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	2002-11-02 16:23:20.000000000 +0100
@@ -66,8 +66,8 @@
     // TODO  clear left and right side of the image if needed
 }
 
-static void draw_func(int x0,int y0, int w,int h,unsigned char* src, unsigned char *srca, int stride){
-    unsigned char* dst;
+static void draw_func(int x0,int y0, int w,int h,unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+    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!
@@ -94,19 +94,19 @@
     switch(vf->priv->dmpi->imgfmt){
     case IMGFMT_BGR15:
     case IMGFMT_RGB15:
-	vo_draw_alpha_rgb15(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,dst,vf->priv->dmpi->stride[0]);
 	break;
     case IMGFMT_BGR16:
     case IMGFMT_RGB16:
-	vo_draw_alpha_rgb16(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,dst,vf->priv->dmpi->stride[0]);
 	break;
     case IMGFMT_BGR24:
     case IMGFMT_RGB24:
-	vo_draw_alpha_rgb24(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,dst,vf->priv->dmpi->stride[0]);
 	break;
     case IMGFMT_BGR32:
     case IMGFMT_RGB32:
-	vo_draw_alpha_rgb32(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,dst,vf->priv->dmpi->stride[0]);
 	break;
     case IMGFMT_YV12:
     case IMGFMT_I420:
@@ -115,13 +115,22 @@
     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,srcp,stride,dst,vf->priv->dmpi->stride[0],
+			   dst2,vf->priv->dmpi->stride[1],
+			   dst3,vf->priv->dmpi->stride[2],
+			   x0 & 0x01, y0 & 0x01);
 	break;
     case IMGFMT_YUY2:
-	vo_draw_alpha_yuy2(w,h,src,srca,stride,dst,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,dst,vf->priv->dmpi->stride[0],x0 & 0x01);
 	break;
     case IMGFMT_UYVY:
-	vo_draw_alpha_yuy2(w,h,src,srca,stride,dst+1,vf->priv->dmpi->stride[0]);
+	vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,dst+1,vf->priv->dmpi->stride[0],x0 & 0x01);
 	break;
     }
 }
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	2002-11-02 17:02:19.000000000 +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 +13,26 @@
 
 extern int verbose; // defined in mplayer.c
 
+#define PALSIZE 9
+/* R,G,B */
+int palette[PALSIZE][3] = {
+    {0, 0, 0}, /* dummy entry */
+    {0, 0, 0},
+    {0, 0, 255},
+    {0, 255, 0},
+    {0, 255, 255},
+    {255, 0, 0},
+    {255, 0, 255},
+    {255, 255, 0},
+    {255, 255, 255},
+};
+
+static unsigned int palette_rgb24[256];
+static unsigned int palette_yuy2[3][256];
+static unsigned int palette_yv12[3][256];
+
+#undef ARCH_X86
+
 #ifdef ARCH_X86
 #define CAN_COMPILE_X86_ASM
 #endif
@@ -52,14 +69,14 @@
 #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
 
@@ -109,139 +126,157 @@
 
 #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, unsigned char *srcp, int srcstride, unsigned char* dstbase1, int dststride1, unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3, int xmod, int ymod){
 #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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 	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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 	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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 	else
-		vo_draw_alpha_yv12_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_X86(w, h, src, srca, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #else
-		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_C(w, h, src, srca, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #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, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #else
-		vo_draw_alpha_yv12_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yv12_C(w, h, src, srca, srcp, srcstride, dstbase1, dststride1, dstbase2, dststride2, dstbase3, dststride3, xmod, ymod);
 #endif
 #endif //!RUNTIME_CPUDETECT
 }
 
-void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride,int xmod){
 #ifdef RUNTIME_CPUDETECT
 #ifdef CAN_COMPILE_X86_ASM
 	// ordered per speed fasterst first
 	if(gCpuCaps.hasMMX2)
-		vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 	else if(gCpuCaps.has3DNow)
-		vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 	else if(gCpuCaps.hasMMX)
-		vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 	else
-		vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #else
-		vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_C(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #endif
 #else //RUNTIME_CPUDETECT
 #ifdef HAVE_MMX2
-		vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #elif defined (HAVE_3DNOW)
-		vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #elif defined (HAVE_MMX)
-		vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #elif defined (ARCH_X86)
-		vo_draw_alpha_yuy2_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #else
-		vo_draw_alpha_yuy2_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_yuy2_C(w, h, src, srca, srcp, srcstride, dstbase, dststride, xmod);
 #endif
 #endif //!RUNTIME_CPUDETECT
 }
 
-void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
 #ifdef RUNTIME_CPUDETECT
 #ifdef CAN_COMPILE_X86_ASM
 	// ordered per speed fasterst first
 	if(gCpuCaps.hasMMX2)
-		vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else if(gCpuCaps.has3DNow)
-		vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else if(gCpuCaps.hasMMX)
-		vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else
-		vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #else
-		vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_C(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #endif
 #else //RUNTIME_CPUDETECT
 #ifdef HAVE_MMX2
-		vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (HAVE_3DNOW)
-		vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (HAVE_MMX)
-		vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (ARCH_X86)
-		vo_draw_alpha_rgb24_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #else
-		vo_draw_alpha_rgb24_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb24_C(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #endif
 #endif //!RUNTIME_CPUDETECT
 }
 
