diff -ur --exclude-from dontdiff vanilla/main/libmpcodecs/vf_expand.c main/libmpcodecs/vf_expand.c
--- vanilla/main/libmpcodecs/vf_expand.c	2002-10-17 09:26:04.000000000 +0200
+++ main/libmpcodecs/vf_expand.c	2002-11-05 18:53:42.000000000 +0100
@@ -67,7 +67,7 @@
 }
 
 static void draw_func(int x0,int y0, int w,int h,unsigned char* src, unsigned char *srca, int stride){
-    unsigned char* dst;
+    unsigned char* dst, *dst2, *dst3;
     if(!vo_osd_changed_flag && vf->priv->dmpi->planes[0]==vf->priv->fb_ptr){
 	// ok, enough to update the area inside the video, leave the black bands
 	// untouched!
@@ -115,7 +115,15 @@
     case IMGFMT_IF09:
     case IMGFMT_Y800:
     case IMGFMT_Y8:
-	vo_draw_alpha_yv12(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	dst2=vf->priv->dmpi->planes[1]+
+	    vf->priv->dmpi->stride[1]*y0+
+	    (vf->priv->dmpi->bpp>>4)*x0;
+	dst3=vf->priv->dmpi->planes[2]+
+	    vf->priv->dmpi->stride[2]*y0+
+	    (vf->priv->dmpi->bpp>>4)*x0;
+	vo_draw_alpha_yv12(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0],
+			   dst2,vf->priv->dmpi->stride[1],
+			   dst3,vf->priv->dmpi->stride[2]);
 	break;
     case IMGFMT_YUY2:
 	vo_draw_alpha_yuy2(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
@@ -137,10 +145,10 @@
 	    remove_func(0,0,vf->priv->exp_w,vf->priv->exp_h);
 	} else {
 	    // partial clear:
-	    vo_remove_text(vf->priv->exp_w,vf->priv->exp_h,remove_func);
+	    vo_remove_text(vf->priv->exp_w,vf->priv->exp_h,vf->priv->dmpi->imgfmt,remove_func);
 	}
     }
-    vo_draw_text(vf->priv->exp_w,vf->priv->exp_h,draw_func);
+    vo_draw_text(vf->priv->exp_w,vf->priv->exp_h,vf->priv->dmpi->imgfmt,draw_func);
     // save buffer pointer for double buffering detection - yes, i know it's
     // ugly method, but note that codecs with DR support does the same...
     vf->priv->fb_ptr=vf->priv->dmpi->planes[0];
@@ -187,7 +195,7 @@
 #ifdef OSD_SUPPORT
     if(vf->priv->osd && (mpi->flags&MP_IMGFLAG_PRESERVE)){
 	// check if we have to render osd!
-	vo_update_osd(vf->priv->exp_w, vf->priv->exp_h);
+	vo_update_osd(vf->priv->exp_w, vf->priv->exp_h, vf->priv->dmpi->imgfmt);
 	if(vo_osd_check_range_update(vf->priv->exp_x,vf->priv->exp_y,
 	    vf->priv->exp_x+mpi->w,vf->priv->exp_y+mpi->h)) return;
     }
diff -ur --exclude-from dontdiff vanilla/main/libvo/osd.c main/libvo/osd.c
--- vanilla/main/libvo/osd.c	2002-05-02 12:46:25.000000000 +0200
+++ main/libvo/osd.c	2002-11-05 18:53:42.000000000 +0100
@@ -2,9 +2,6 @@
 // These are "reference implementations", should be optimized later (MMX, etc)
 // Templating Code from Michael Niedermayer (michaelni@gmx.at) is under GPL
 
-//#define FAST_OSD
-//#define FAST_OSD_TABLE
-
 #include "config.h"
 #include "osd.h"
 #include "mp_msg.h"
@@ -16,6 +13,8 @@
 
 extern int verbose; // defined in mplayer.c
 
+#undef ARCH_X86
+
 #ifdef ARCH_X86
 #define CAN_COMPILE_X86_ASM
 #endif
@@ -28,9 +27,9 @@
 
 //Note: we have C, X86-nommx, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
 //Plain C versions
-#if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
+//#if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
 #define COMPILE_C
-#endif
+//#endif
 
 #ifdef CAN_COMPILE_X86_ASM
 
@@ -52,14 +51,14 @@
 #undef HAVE_3DNOW
 #undef ARCH_X86
 
-#ifdef COMPILE_C
+//#ifdef COMPILE_C
 #undef HAVE_MMX
 #undef HAVE_MMX2
 #undef HAVE_3DNOW
 #undef ARCH_X86
 #define RENAME(a) a ## _C
 #include "osd_template.c"
-#endif
+//#endif
 
 #ifdef CAN_COMPILE_X86_ASM
 
@@ -109,32 +108,32 @@
 
 #endif //CAN_COMPILE_X86_ASM
 
-void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase1, int dststride1, unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3){
 #ifdef RUNTIME_CPUDETECT
 #ifdef CAN_COMPILE_X86_ASM
 	// ordered per speed fasterst first
 	if(gCpuCaps.hasMMX2)
-		vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 	else if(gCpuCaps.has3DNow)
-		vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 	else if(gCpuCaps.hasMMX)
-		vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 	else
-		vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #else
-		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #endif
 #else //RUNTIME_CPUDETECT
 #ifdef HAVE_MMX2
-		vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_MMX2(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #elif defined (HAVE_3DNOW)
-		vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_3DNow(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #elif defined (HAVE_MMX)
-		vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_MMX(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #elif defined (ARCH_X86)
-		vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #else
-		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3);
 #endif
 #endif //!RUNTIME_CPUDETECT
 }
@@ -229,19 +228,7 @@
 #endif //!RUNTIME_CPUDETECT
 }
 
