Opie Player2 XINE update:

	-update xine code alphablend, yuv2rgb to be up to date
	-XINELib:
		-use the XINE infrastructure to change the gui width/height
		-use the XINE infrastructure to change the brightness
	-Update the NULL Video Driver:
		-make use of frame_dest_cb and and dest_size_cb (callback)
		 to update the size of the gui
		-use alphablend
		-merge the video_out_fb xine driver

	To Alwin: your testvideo now lockups op2. with valgrind I see
	it has to do with the yuv2rgb conversion methods

8 files changed, 1893 insertions, 689 deletions

diff --git a/noncore/multimedia/opieplayer2/alphablend.c b/noncore/multimedia/opieplayer2/alphablend.c
index cdd7b28..ab80531 100644
--- a/noncore/multimedia/opieplayer2/alphablend.c
+++ b/noncore/multimedia/opieplayer2/alphablend.c
@@ -1,13 +1,10 @@
-//TOAST_SPU will define ALL spu entries - no matter the tranparency
-//#define TOAST_SPU
-/* #define PRIV_CLUT */
-/* Currently only blend_yuv(..) works */
 /*
  *
  * Copyright (C) James Courtier-Dutton James@superbug.demon.co.uk - July 2001
  * 
  * Copyright (C) 2000  Thomas Mirlacher
+ *               2002-2004 the xine project
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
  * the Free Software Foundation; either version 2 of the License, or
@@ -29,19 +26,19 @@
  */
 
 /*
 #define LOG_BLEND_YUV
+#define LOG_BLEND_RGB16
 */
 
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>
-#include <inttypes.h>
 
-#include <xine.h>
 #include <xine/xine_internal.h>
 #include <xine/video_out.h>
 #include "alphablend.h"
+#include "bswap.h"
 
 
 #define BLEND_COLOR(dst, src, mask, o) ((((src&mask)*o + ((dst&mask)*(0x0f-o)))/0xf) & mask)
 
@@ -58,9 +55,9 @@ static void mem_blend16(uint16_t *mem, uint16_t clr, uint8_t o, int len) {
   }
 }
 
 static void mem_blend24(uint8_t *mem, uint8_t r, uint8_t g, uint8_t b,
- uint8_t o, int len) {
+                        uint8_t o, int len) {
   uint8_t *limit = mem + len*3;
   while (mem < limit) {
     *mem = BLEND_BYTE(*mem, r, o);
     mem++;
@@ -70,21 +67,8 @@ static void mem_blend24(uint8_t *mem, uint8_t r, uint8_t g, uint8_t b,
     mem++;
   }
 }
 
-static void mem_blend24_32(uint8_t *mem, uint8_t r, uint8_t g, uint8_t b,
- uint8_t o, int len) {
-  uint8_t *limit = mem + len*4;
-  while (mem < limit) {
-    *mem = BLEND_BYTE(*mem, r, o);
-    mem++;
-    *mem = BLEND_BYTE(*mem, g, o);
-    mem++;
-    *mem = BLEND_BYTE(*mem, b, o);
-    mem += 2;
-  }
-}
-
 static void mem_blend32(uint8_t *mem, uint8_t *src, uint8_t o, int len) {
   uint8_t *limit = mem + len*4;
   while (mem < limit) {
     *mem = BLEND_BYTE(*mem, src[0], o);
@@ -97,9 +81,8 @@ static void mem_blend32(uint8_t *mem, uint8_t *src, uint8_t o, int len) {
     mem++;
   }
 }
 
-
 /*
  * Some macros for fixed point arithmetic.
  *
  * The blend_rgb* routines perform rle image scaling using
@@ -126,112 +109,416 @@ rle_img_advance_line(rle_elem_t *rle, rle_elem_t *rle_limit, int w)
   }
   return rle;
 }
 