-void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
 #ifdef RUNTIME_CPUDETECT
 #ifdef CAN_COMPILE_X86_ASM
 	// ordered per speed fasterst first
 	if(gCpuCaps.hasMMX2)
-		vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else if(gCpuCaps.has3DNow)
-		vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else if(gCpuCaps.hasMMX)
-		vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 	else
-		vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #else
-		vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_C(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #endif
 #else //RUNTIME_CPUDETECT
 #ifdef HAVE_MMX2
-		vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_MMX2(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (HAVE_3DNOW)
-		vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_3DNow(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (HAVE_MMX)
-		vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_MMX(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #elif defined (ARCH_X86)
-		vo_draw_alpha_rgb32_X86(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_X86(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #else
-		vo_draw_alpha_rgb32_C(w, h, src, srca, srcstride, dstbase, dststride);
+		vo_draw_alpha_rgb32_C(w, h, src, srca, srcp, srcstride, dstbase, dststride);
 #endif
 #endif //!RUNTIME_CPUDETECT
 }
 
-#ifdef FAST_OSD_TABLE
-static unsigned short fast_osd_15bpp_table[256];
-static unsigned short fast_osd_16bpp_table[256];
-#endif
+#define MIN(a, b)       ((a)<(b)?(a):(b))
+#define MAX(a, b)       ((a)>(b)?(a):(b))
 
-void vo_draw_alpha_init(){
-#ifdef FAST_OSD_TABLE
+void vo_draw_alpha_generate_palettes()
+{
     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);
+    int r, g, b, y, u, v;
+    
+    for (i = 0; i < PALSIZE; i++) {
+	r = palette[i][0];
+	g = palette[i][1];
+	b = palette[i][2];
+	y = MIN(MAX((int)(0.1494 * r + 0.6061 * g + 0.2445 * b), 0), 0xff);
+	u = MIN(MAX((int)(0.6066 * r - 0.4322 * g - 0.1744 * b) + 128, 0), 0xff);
+	v = MIN(MAX((int)(-0.08435 * r - 0.3422 * g + 0.4266 * b) + 128, 0), 0xff);
+
+	palette_yv12[0][i] = y;
+	palette_yv12[1][i] = u;
+	palette_yv12[2][i] = v;
+
+//	palette_yuy2[i] = (u >> 4) << 12 | (y >> 4) << 8 | (v >> 4) << 4 | (y & 0x0f);
+	palette_yuy2[0][i] = y;
+	palette_yuy2[1][i] = u;
+	palette_yuy2[2][i] = v;
+
+	palette_rgb24[i] = b << 16 | g << 8 | r;
     }
-#endif
+}
+
+
+void vo_draw_alpha_init(){
 //FIXME the optimized stuff is a lie for 15/16bpp as they arent optimized yet
 	if(verbose)
 	{
@@ -273,31 +308,23 @@
 #endif
 #endif //!RUNTIME_CPUDETECT
 	}
+	vo_draw_alpha_generate_palettes();
 }
 
-void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
     int y;
     for(y=0;y<h;y++){
         register unsigned short *dst = (unsigned short*) dstbase;
         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[x]*(palette_rgb24[srcp[x]]>>16 & 0xff)>>8))>>3;
+                g=(((g*srca[x])>>5)+(src[x]*(palette_rgb24[srcp[x]]>>8 & 0xff)>>8))>>3;
+                b=(((b*srca[x])>>5)+(src[x]*(palette_rgb24[srcp[x]] & 0xff)>>8))>>3;
                 dst[x]=(b<<10)|(g<<5)|r;
-#endif
             }
         }
         src+=srcstride;
@@ -307,28 +334,20 @@
     return;
 }
 
-void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
     int y;
     for(y=0;y<h;y++){
         register unsigned short *dst = (unsigned short*) dstbase;
         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[x]*(palette_rgb24[srcp[x]]>>16 & 0xff)>>8))>>3;
+                g=(((g*srca[x])>>6)+(src[x]*(palette_rgb24[srcp[x]]>>8 & 0xff)>>8))>>2;
+                b=(((b*srca[x])>>5)+(src[x]*(palette_rgb24[srcp[x]] & 0xff)>>8))>>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	2002-11-02 16:20:37.000000000 +0100
@@ -7,12 +7,12 @@
 
 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_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);
-extern void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride);
-extern void vo_draw_alpha_rgb16(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, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride1,unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3, int xmod, int ymod);
+extern void vo_draw_alpha_yuy2(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride,int xmod);
+extern void vo_draw_alpha_rgb24(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride);
+extern void vo_draw_alpha_rgb32(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride);
+extern void vo_draw_alpha_rgb15(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride);
+extern void vo_draw_alpha_rgb16(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride);
 
