]> git.sesse.net Git - vlc/blobdiff - modules/video_filter/blend.c
Add a "(" to fix previous commit.
[vlc] / modules / video_filter / blend.c
index 94a9ac3ac0d1b1cd58490a3047908f0255233d10..89d7949e476e59902596c73f051f5eb9098d4fee 100644 (file)
@@ -57,6 +57,19 @@ static void BlendR24( filter_t *, picture_t *, picture_t *, picture_t *,
 static void BlendYUVPacked( filter_t *, picture_t *, picture_t *, picture_t *,
                             int, int, int, int, int );
 
+/* I420, YV12 */
+static void BlendI420I420( filter_t *, picture_t *, picture_t *, picture_t *,
+                           int, int, int, int, int );
+static void BlendI420I420_no_alpha(
+                           filter_t *, picture_t *, picture_t *, picture_t *,
+                           int, int, int, int );
+static void BlendI420R16( filter_t *, picture_t *, picture_t *, picture_t *,
+                           int, int, int, int, int );
+static void BlendI420R24( filter_t *, picture_t *, picture_t *, picture_t *,
+                          int, int, int, int, int );
+static void BlendI420YUVPacked( filter_t *, picture_t *, picture_t *,
+                                picture_t *, int, int, int, int, int );
+
 /* YUVP */
 static void BlendPalI420( filter_t *, picture_t *, picture_t *, picture_t *,
                           int, int, int, int, int );
@@ -68,6 +81,8 @@ static void BlendPalRV( filter_t *, picture_t *, picture_t *, picture_t *,
 /* RGBA */
 static void BlendRGBAI420( filter_t *, picture_t *, picture_t *, picture_t *,
                            int, int, int, int, int );
+static void BlendRGBAYUVPacked( filter_t *, picture_t *, picture_t *,
+                                picture_t *, int, int, int, int, int );
 static void BlendRGBAR16( filter_t *, picture_t *, picture_t *, picture_t *,
                           int, int, int, int, int );
 static void BlendRGBAR24( filter_t *, picture_t *, picture_t *, picture_t *,
@@ -94,9 +109,11 @@ static int OpenFilter( vlc_object_t *p_this )
      * We could try to use a chroma filter if we can't. */
     int in_chroma = p_filter->fmt_in.video.i_chroma;
     int out_chroma = p_filter->fmt_out.video.i_chroma;
-    if(
-      ( ( in_chroma  != VLC_FOURCC('Y','U','V','A') &&
-          in_chroma  != VLC_FOURCC('Y','U','V','P') ) ||
+    if( ( in_chroma  != VLC_FOURCC('Y','U','V','A') &&
+          in_chroma  != VLC_FOURCC('I','4','2','0') &&
+          in_chroma  != VLC_FOURCC('Y','V','1','2') &&
+          in_chroma  != VLC_FOURCC('Y','U','V','P') &&
+          in_chroma  != VLC_FOURCC('R','G','B','A') ) ||
         ( out_chroma != VLC_FOURCC('I','4','2','0') &&
           out_chroma != VLC_FOURCC('Y','U','Y','2') &&
           out_chroma != VLC_FOURCC('Y','V','1','2') &&
@@ -105,14 +122,6 @@ static int OpenFilter( vlc_object_t *p_this )
           out_chroma != VLC_FOURCC('R','V','1','6') &&
           out_chroma != VLC_FOURCC('R','V','2','4') &&
           out_chroma != VLC_FOURCC('R','V','3','2') ) )
-      &&
-      ( ( in_chroma  != VLC_FOURCC('R','G','B','A') ) ||
-        ( out_chroma != VLC_FOURCC('I','4','2','0') &&
-          out_chroma != VLC_FOURCC('Y','V','1','2') &&
-          out_chroma != VLC_FOURCC('R','V','1','6') &&
-          out_chroma != VLC_FOURCC('R','V','2','4') &&
-          out_chroma != VLC_FOURCC('R','V','3','2' ) ) )
-      )
     {
         return VLC_EGENERIC;
     }
@@ -158,6 +167,8 @@ static void Blend( filter_t *p_filter, picture_t *p_dst,
 {
     int i_width, i_height;
 
+    if( i_alpha == 0 ) return;
+
     i_width = __MIN((int)p_filter->fmt_out.video.i_visible_width - i_x_offset,
                     (int)p_filter->fmt_in.video.i_visible_width);
 