-
 void blend_rgb16 (uint8_t * img, vo_overlay_t * img_overl,
-                  int img_width, int img_height,
-                  int dst_width, int dst_height)
+                  int img_width, int img_height,
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data)
 {
   uint8_t *trans;
-  clut_t* clut = (clut_t*) img_overl->clip_color;
+  clut_t *clut;
 
   int src_width = img_overl->width;
   int src_height = img_overl->height;
   rle_elem_t *rle = img_overl->rle;
+  rle_elem_t *rle_start = img_overl->rle;
   rle_elem_t *rle_limit = rle + img_overl->num_rle;
   int x, y, x1_scaled, x2_scaled;
-          int dy, dy_step, x_scale;/* scaled 2**SCALE_SHIFT */
+  int dy, dy_step, x_scale;     /* scaled 2**SCALE_SHIFT */
   int clip_right;
   uint16_t *img_pix;
+  int rlelen;
+  int rle_this_bite;
+  int rle_remainder;
+  int zone_state=0;
+  uint8_t clr_next,clr;
+  uint16_t o;
+  double img_offset;
+  int stripe_height;
+/*
+ * Let zone_state keep state.
+ * 0 = Starting.
+ * 1 = Above button.
+ * 2 = Left of button.
+ * 3 = Inside of button.
+ * 4 = Right of button.
+ * 5 = Below button.
+ * 6 = Finished.
+ *
+ * Each time round the loop, update the state.
+ * We can do this easily and cheaply(fewer IF statements per cycle) as we are testing rle end position anyway.
+ * Possible optimization is to ensure that rle never overlaps from outside to inside a button.
+ * Possible optimization is to pre-scale the RLE overlay, so that no scaling is needed here.
+ */
 
-  dy_step = INT_TO_SCALED(dst_height) / img_height;
-  x_scale = INT_TO_SCALED(img_width)  / dst_width;
+#ifdef LOG_BLEND_RGB16
+  printf("blend_rgb16: img_height=%i, dst_height=%i\n", img_height, dst_height);
+  printf("blend_rgb16: img_width=%i, dst_width=%i\n", img_width, dst_width);
+  if (img_width & 1) { printf("blend_rgb16s: odd\n");}
+  else { printf("blend_rgb16s: even\n");}
 
-  img_pix = (uint16_t *) img
+#endif
+/* stripe_height is used in yuv2rgb scaling, so use the same scale factor here for overlays. */
+  stripe_height = 16 * img_height / dst_height;
+/*  dy_step = INT_TO_SCALED(dst_height) / img_height; */
+  dy_step = INT_TO_SCALED(16) / stripe_height;
+  x_scale = INT_TO_SCALED(img_width)  / dst_width;
+#ifdef LOG_BLEND_RGB16
+  printf("blend_rgb16: dy_step=%i, x_scale=%i\n", dy_step, x_scale);
+#endif
+  if (img_width & 1) img_width++;
+  img_offset = ( ( (img_overl->y * img_height) / dst_height) * img_width) 
+             + ( (img_overl->x * img_width) / dst_width);
+#ifdef LOG_BLEND_RGB16
+  printf("blend_rgb16: x=%i, y=%i, w=%i, h=%i, img_offset=%lf\n", img_overl->x, img_overl->y,
+    img_overl->width,
+    img_overl->height,
+    img_offset);
+#endif
+  img_pix = (uint16_t *) img + (int)img_offset;
+/* 
       + (img_overl->y * img_height / dst_height) * img_width
       + (img_overl->x * img_width / dst_width);
-
-  trans = img_overl->clip_trans;
+*/
 
   /* avoid wraping overlay if drawing to small image */
   if( (img_overl->x + img_overl->clip_right) < dst_width )
     clip_right = img_overl->clip_right;
   else
-    clip_right = dst_width - 1 - img_overl->x;
+    clip_right = dst_width - img_overl->x;
 
   /* avoid buffer overflow */
   if( (src_height + img_overl->y) >= dst_height )
-    src_height = dst_height - 1 - img_overl->y;
+    src_height = dst_height - img_overl->y;
 
-  for (y = dy = 0; y < src_height && rle < rle_limit;) {
-    int mask = !(img_overl->clip_top > y || img_overl->clip_bottom < y);
-    rle_elem_t *rle_start = rle;
+  rlelen = rle_remainder = rle_this_bite = 0;
+  rle_remainder = rlelen = rle->len;
+  clr_next = rle->color;
+  rle++;
+  y = dy = 0;
+  x = x1_scaled = x2_scaled = 0;
 