-#ifdef FAST_OSD_TABLE
-static unsigned short fast_osd_15bpp_table[256];
-static unsigned short fast_osd_16bpp_table[256];
-#endif
-
 void vo_draw_alpha_init(){
-#ifdef FAST_OSD_TABLE
-    int i;
-    for(i=0;i<256;i++){
-        fast_osd_15bpp_table[i]=((i>>3)<<10)|((i>>3)<<5)|(i>>3);
-        fast_osd_16bpp_table[i]=((i>>3)<<11)|((i>>2)<<5)|(i>>3);
-    }
-#endif
 //FIXME the optimized stuff is a lie for 15/16bpp as they arent optimized yet
 	if(verbose)
 	{
@@ -282,22 +269,13 @@
         register int x;
         for(x=0;x<w;x++){
             if(srca[x]){
-#ifdef FAST_OSD
-#ifdef FAST_OSD_TABLE
-                dst[x]=fast_osd_15bpp_table[src[x]];
-#else
-		register unsigned int a=src[x]>>3;
-                dst[x]=(a<<10)|(a<<5)|a;
-#endif
-#else
                 unsigned char r=dst[x]&0x1F;
                 unsigned char g=(dst[x]>>5)&0x1F;
                 unsigned char b=(dst[x]>>10)&0x1F;
-                r=(((r*srca[x])>>5)+src[x])>>3;
-                g=(((g*srca[x])>>5)+src[x])>>3;
-                b=(((b*srca[x])>>5)+src[x])>>3;
+                r=(((r*srca[x])>>5)+src[3*x])>>3;
+                g=(((g*srca[x])>>5)+src[3*x+1])>>3;
+                b=(((b*srca[x])>>5)+src[3*x+2])>>3;
                 dst[x]=(b<<10)|(g<<5)|r;
-#endif
             }
         }
         src+=srcstride;
@@ -314,21 +292,13 @@
         register int x;
         for(x=0;x<w;x++){
             if(srca[x]){
-#ifdef FAST_OSD
-#ifdef FAST_OSD_TABLE
-                dst[x]=fast_osd_16bpp_table[src[x]];
-#else
-                dst[x]=((src[x]>>3)<<11)|((src[x]>>2)<<5)|(src[x]>>3);
-#endif
-#else
                 unsigned char r=dst[x]&0x1F;
                 unsigned char g=(dst[x]>>5)&0x3F;
                 unsigned char b=(dst[x]>>11)&0x1F;
-                r=(((r*srca[x])>>5)+src[x])>>3;
-                g=(((g*srca[x])>>6)+src[x])>>2;
-                b=(((b*srca[x])>>5)+src[x])>>3;
+                r=(((r*srca[x])>>5)+src[3*x])>>3;
+                g=(((g*srca[x])>>6)+src[3*x+1])>>2;
+                b=(((b*srca[x])>>5)+src[3*x+2])>>3;
                 dst[x]=(b<<11)|(g<<5)|r;
-#endif
             }
         }
         src+=srcstride;
diff -ur --exclude-from dontdiff vanilla/main/libvo/osd.h main/libvo/osd.h
--- vanilla/main/libvo/osd.h	2001-06-11 00:25:09.000000000 +0200
+++ main/libvo/osd.h	2002-11-05 18:53:42.000000000 +0100
@@ -7,7 +7,7 @@
 
 extern void vo_draw_alpha_init(); // build tables
 
-extern void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride);
+extern void vo_draw_alpha_yv12(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride1,unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3);
 extern void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride);
 extern void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride);
 extern void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride);
diff -ur --exclude-from dontdiff vanilla/main/libvo/osd_template.c main/libvo/osd_template.c
--- vanilla/main/libvo/osd_template.c	2002-03-16 18:12:19.000000000 +0100
+++ main/libvo/osd_template.c	2002-11-05 18:53:42.000000000 +0100
@@ -27,11 +27,9 @@
 #define EMMS     "emms"
 #endif
 
-static inline void RENAME(vo_draw_alpha_yv12)(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+static inline void RENAME(vo_draw_alpha_yv12)(int w,int h, unsigned char* src, unsigned char *srca1, int srcstride, unsigned char* dstbase1,int dststride1, unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3){
     int y;
-#if defined(FAST_OSD) && !defined(HAVE_MMX)
-    w=w>>1;
-#endif
+    unsigned char *srca = srca1;
 PROFILE_START();
     for(y=0;y<h;y++){
         register int x;
@@ -45,7 +43,7 @@
 	"movq %%mm5, %%mm4\n\t"
 	"psllw $8, %%mm5\n\t" //FF00FF00FF00
 	"psrlw $8, %%mm4\n\t" //00FF00FF00FF
-	::"m"(*dstbase),"m"(*srca),"m"(*src):"memory");
+	::"m"(*dstbase1),"m"(*srca),"m"(*src):"memory");
     for(x=0;x<w;x+=8){
 	asm volatile(
 		"movl %1, %%eax\n\t"
@@ -71,23 +69,43 @@
 		"paddb	%2, %%mm0\n\t"
 		"movq	%%mm0, %0\n\t"
 		"1:\n\t"
-		:: "m" (dstbase[x]), "m" (srca[x]), "m" (src[x])
+		:: "m" (dstbase1[x]), "m" (srca[x]), "m" (src[x])
 		: "%eax");
 	}
 #else
         for(x=0;x<w;x++){
-#ifdef FAST_OSD
-            if(srca[2*x+0]) dstbase[2*x+0]=src[2*x+0];
-            if(srca[2*x+1]) dstbase[2*x+1]=src[2*x+1];
-#else
-            if(srca[x]) dstbase[x]=((dstbase[x]*srca[x])>>8)+src[x];
-#endif
+            if(srca[x]) dstbase1[x]=((dstbase1[x]*srca[x])>>8)+src[x];
         }
 #endif
         src+=srcstride;
         srca+=srcstride;
-        dstbase+=dststride;
+        dstbase1+=dststride1;
+    }
+
+    srca = srca1;
+    for(y=0;y<h/2;y++){
+        register int x;
+        for(x=0;x<w/2;x++){
+            if(srca[2*x] && srca[2*x] < 128) 
+		dstbase2[x]=src[x];
+	}
+        dstbase2+=dststride2;
+	src+=srcstride/2;
+        srca+=srcstride*2;
+    }
+
+    srca = srca1;
+    for(y=0;y<h/2;y++){
+        register int x;
+        for(x=0;x<w/2;x++){
+            if(srca[2*x] && srca[2*x] < 128)
+		dstbase3[x]=src[x];
+	}
+        dstbase3+=dststride3;
+	src+=srcstride/2;
+        srca+=srcstride*2;
     }
+
 #ifdef HAVE_MMX
 	asm volatile(EMMS:::"memory");
 #endif
@@ -97,10 +115,10 @@
 