 #endif
 
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	2002-11-02 18:03:27.000000000 +0100
@@ -27,12 +27,39 @@
 #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 calc_pal(unsigned char *srcp, unsigned char *dst, int srcstride, unsigned int *pal)
+{
+    int cnt = 0;
+    int pal_idx=0;
+    int s,d;
+    
+    
+    if (srcp[0]) { pal_idx = srcp[0]; cnt++; }
+    if (srcp[1]) { pal_idx = srcp[1]; cnt++; }
+    if (srcp[srcstride]) { pal_idx = srcp[srcstride]; cnt++; }
+    if (srcp[srcstride+1]) { pal_idx = srcp[srcstride+1]; cnt++; }
+
+    s = pal[pal_idx];
+    d = *dst;
+    
+    d -= 128;
+    s -= 128;
+    
+    d = (d*(4-cnt)+s*cnt)/4;
+    
+    *dst = d+128;
+    
+//    *dst = (*dst*(4-cnt) + pal[pal_idx]*cnt)/4;
+}
+
+static inline void RENAME(vo_draw_alpha_yv12)(int w,int h, unsigned char* src1, unsigned char *srca1, unsigned char *srcp1, int srcstride, unsigned char* dstbase1,int dststride1, unsigned char* dstbase2, int dststride2, unsigned char* dstbase3, int dststride3, int xmod, int ymod){
     int y;
-#if defined(FAST_OSD) && !defined(HAVE_MMX)
-    w=w>>1;
-#endif
+    unsigned char *src = src1;
+    unsigned char *srca = srca1;
+    unsigned char *srcp;
+    unsigned char *pal;
 PROFILE_START();
+    srcp = srcp1;
     for(y=0;y<h;y++){
         register int x;
 #ifdef HAVE_MMX
@@ -45,7 +72,7 @@
 	"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"
@@ -71,22 +98,40 @@
 		"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]*palette_yv12[0][srcp[x]])>>8);
         }
 #endif
         src+=srcstride;
         srca+=srcstride;
-        dstbase+=dststride;
+	srcp+=srcstride;
+        dstbase1+=dststride1;
+    }
+    srcp = srcp1-srcstride*ymod-xmod;
+    dstbase2+=dststride2;
+    srcp+=srcstride*2;
+    for(y=1;y<h/2-1;y++){
+        register int x;
+        for(x=1;x<w/2-1;x++){
+	    calc_pal(srcp+2*x, dstbase2+x, srcstride, palette_yv12[1]);
+	}
+        dstbase2+=dststride2;
+	srcp+=srcstride*2;
+    }
+    srcp = srcp1-srcstride*ymod-xmod;
+    dstbase3+=dststride3;
+    srcp+=srcstride*2;
+    for(y=1;y<h/2-1;y++){
+        register int x;
+        for(x=1;x<w/2-1;x++){
+	    calc_pal(srcp+2*x, dstbase3+x, srcstride, palette_yv12[2]);
+	}
+        dstbase3+=dststride3;
+	srcp+=srcstride*2;
     }
 #ifdef HAVE_MMX
 	asm volatile(EMMS:::"memory");
@@ -95,12 +140,37 @@
     return;
 }
 
-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){
+static inline void calc_pal_yuy2(unsigned char *srcp, unsigned char *dst, unsigned int *pal)
+{
+    int cnt = 0;
+    int pal_idx=0;
+
+    int d, s;
+    
+    
+    if (srcp[0]) { pal_idx = srcp[0]; cnt++; }
+    if (srcp[1]) { pal_idx = srcp[1]; cnt++; }
+
+    s = pal[pal_idx];
+    d = *dst;
+    
+    d -= 128;
+    s -= 128;
+    
+    d = (d*(2-cnt)+s*cnt)/2;
+    
+
+    *dst = (*dst*(2-cnt) + pal[pal_idx]*cnt)/2;
+    *dst = d+128;
+    
+}
+
+static inline void RENAME(vo_draw_alpha_yuy2)(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride,int xmod){
     int y;
-#if defined(FAST_OSD) && !defined(HAVE_MMX)
-    w=w>>1;
-#endif
+    int pal;
 PROFILE_START();
+
+    srcp -= xmod;
     for(y=0;y<h;y++){
         register int x;
 #ifdef HAVE_MMX
@@ -141,20 +211,21 @@
 		: "%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
+        pal = xmod;
+        for(x=1;x<w;x++){
             if(srca[x]) {
-               dstbase[2*x]=((dstbase[2*x]*srca[x])>>8)+src[x];
+		dstbase[2*x]=((dstbase[2*x]*srca[x])>>8)+((src[x]*palette_yuy2[0][srcp[x]])>>8);
+//	       if (x & 1) dstbase[2*x+1] = 0;
+//	       else dstbase[2*x+1] = 128;
                dstbase[2*x+1]=((((signed)dstbase[2*x+1]-128)*srca[x])>>8)+128;
+	       calc_pal_yuy2(srcp+(x&(~1)), dstbase+2*x+1, palette_yuy2[pal+1]);
            }
-#endif
+	    pal = pal ^ 1;
         }
 #endif
 	src+=srcstride;
         srca+=srcstride;
+        srcp+=srcstride;
         dstbase+=dststride;
     }
 #ifdef HAVE_MMX
@@ -164,7 +235,7 @@
     return;
 }
 