-    for (x = x1_scaled = 0; x < src_width;) {
-      uint8_t clr;
-      uint16_t o;
-      int rlelen;
+#ifdef LOG_BLEND_RGB16
+  printf("blend_rgb16 started\n"); 
+#endif
+
+  while (zone_state != 6) {
+    clr = clr_next;
+    switch (zone_state) {
+    case 0:  /* Starting */
+      /* FIXME: Get libspudec to set clip_top to -1 if no button */
+      if (img_overl->clip_top < 0) {
+#ifdef LOG_BLEND_RGB16
+        printf("blend_rgb16: No button clip area\n");
+#endif
 
-      clr = rle->color;
+        zone_state = 7;
+        break;
+      }
+#ifdef LOG_BLEND_RGB16
+      printf("blend_rgb16: Button clip area found. (%d,%d) .. (%d,%d)\n",
+        img_overl->clip_left,
+        img_overl->clip_top,
+        img_overl->clip_right,
+        img_overl->clip_bottom);
+#endif
+      if (y < img_overl->clip_top) {
+        zone_state = 1;
+        break;
+      } else if (y >= img_overl->clip_bottom) {
+        zone_state = 5;
+        break;
+      } else if (x < img_overl->clip_left) {
+        zone_state = 2;
+        break;
+      } else if (x >= img_overl->clip_right) {
+        zone_state = 4;
+        break;
+      } else {
+        zone_state = 3;
+        break;
+      }
+      break;
+    case 1:  /* Above clip area */
+      clut = (clut_t*) img_overl->color;
+      trans = img_overl->trans;
       o   = trans[clr];
-      rlelen = rle->len;
-
-      if (o && mask) {
-        /* threat cases where clipping border is inside rle->len pixels */
-        if ( img_overl->clip_left > x ) {
-          if( img_overl->clip_left < x + rlelen ) {
-            x1_scaled = SCALED_TO_INT( img_overl->clip_left * x_scale );
-            rlelen -= img_overl->clip_left - x;
-            x += img_overl->clip_left - x;
-          } else {
-            o = 0;
+      rle_this_bite = rle_remainder;
+      rle_remainder = 0;
+      rlelen -= rle_this_bite;
+      /*printf("(x,y) = (%03i,%03i), clr=%03x, len=%03i, zone=%i\n", x, y, clr, rle_this_bite, zone_state); */
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      }
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+ 
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
           }
-        } else if( clip_right < x + rlelen ) {
-          if( clip_right > x ) {
-            x2_scaled = SCALED_TO_INT( clip_right * x_scale);
-            mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o,
-                        x2_scaled-x1_scaled);
-            o = 0;            
-          } else {
-            o = 0;
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+      } 
+      rle_remainder = rlelen = rle->len;
+      clr_next = rle->color;
+      rle++;
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      if (y >= img_overl->clip_top) {
+        zone_state = 2;
+#ifdef LOG_BLEND_RGB16
+        printf("blend_rgb16: Button clip top reached. y=%i, top=%i\n",
+                y, img_overl->clip_top);
+#endif
+        if (x >= img_overl->clip_left) {
+          zone_state = 3;
+          if (x >= img_overl->clip_right) {
+            zone_state = 4;
           }
-        } 
+        }
       }
-      
-      x2_scaled = SCALED_TO_INT((x + rlelen) * x_scale);
-      if (o && mask) {
-        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      break;
+    case 2:  /* Left of button */
+      clut = (clut_t*) img_overl->color;
+      trans = img_overl->trans;
+      o   = trans[clr];
+      if (x + rle_remainder <= img_overl->clip_left) {
+        rle_this_bite = rle_remainder;
+        rle_remainder = rlelen = rle->len;
+        clr_next = rle->color;
+        rle++;
+      } else {
+        rle_this_bite = img_overl->clip_left - x;
+        rle_remainder -= rle_this_bite;
+        zone_state = 3;
       }
-
-      x1_scaled = x2_scaled;
-      x += rlelen;
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      }
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
+          }
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+        if (y >= img_overl->clip_bottom) {
+          zone_state = 5;
+          break;
+        }
+      }
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      break;
+    case 3:  /* In button */
+      clut = (clut_t*) img_overl->clip_color;
+      trans = img_overl->clip_trans;
+      o   = trans[clr];
+      if (x + rle_remainder <= img_overl->clip_right) {
+        rle_this_bite = rle_remainder;
+        rle_remainder = rlelen = rle->len;
+        clr_next = rle->color;
+        rle++;
+      } else {
+        rle_this_bite = img_overl->clip_right - x;
+        rle_remainder -= rle_this_bite;
+        zone_state = 4;
+      }
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      }
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
+          }
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+        if (y >= img_overl->clip_bottom) {
+          zone_state = 5;
+          break;
+        }
+      } 
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      break;
+    case 4:  /* Right of button */
+      clut = (clut_t*) img_overl->color;
+      trans = img_overl->trans;
+      o   = trans[clr];
+      if (x + rle_remainder <= src_width) {
+        rle_this_bite = rle_remainder;
+        rle_remainder = rlelen = rle->len;
+        clr_next = rle->color;
+        rle++;
+      } else {
+        rle_this_bite = src_width - x;
+        rle_remainder -= rle_this_bite;
+        zone_state = 2;
+      }
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      }
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
+          }
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+        if (y >= img_overl->clip_bottom) {
+          zone_state = 5;
+          break;
+        }
+      } 
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      break;
+    case 5:  /* Below button */
+      clut = (clut_t*) img_overl->color;
+      trans = img_overl->trans;
+      o   = trans[clr];
+      rle_this_bite = rle_remainder;
+      rle_remainder = 0;
+      rlelen -= rle_this_bite;
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
+      }
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
+          }
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+      } 
+      rle_remainder = rlelen = rle->len;
+      clr_next = rle->color;
       rle++;
-      if (rle >= rle_limit) break;
-    }
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      break;
+    case 6:  /* Finished */
+      _x_abort();
 