 static inline void RENAME(vo_draw_alpha_yuy2)(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
     int y;
-#if defined(FAST_OSD) && !defined(HAVE_MMX)
-    w=w>>1;
-#endif
+    return;
+    
 PROFILE_START();
+
     for(y=0;y<h;y++){
         register int x;
 #ifdef HAVE_MMX
@@ -141,16 +159,11 @@
 		: "%eax");
 	}
 #else
-        for(x=0;x<w;x++){
-#ifdef FAST_OSD
-            if(srca[2*x+0]) dstbase[4*x+0]=src[2*x+0];
-            if(srca[2*x+1]) dstbase[4*x+2]=src[2*x+1];
-#else
+        for(x=1;x<w;x++){
             if(srca[x]) {
-               dstbase[2*x]=((dstbase[2*x]*srca[x])>>8)+src[x];
-               dstbase[2*x+1]=((((signed)dstbase[2*x+1]-128)*srca[x])>>8)+128;
+		dstbase[2*x]=((dstbase[2*x]*srca[x])>>8)+src[2*x];
+		dstbase[2*x+1]=((((signed)dstbase[2*x+1]-128)*srca[x])>>8)+128+src[2*x+1];
            }
-#endif
         }
 #endif
 	src+=srcstride;
@@ -245,13 +258,9 @@
 #else /*non x86 arch*/
         for(x=0;x<w;x++){
             if(srca[x]){
-#ifdef FAST_OSD
-		dst[0]=dst[1]=dst[2]=src[x];
-#else
-		dst[0]=((dst[0]*srca[x])>>8)+src[x];
-		dst[1]=((dst[1]*srca[x])>>8)+src[x];
-		dst[2]=((dst[2]*srca[x])>>8)+src[x];
-#endif
+		dst[0]=((dst[0]*srca[x])>>8)+src[3*x];
+		dst[1]=((dst[1]*srca[x])>>8)+src[3*x+1];
+		dst[2]=((dst[2]*srca[x])>>8)+src[3*x+2];
             }
             dst+=3; // 24bpp
         }
@@ -400,17 +409,13 @@
 #else /*non x86 arch*/
         for(x=0;x<w;x++){
             if(srca[x]){
-#ifdef FAST_OSD
-		dstbase[4*x+0]=dstbase[4*x+1]=dstbase[4*x+2]=src[x];
-#else
-		dstbase[4*x+0]=((dstbase[4*x+0]*srca[x])>>8)+src[x];
-		dstbase[4*x+1]=((dstbase[4*x+1]*srca[x])>>8)+src[x];
-		dstbase[4*x+2]=((dstbase[4*x+2]*srca[x])>>8)+src[x];
-#endif
+		dstbase[4*x+0]=((dstbase[4*x+0]*srca[x])>>8)+src[3*x];
+		dstbase[4*x+1]=((dstbase[4*x+1]*srca[x])>>8)+src[3*x+1];
+		dstbase[4*x+2]=((dstbase[4*x+2]*srca[x])>>8)+src[3*x+2];
             }
         }
 #endif /* arch_x86 */
-        src+=srcstride;
+        src+=srcstride*3;
         srca+=srcstride;
         dstbase+=dststride;
     }
diff -ur --exclude-from dontdiff vanilla/main/libvo/sub.c main/libvo/sub.c
--- vanilla/main/libvo/sub.c	2002-10-06 23:04:28.000000000 +0200
+++ main/libvo/sub.c	2002-11-05 19:11:22.000000000 +0100
@@ -43,14 +43,31 @@
     return h;
 }
 
+static int xpal = 0;
+
+#define PALSIZE 9
+/* R,G,B */
+static int palette[PALSIZE][3] = {
+    {128, 128, 128},
+    {128, 128, 255},
+    {128, 255, 128},
+    {128, 255, 255},
+    {255, 128, 128},
+    {255, 128, 255},
+    {255, 255, 128},
+    {255, 255, 255},
+};
+
+int vo_convert_osd_format(mp_osd_obj_t* obj, uint32_t format);
+
 // renders char to a big per-object buffer where alpha and bitmap are separated
-static void draw_alpha_buf(mp_osd_obj_t* obj, int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_buf(mp_osd_obj_t* obj, int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride, unsigned char pal_idx)
 {
     int dststride = obj->stride;
     int dstskip = obj->stride-w;
     int srcskip = stride-w;
     int i, j;
-    unsigned char *b = obj->bitmap_buffer + (y0-obj->bbox.y1)*dststride + (x0-obj->bbox.x1);
+    unsigned char *b = obj->rgb_buffer + (y0-obj->bbox.y1)*dststride*3 + (x0-obj->bbox.x1)*3;
     unsigned char *a = obj->alpha_buffer  + (y0-obj->bbox.y1)*dststride + (x0-obj->bbox.x1);
     unsigned char *bs = src;
     unsigned char *as = srca;
@@ -63,17 +80,26 @@
     }
     
     for (i = 0; i < h; i++) {
-	for (j = 0; j < w; j++, b++, a++, bs++, as++) {
-	    if (*b < *bs) *b = *bs;
+	for (j = 0; j < w; j++, a++, b+=3, bs++, as++) {
+	    if (*b < *bs) {
+		b[0] = *bs*palette[xpal][0] >> 8;
+		b[1] = *bs*palette[xpal][1] >> 8;
+		b[2] = *bs*palette[xpal][2] >> 8;
+	    }
 	    if (*as) {
-		if (*a == 0 || *a > *as) *a = *as;
+		if (*a == 0 || *a > *as) {
+		    *a = *as;
+		}
 	    }
 	}
-	b+= dstskip;
+	b+= dstskip*3;
 	a+= dstskip;
 	bs+= srcskip;
 	as+= srcskip;
     }
+    xpal++;
+    
+    if (xpal > 7) xpal = 0;
 }
 