-static inline void RENAME(vo_draw_alpha_rgb24)(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+static inline void RENAME(vo_draw_alpha_rgb24)(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
     int y;
     for(y=0;y<h;y++){
         register unsigned char *dst = dstbase;
@@ -245,19 +316,16 @@
 #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[x]*(palette_rgb24[srcp[x]]>>16 & 0xff)>>8);
+		dst[1]=((dst[1]*srca[x])>>8)+(src[x]*(palette_rgb24[srcp[x]]>>8 & 0xff)>>8);;
+		dst[2]=((dst[2]*srca[x])>>8)+(src[x]*(palette_rgb24[srcp[x]] & 0xff)>>8);
             }
             dst+=3; // 24bpp
         }
 #endif /* arch_x86 */
         src+=srcstride;
         srca+=srcstride;
+        srcp+=srcstride;
         dstbase+=dststride;
     }
 #ifdef HAVE_MMX
@@ -266,7 +334,7 @@
     return;
 }
 
-static inline void RENAME(vo_draw_alpha_rgb32)(int w,int h, unsigned char* src, unsigned char *srca, int srcstride, unsigned char* dstbase,int dststride){
+static inline void RENAME(vo_draw_alpha_rgb32)(int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int srcstride, unsigned char* dstbase,int dststride){
     int y;
 PROFILE_START();
     for(y=0;y<h;y++){
@@ -400,18 +468,15 @@
 #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[x]*(palette_rgb24[srcp[x]]>>16 & 0xff)>>8);
+		dstbase[4*x+1]=((dstbase[4*x+1]*srca[x])>>8)+(src[x]*(palette_rgb24[srcp[x]]>>8 & 0xff)>>8);
+		dstbase[4*x+2]=((dstbase[4*x+2]*srca[x])>>8)+(src[x]*(palette_rgb24[srcp[x]] & 0xff)>>8);
             }
         }
 #endif /* arch_x86 */
         src+=srcstride;
         srca+=srcstride;
+        srcp+=srcstride;
         dstbase+=dststride;
     }
 #ifdef HAVE_MMX
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	2002-11-02 17:05:13.000000000 +0100
@@ -43,8 +43,10 @@
     return h;
 }
 
+static int xpal = 0;
+
 // 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;
@@ -52,6 +54,7 @@
     int i, j;
     unsigned char *b = obj->bitmap_buffer + (y0-obj->bbox.y1)*dststride + (x0-obj->bbox.x1);
     unsigned char *a = obj->alpha_buffer  + (y0-obj->bbox.y1)*dststride + (x0-obj->bbox.x1);
+    unsigned char *p = obj->palette_buffer  + (y0-obj->bbox.y1)*dststride + (x0-obj->bbox.x1);
     unsigned char *bs = src;
     unsigned char *as = srca;
 
@@ -63,17 +66,26 @@
     }
     
     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++, b++, a++, p++, bs++, as++) {
+	    if (*b < *bs) {
+		*b = *bs;
+		*p = xpal;
+	    }
 	    if (*as) {
-		if (*a == 0 || *a > *as) *a = *as;
+		if (*a == 0 || *a > *as) {
+		    *a = *as;
+		}
 	    }
 	}
 	b+= dstskip;
 	a+= dstskip;
+	p+= dstskip;
 	bs+= srcskip;
 	as+= srcskip;
     }
+    xpal++;
+    
+    if (xpal > 8) xpal = 0;
 }
 
 // allocates/enlarges the alpha/bitmap buffer
@@ -87,22 +99,26 @@
     if (obj->allocated<len) {
 	obj->allocated = len;
 	free(obj->bitmap_buffer);
+	free(obj->palette_buffer);
 	free(obj->alpha_buffer);
 	obj->bitmap_buffer = (unsigned char *)memalign(16, len);
 	obj->alpha_buffer = (unsigned char *)memalign(16, len);
+	obj->palette_buffer = (unsigned char *)memalign(16, len);
     }
     memset(obj->bitmap_buffer, 0, len);
     memset(obj->alpha_buffer, 0, len);
+    memset(obj->palette_buffer, 0, len);
 }
 
 // renders the buffer
-inline static void vo_draw_text_from_buffer(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_text_from_buffer(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)){
     if (obj->allocated > 0) {
 	draw_alpha(obj->bbox.x1,obj->bbox.y1,
 		   obj->bbox.x2-obj->bbox.x1,
 		   obj->bbox.y2-obj->bbox.y1,
 		   obj->bitmap_buffer,
 		   obj->alpha_buffer,
+		   obj->palette_buffer,
 		   obj->stride);
     }
 }
@@ -139,7 +155,8 @@
 			   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;
         }
 }
@@ -230,7 +247,7 @@
               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 +257,7 @@
               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 +268,7 @@
 	   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 +281,7 @@
 	   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);
 	}
@@ -276,7 +293,7 @@
               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;
 
     }
@@ -421,7 +438,7 @@
 			     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;
 	 }
@@ -443,7 +460,7 @@
   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);
 }
@@ -590,7 +607,7 @@
     }
 }
 