@@ -226,6 +237,41 @@ static void Blend( filter_t *p_filter, picture_t *p_dst,
                                 i_width, i_height, i_alpha );
                     return;
             }
+        case VLC_FOURCC('Y','V','1','2'):
+        case VLC_FOURCC('I','4','2','0'):
+            switch( p_filter->fmt_out.video.i_chroma )
+            {
+                case VLC_FOURCC('I','4','2','0'):
+                case VLC_FOURCC('Y','V','1','2'):
+                    if( i_alpha == 0xff )
+                        BlendI420I420_no_alpha(
+                                   p_filter, p_dst, p_dst_orig, p_src,
+                                   i_x_offset, i_y_offset,
+                                   i_width, i_height );
+                    else
+                        BlendI420I420( p_filter, p_dst, p_dst_orig, p_src,
+                                       i_x_offset, i_y_offset,
+                                       i_width, i_height, i_alpha );
+                    return;
+                case VLC_FOURCC('Y','U','Y','2'):
+                case VLC_FOURCC('U','Y','V','Y'):
+                case VLC_FOURCC('Y','V','Y','U'):
+                    BlendI420YUVPacked( p_filter, p_dst, p_dst_orig, p_src,
+                                        i_x_offset, i_y_offset,
+                                        i_width, i_height, i_alpha );
+                    return;
+                case VLC_FOURCC('R','V','1','6'):
+                    BlendI420R16( p_filter, p_dst, p_dst_orig, p_src,
+                                  i_x_offset, i_y_offset,
+                                  i_width, i_height, i_alpha );
+                    return;
+                case VLC_FOURCC('R','V','2','4'):
+                case VLC_FOURCC('R','V','3','2'):
+                    BlendI420R24( p_filter, p_dst, p_dst_orig, p_src,
+                                  i_x_offset, i_y_offset,
+                                  i_width, i_height, i_alpha );
+                    return;
+            }
         case VLC_FOURCC('R','G','B','A'):
             switch( p_filter->fmt_out.video.i_chroma )
             {
@@ -235,6 +281,13 @@ static void Blend( filter_t *p_filter, picture_t *p_dst,
                                    i_x_offset, i_y_offset,
                                    i_width, i_height, i_alpha );
                     return;
+                case VLC_FOURCC('Y','U','Y','2'):
+                case VLC_FOURCC('U','Y','V','Y'):
+                case VLC_FOURCC('Y','V','Y','U'):
+                    BlendRGBAYUVPacked( p_filter, p_dst, p_dst_orig, p_src,
+                                        i_x_offset, i_y_offset,
+                                        i_width, i_height, i_alpha );
+                    return;
                 case VLC_FOURCC('R','V','2','4'):
                 case VLC_FOURCC('R','V','3','2'):
                     BlendRGBAR24( p_filter, p_dst, p_dst_orig, p_src,
@@ -302,7 +355,7 @@ static void BlendI420( filter_t *p_filter, picture_t *p_dst,
     uint8_t *p_src1_u, *p_src2_u, *p_dst_u;
     uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
     uint8_t *p_trans;
-    int i_x, i_y, i_trans;
+    int i_x, i_y, i_trans = 0;
     vlc_bool_t b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
@@ -367,7 +420,8 @@ static void BlendI420( filter_t *p_filter, picture_t *p_dst,
         /* Draw until we reach the end of the line */
         for( i_x = 0; i_x < i_width; i_x++ )
         {
-            i_trans = ( p_trans[i_x] * i_alpha ) / 255;
+            if( p_trans )
+                i_trans = ( p_trans[i_x] * i_alpha ) / 255;
             if( !i_trans )
             {
                 /* Completely transparent. Don't change pixel */
@@ -378,46 +432,554 @@ static void BlendI420( filter_t *p_filter, picture_t *p_dst,
                 /* Completely opaque. Completely overwrite underlying pixel */
                 p_dst_y[i_x] = p_src2_y[i_x];
 