 // allocates/enlarges the alpha/bitmap buffer
@@ -82,17 +108,22 @@
     int len;
     if (obj->bbox.x2 < obj->bbox.x1) obj->bbox.x2 = obj->bbox.x1;
     if (obj->bbox.y2 < obj->bbox.y1) obj->bbox.y2 = obj->bbox.y1;
+    if ((obj->bbox.y2-obj->bbox.y1) & 1) obj->bbox.y1--;
     obj->stride = ((obj->bbox.x2-obj->bbox.x1)+7)&(~7);
     len = obj->stride*(obj->bbox.y2-obj->bbox.y1);
+    fprintf(stderr, "alloc: %d %d\n", obj->stride, obj->bbox.y2-obj->bbox.y1);
     if (obj->allocated<len) {
 	obj->allocated = len;
-	free(obj->bitmap_buffer);
+	free(obj->rgb_buffer);
+	free(obj->dest_buffer);
 	free(obj->alpha_buffer);
-	obj->bitmap_buffer = (unsigned char *)memalign(16, len);
 	obj->alpha_buffer = (unsigned char *)memalign(16, len);
+	obj->rgb_buffer = (unsigned char *)memalign(16, len*3);
+	obj->dest_buffer = (unsigned char *)memalign(16, len*4);
     }
-    memset(obj->bitmap_buffer, 0, len);
     memset(obj->alpha_buffer, 0, len);
+    memset(obj->rgb_buffer, 0, len*3);
+    memset(obj->dest_buffer, 0, len*4);
 }
 
 // renders the buffer
@@ -101,13 +132,13 @@
 	draw_alpha(obj->bbox.x1,obj->bbox.y1,
 		   obj->bbox.x2-obj->bbox.x1,
 		   obj->bbox.y2-obj->bbox.y1,
-		   obj->bitmap_buffer,
+		   obj->dest_buffer,
 		   obj->alpha_buffer,
 		   obj->stride);
     }
 }
 
-inline static void vo_update_text_osd(mp_osd_obj_t* obj,int dxs,int dys){
+inline static void vo_update_text_osd(mp_osd_obj_t* obj,int dxs,int dys,uint32_t format){
 	unsigned char *cp=vo_osd_text;
 	int x=20;
 	int h=0;
@@ -139,9 +170,12 @@
 			   vo_font->pic_a[font]->h,
 			   vo_font->pic_b[font]->bmp+vo_font->start[c],
 			   vo_font->pic_a[font]->bmp+vo_font->start[c],
-			   vo_font->pic_a[font]->w);
+			   vo_font->pic_a[font]->w,
+			   1);
           x+=vo_font->width[c]+vo_font->charspace;
         }
+
+	vo_convert_osd_format(obj, format);
 }
 
 int vo_osd_progbar_type=-1;
@@ -155,7 +189,7 @@
 // 
 //  the above schema is rescalled to n=elems bars
 
-inline static void vo_update_text_progbar(mp_osd_obj_t* obj,int dxs,int dys){
+inline static void vo_update_text_progbar(mp_osd_obj_t* obj,int dxs,int dys,uint32_t format){
 
     obj->flags|=OSDFLAG_CHANGED|OSDFLAG_VISIBLE;
     
@@ -230,7 +264,7 @@
               vo_font->pic_a[font]->h,
               vo_font->pic_b[font]->bmp+vo_font->start[c],
               vo_font->pic_a[font]->bmp+vo_font->start[c],
-              vo_font->pic_a[font]->w);
+              vo_font->pic_a[font]->w, 1);
 	}
    
         c=OSD_PB_START;
@@ -240,7 +274,7 @@
               vo_font->pic_a[font]->h,
               vo_font->pic_b[font]->bmp+vo_font->start[c],
               vo_font->pic_a[font]->bmp+vo_font->start[c],
-              vo_font->pic_a[font]->w);
+              vo_font->pic_a[font]->w, 1);
         x+=vo_font->width[c]+vo_font->charspace;
 
    	c=OSD_PB_0;
@@ -251,7 +285,7 @@
 	   sa=vo_font->pic_a[font]->bmp+vo_font->start[c];
 	   st=vo_font->pic_a[font]->w;
 	   if ((i=mark)) do {
-	       draw_alpha_buf(obj,x,y,w,h,s,sa,st);
+	       draw_alpha_buf(obj,x,y,w,h,s,sa,st,1);
 	       x+=charw;
 	   } while(--i);
 	}
@@ -264,7 +298,7 @@
 	   sa=vo_font->pic_a[font]->bmp+vo_font->start[c];
 	   st=vo_font->pic_a[font]->w;
 	   if ((i=elems-mark)) do {
-	       draw_alpha_buf(obj,x,y,w,h,s,sa,st);
+	       draw_alpha_buf(obj,x,y,w,h,s,sa,st,1);
 	       x+=charw;
 	   } while(--i);
 	}
@@ -276,19 +310,21 @@
               vo_font->pic_a[font]->h,
               vo_font->pic_b[font]->bmp+vo_font->start[c],
               vo_font->pic_a[font]->bmp+vo_font->start[c],
-              vo_font->pic_a[font]->w);
+              vo_font->pic_a[font]->w, 1);
 //        x+=vo_font->width[c]+vo_font->charspace;
 
     }
 //        vo_osd_progbar_value=(vo_osd_progbar_value+1)&0xFF;
 
+	vo_convert_osd_format(obj, format);
+
 }
 
 subtitle* vo_sub=NULL;
 
 // vo_draw_text_sub(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride))
 
-inline static void vo_update_text_sub(mp_osd_obj_t* obj,int dxs,int dys){
+inline static void vo_update_text_sub(mp_osd_obj_t* obj,int dxs,int dys,uint32_t format){
    unsigned char *t;
    int c,i,j,l,x,y,font,prevc;
    int len;
@@ -421,17 +457,18 @@
 			     vo_font->pic_a[font]->h+y<obj->dys ? vo_font->pic_a[font]->h : obj->dys-y,
 			     vo_font->pic_b[font]->bmp+vo_font->start[c],
 			     vo_font->pic_a[font]->bmp+vo_font->start[c],
-			     vo_font->pic_a[font]->w);
+			     vo_font->pic_a[font]->w, 1);
 	       x+=vo_font->width[c]+vo_font->charspace;
                prevc = c;
 	 }
          if (!--l) break;
          y+=vo_font->height;
     }
-    
+
+    vo_convert_osd_format(obj, format);
 }
 
-inline static void vo_update_spudec_sub(mp_osd_obj_t* obj, int dxs, int dys)
+inline static void vo_update_spudec_sub(mp_osd_obj_t* obj, int dxs, int dys, uint32_t format)
 {
   unsigned int bbox[4];
   int i;
@@ -443,7 +480,7 @@
   obj->flags |= OSDFLAG_BBOX;
 }
 