-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,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)){
     mp_osd_obj_t* obj=vo_osd_list;
     vo_update_osd(dxs,dys);
     while(obj){
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	2002-11-02 10:41:28.000000000 +0100
@@ -44,6 +44,7 @@
 
     int allocated;
     unsigned char *alpha_buffer;
+    unsigned char *palette_buffer;
     unsigned char *bitmap_buffer;
 } mp_osd_obj_t;
 
@@ -102,7 +103,7 @@
 //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_draw_text(int dxs,int dys,void (*draw_alpha)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride));
 extern void vo_remove_text(int dxs,int dys,void (*remove)(int x0,int y0, int w,int h));
 
 void vo_init_osd();
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	2002-11-02 16:22:22.000000000 +0100
@@ -227,7 +227,7 @@
   }	
 }
 
-static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
   UNUSED(x0);
   UNUSED(y0);
@@ -235,42 +235,46 @@
   UNUSED(h);
   UNUSED(src);
   UNUSED(srca);
+  UNUSED(srcp);
   UNUSED(stride);
 }
 
-static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
     uint32_t bespitch = /*(*/mga_vid_config.src_width;// + 15) & ~15;
     switch(mga_vid_config.format){
     case IMGFMT_BGR15:
     case IMGFMT_RGB15:
-	vo_draw_alpha_rgb15(w,h,src,srca,stride,lvo_mem+2*(y0*bespitch+x0),2*bespitch);
+	vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,lvo_mem+2*(y0*bespitch+x0),2*bespitch);
 	break;
     case IMGFMT_BGR16:
     case IMGFMT_RGB16:
-	vo_draw_alpha_rgb16(w,h,src,srca,stride,lvo_mem+2*(y0*bespitch+x0),2*bespitch);
+	vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,lvo_mem+2*(y0*bespitch+x0),2*bespitch);
 	break;
     case IMGFMT_BGR24:
     case IMGFMT_RGB24:
-	vo_draw_alpha_rgb24(w,h,src,srca,stride,lvo_mem+3*(y0*bespitch+x0),3*bespitch);
+	vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,lvo_mem+3*(y0*bespitch+x0),3*bespitch);
 	break;
     case IMGFMT_BGR32:
     case IMGFMT_RGB32:
-	vo_draw_alpha_rgb32(w,h,src,srca,stride,lvo_mem+4*(y0*bespitch+x0),4*bespitch);
+	vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,lvo_mem+4*(y0*bespitch+x0),4*bespitch);
 	break;
     case IMGFMT_YV12:
     case IMGFMT_IYUV:
     case IMGFMT_I420:
-        vo_draw_alpha_yv12(w,h,src,srca,stride,lvo_mem+bespitch*y0+x0,bespitch);
+        vo_draw_alpha_yv12(w,h,src,srca,srcp,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,
+			   x0 & 0x01, y0 & 0x01);
         break;
     case IMGFMT_YUY2:
-        vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+2*(bespitch*y0+x0),bespitch);
+        vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,lvo_mem+2*(bespitch*y0+x0),bespitch,x0 & 0x01);
         break;
     case IMGFMT_UYVY:
-        vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+2*(bespitch*y0+x0)+1,bespitch);
+        vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,lvo_mem+2*(bespitch*y0+x0)+1,bespitch,x0 & 0x01);
         break;
     default:
-        draw_alpha_null(x0,y0,w,h,src,srca,stride);
+        draw_alpha_null(x0,y0,w,h,src,srca,srcp,stride);
     }
 }
 
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_dga.c main/libvo/vo_dga.c
--- vanilla/main/libvo/vo_dga.c	2002-11-01 19:44:47.000000000 +0100
+++ main/libvo/vo_dga.c	2002-11-02 10:47:02.000000000 +0100
@@ -282,7 +282,7 @@
 
 //---------------------------------------------------------
 
-static void draw_alpha( int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride ){
+static void draw_alpha( int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride ){
 
   char *d;
   unsigned int offset;
@@ -295,16 +295,16 @@
   switch( HW_MODE.vdm_mplayer_depth ){
 
   case 32: 
-    vo_draw_alpha_rgb32(w,h,src,srca,stride, d+4*offset , 4*buffer_stride); 
+    vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride, d+4*offset , 4*buffer_stride); 
     break;
   case 24: 
-    vo_draw_alpha_rgb24(w,h,src,srca,stride, d+3*offset , 3*buffer_stride);
+    vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride, d+3*offset , 3*buffer_stride);
     break;
   case 15:
-    vo_draw_alpha_rgb15(w,h,src,srca,stride, d+2*offset , 2*buffer_stride);
+    vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride, d+2*offset , 2*buffer_stride);
     break;
   case 16:        
-    vo_draw_alpha_rgb16(w,h,src,srca,stride, d+2*offset , 2*buffer_stride);
+    vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride, d+2*offset , 2*buffer_stride);
     break;
   }
 }
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	2002-11-02 11:02:07.000000000 +0100
@@ -161,9 +161,9 @@
 }
 