-                if( b_even_scanline && i_x % 2 == 0 )
-                {
-                    p_dst_u[i_x/2] = p_src2_u[i_x];
-                    p_dst_v[i_x/2] = p_src2_v[i_x];
-                }
-                continue;
-            }
+                if( b_even_scanline && i_x % 2 == 0 )
+                {
+                    p_dst_u[i_x/2] = p_src2_u[i_x];
+                    p_dst_v[i_x/2] = p_src2_v[i_x];
+                }
+                continue;
+            }
+
+            /* Blending */
+            p_dst_y[i_x] = ( (uint16_t)p_src2_y[i_x] * i_trans +
+                (uint16_t)p_src1_y[i_x] * (MAX_TRANS - i_trans) )
+                >> TRANS_BITS;
+
+            if( b_even_scanline && i_x % 2 == 0 )
+            {
+                p_dst_u[i_x/2] = ( (uint16_t)p_src2_u[i_x] * i_trans +
+                (uint16_t)p_src1_u[i_x/2] * (MAX_TRANS - i_trans) )
+                >> TRANS_BITS;
+                p_dst_v[i_x/2] = ( (uint16_t)p_src2_v[i_x] * i_trans +
+                (uint16_t)p_src1_v[i_x/2] * (MAX_TRANS - i_trans) )
+                >> TRANS_BITS;
+            }
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}
+
+static void BlendR16( filter_t *p_filter, picture_t *p_dst_pic,
+                      picture_t *p_dst_orig, picture_t *p_src,
+                      int i_x_offset, int i_y_offset,
+                      int i_width, int i_height, int i_alpha )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch;
+    uint8_t *p_dst, *p_src1, *p_src2_y;
+    uint8_t *p_src2_u, *p_src2_v;
+    uint8_t *p_trans;
+    int i_x, i_y, i_pix_pitch, i_trans = 0;
+    int r, g, b;
+
+    i_pix_pitch = p_dst_pic->p->i_pixel_pitch;
+    i_dst_pitch = p_dst_pic->p->i_pitch;
+    p_dst = p_dst_pic->p->p_pixels + i_x_offset * i_pix_pitch +
+            p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+            p_dst_pic->p->i_pitch *
+            ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1 = p_dst_orig->p->p_pixels + i_x_offset * i_pix_pitch +
+               p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+               p_dst_orig->p->i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src2_pitch = p_src->p[Y_PLANE].i_pitch;
+    p_src2_y = p_src->p[Y_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset +
+               p_src->p[Y_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+    p_src2_u = p_src->p[U_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[U_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+    p_src2_v = p_src->p[V_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+
+    p_trans = p_src->p[A_PLANE].p_pixels +
+              p_filter->fmt_in.video.i_x_offset +
+              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+
+#define MAX_TRANS 255
+#define TRANS_BITS  8
+
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+         p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
+         p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
+         p_src2_v += i_src2_pitch )
+    {
+        /* Draw until we reach the end of the line */
+        for( i_x = 0; i_x < i_width; i_x++ )
+        {
+            if( p_trans )
+                i_trans = ( p_trans[i_x] * i_alpha ) / 255;
+            if( !i_trans )
+            {
+                /* Completely transparent. Don't change pixel */
+                continue;
+            }
+            else if( i_trans == MAX_TRANS )
+            {
+                /* Completely opaque. Completely overwrite underlying pixel */
+                yuv_to_rgb( &r, &g, &b,
+                            p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+
+    ((uint16_t *)(&p_dst[i_x * i_pix_pitch]))[0] = ((r >> 3) << 11) | ((g >> 2) << 5) | (b >> 3);
+                continue;
+            }
+
+            /* Blending */
+            /* FIXME: do the blending */
+            yuv_to_rgb( &r, &g, &b,
+                        p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+
+    ((uint16_t *)(&p_dst[i_x * i_pix_pitch]))[0] = ((r >> 3) << 11) | ((g >> 2) << 5) | (b >> 3);
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}
+
+static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
+                      picture_t *p_dst_orig, picture_t *p_src,
+                      int i_x_offset, int i_y_offset,
+                      int i_width, int i_height, int i_alpha )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch;
+    uint8_t *p_dst, *p_src1, *p_src2_y;
+    uint8_t *p_src2_u, *p_src2_v;
+    uint8_t *p_trans;
+    int i_x, i_y, i_pix_pitch, i_trans = 0;
+    int r, g, b;
+
+    i_pix_pitch = p_dst_pic->p->i_pixel_pitch;
+    i_dst_pitch = p_dst_pic->p->i_pitch;
+    p_dst = p_dst_pic->p->p_pixels + i_x_offset * i_pix_pitch +
+            p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+            p_dst_pic->p->i_pitch *
+            ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1 = p_dst_orig->p->p_pixels + i_x_offset * i_pix_pitch +
+               p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+               p_dst_orig->p->i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src2_pitch = p_src->p[Y_PLANE].i_pitch;
+    p_src2_y = p_src->p[Y_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset +
+               p_src->p[Y_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+    p_src2_u = p_src->p[U_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[U_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+    p_src2_v = p_src->p[V_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+
+    p_trans = p_src->p[A_PLANE].p_pixels +
+              p_filter->fmt_in.video.i_x_offset +
+              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+
+#define MAX_TRANS 255
+#define TRANS_BITS  8
+
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+         p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
+         p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
+         p_src2_v += i_src2_pitch )
+    {
+        /* Draw until we reach the end of the line */
+        for( i_x = 0; i_x < i_width; i_x++ )
+        {
+            if( p_trans )
+                i_trans = ( p_trans[i_x] * i_alpha ) / 255;
+            if( !i_trans )
+            {
+                /* Completely transparent. Don't change pixel */
+                continue;
+            }
+            else if( i_trans == MAX_TRANS )
+            {
+                /* Completely opaque. Completely overwrite underlying pixel */
+                yuv_to_rgb( &r, &g, &b,
+                            p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+
+                p_dst[i_x * i_pix_pitch]     = r;
+                p_dst[i_x * i_pix_pitch + 1] = g;
+                p_dst[i_x * i_pix_pitch + 2] = b;
+                continue;
+            }
+
+            /* Blending */
+            yuv_to_rgb( &r, &g, &b,
+                        p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+
+            p_dst[i_x * i_pix_pitch]     = ( r * i_trans +
+                (uint16_t)p_src1[i_x * i_pix_pitch] *
+                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
+            p_dst[i_x * i_pix_pitch + 1] = ( g * i_trans +
+                (uint16_t)p_src1[i_x * i_pix_pitch + 1] *
+                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
+            p_dst[i_x * i_pix_pitch + 2] = ( b * i_trans +
+                (uint16_t)p_src1[i_x * i_pix_pitch + 2] *
+                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}
+
+static void BlendYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
+                            picture_t *p_dst_orig, picture_t *p_src,
+                            int i_x_offset, int i_y_offset,
+                            int i_width, int i_height, int i_alpha )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch;
+    uint8_t *p_dst, *p_src1, *p_src2_y;
+    uint8_t *p_src2_u, *p_src2_v;
+    uint8_t *p_trans;
+    int i_x, i_y, i_pix_pitch, i_trans = 0;
+    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
+
+    if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','U','Y','2') )
+    {
+        i_l_offset = 0;
+        i_u_offset = 1;
+        i_v_offset = 3;
+    }
+    else if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('U','Y','V','Y') )
+    {
+        i_l_offset = 1;
+        i_u_offset = 0;
+        i_v_offset = 2;
+    }
+    else if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','V','Y','U') )
+    {
+        i_l_offset = 0;
+        i_u_offset = 3;
+        i_v_offset = 1;
+    }
+
+    i_pix_pitch = 2;
+    i_dst_pitch = p_dst_pic->p->i_pitch;
+    p_dst = p_dst_pic->p->p_pixels + i_x_offset * i_pix_pitch +
+            p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+            p_dst_pic->p->i_pitch *
+            ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1 = p_dst_orig->p->p_pixels + i_x_offset * i_pix_pitch +
+               p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+               p_dst_orig->p->i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src2_pitch = p_src->p[Y_PLANE].i_pitch;
+    p_src2_y = p_src->p[Y_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset +
+               p_src->p[Y_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+    p_src2_u = p_src->p[U_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[U_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+    p_src2_v = p_src->p[V_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+
+    p_trans = p_src->p[A_PLANE].p_pixels +
+              p_filter->fmt_in.video.i_x_offset +