-inline static void vo_draw_spudec_sub(mp_osd_obj_t* obj, void (*draw_alpha)(int x0, int y0, int w, int h, unsigned char* src, unsigned char* srca, int stride))
+inline static void vo_draw_spudec_sub(mp_osd_obj_t* obj, void (*draw_alpha)(int x0, int y0, int w, int h, unsigned char* src, unsigned char* srca, unsigned char* srcp, int stride))
 {
   spudec_draw_scaled(vo_spudec, obj->dxs, obj->dys, draw_alpha);
 }
@@ -464,7 +501,8 @@
     vo_osd_list=osd;
     osd->type=type;
     osd->alpha_buffer = NULL;
-    osd->bitmap_buffer = NULL;
+    osd->rgb_buffer = NULL;
+    osd->dest_buffer = NULL;
     osd->allocated = -1;
     return osd;
 }
@@ -474,14 +512,51 @@
     while(obj){
 	mp_osd_obj_t* next=obj->next;
 	if (obj->alpha_buffer) free(obj->alpha_buffer);
-	if (obj->bitmap_buffer) free(obj->bitmap_buffer);
+	if (obj->rgb_buffer) free(obj->rgb_buffer);
+	if (obj->dest_buffer) free(obj->dest_buffer);
 	free(obj);
 	obj=next;
     }
     vo_osd_list=NULL;
 }
 
-int vo_update_osd(int dxs,int dys){
+int vo_convert_osd_format(mp_osd_obj_t* obj, uint32_t format) 
+{
+    switch (format) {
+    case IMGFMT_YV12:
+    case IMGFMT_I420:
+    case IMGFMT_IYUV:
+//	if (obj->dest_buffer)
+	fprintf(stderr, "rgb2yuv: %d %d\n", obj->stride, obj->bbox.y2-obj->bbox.y1);
+	if (obj->stride >= 8)
+	    rgb24toyv12(obj->rgb_buffer, obj->dest_buffer,
+			 obj->dest_buffer+obj->stride*(obj->bbox.y2-obj->bbox.y1),
+			 obj->dest_buffer+obj->stride*(obj->bbox.y2-obj->bbox.y1)*5/4,
+			 obj->stride, obj->bbox.y2-obj->bbox.y1,
+			 obj->stride, obj->stride >> 1, obj->stride*3);
+
+//	    memset(obj->dest_buffer, 0, obj->stride*(obj->bbox.y2-obj->bbox.y1)*3/2);
+	    
+	break;
+    case IMGFMT_YUY2:
+    case IMGFMT_YVYU:
+    case IMGFMT_UYVY:
+	break;
+    case IMGFMT_RGB15:
+    case IMGFMT_BGR15:
+    case IMGFMT_RGB16:
+    case IMGFMT_BGR16:
+    case IMGFMT_RGB24:
+    case IMGFMT_BGR24:
+    case IMGFMT_RGB32:
+    case IMGFMT_BGR32:
+	fprintf(stderr, "rgbcopy: %d %d\n", obj->stride, obj->bbox.y2-obj->bbox.y1);
+	memcpy(obj->dest_buffer, obj->rgb_buffer, obj->stride*(obj->bbox.y2-obj->bbox.y1)*3);
+	break;
+    }
+}
+
+int vo_update_osd(int dxs, int dys, uint32_t format){
     mp_osd_obj_t* obj=vo_osd_list;
     int chg=0;
 
@@ -499,14 +574,14 @@
 	obj->flags&=~OSDFLAG_BBOX;
 	switch(obj->type){
 	case OSDTYPE_SUBTITLE:
-	    vo_update_text_sub(obj,dxs,dys);
+	    vo_update_text_sub(obj,dxs,dys,format);
 	    break;
 	case OSDTYPE_PROGBAR:
-	    vo_update_text_progbar(obj,dxs,dys);
+	    vo_update_text_progbar(obj,dxs,dys,format);
 	    break;
 	case OSDTYPE_SPU:
 	    if(sub_visibility && vo_spudec && spudec_visible(vo_spudec)){
-	        vo_update_spudec_sub(obj, dxs, dys);
+	        vo_update_spudec_sub(obj, dxs, dys,format);
 		obj->flags|=OSDFLAG_VISIBLE|OSDFLAG_CHANGED;
 	    }
 	    else
@@ -514,7 +589,7 @@
 	    break;
 	case OSDTYPE_OSD:
 	    if(vo_font && vo_osd_text && vo_osd_text[0]){
-		vo_update_text_osd(obj,dxs,dys); // update bbox
+		vo_update_text_osd(obj,dxs,dys,format); // update bbox
 		obj->flags|=OSDFLAG_VISIBLE|OSDFLAG_CHANGED;
 	    } else
 		obj->flags&=~OSDFLAG_VISIBLE;
@@ -572,9 +647,9 @@
 
 int vo_osd_changed_flag=0;
 
-void vo_remove_text(int dxs,int dys,void (*remove)(int x0,int y0, int w,int h)){
+void vo_remove_text(int dxs,int dys,uint32_t format,void (*remove)(int x0,int y0, int w,int h)){
     mp_osd_obj_t* obj=vo_osd_list;
-    vo_update_osd(dxs,dys);
+    vo_update_osd(dxs,dys,format);
     while(obj){
       if(((obj->flags&OSDFLAG_CHANGED) || (obj->flags&OSDFLAG_VISIBLE)) && 
          (obj->flags&OSDFLAG_OLD_BBOX)){
@@ -590,9 +665,9 @@
     }
 }
 
-void vo_draw_text(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)){
+void vo_draw_text(int dxs,int dys,uint32_t format,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)){
     mp_osd_obj_t* obj=vo_osd_list;
-    vo_update_osd(dxs,dys);
+    vo_update_osd(dxs,dys,format);
     while(obj){
       if(obj->flags&OSDFLAG_VISIBLE){
 	vo_osd_changed_flag=obj->flags&OSDFLAG_CHANGED;	// temp hack
diff -ur --exclude-from dontdiff vanilla/main/libvo/sub.h main/libvo/sub.h
--- vanilla/main/libvo/sub.h	2002-10-06 23:04:28.000000000 +0200
+++ main/libvo/sub.h	2002-11-05 18:53:42.000000000 +0100
@@ -1,7 +1,8 @@
-
 #ifndef __MPLAYER_SUB_H
 #define __MPLAYER_SUB_H
 
