         printf("vo_xvmc: PutSurface failer, critical error!\n");
+         assert(0);
+      }
+   }
 
-   assert(rez==Success);
-   
    p_render_surface_visible = p_render_surface_to_show;
    p_render_surface_to_show = NULL;
 }
@@ -595,7 +1128,7 @@ int e=vo_x11_check_events(mDisplay);
    if(e&VO_EVENT_RESIZE)
    {
       e |= VO_EVENT_EXPOSE;
-      
+
       XGetGeometry( mDisplay,vo_window,&mRoot,&drwX,&drwY,&vo_dwidth,&vo_dheight,
                    &drwBorderWidth,&drwDepth );
       drwX = drwY = 0;
@@ -615,9 +1148,9 @@ int e=vo_x11_check_events(mDisplay);
    }
    if ( e & VO_EVENT_EXPOSE )
    {
-      mDrawColorKey(drwX,drwY,vo_dwidth,vo_dheight);     
-     if(p_render_surface_visible != NULL)
-      XvMCPutSurface(mDisplay, p_render_surface_visible->p_surface,vo_window,
+      mDrawColorKey(drwX,drwY,vo_dwidth,vo_dheight);
+      if(p_render_surface_visible != NULL)
+         XvMCPutSurface(mDisplay, p_render_surface_visible->p_surface,vo_window,
                      0, 0, image_width, image_height,
                      drwX,drwY,vo_dwidth,vo_dheight,
                      3);//,p_render_surface_visible->display_flags);!!
@@ -626,10 +1159,22 @@ int e=vo_x11_check_events(mDisplay);
 
 static void xvmc_free(void){
 int i;
+   if( subpicture_alloc ){
+
+      XvMCDestroySubpicture(mDisplay,&subpicture);
+      deallocate_xvimage();
+
+      subpicture_alloc = 0;
+
+      if(verbose > 3)
+         printf("vo_xvmc: subpicture destroyed\n");
+   }
 
    if( number_of_surfaces ){
+
       XvMCDestroyMacroBlocks(mDisplay,&mv_blocks);
       XvMCDestroyBlocks(mDisplay,&data_blocks);
+
       for(i=0; i<number_of_surfaces; i++)
       {
          XvMCHideSurface(mDisplay,&surface_array[i]);//it doesn't hurt, I hope
@@ -640,17 +1185,20 @@ int i;
             printf("vo_xvmc::uninit surface_render[%d].status=%d\n",i,
                     surface_render[i].state); 
       }
-      
+
       free(surface_render);surface_render=NULL;
 
       XvMCDestroyContext(mDisplay,&ctx);
-      if( verbose > 3) printf("vo_xvmc: Context sucessfuly freed\n");
       number_of_surfaces = 0;
+
+      if(verbose > 3) printf("vo_xvmc: Context sucessfuly freed\n");
    }
+
+
    if( xv_port !=0 ){
       XvUngrabPort(mDisplay,xv_port,CurrentTime);
       xv_port = 0;
-      if( verbose > 3) printf("vo_xvmc: xv_port sucessfuly ungrabed\n");
+      if(verbose > 3) printf("vo_xvmc: xv_port sucessfuly ungrabed\n");
    }
 }
 
@@ -675,13 +1223,14 @@ int mode_id;
    mode_id = xvmc_find_surface_by_format(format, 16, 16, &qsurface_info, 1);//true=1 - quering
 
    if( mode_id == 0 ) return 0;
-   
+
    flags = VFCAP_CSP_SUPPORTED |