+              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+
+    i_width = (i_width >> 1) << 1; /* Needs to be a multiple of 2 */
+
+#define MAX_TRANS 255
+#define TRANS_BITS  8
+
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+         p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
+         p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
+         p_src2_v += i_src2_pitch )
+    {
+        /* Draw until we reach the end of the line */
+        for( i_x = 0; i_x < i_width; i_x++, b_even = !b_even )
+        {
+            i_trans = ( p_trans[i_x] * i_alpha ) / 255;
+            if( !i_trans )
+            {
+                /* Completely transparent. Don't change pixel */
+            }
+            else if( i_trans == MAX_TRANS )
+            {
+                /* Completely opaque. Completely overwrite underlying pixel */
+                p_dst[i_x * 2 + i_l_offset]     = p_src2_y[i_x];
+
+                if( b_even )
+                {
+                    if( p_trans[i_x+1] > 0xaa )
+                    {
+                        p_dst[i_x * 2 + i_u_offset] = (p_src2_u[i_x]+p_src2_u[i_x+1])>>1;
+                        p_dst[i_x * 2 + i_v_offset] = (p_src2_v[i_x]+p_src2_v[i_x+1])>>1;
+                    }
+                    else
+                    {
+                        p_dst[i_x * 2 + i_u_offset] = p_src2_u[i_x];
+                        p_dst[i_x * 2 + i_v_offset] = p_src2_v[i_x];
+                    }
+                }
+            }
+            else
+            {
+                /* Blending */
+                p_dst[i_x * 2 + i_l_offset]     = ( (uint16_t)p_src2_y[i_x] * i_trans +
+                    (uint16_t)p_src1[i_x * 2 + i_l_offset] * (MAX_TRANS - i_trans) )
+                    >> TRANS_BITS;
+
+                if( b_even )
+                {
+                    uint16_t i_u = 0;
+                    uint16_t i_v = 0;
+                    if( p_trans[i_x+1] > 0xaa )
+                    {
+                        i_u = (p_src2_u[i_x]+p_src2_u[i_x+1])>>1;
+                        i_v = (p_src2_v[i_x]+p_src2_v[i_x+1])>>1;
+                    }
+                    else
+                    {
+                        i_u = p_src2_u[i_x];
+                        i_v = p_src2_v[i_x];
+                    }
+                    p_dst[i_x * 2 + i_u_offset] = ( (uint16_t)i_u * i_trans +
+                        (uint16_t)p_src1[i_x * 2 + i_u_offset] * (MAX_TRANS - i_trans) )
+                        >> TRANS_BITS;
+                    p_dst[i_x * 2 + i_v_offset] = ( (uint16_t)i_v * i_trans +
+                        (uint16_t)p_src1[i_x * 2 + i_v_offset] * (MAX_TRANS - i_trans) )
+                        >> TRANS_BITS;
+                }
+            }
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}
+/***********************************************************************
+ * I420, YV12
+ ***********************************************************************/
+static void BlendI420I420( filter_t *p_filter, picture_t *p_dst,
+                           picture_t *p_dst_orig, picture_t *p_src,
+                           int i_x_offset, int i_y_offset,
+                           int i_width, int i_height, int i_alpha )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch;
+    uint8_t *p_src1_y, *p_src2_y, *p_dst_y;
+    uint8_t *p_src1_u, *p_src2_u, *p_dst_u;
+    uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
+    int i_x, i_y;
+    vlc_bool_t b_even_scanline = i_y_offset % 2;
+
+    i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
+    p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
+              p_filter->fmt_out.video.i_x_offset +
+              p_dst->p[Y_PLANE].i_pitch *
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+    p_dst_u = p_dst->p[U_PLANE].p_pixels + i_x_offset/2 +
+              p_filter->fmt_out.video.i_x_offset/2 +
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+              p_dst->p[U_PLANE].i_pitch;
+    p_dst_v = p_dst->p[V_PLANE].p_pixels + i_x_offset/2 +
+              p_filter->fmt_out.video.i_x_offset/2 +
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+              p_dst->p[V_PLANE].i_pitch;
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1_y = p_dst_orig->p[Y_PLANE].p_pixels + i_x_offset +
+               p_filter->fmt_out.video.i_x_offset +
+               p_dst_orig->p[Y_PLANE].i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+    p_src1_u = p_dst_orig->p[U_PLANE].p_pixels + i_x_offset/2 +
+               p_filter->fmt_out.video.i_x_offset/2 +
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+               p_dst_orig->p[U_PLANE].i_pitch;
+    p_src1_v = p_dst_orig->p[V_PLANE].p_pixels + i_x_offset/2 +
+               p_filter->fmt_out.video.i_x_offset/2 +
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+               p_dst_orig->p[V_PLANE].i_pitch;
+
+    i_src2_pitch = p_src->p[Y_PLANE].i_pitch;
+    p_src2_y = p_src->p[Y_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset +
+               p_src->p[Y_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+    p_src2_u = p_src->p[U_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[U_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+    p_src2_v = p_src->p[V_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+
+#define MAX_TRANS 255
+#define TRANS_BITS  8
+
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++,
+         p_dst_y += i_dst_pitch, p_src1_y += i_src1_pitch,
+         p_src2_y += i_src2_pitch )
+    {
+        if( b_even_scanline )
+        {
+            p_dst_u  += i_dst_pitch/2;
+            p_dst_v  += i_dst_pitch/2;
+            p_src1_u += i_src1_pitch/2;
+            p_src1_v += i_src1_pitch/2;
+        }
+        b_even_scanline = !b_even_scanline;
+
+        /* Draw until we reach the end of the line */
+        for( i_x = 0; i_x < i_width; i_x++ )
+        {
+            if( i_alpha == MAX_TRANS )
+            {
+                /* Completely opaque. Completely overwrite underlying pixel */
+                p_dst_y[i_x] = p_src2_y[i_x];
+
+                if( b_even_scanline && i_x % 2 == 0 )
+                {
+                    p_dst_u[i_x/2] = p_src2_u[i_x/2];
+                    p_dst_v[i_x/2] = p_src2_v[i_x/2];
+                }
+                continue;
+            }
+
+            /* Blending */
+            p_dst_y[i_x] = ( (uint16_t)p_src2_y[i_x] * i_alpha +
+                (uint16_t)p_src1_y[i_x] * (MAX_TRANS - i_alpha) )
+                >> TRANS_BITS;
+
+            if( b_even_scanline && i_x % 2 == 0 )
+            {
+                p_dst_u[i_x/2] = ( (uint16_t)p_src2_u[i_x/2] * i_alpha +
+                (uint16_t)p_src1_u[i_x/2] * (MAX_TRANS - i_alpha) )
+                >> TRANS_BITS;
+                p_dst_v[i_x/2] = ( (uint16_t)p_src2_v[i_x/2] * i_alpha +
+                (uint16_t)p_src1_v[i_x/2] * (MAX_TRANS - i_alpha) )
+                >> TRANS_BITS;
+            }
+        }
+        if( i_y%2 == 1 )
+        {
+            p_src2_u += i_src2_pitch/2;
+            p_src2_v += i_src2_pitch/2;
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}
+static void BlendI420I420_no_alpha( filter_t *p_filter, picture_t *p_dst,
+                                    picture_t *p_dst_orig, picture_t *p_src,
+                                    int i_x_offset, int i_y_offset,
+                                    int i_width, int i_height )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch;
+    uint8_t *p_src1_y, *p_src2_y, *p_dst_y;
+    uint8_t *p_src1_u, *p_src2_u, *p_dst_u;
+    uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
+    int i_y;
+    vlc_bool_t b_even_scanline = i_y_offset % 2;
+
+    i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
+    p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
+              p_filter->fmt_out.video.i_x_offset +
+              p_dst->p[Y_PLANE].i_pitch *
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+    p_dst_u = p_dst->p[U_PLANE].p_pixels + i_x_offset/2 +
+              p_filter->fmt_out.video.i_x_offset/2 +
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+              p_dst->p[U_PLANE].i_pitch;
+    p_dst_v = p_dst->p[V_PLANE].p_pixels + i_x_offset/2 +
+              p_filter->fmt_out.video.i_x_offset/2 +
+              ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+              p_dst->p[V_PLANE].i_pitch;
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1_y = p_dst_orig->p[Y_PLANE].p_pixels + i_x_offset +
+               p_filter->fmt_out.video.i_x_offset +
+               p_dst_orig->p[Y_PLANE].i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+    p_src1_u = p_dst_orig->p[U_PLANE].p_pixels + i_x_offset/2 +
+               p_filter->fmt_out.video.i_x_offset/2 +
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+               p_dst_orig->p[U_PLANE].i_pitch;
+    p_src1_v = p_dst_orig->p[V_PLANE].p_pixels + i_x_offset/2 +
+               p_filter->fmt_out.video.i_x_offset/2 +
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset ) / 2 *
+               p_dst_orig->p[V_PLANE].i_pitch;
+
+    i_src2_pitch = p_src->p[Y_PLANE].i_pitch;
+    p_src2_y = p_src->p[Y_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset +
+               p_src->p[Y_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
+    p_src2_u = p_src->p[U_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[U_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
+    p_src2_v = p_src->p[V_PLANE].p_pixels +
+               p_filter->fmt_in.video.i_x_offset/2 +
+               p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
 