+#include <inttypes.h>
+
 typedef struct mp_osd_bbox_s {
     int x1,y1,x2,y2;
 } mp_osd_bbox_t;
@@ -44,7 +45,8 @@
 
     int allocated;
     unsigned char *alpha_buffer;
-    unsigned char *bitmap_buffer;
+    unsigned char *rgb_buffer;
+    unsigned char *dest_buffer;
 } mp_osd_obj_t;
 
 
@@ -102,11 +104,11 @@
 //extern void vo_draw_text_osd(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
 //extern void vo_draw_text_progbar(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
 //extern void vo_draw_text_sub(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
-extern void vo_draw_text(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
-extern void vo_remove_text(int dxs,int dys,void (*remove)(int x0,int y0, int w,int h));
+extern void vo_draw_text(int dxs,int dys,uint32_t format,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
+extern void vo_remove_text(int dxs,int dys,uint32_t format,void (*remove)(int x0,int y0, int w,int h));
 
 void vo_init_osd();
-int vo_update_osd(int dxs,int dys);
+int vo_update_osd(int dxs,int dys,uint32_t format);
 int vo_osd_changed(int new_value);
 int vo_osd_check_range_update(int,int,int,int);
 
diff -ur --exclude-from dontdiff vanilla/main/libvo/vesa_lvo.c main/libvo/vesa_lvo.c
--- vanilla/main/libvo/vesa_lvo.c	2002-07-26 18:20:58.000000000 +0200
+++ main/libvo/vesa_lvo.c	2002-11-05 18:53:42.000000000 +0100
@@ -261,12 +261,18 @@
     case IMGFMT_YV12:
     case IMGFMT_IYUV:
     case IMGFMT_I420:
-        vo_draw_alpha_yv12(w,h,src,srca,stride,lvo_mem+bespitch*y0+x0,bespitch);
+	x0 &= ~1;
+	y0 &= ~1;
+        vo_draw_alpha_yv12(w,h,src,srca,stride,lvo_mem+bespitch*y0+x0,bespitch,
+			   lvo_mem+bespitch*y0/4+x0/2+mga_vid_config.src_width*mga_vid_config.src_height,bespitch,
+			   lvo_mem+bespitch*y0/4+x0/2+mga_vid_config.src_width*mga_vid_config.src_height*5/4,bespitch);
         break;
     case IMGFMT_YUY2:
+	x0 &= ~1;
         vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+2*(bespitch*y0+x0),bespitch);
         break;
     case IMGFMT_UYVY:
+	x0 &= ~1;
         vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+2*(bespitch*y0+x0)+1,bespitch);
         break;
     default:
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_aa.c main/libvo/vo_aa.c
--- vanilla/main/libvo/vo_aa.c	2002-10-06 23:04:28.000000000 +0200
+++ main/libvo/vo_aa.c	2002-11-05 18:53:42.000000000 +0100
@@ -569,8 +569,8 @@
     vo_osd_text=NULL;
     vo_osd_progbar_type_save=vo_osd_progbar_type;
     vo_osd_progbar_type=-1;
-    vo_remove_text(aa_scrwidth(c), aa_scrheight(c),clear_alpha);
-    vo_draw_text(aa_scrwidth(c), aa_scrheight(c), draw_alpha);
+    vo_remove_text(aa_scrwidth(c), aa_scrheight(c),0,clear_alpha);
+    vo_draw_text(aa_scrwidth(c), aa_scrheight(c),0,draw_alpha);
     vo_osd_text=vo_osd_text_save;
     vo_osd_progbar_type=vo_osd_progbar_type_save;
 #endif
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_dga.c main/libvo/vo_dga.c
--- vanilla/main/libvo/vo_dga.c	2002-11-03 17:50:25.000000000 +0100
+++ main/libvo/vo_dga.c	2002-11-05 18:53:42.000000000 +0100
@@ -408,7 +408,7 @@
 #include "sub.h"
 
 static void draw_osd(void)
-{ vo_draw_text(vo_dga_src_width,vo_dga_src_height,draw_alpha); }
+{ vo_draw_text(vo_dga_src_width,vo_dga_src_height,vo_dga_src_format,draw_alpha); }
 
 static void switch_video_buffers(void)
 {
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_fbdev.c main/libvo/vo_fbdev.c
--- vanilla/main/libvo/vo_fbdev.c	2002-10-30 21:39:27.000000000 +0100
+++ main/libvo/vo_fbdev.c	2002-11-05 18:53:42.000000000 +0100
@@ -1239,7 +1239,7 @@
 
 static void draw_osd(void)
 {
-	vo_draw_text(in_width, in_height, draw_alpha);
+	vo_draw_text(in_width, in_height, pixel_format, draw_alpha);
 }
 
 static void flip_page(void)
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_gif89a.c main/libvo/vo_gif89a.c
--- vanilla/main/libvo/vo_gif89a.c	2002-09-22 20:21:45.000000000 +0200
+++ main/libvo/vo_gif89a.c	2002-11-05 18:53:42.000000000 +0100
@@ -170,7 +170,7 @@
 static void draw_osd(void)
 {
 #ifdef USE_OSD
-  vo_draw_text(image_width, image_height, draw_alpha);
+  vo_draw_text(image_width, image_height, IMGFMT_BGR24, draw_alpha);
 #endif
 }
 
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_gl2.c main/libvo/vo_gl2.c
--- vanilla/main/libvo/vo_gl2.c	2002-10-29 23:20:10.000000000 +0100
+++ main/libvo/vo_gl2.c	2002-11-05 18:53:42.000000000 +0100
@@ -1040,7 +1040,7 @@
 }
 
 static void draw_osd(void)
-{ vo_draw_text(image_width,image_height,draw_alpha_fnc); }
+{ vo_draw_text(image_width,image_height,image_format,draw_alpha_fnc); }
 
 static void
 flip_page(void)
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_jpeg.c main/libvo/vo_jpeg.c
--- vanilla/main/libvo/vo_jpeg.c	2002-09-22 20:21:45.000000000 +0200
+++ main/libvo/vo_jpeg.c	2002-11-05 18:53:42.000000000 +0100
@@ -175,7 +175,7 @@
 
 static void draw_osd(void)
 {
- vo_draw_text(image_width, image_height, draw_alpha);
+ vo_draw_text(image_width, image_height, IMGFMT_BGR24, draw_alpha);
 }
 