-    img_pix += img_width;
-    dy += dy_step;
-    if (dy >= INT_TO_SCALED(1)) {
-      dy -= INT_TO_SCALED(1);
-      ++y;
-      while (dy >= INT_TO_SCALED(1)) {
-        rle = rle_img_advance_line(rle, rle_limit, src_width);
-        dy -= INT_TO_SCALED(1);
-        ++y;
+    case 7:  /* No button */
+      clut = (clut_t*) img_overl->color;
+      trans = img_overl->trans;
+      o   = trans[clr];
+      rle_this_bite = rle_remainder;
+      rle_remainder = 0;
+      rlelen -= rle_this_bite;
+      if (o) {
+        x1_scaled = SCALED_TO_INT( x * x_scale );
+        x2_scaled = SCALED_TO_INT( (x + rle_this_bite) * x_scale);
+        mem_blend16(img_pix+x1_scaled, *((uint16_t *)&clut[clr]), o, x2_scaled-x1_scaled);
       }
-    } else {
-              rle = rle_start;  /* y-scaling, reuse the last rle encoded line */
+      x += rle_this_bite;
+      if (x >= src_width ) { 
+        x -= src_width;
+        img_pix += img_width;
+        dy += dy_step;
+        if (dy >= INT_TO_SCALED(1)) {
+          dy -= INT_TO_SCALED(1);
+          ++y;
+          while (dy >= INT_TO_SCALED(1)) {
+            rle = rle_img_advance_line(rle, rle_limit, src_width);
+            dy -= INT_TO_SCALED(1);
+            ++y;
+          }
+          rle_start = rle;
+        } else {
+          rle = rle_start;          /* y-scaling, reuse the last rle encoded line */
+        }
+      } 
+      rle_remainder = rlelen = rle->len;
+      clr_next = rle->color;
+      rle++;
+      if (rle >= rle_limit) {
+        zone_state = 6;
+      }
+      break;
+    default:
+      ;
     }
   }
+#ifdef LOG_BLEND_RGB16
+  printf("blend_rgb16 ended\n");
+#endif
+ 
 }
 
 void blend_rgb24 (uint8_t * img, vo_overlay_t * img_overl,
                   int img_width, int img_height,
-                  int dst_width, int dst_height)
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data)
 {
-  clut_t* clut = (clut_t*) img_overl->clip_color;
-  uint8_t *trans;
   int src_width = img_overl->width;
   int src_height = img_overl->height;
   rle_elem_t *rle = img_overl->rle;
   rle_elem_t *rle_limit = rle + img_overl->num_rle;
@@ -245,67 +532,108 @@ void blend_rgb24 (uint8_t * img, vo_overlay_t * img_overl,
 
   img_pix = img + 3 * (  (img_overl->y * img_height / dst_height) * img_width
                        + (img_overl->x * img_width  / dst_width));
 