-            /* Blending */
-            p_dst_y[i_x] = ( (uint16_t)p_src2_y[i_x] * i_trans +
-                (uint16_t)p_src1_y[i_x] * (MAX_TRANS - i_trans) )
-                >> TRANS_BITS;
+    i_width &= ~1;
 
-            if( b_even_scanline && i_x % 2 == 0 )
-            {
-                p_dst_u[i_x/2] = ( (uint16_t)p_src2_u[i_x] * i_trans +
-                (uint16_t)p_src1_u[i_x/2] * (MAX_TRANS - i_trans) )
-                >> TRANS_BITS;
-                p_dst_v[i_x/2] = ( (uint16_t)p_src2_v[i_x] * i_trans +
-                (uint16_t)p_src1_v[i_x/2] * (MAX_TRANS - i_trans) )
-                >> TRANS_BITS;
-            }
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++,
+         p_dst_y += i_dst_pitch, p_src1_y += i_src1_pitch,
+         p_src2_y += i_src2_pitch )
+    {
+        /* Completely opaque. Completely overwrite underlying pixel */
+        p_filter->p_libvlc->pf_memcpy( p_dst_y, p_src2_y, i_width );
+        if( b_even_scanline )
+        {
+            p_dst_u  += i_dst_pitch/2;
+            p_dst_v  += i_dst_pitch/2;
+            p_src1_u += i_src1_pitch/2;
+            p_src1_v += i_src1_pitch/2;
+        }
+        else
+        {
+            p_filter->p_libvlc->pf_memcpy( p_dst_u, p_src2_u, i_width/2 );
+            p_filter->p_libvlc->pf_memcpy( p_dst_v, p_src2_v, i_width/2 );
+        }
+        b_even_scanline = !b_even_scanline;
+        if( i_y%2 == 1 )
+        {
+            p_src2_u += i_src2_pitch/2;
+            p_src2_v += i_src2_pitch/2;
         }
     }
 
-#undef MAX_TRANS
-#undef TRANS_BITS
-
     return;
 }
 
-static void BlendR16( filter_t *p_filter, picture_t *p_dst_pic,
-                      picture_t *p_dst_orig, picture_t *p_src,
-                      int i_x_offset, int i_y_offset,
-                      int i_width, int i_height, int i_alpha )
+static void BlendI420R16( filter_t *p_filter, picture_t *p_dst_pic,
+                          picture_t *p_dst_orig, picture_t *p_src,
+                          int i_x_offset, int i_y_offset,
+                          int i_width, int i_height, int i_alpha )
 {
     int i_src1_pitch, i_src2_pitch, i_dst_pitch;
     uint8_t *p_dst, *p_src1, *p_src2_y;
     uint8_t *p_src2_u, *p_src2_v;
-    uint8_t *p_trans;
     int i_x, i_y, i_pix_pitch;
     int r, g, b;
 