 #ifdef USE_OSD
-static void draw_alpha(int x0, int y0, int w, int h, unsigned char *src, unsigned char *srca, int stride)
+static void draw_alpha(int x0, int y0, int w, int h, unsigned char *src, unsigned char *srca, unsigned char *srcp, int stride)
 {
-  vo_draw_alpha_rgb24(w, h, src, srca, stride, image_data + 3 * (y0 * image_width + x0), 3 * image_width);
+  vo_draw_alpha_rgb24(w, h, src, srca, srcp, stride, image_data + 3 * (y0 * image_width + x0), 3 * image_width);
 }
 #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	2002-11-02 11:01:05.000000000 +0100
@@ -99,7 +99,7 @@
 static int      used_s=0, used_r=0, used_b=0, used_info_done=0;
 
 static void (*draw_alpha_fnc)
-                 (int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+                 (int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride);
 
 /* The squares that are tiled to make up the game screen polygon */
 
@@ -579,23 +579,23 @@
   glLoadIdentity();
 }
 
-static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb32(w,h,src,srca,stride,ImageData+4*(y0*image_width+x0),4*image_width);
+static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,ImageData+4*(y0*image_width+x0),4*image_width);
 }
 
-static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb24(w,h,src,srca,stride,ImageData+3*(y0*image_width+x0),3*image_width);
+static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,ImageData+3*(y0*image_width+x0),3*image_width);
 }
 
-static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb16(w,h,src,srca,stride,ImageData+2*(y0*image_width+x0),2*image_width);
+static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,ImageData+2*(y0*image_width+x0),2*image_width);
 }
 
-static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb15(w,h,src,srca,stride,ImageData+2*(y0*image_width+x0),2*image_width);
+static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,ImageData+2*(y0*image_width+x0),2*image_width);
 }
 
-static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
 }
 
 static int choose_glx_visual(Display *dpy, int scr, XVisualInfo *res_vi)
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	2002-11-02 11:01:39.000000000 +0100
@@ -54,9 +54,9 @@
 static int cspace=RGB;
 static int framenum=0;
 
-static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
- vo_draw_alpha_rgb24(w, h, src, srca, stride, image_data + 3 * (y0 * image_width + x0), 3 * image_width);
+ vo_draw_alpha_rgb24(w, h, src, srca, srcp, stride, image_data + 3 * (y0 * image_width + x0), 3 * image_width);
 }
 
 static uint32_t config(uint32_t width, uint32_t height, uint32_t d_width, uint32_t d_height, uint32_t fullscreen, char *title, uint32_t format)
diff -ur --exclude-from dontdiff vanilla/main/libvo/vo_sdl.c main/libvo/vo_sdl.c
--- vanilla/main/libvo/vo_sdl.c	2002-10-10 10:22:23.000000000 +0200
+++ main/libvo/vo_sdl.c	2002-11-02 17:11:47.000000000 +0100
@@ -307,7 +307,7 @@
  *
  **/
 
-static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
 	struct sdl_priv_s *priv = &sdl_priv;
 	
     if(priv->osd_has_changed) {
@@ -348,16 +348,20 @@
 		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]);
+            vo_draw_alpha_yv12(w,h,src,srca,srcp,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],
+			       x0 & 0x01, y0 & 0x01);
 		break;
 		case IMGFMT_YUY2:
         	case IMGFMT_YVYU:
                 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]);
+    			vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,((uint8_t *) *(priv->overlay->pixels))+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0],x0 & 0x01);
 		break;	
         	case IMGFMT_UYVY:
                 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]);
+    			vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,((uint8_t *) *(priv->overlay->pixels))+priv->overlay->pitches[0]*y0+x0,priv->overlay->pitches[0],x0 & 0x01);
 		break;
 
 		default:
@@ -366,19 +370,19 @@
 		switch(priv->format) {
 		case IMGFMT_RGB15:
 		case IMGFMT_BGR15:
-    			vo_draw_alpha_rgb15(w,h,src,srca,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
+    			vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
 		break;
 		case IMGFMT_RGB16:
 		case IMGFMT_BGR16:
-    			vo_draw_alpha_rgb16(w,h,src,srca,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
+    			vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
 		break;
 		case IMGFMT_RGB24:
 		case IMGFMT_BGR24:
-    			vo_draw_alpha_rgb24(w,h,src,srca,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
+    			vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
 		break;
 		case IMGFMT_RGB32:
 		case IMGFMT_BGR32:
-    			vo_draw_alpha_rgb32(w,h,src,srca,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
+    			vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,((uint8_t *) priv->surface->pixels)+y0*priv->surface->pitch+x0,priv->surface->pitch);
 		break;
 		}
         }
@@ -387,19 +391,19 @@
 		switch(priv->format) {
 		case IMGFMT_RGB15:
 		case IMGFMT_BGR15:
-    			vo_draw_alpha_rgb15(w,h,src,srca,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
+    			vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
 		break;
 		case IMGFMT_RGB16:
 		case IMGFMT_BGR16:
-    			vo_draw_alpha_rgb16(w,h,src,srca,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
+    			vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
 		break;
 		case IMGFMT_RGB24:
 		case IMGFMT_BGR24:
-    			vo_draw_alpha_rgb24(w,h,src,srca,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
+    			vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
 		break;
 		case IMGFMT_RGB32:
 		case IMGFMT_BGR32:
-    			vo_draw_alpha_rgb32(w,h,src,srca,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
+    			vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,((uint8_t *) priv->rgbsurface->pixels)+y0*priv->rgbsurface->pitch+x0,priv->rgbsurface->pitch);
 		break;
 		}
         }
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	2002-11-02 16:22:59.000000000 +0100
@@ -289,7 +289,7 @@
   }	
 }
 