-  trans = img_overl->clip_trans;
-
   /* avoid wraping overlay if drawing to small image */
-  if( (img_overl->x + img_overl->clip_right) < dst_width )
+  if( (img_overl->x + img_overl->clip_right) <= dst_width )
     clip_right = img_overl->clip_right;
   else
-    clip_right = dst_width - 1 - img_overl->x;
+    clip_right = dst_width - img_overl->x;
 
   /* avoid buffer overflow */
-  if( (src_height + img_overl->y) >= dst_height )
-    src_height = dst_height - 1 - img_overl->y;
+  if( (src_height + img_overl->y) > dst_height )
+    src_height = dst_height - img_overl->y;
 
   for (dy = y = 0; y < src_height && rle < rle_limit; ) {
-    int mask = !(img_overl->clip_top > y || img_overl->clip_bottom < y);
+    int mask = !(y < img_overl->clip_top || y >= img_overl->clip_bottom);
     rle_elem_t *rle_start = rle;
 
+    int rlelen = 0;
+    uint8_t clr = 0;
+
     for (x = x1_scaled = 0; x < src_width;) {
-      uint8_t clr;
+      int rle_bite;
+      clut_t *colors;
+      uint8_t *trans;
       uint16_t o;
-      int rlelen;
 
-      clr = rle->color;
-      o   = trans[clr];
-      rlelen = rle->len;
-
-      if (o && mask) {
-        /* threat cases where clipping border is inside rle->len pixels */
-        if ( img_overl->clip_left > x ) {
-          if( img_overl->clip_left < x + rlelen ) {
-            x1_scaled = SCALED_TO_INT( img_overl->clip_left * x_scale );
-            rlelen -= img_overl->clip_left - x;
-            x += img_overl->clip_left - x;
+      /* take next element from rle list everytime an element is finished */
+      if (rlelen <= 0) {
+        if (rle >= rle_limit)
+          break;
+        
+        rlelen = rle->len;
+        clr = rle->color;
+        rle++;
+      }
+
+      if (!mask) {
+        /* above or below clipping area */
+        
+        rle_bite = rlelen;
+        /* choose palette for surrounding area */
+        colors = (clut_t*)img_overl->color;
+        trans = img_overl->trans;
+      } else {
+        /* treat cases where clipping border is inside rle->len pixels */
+        if ( x < img_overl->clip_left ) {
+          /* starts left */
+          if( x + rlelen > img_overl->clip_left ) {
+            /* ends not left */
+
+            /* choose the largest "bite" up to palette change */
+            rle_bite = img_overl->clip_left - x;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           } else {
-            o = 0;
+            /* ends left */
+            
+            rle_bite = rlelen;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           }
-        } else if( clip_right < x + rlelen ) {
-          if( clip_right > x ) {
-            x2_scaled = SCALED_TO_INT( clip_right * x_scale);
-            mem_blend24(img_pix + x1_scaled*3, clut[clr].cb,
-                    clut[clr].cr, clut[clr].y,
-                    o, x2_scaled-x1_scaled);
-            o = 0;            
+        } else if( x + rlelen > clip_right ) {
+          /* ends right */
+          if( x < clip_right ) {
+            /* starts not right */
+
+            /* choose the largest "bite" up to palette change */
+            rle_bite = clip_right - x;
+            /* we're in the center area so choose clip palette */
+            colors = (clut_t*)img_overl->clip_color;
+            trans = img_overl->clip_trans;
           } else {
-            o = 0;
+            /* starts right */
+            
+            rle_bite = rlelen;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           }
-        } 
+        } else {
+          /* starts not left and ends not right */
+          
+          rle_bite = rlelen;
+          /* we're in the center area so choose clip palette */
+          colors = (clut_t*)img_overl->clip_color;
+          trans = img_overl->clip_trans;
+        }
       }
       