@@ -445,43 +1007,39 @@ static void BlendR16( filter_t *p_filter, picture_t *p_dst_pic,
                p_filter->fmt_in.video.i_x_offset/2 +
                p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
 
-    p_trans = p_src->p[A_PLANE].p_pixels +
-              p_filter->fmt_in.video.i_x_offset +
-              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
-
 #define MAX_TRANS 255
 #define TRANS_BITS  8
 
     /* Draw until we reach the bottom of the subtitle */
-    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+    for( i_y = 0; i_y < i_height; i_y++,
          p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
-         p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
-         p_src2_v += i_src2_pitch )
+         p_src2_y += i_src2_pitch )
     {
         /* Draw until we reach the end of the line */
         for( i_x = 0; i_x < i_width; i_x++ )
         {
-            if( !p_trans[i_x] )
-            {
-                /* Completely transparent. Don't change pixel */
-                continue;
-            }
-            else if( p_trans[i_x] == MAX_TRANS )
+            if( i_alpha == MAX_TRANS )
             {
                 /* Completely opaque. Completely overwrite underlying pixel */
                 yuv_to_rgb( &r, &g, &b,
-                            p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+                            p_src2_y[i_x], p_src2_u[i_x/2], p_src2_v[i_x/2] );
 
     ((uint16_t *)(&p_dst[i_x * i_pix_pitch]))[0] = ((r >> 3) << 11) | ((g >> 2) << 5) | (b >> 3);
                 continue;
             }
 
             /* Blending */
+            /* FIXME: do the blending */
             yuv_to_rgb( &r, &g, &b,
-                        p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+                        p_src2_y[i_x], p_src2_u[i_x/2], p_src2_v[i_x/2] );
 
     ((uint16_t *)(&p_dst[i_x * i_pix_pitch]))[0] = ((r >> 3) << 11) | ((g >> 2) << 5) | (b >> 3);
         }
+        if( i_y%2 == 1 )
+        {
+            p_src2_u += i_src2_pitch/2;
+            p_src2_v += i_src2_pitch/2;
+        }
     }
 
 #undef MAX_TRANS
@@ -490,16 +1048,15 @@ static void BlendR16( filter_t *p_filter, picture_t *p_dst_pic,
     return;
 }
 
-static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
-                      picture_t *p_dst_orig, picture_t *p_src,
-                      int i_x_offset, int i_y_offset,
-                      int i_width, int i_height, int i_alpha )
+static void BlendI420R24( filter_t *p_filter, picture_t *p_dst_pic,
+                          picture_t *p_dst_orig, picture_t *p_src,
+                          int i_x_offset, int i_y_offset,
+                          int i_width, int i_height, int i_alpha )
 {
     int i_src1_pitch, i_src2_pitch, i_dst_pitch;
     uint8_t *p_dst, *p_src1, *p_src2_y;
     uint8_t *p_src2_u, *p_src2_v;
-    uint8_t *p_trans;
-    int i_x, i_y, i_pix_pitch, i_trans;
+    int i_x, i_y, i_pix_pitch;
     int r, g, b;
 
     i_pix_pitch = p_dst_pic->p->i_pixel_pitch;
@@ -526,15 +1083,11 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
                p_filter->fmt_in.video.i_x_offset/2 +
                p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
 
-    p_trans = p_src->p[A_PLANE].p_pixels +
-              p_filter->fmt_in.video.i_x_offset +
-              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
-
 #define MAX_TRANS 255
 #define TRANS_BITS  8
 
     /* Draw until we reach the bottom of the subtitle */
-    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+    for( i_y = 0; i_y < i_height; i_y++,
          p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
          p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
          p_src2_v += i_src2_pitch )
@@ -542,17 +1095,11 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
         /* Draw until we reach the end of the line */
         for( i_x = 0; i_x < i_width; i_x++ )
         {
-            i_trans = ( p_trans[i_x] * i_alpha ) / 255;
-            if( !i_trans )
-            {
-                /* Completely transparent. Don't change pixel */
-                continue;
-            }
-            else if( i_trans == MAX_TRANS )
+            if( i_alpha == MAX_TRANS )
             {
                 /* Completely opaque. Completely overwrite underlying pixel */
                 yuv_to_rgb( &r, &g, &b,
-                            p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+                            p_src2_y[i_x], p_src2_u[i_x/2], p_src2_v[i_x/2] );
 
                 p_dst[i_x * i_pix_pitch]     = r;
                 p_dst[i_x * i_pix_pitch + 1] = g;
@@ -562,17 +1109,22 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
 
             /* Blending */
             yuv_to_rgb( &r, &g, &b,
-                        p_src2_y[i_x], p_src2_u[i_x], p_src2_v[i_x] );
+                        p_src2_y[i_x], p_src2_u[i_x/2], p_src2_v[i_x/2] );
 
-            p_dst[i_x * i_pix_pitch]     = ( r * i_trans +
+            p_dst[i_x * i_pix_pitch]     = ( r * i_alpha +
                 (uint16_t)p_src1[i_x * i_pix_pitch] *
-                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
-            p_dst[i_x * i_pix_pitch + 1] = ( g * i_trans +
+                (MAX_TRANS - i_alpha) ) >> TRANS_BITS;
+            p_dst[i_x * i_pix_pitch + 1] = ( g * i_alpha +
                 (uint16_t)p_src1[i_x * i_pix_pitch + 1] *
-                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
-            p_dst[i_x * i_pix_pitch + 2] = ( b * i_trans +
+                (MAX_TRANS - i_alpha) ) >> TRANS_BITS;
+            p_dst[i_x * i_pix_pitch + 2] = ( b * i_alpha +
                 (uint16_t)p_src1[i_x * i_pix_pitch + 2] *
-                (MAX_TRANS - i_trans) ) >> TRANS_BITS;
+                (MAX_TRANS - i_alpha) ) >> TRANS_BITS;
+        }
+        if( i_y%2 == 1 )
+        {
+            p_src2_u += i_src2_pitch/2;
+            p_src2_v += i_src2_pitch/2;
         }
     }
 