-static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
     uint32_t apitch,bespitch;
     void *lvo_mem;
@@ -304,35 +304,38 @@
     case IMGFMT_Y8:
     case IMGFMT_Y800:
 	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,srcp,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,
+			   x0&1, y0&1);
         break;
     case IMGFMT_YUY2:
 	bespitch = (vidix_play.src.w*2 + apitch) & (~apitch);
-        vo_draw_alpha_yuy2(w,h,src,srca,stride,lvo_mem+bespitch*y0+2*x0,bespitch);
+        vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,lvo_mem+bespitch*y0+2*x0,bespitch,x0 & 0x01);
         break;
     case IMGFMT_UYVY:
 	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);
+        vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,lvo_mem+bespitch*y0+2*x0+1,bespitch,x0 & 0x01);
         break;
     case IMGFMT_RGB32:
     case IMGFMT_BGR32:
 	bespitch = (vidix_play.src.w*4 + apitch) & (~apitch);
-	vo_draw_alpha_rgb32(w,h,src,srca,stride,lvo_mem+y0*bespitch+4*x0,bespitch);
+	vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,lvo_mem+y0*bespitch+4*x0,bespitch);
         break;
     case IMGFMT_RGB24:
     case IMGFMT_BGR24:
 	bespitch = (vidix_play.src.w*3 + apitch) & (~apitch);
-	vo_draw_alpha_rgb24(w,h,src,srca,stride,lvo_mem+y0*bespitch+3*x0,bespitch);
+	vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,lvo_mem+y0*bespitch+3*x0,bespitch);
         break;
     case IMGFMT_RGB16:
     case IMGFMT_BGR16:
 	bespitch = (vidix_play.src.w*2 + apitch) & (~apitch);
-	vo_draw_alpha_rgb16(w,h,src,srca,stride,lvo_mem+y0*bespitch+2*x0,bespitch);
+	vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,lvo_mem+y0*bespitch+2*x0,bespitch);
         break;
     case IMGFMT_RGB15:
     case IMGFMT_BGR15:
 	bespitch = (vidix_play.src.w*2 + apitch) & (~apitch);
-	vo_draw_alpha_rgb15(w,h,src,srca,stride,lvo_mem+y0*bespitch+2*x0,bespitch);
+	vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,lvo_mem+y0*bespitch+2*x0,bespitch);
         break;
     default:
 	return;
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	2002-11-02 11:03:12.000000000 +0100
@@ -103,7 +103,7 @@
 static unsigned video_mode; /* selected video mode for playback */
 static struct VesaModeInfoBlock video_mode_info;
 static int flip_trigger = 0;
-static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride);
 
 /* multibuffering */
 uint8_t*  video_base; /* should be never changed */
@@ -281,7 +281,7 @@
 /* Please comment it out if you want have OSD within movie */
 /*#define OSD_OUTSIDE_MOVIE 1*/
 
-static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
    unsigned int dstride=HAS_DGA()?video_mode_info.XResolution:dstW;
 #ifndef OSD_OUTSIDE_MOVIE
@@ -291,10 +291,10 @@
 	y0 += y_offset;
    }
 #endif
-   vo_draw_alpha_rgb32(w,h,src,srca,stride,dga_buffer+4*(y0*dstride+x0),4*dstride);
+   vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,dga_buffer+4*(y0*dstride+x0),4*dstride);
 }
 
-static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
    unsigned int dstride=HAS_DGA()?video_mode_info.XResolution:dstW;
 #ifndef OSD_OUTSIDE_MOVIE
@@ -304,10 +304,10 @@
 	y0 += y_offset;
    }
 #endif
-   vo_draw_alpha_rgb24(w,h,src,srca,stride,dga_buffer+3*(y0*dstride+x0),3*dstride);
+   vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,dga_buffer+3*(y0*dstride+x0),3*dstride);
 }
 
-static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
    unsigned int dstride=HAS_DGA()?video_mode_info.XResolution:dstW;
 #ifndef OSD_OUTSIDE_MOVIE
@@ -317,10 +317,10 @@
 	y0 += y_offset;
    }
 #endif
-   vo_draw_alpha_rgb16(w,h,src,srca,stride,dga_buffer+2*(y0*dstride+x0),2*dstride);
+   vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,dga_buffer+2*(y0*dstride+x0),2*dstride);
 }
 
-static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
    unsigned int dstride=HAS_DGA()?video_mode_info.XResolution:dstW;
 #ifndef OSD_OUTSIDE_MOVIE
@@ -330,10 +330,10 @@
 	y0 += y_offset;
    }
 #endif
-   vo_draw_alpha_rgb15(w,h,src,srca,stride,dga_buffer+2*(y0*dstride+x0),2*dstride);
+   vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,dga_buffer+2*(y0*dstride+x0),2*dstride);
 }
 