 static void flip_page (void)
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_sdl.c main/libvo/vo_sdl.c
--- vanilla/main/libvo/vo_sdl.c	2002-11-03 12:16:12.000000000 +0100
+++ main/libvo/vo_sdl.c	2002-11-05 18:53:42.000000000 +0100
@@ -348,14 +348,21 @@
 		case IMGFMT_YV12:  
 		case IMGFMT_I420:
         	case IMGFMT_IYUV:
-            vo_draw_alpha_yv12(w,h,src,srca,stride,((uint8_t *) *(priv->overlay->pixels))+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0]);
+		    x0 &= ~1;
+		    y0 &= ~1;
+		    vo_draw_alpha_yv12(w,h,src,srca,stride,
+				       priv->overlay->pixels[0]+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0],
+				       priv->overlay->pixels[1]+priv->overlay->pitches[1]*(y0>>1)+x0/2,priv->overlay->pitches[1],
+				       priv->overlay->pixels[2]+priv->overlay->pitches[2]*(y0>>1)+x0/2,priv->overlay->pitches[2]);
 		break;
 		case IMGFMT_YUY2:
         	case IMGFMT_YVYU:
+		    x0 &= ~1;
                 x0 *= 2;
     			vo_draw_alpha_yuy2(w,h,src,srca,stride,((uint8_t *) *(priv->overlay->pixels))+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0]);
 		break;	
         	case IMGFMT_UYVY:
+		    x0 &= ~1;
                 x0 *= 2;
     			vo_draw_alpha_yuy2(w,h,src,srca,stride,((uint8_t *) *(priv->overlay->pixels))+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0]);
 		break;
@@ -1455,12 +1462,12 @@
  
 	/* update osd/subtitles */
     if(priv->mode == YUV)
-        vo_draw_text(priv->overlay->w, priv->overlay->h, draw_alpha);
+        vo_draw_text(priv->overlay->w, priv->overlay->h, priv->format, draw_alpha);
     else {
         if(priv->dblit)
-            vo_draw_text(priv->surface->w, priv->surface->h, draw_alpha);
+            vo_draw_text(priv->surface->w, priv->surface->h, priv->format, draw_alpha);
         else
-            vo_draw_text(priv->rgbsurface->w, priv->rgbsurface->h, draw_alpha);
+            vo_draw_text(priv->rgbsurface->w, priv->rgbsurface->h, priv->format, draw_alpha);
     }
 }
 
diff -ur --exclude-from dontdiff vanilla/main/libvo/vosub_vidix.c main/libvo/vosub_vidix.c
--- vanilla/main/libvo/vosub_vidix.c	2002-10-10 10:22:23.000000000 +0200
+++ main/libvo/vosub_vidix.c	2002-11-05 18:53:42.000000000 +0100
@@ -303,14 +303,20 @@
     case IMGFMT_IF09:
     case IMGFMT_Y8:
     case IMGFMT_Y800:
+	x0 &= ~1;
+	y0 &= ~1;
 	bespitch = (vidix_play.src.w + apitch) & (~apitch);
-        vo_draw_alpha_yv12(w,h,src,srca,stride,lvo_mem+bespitch*y0+x0,bespitch);
+        vo_draw_alpha_yv12(w,h,src,srca,stride,lvo_mem+bespitch*y0+x0,bespitch,
+			   lvo_mem+bespitch*(y0/2)+x0/2+vidix_play.offset.v,bespitch/2,
+			   lvo_mem+bespitch*(y0/2)+x0/2+vidix_play.offset.u,bespitch/2);
         break;
     case IMGFMT_YUY2:
+	x0 &= ~1;
 	bespitch = (vidix_play.src.w*2 + apitch) & (~apitch);
         vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+bespitch*y0+2*x0,bespitch);
         break;
     case IMGFMT_UYVY:
+	x0 &= ~1;
 	bespitch = (vidix_play.src.w*2 + apitch) & (~apitch);
         vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+bespitch*y0+2*x0+1,bespitch);
         break;
@@ -343,7 +349,7 @@
 {
   if(verbose > 1) printf("vosub_vidix: vidix_draw_osd() was called\n");
   /* TODO: hw support */
-  vo_draw_text(vidix_play.src.w,vidix_play.src.h,draw_alpha);
+  vo_draw_text(vidix_play.src.w,vidix_play.src.h,vidix_play.fourcc,draw_alpha);
 }
 
 uint32_t vidix_query_fourcc(uint32_t format)
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_vesa.c main/libvo/vo_vesa.c
--- vanilla/main/libvo/vo_vesa.c	2002-10-10 10:22:23.000000000 +0200
+++ main/libvo/vo_vesa.c	2002-11-05 18:53:42.000000000 +0100
@@ -358,7 +358,7 @@
    w = dstW;
    h = dstH;
 #endif
-   if(dga_buffer) vo_draw_text(w,h,draw_alpha_fnc);
+   if(dga_buffer) vo_draw_text(w,h,dstFourcc,draw_alpha_fnc);
  }
 }
 
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_x11.c main/libvo/vo_x11.c
--- vanilla/main/libvo/vo_x11.c	2002-10-29 23:20:10.000000000 +0100
+++ main/libvo/vo_x11.c	2002-11-05 18:53:42.000000000 +0100
@@ -456,7 +456,7 @@
 }
 
 static void draw_osd(void)
-{ vo_draw_text(image_width,image_height,draw_alpha_fnc); }
+{ vo_draw_text(image_width,image_height,out_format,draw_alpha_fnc); }
 
 static void flip_page( void ){
     Display_Image( myximage,ImageData );
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_xv.c main/libvo/vo_xv.c
--- vanilla/main/libvo/vo_xv.c	2002-10-23 19:59:22.000000000 +0200
+++ main/libvo/vo_xv.c	2002-11-05 18:53:42.000000000 +0100
@@ -90,13 +90,24 @@
 