@@ -582,16 +1134,15 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
     return;
 }
 
-static void BlendYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
-                            picture_t *p_dst_orig, picture_t *p_src,
-                            int i_x_offset, int i_y_offset,
-                            int i_width, int i_height, int i_alpha )
+static void BlendI420YUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
+                                picture_t *p_dst_orig, picture_t *p_src,
+                                int i_x_offset, int i_y_offset,
+                                int i_width, int i_height, int i_alpha )
 {
     int i_src1_pitch, i_src2_pitch, i_dst_pitch;
     uint8_t *p_dst, *p_src1, *p_src2_y;
     uint8_t *p_src2_u, *p_src2_v;
-    uint8_t *p_trans;
-    int i_x, i_y, i_pix_pitch, i_trans;
+    int i_x, i_y, i_pix_pitch;
     vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
     int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
 
@@ -638,17 +1189,13 @@ static void BlendYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
                p_filter->fmt_in.video.i_x_offset/2 +
                p_src->p[V_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset/2;
 
-    p_trans = p_src->p[A_PLANE].p_pixels +
-              p_filter->fmt_in.video.i_x_offset +
-              p_src->p[A_PLANE].i_pitch * p_filter->fmt_in.video.i_y_offset;
-
-    i_width = (i_width >> 1) << 1; /* Needs to be a multiple of 2 */
+    i_width &= ~1; /* Needs to be a multiple of 2 */
 
 #define MAX_TRANS 255
 #define TRANS_BITS  8
 
     /* Draw until we reach the bottom of the subtitle */
-    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+    for( i_y = 0; i_y < i_height; i_y++,
          p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
          p_src2_y += i_src2_pitch, p_src2_u += i_src2_pitch,
          p_src2_v += i_src2_pitch )
@@ -656,60 +1203,42 @@ static void BlendYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
         /* Draw until we reach the end of the line */
         for( i_x = 0; i_x < i_width; i_x++, b_even = !b_even )
         {
-            i_trans = ( p_trans[i_x] * i_alpha ) / 255;
-            if( !i_trans )
-            {
-                /* Completely transparent. Don't change pixel */
-            }
-            else if( i_trans == MAX_TRANS )
+            if( i_alpha == MAX_TRANS )
             {
                 /* Completely opaque. Completely overwrite underlying pixel */
                 p_dst[i_x * 2 + i_l_offset]     = p_src2_y[i_x];
 
                 if( b_even )
                 {
-                    if( p_trans[i_x+1] > 0xaa )
-                    {
-                        p_dst[i_x * 2 + i_u_offset] = (p_src2_u[i_x]+p_src2_u[i_x+1])>>1;
-                        p_dst[i_x * 2 + i_v_offset] = (p_src2_v[i_x]+p_src2_v[i_x+1])>>1;
-                    }
-                    else
-                    {
-                        p_dst[i_x * 2 + i_u_offset] = p_src2_u[i_x];
-                        p_dst[i_x * 2 + i_v_offset] = p_src2_v[i_x];
-                    }
+                    p_dst[i_x * 2 + i_u_offset] = p_src2_u[i_x/2];
+                    p_dst[i_x * 2 + i_v_offset] = p_src2_v[i_x/2];
                 }
             }
             else
             {
                 /* Blending */
-                p_dst[i_x * 2 + i_l_offset]     = ( (uint16_t)p_src2_y[i_x] * i_trans +
-                    (uint16_t)p_src1[i_x * 2 + i_l_offset] * (MAX_TRANS - i_trans) )
+                p_dst[i_x * 2 + i_l_offset]     = ( (uint16_t)p_src2_y[i_x] * i_alpha +
+                    (uint16_t)p_src1[i_x * 2 + i_l_offset] * (MAX_TRANS - i_alpha) )
                     >> TRANS_BITS;
 
                 if( b_even )
                 {
-                    uint16_t i_u = 0;
-                    uint16_t i_v = 0;
-                    if( p_trans[i_x+1] > 0xaa )
-                    {
-                        i_u = (p_src2_u[i_x]+p_src2_u[i_x+1])>>1;
-                        i_v = (p_src2_v[i_x]+p_src2_v[i_x+1])>>1;
-                    }
-                    else
-                    {
-                        i_u = p_src2_u[i_x];
-                        i_v = p_src2_v[i_x];
-                    }
-                    p_dst[i_x * 2 + i_u_offset] = ( (uint16_t)i_u * i_trans +
-                        (uint16_t)p_src1[i_x * 2 + i_u_offset] * (MAX_TRANS - i_trans) )
+                    uint16_t i_u = p_src2_u[i_x/2];
+                    uint16_t i_v = p_src2_v[i_x/2];
+                    p_dst[i_x * 2 + i_u_offset] = ( (uint16_t)i_u * i_alpha +
+                        (uint16_t)p_src1[i_x * 2 + i_u_offset] * (MAX_TRANS - i_alpha) )
                         >> TRANS_BITS;
-                    p_dst[i_x * 2 + i_v_offset] = ( (uint16_t)i_v * i_trans +
-                        (uint16_t)p_src1[i_x * 2 + i_v_offset] * (MAX_TRANS - i_trans) )
+                    p_dst[i_x * 2 + i_v_offset] = ( (uint16_t)i_v * i_alpha +
+                        (uint16_t)p_src1[i_x * 2 + i_v_offset] * (MAX_TRANS - i_alpha) )
                         >> TRANS_BITS;
                 }
             }
         }
+        if( i_y%2 == 1 )
+        {
+            p_src2_u += i_src2_pitch/2;
+            p_src2_v += i_src2_pitch/2;
+        }
     }
 