-      x2_scaled = SCALED_TO_INT((x + rlelen) * x_scale);
-      if (o && mask) {
-        mem_blend24(img_pix + x1_scaled*3, clut[clr].cb,
-                    clut[clr].cr, clut[clr].y,
+      x2_scaled = SCALED_TO_INT((x + rle_bite) * x_scale);
+
+      o = trans[clr];
+      if (o) {
+        mem_blend24(img_pix + x1_scaled*3,
+                    colors[clr].cb, colors[clr].cr, colors[clr].y,
                     o, x2_scaled-x1_scaled);
       }
 
       x1_scaled = x2_scaled;
-      x += rlelen;
-      rle++;
-      if (rle >= rle_limit) break;
+      x += rle_bite;
+      rlelen -= rle_bite;
     }
 
     img_pix += img_width * 3;
     dy += dy_step;
@@ -324,12 +652,11 @@ void blend_rgb24 (uint8_t * img, vo_overlay_t * img_overl,
 }
 
 void blend_rgb32 (uint8_t * img, vo_overlay_t * img_overl,
                   int img_width, int img_height,
-                  int dst_width, int dst_height)
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data)
 {
-  clut_t* clut = (clut_t*) img_overl->clip_color;
-  uint8_t *trans;
   int src_width = img_overl->width;
   int src_height = img_overl->height;
   rle_elem_t *rle = img_overl->rle;
   rle_elem_t *rle_limit = rle + img_overl->num_rle;
@@ -343,67 +670,106 @@ void blend_rgb32 (uint8_t * img, vo_overlay_t * img_overl,
 
   img_pix = img + 4 * (  (img_overl->y * img_height / dst_height) * img_width
                        + (img_overl->x * img_width / dst_width));
 
-  trans = img_overl->clip_trans;
-
   /* avoid wraping overlay if drawing to small image */
-  if( (img_overl->x + img_overl->clip_right) < dst_width )
+  if( (img_overl->x + img_overl->clip_right) <= dst_width )
     clip_right = img_overl->clip_right;
   else
-    clip_right = dst_width - 1 - img_overl->x;
+    clip_right = dst_width - img_overl->x;
 
   /* avoid buffer overflow */
-  if( (src_height + img_overl->y) >= dst_height )
-    src_height = dst_height - 1 - img_overl->y;
+  if( (src_height + img_overl->y) > dst_height )
+    src_height = dst_height - img_overl->y;
 
   for (y = dy = 0; y < src_height && rle < rle_limit; ) {
-    int mask = !(img_overl->clip_top > y || img_overl->clip_bottom < y);
+    int mask = !(y < img_overl->clip_top || y >= img_overl->clip_bottom);
     rle_elem_t *rle_start = rle;
 
+    int rlelen = 0;
+    uint8_t clr = 0;
+    
     for (x = x1_scaled = 0; x < src_width;) {
-      uint8_t clr;
+      int rle_bite;
+      clut_t *colors;
+      uint8_t *trans;
       uint16_t o;
-      int rlelen;
+      
+      /* take next element from rle list everytime an element is finished */
+      if (rlelen <= 0) {
+        if (rle >= rle_limit)
+          break;
+        
+        rlelen = rle->len;
+        clr = rle->color;
+        rle++;
+      }
 