-static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride)
+static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride)
 {
   UNUSED(x0);
   UNUSED(y0);
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	2002-11-02 10:44:59.000000000 +0100
@@ -43,7 +43,7 @@
 
 /* private prototypes */
 static void Display_Image ( XImage * myximage,unsigned char *ImageData );
-static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride);
 
 /* local data */
 static unsigned char *ImageData;
@@ -93,23 +93,23 @@
   }
 }
 
-static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb32(w,h,src,srca,stride,ImageData+4*(y0*image_width+x0),4*image_width);
+static void draw_alpha_32(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb32(w,h,src,srca,srcp,stride,ImageData+4*(y0*image_width+x0),4*image_width);
 }
 
-static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb24(w,h,src,srca,stride,ImageData+3*(y0*image_width+x0),3*image_width);
+static void draw_alpha_24(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb24(w,h,src,srca,srcp,stride,ImageData+3*(y0*image_width+x0),3*image_width);
 }
 
-static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb16(w,h,src,srca,stride,ImageData+2*(y0*image_width+x0),2*image_width);
+static void draw_alpha_16(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb16(w,h,src,srca,srcp,stride,ImageData+2*(y0*image_width+x0),2*image_width);
 }
 
-static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
-   vo_draw_alpha_rgb15(w,h,src,srca,stride,ImageData+2*(y0*image_width+x0),2*image_width);
+static void draw_alpha_15(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
+   vo_draw_alpha_rgb15(w,h,src,srca,srcp,stride,ImageData+2*(y0*image_width+x0),2*image_width);
 }
 
-static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
 }
 
 static SwsContext *swsContext=NULL;
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	2002-11-02 16:21:52.000000000 +0100
@@ -86,30 +86,38 @@
 static uint32_t               drwX,drwY,drwBorderWidth,drwDepth;
 static uint32_t               dwidth,dheight;
 
-static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride);
+static void (*draw_alpha_fnc)(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride);
 
-static void draw_alpha_yv12(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_yv12(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
    x0+=(vo_panscan_x>>2);
-   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]);
+
+   vo_draw_alpha_yv12(w,h,src,srca,srcp,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[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],
+		      x0 & 0x01, y0 & 0x01
+       );
+
 }
 
-static void draw_alpha_yuy2(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_yuy2(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
    x0+=(vo_panscan_x>>2);
-   vo_draw_alpha_yuy2(w,h,src,srca,stride,
+   vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,
        xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
-       xvimage[current_buf]->pitches[0]*y0+2*x0,xvimage[current_buf]->pitches[0]);
+       xvimage[current_buf]->pitches[0]*y0+2*x0,xvimage[current_buf]->pitches[0],x0 & 0x01);
 }
 
-static void draw_alpha_uyvy(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_uyvy(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
    x0+=(vo_panscan_x>>2);
-   vo_draw_alpha_yuy2(w,h,src,srca,stride,
+   vo_draw_alpha_yuy2(w,h,src,srca,srcp,stride,
        xvimage[current_buf]->data+xvimage[current_buf]->offsets[0]+
-       xvimage[current_buf]->pitches[0]*y0+2*x0+1,xvimage[current_buf]->pitches[0]);
+       xvimage[current_buf]->pitches[0]*y0+2*x0+1,xvimage[current_buf]->pitches[0],x0 & 0x01);
 }
 
-static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, int stride){
+static void draw_alpha_null(int x0,int y0, int w,int h, unsigned char* src, unsigned char *srca, unsigned char *srcp, int stride){
 }
 
 static int xv_set_eq(char *name, int value)
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	2002-11-02 10:40:59.000000000 +0100
@@ -158,24 +158,28 @@
 }
 
 static void draw_alpha(int x0, int y0, int w, int h, unsigned char *src,
-                       unsigned char *srca, int stride) {
+                       unsigned char *srca, unsigned char *srcp, int stride) {
 	switch (using_format)
 	{
     	case IMGFMT_YV12:
-	    	vo_draw_alpha_yv12(w, h, src, srca, stride, 
-				       image + y0 * image_width + x0, image_width);
+	    	vo_draw_alpha_yv12(w, h, src, srca, srcp, stride, 
+				   image + y0 * image_width + x0, 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,
+				   x0 & 0x01, y0 & 0x01
+				   );
 			break;
 		
 		case IMGFMT_BGR|24:
 		case IMGFMT_RGB|24:
 			if (config_interlace != Y4M_ILACE_BOTTOM_FIRST)
-				vo_draw_alpha_rgb24(w, h, src, srca, stride,
+				vo_draw_alpha_rgb24(w, h, src, srca, srcp, stride,
 						rgb_buffer + (y0 * image_width + x0) * 3, image_width * 3);
 			else
 			{
 				swap_fields (rgb_buffer, image_height, image_width * 3);
 
-				vo_draw_alpha_rgb24(w, h, src, srca, stride,
+				vo_draw_alpha_rgb24(w, h, src, srca, srcp, stride,
 						rgb_buffer + (y0 * image_width  + x0) * 3, image_width * 3);
 				
 				swap_fields (rgb_buffer, image_height, image_width * 3);
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	2002-11-02 11:44:01.000000000 +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	2002-11-02 11:43:36.000000000 +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