 #undef MAX_TRANS
@@ -1169,7 +1698,7 @@ static void BlendRGBAI420( filter_t *p_filter, picture_t *p_dst,
     return;
 }
 
-static void BlendRGBAR16( filter_t *p_filter, picture_t *p_dst_pic,
+static void BlendRGBAR24( filter_t *p_filter, picture_t *p_dst_pic,
                           picture_t *p_dst_orig, picture_t *p_src,
                           int i_x_offset, int i_y_offset,
                           int i_width, int i_height, int i_alpha )
@@ -1247,7 +1776,7 @@ static void BlendRGBAR16( filter_t *p_filter, picture_t *p_dst_pic,
     return;
 }
 
-static void BlendRGBAR24( filter_t *p_filter, picture_t *p_dst_pic,
+static void BlendRGBAR16( filter_t *p_filter, picture_t *p_dst_pic,
                           picture_t *p_dst_orig, picture_t *p_src,
                           int i_x_offset, int i_y_offset,
                           int i_width, int i_height, int i_alpha )
@@ -1325,3 +1854,114 @@ static void BlendRGBAR24( filter_t *p_filter, picture_t *p_dst_pic,
 
     return;
 }
+
+static void BlendRGBAYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
+                                picture_t *p_dst_orig, picture_t *p_src,
+                                int i_x_offset, int i_y_offset,
+                                int i_width, int i_height, int i_alpha )
+{
+    int i_src1_pitch, i_src2_pitch, i_dst_pitch, i_src_pix_pitch;
+    uint8_t *p_dst, *p_src1, *p_src2;
+    uint8_t *p_trans;
+    int i_x, i_y, i_pix_pitch, i_trans;
+    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
+    uint8_t y, u, v;
+
+    if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','U','Y','2') )
+    {
+        i_l_offset = 0;
+        i_u_offset = 1;
+        i_v_offset = 3;
+    }
+    else if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('U','Y','V','Y') )
+    {
+        i_l_offset = 1;
+        i_u_offset = 0;
+        i_v_offset = 2;
+    }
+    else if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','V','Y','U') )
+    {
+        i_l_offset = 0;
+        i_u_offset = 3;
+        i_v_offset = 1;
+    }
+
+    i_pix_pitch = 2;
+    i_dst_pitch = p_dst_pic->p->i_pitch;
+    p_dst = p_dst_pic->p->p_pixels + i_x_offset * i_pix_pitch +
+            p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+            p_dst_pic->p->i_pitch *
+            ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src1_pitch = p_dst_orig->p[Y_PLANE].i_pitch;
+    p_src1 = p_dst_orig->p->p_pixels + i_x_offset * i_pix_pitch +
+               p_filter->fmt_out.video.i_x_offset * i_pix_pitch +
+               p_dst_orig->p->i_pitch *
+               ( i_y_offset + p_filter->fmt_out.video.i_y_offset );
+
+    i_src_pix_pitch = p_src->p->i_pixel_pitch;
+    i_src2_pitch = p_src->p->i_pitch;
+    p_src2 = p_src->p->p_pixels +
+             p_filter->fmt_in.video.i_x_offset * i_src2_pitch +
+             p_src->p->i_pitch * p_filter->fmt_in.video.i_y_offset;
+
+    i_width = (i_width >> 1) << 1; /* Needs to be a multiple of 2 */
+
+#define MAX_TRANS 255
+#define TRANS_BITS  8
+
+    /* Draw until we reach the bottom of the subtitle */
+    for( i_y = 0; i_y < i_height; i_y++, p_trans += i_src2_pitch,
+         p_dst += i_dst_pitch, p_src1 += i_src1_pitch,
+         p_src2 += i_src2_pitch )
+    {
+        /* Draw until we reach the end of the line */
+        for( i_x = 0; i_x < i_width; i_x++, b_even = !b_even )
+        {
+#define     R         ( p_src2[i_x * i_src_pix_pitch + 0] )
+#define     G         ( p_src2[i_x * i_src_pix_pitch + 1] )
+#define     B         ( p_src2[i_x * i_src_pix_pitch + 2] )
+            i_trans = ( p_src2[i_x * i_src_pix_pitch + 3] * i_alpha ) / 255;
+            if( !i_trans )
+            {
+                /* Completely transparent. Don't change pixel */
+            }
+            else if( i_trans == MAX_TRANS )
+            {
+                /* Completely opaque. Completely overwrite underlying pixel */
+                rgb_to_yuv( &y, &u, &v, R, G, B );
+                p_dst[i_x * 2 + i_l_offset] = y;
+
+                if( b_even )
+                {
+                    p_dst[i_x * 2 + i_u_offset] = u;
+                    p_dst[i_x * 2 + i_v_offset] = v;
+                }
+            }
+            else
+            {
+                /* Blending */
+                rgb_to_yuv( &y, &u, &v, R, G, B );
+                p_dst[i_x * 2 + i_l_offset]     = ( (uint16_t)y * i_trans +
+                    (uint16_t)p_src1[i_x * 2 + i_l_offset] * (MAX_TRANS - i_trans) )
+                    >> TRANS_BITS;
+
+                if( b_even )
+                {
+                    p_dst[i_x * 2 + i_u_offset] = ( (uint16_t)u * i_trans +
+                        (uint16_t)p_src1[i_x * 2 + i_u_offset] * (MAX_TRANS - i_trans) )
+                        >> TRANS_BITS;
+                    p_dst[i_x * 2 + i_v_offset] = ( (uint16_t)v * i_trans +
+                        (uint16_t)p_src1[i_x * 2 + i_v_offset] * (MAX_TRANS - i_trans) )
+                        >> TRANS_BITS;
+                }
+            }
+        }
+    }
+
+#undef MAX_TRANS
+#undef TRANS_BITS
+
+    return;
+}