 static void draw_alpha_yv12(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
    x0+=(vo_panscan_x>>2);
+
+   x0 &= ~1;
+   y0 &= ~1;
+
    vo_draw_alpha_yv12(w,h,src,srca,stride,
-       xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
-       xvimage[current_buf]->pitches[0]*y0+x0,xvimage[current_buf]->pitches[0]);
+		      xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
+		      xvimage[current_buf]->pitches[0]*y0+x0,xvimage[current_buf]->pitches[0],
+		      xvimage[current_buf]->data+xvimage[current_buf]->offsets[1]+
+		      xvimage[current_buf]->pitches[1]*(y0>>1)+(x0>>1),xvimage[current_buf]->pitches[1],
+		      xvimage[current_buf]->data+xvimage[current_buf]->offsets[2]+
+		      xvimage[current_buf]->pitches[2]*(y0>>1)+(x0>>1),xvimage[current_buf]->pitches[2]
+       );
+
 }
 
 static void draw_alpha_yuy2(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
    x0+=(vo_panscan_x>>2);
+   x0 &= ~1;
    vo_draw_alpha_yuy2(w,h,src,srca,stride,
        xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
        xvimage[current_buf]->pitches[0]*y0+2*x0,xvimage[current_buf]->pitches[0]);
@@ -104,6 +115,7 @@
 
 static void draw_alpha_uyvy(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
    x0+=(vo_panscan_x>>2);
+   x0 &= ~1;
    vo_draw_alpha_yuy2(w,h,src,srca,stride,
        xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
        xvimage[current_buf]->pitches[0]*y0+2*x0+1,xvimage[current_buf]->pitches[0]);
@@ -552,7 +564,7 @@
 }
 
 static void draw_osd(void)
-{ vo_draw_text(image_width-(vo_panscan_x>>1),image_height,draw_alpha_fnc);}
+{ vo_draw_text(image_width-(vo_panscan_x>>1),image_height,image_format,draw_alpha_fnc);}
 
 static void flip_page(void)
 {
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_yuv4mpeg.c main/libvo/vo_yuv4mpeg.c
--- vanilla/main/libvo/vo_yuv4mpeg.c	2002-10-10 10:22:23.000000000 +0200
+++ main/libvo/vo_yuv4mpeg.c	2002-11-05 18:53:42.000000000 +0100
@@ -163,7 +163,10 @@
 	{
     	case IMGFMT_YV12:
 	    	vo_draw_alpha_yv12(w, h, src, srca, stride, 
-				       image + y0 * image_width + x0, image_width);
+				   image + (y0&~1) * image_width + (x0&~1), image_width,
+				   image + y0 * image_width/4 + x0/2 + image_width*image_height, image_width/2,
+				   image + y0 * image_width/4 + x0/2 + image_width*image_height*5/4, image_width/2
+				   );
 			break;
 		
 		case IMGFMT_BGR|24:
@@ -186,7 +189,7 @@
 
 static void draw_osd(void)
 {
-    vo_draw_text(image_width, image_height, draw_alpha);
+    vo_draw_text(image_width, image_height, using_format, draw_alpha);
 }
 
 static void deinterleave_fields(uint8_t *ptr, const int stride,
Only in main: mpconf
Only in main: Solyaris.CD1of2.srt
diff -ur --exclude-from dontdiff vanilla/main/spudec.c main/spudec.c
--- vanilla/main/spudec.c	2002-10-15 16:42:45.000000000 +0200
+++ main/spudec.c	2002-11-05 18:53:42.000000000 +0100
@@ -508,13 +508,13 @@
     return ret;
 }
 
-void spudec_draw(void *this, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride))
+void spudec_draw(void *this, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride))
 {
     spudec_handle_t *spu = (spudec_handle_t *)this;
     if (spu->start_pts <= spu->now_pts && spu->now_pts < spu->end_pts && spu->image)
     {
 	draw_alpha(spu->start_col, spu->start_row, spu->width, spu->height,
-		   spu->image, spu->aimage, spu->stride);
+		   spu->image, spu->aimage, NULL, spu->stride); /* FIXME */
 	spu->spu_changed = 0;
     }
 }
@@ -607,7 +607,7 @@
 }
 #endif
 
-void spudec_draw_scaled(void *me, unsigned int dxs, unsigned int dys, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride))
+void spudec_draw_scaled(void *me, unsigned int dxs, unsigned int dys, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride))
 {
   spudec_handle_t *spu = (spudec_handle_t *)me;
   scale_pixel *table_x;
@@ -618,7 +618,7 @@
       if (spu->image)
       {
 	draw_alpha(spu->start_col, spu->start_row, spu->width, spu->height,
-		   spu->image, spu->aimage, spu->stride);
+		   spu->image, spu->aimage, NULL, spu->stride); /* FIXME */
 	spu->spu_changed = 0;
       }
     }
@@ -902,7 +902,7 @@
 	spu->scaled_start_row = dys - spu->scaled_height - 1;
 #endif
 	draw_alpha(spu->scaled_start_col, spu->scaled_start_row, spu->scaled_width, spu->scaled_height,
-		   spu->scaled_image, spu->scaled_aimage, spu->scaled_stride);
+		   spu->scaled_image, spu->scaled_aimage, NULL, spu->scaled_stride); /* FIXME */
 	spu->spu_changed = 0;
       }
     }
diff -ur --exclude-from dontdiff vanilla/main/spudec.h main/spudec.h
--- vanilla/main/spudec.h	2002-08-15 18:52:52.000000000 +0200
+++ main/spudec.h	2002-11-05 18:53:42.000000000 +0100
@@ -3,8 +3,8 @@
 
 void spudec_heartbeat(void *this, unsigned int pts100);
 void spudec_assemble(void *this, unsigned char *packet, unsigned int len, unsigned int pts100);
-void spudec_draw(void *this, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
-void spudec_draw_scaled(void *this, unsigned int dxs, unsigned int dys, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
+void spudec_draw(void *this, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride));
+void spudec_draw_scaled(void *this, unsigned int dxs, unsigned int dys, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride));
 void spudec_update_palette(void *this, unsigned int *palette);
 void *spudec_new_scaled(unsigned int *palette, unsigned int frame_width, unsigned int frame_height);
 void *spudec_new_scaled_vobsub(unsigned int *palette, unsigned int *cuspal, unsigned int custom, unsigned int frame_width, unsigned int frame_height);
@@ -16,6 +16,5 @@
 void spudec_set_hw_spu(void *this, vo_functions_t *hw_spu);
 int spudec_changed(void *this);
 void spudec_calc_bbox(void *me, unsigned int dxs, unsigned int dys, unsigned int* bbox);
-void spudec_draw_scaled(void *me, unsigned int dxs, unsigned int dys, void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride));
 #endif