-      clr = rle->color;
-      o   = trans[clr];
-      rlelen = rle->len;
-
-      if (o && mask) {
-        /* threat cases where clipping border is inside rle->len pixels */
-        if ( img_overl->clip_left > x ) {
-          if( img_overl->clip_left < x + rlelen ) {
-            x1_scaled = SCALED_TO_INT( img_overl->clip_left * x_scale );
-            rlelen -= img_overl->clip_left - x;
-            x += img_overl->clip_left - x;
+      if (!mask) {
+        /* above or below clipping area */
+        
+        rle_bite = rlelen;
+        /* choose palette for surrounding area */
+        colors = (clut_t*)img_overl->color;
+        trans = img_overl->trans;
+      } else {
+        /* treat cases where clipping border is inside rle->len pixels */
+        if ( x < img_overl->clip_left ) {
+          /* starts left */
+          if( x + rlelen > img_overl->clip_left ) {
+            /* ends not left */
+
+            /* choose the largest "bite" up to palette change */
+            rle_bite = img_overl->clip_left - x;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           } else {
-            o = 0;
+            /* ends left */
+            
+            rle_bite = rlelen;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           }
-        } else if( clip_right < x + rlelen ) {
-          if( clip_right > x ) {
-            x2_scaled = SCALED_TO_INT( clip_right * x_scale);
-            mem_blend24_32(img_pix + x1_scaled*4, clut[clr].cb,
-                    clut[clr].cr, clut[clr].y,
-                    o, x2_scaled-x1_scaled);
-            o = 0;            
+        } else if( x + rlelen > clip_right ) {
+          /* ends right */
+          if( x < clip_right ) {
+            /* starts not right */
+
+            /* choose the largest "bite" up to palette change */
+            rle_bite = clip_right - x;
+            /* we're in the center area so choose clip palette */
+            colors = (clut_t*)img_overl->clip_color;
+            trans = img_overl->clip_trans;
           } else {
-            o = 0;
+            /* starts right */
+            
+            rle_bite = rlelen;
+            /* choose palette for surrounding area */
+            colors = (clut_t*)img_overl->color;
+            trans = img_overl->trans;
           }
-        } 
+        } else {
+          /* starts not left and ends not right */
+          
+          rle_bite = rlelen;
+          /* we're in the center area so choose clip palette */
+          colors = (clut_t*)img_overl->clip_color;
+          trans = img_overl->clip_trans;
+        }
       }
+      
+      x2_scaled = SCALED_TO_INT((x + rle_bite) * x_scale);
 
-      x2_scaled = SCALED_TO_INT((x + rlelen) * x_scale);
-      if (o && mask) {
-        mem_blend24_32(img_pix + x1_scaled*4, clut[clr].cb,
-                    clut[clr].cr, clut[clr].y,
-                    o, x2_scaled-x1_scaled);
+      o = trans[clr];
+      if (o) {
+        mem_blend32(img_pix + x1_scaled*4, (uint8_t *)&colors[clr], o, x2_scaled-x1_scaled);
       }
 
       x1_scaled = x2_scaled;
-      x += rlelen;
-      rle++;
-      if (rle >= rle_limit) break;
+      x += rle_bite;
+      rlelen -= rle_bite;
     }
 
     img_pix += img_width * 4;
     dy += dy_step;
@@ -429,11 +795,111 @@ static void mem_blend8(uint8_t *mem, uint8_t val, uint8_t o, size_t sz)
     mem++;
   }
 }
 
+static void blend_yuv_exact(uint8_t *dst_cr, uint8_t *dst_cb,
+                            int src_width, int x_odd,
+                            uint8_t *(*blend_yuv_data)[ 3 ][ 2 ])
+{
+  int x;
+  
+  for (x = 0; x < src_width; x += 2) {
+    /* get opacity of the 4 pixels that share chroma */
+    int o00 = (*blend_yuv_data)[ 0 ][ 0 ][ x + 0 ];
+    int o01 = (*blend_yuv_data)[ 0 ][ 0 ][ x + 1 ];
+    int o10 = (*blend_yuv_data)[ 0 ][ 1 ][ x + 0 ];
+    int o11 = (*blend_yuv_data)[ 0 ][ 1 ][ x + 1 ];
+
+    /* are there any pixels a little bit opaque? */
+    if (o00 || o01 || o10 || o11) {
+      /* get the chroma components of the 4 pixels */
+      int cr00 = -128 + (*blend_yuv_data)[ 1 ][ 0 ][ x + 0 ];
+      int cr01 = -128 + (*blend_yuv_data)[ 1 ][ 0 ][ x + 1 ];
+      int cr10 = -128 + (*blend_yuv_data)[ 1 ][ 1 ][ x + 0 ];
+      int cr11 = -128 + (*blend_yuv_data)[ 1 ][ 1 ][ x + 1 ];
+          
+      int cb00 = -128 + (*blend_yuv_data)[ 2 ][ 0 ][ x + 0 ];
+      int cb01 = -128 + (*blend_yuv_data)[ 2 ][ 0 ][ x + 1 ];
+      int cb10 = -128 + (*blend_yuv_data)[ 2 ][ 1 ][ x + 0 ];
+      int cb11 = -128 + (*blend_yuv_data)[ 2 ][ 1 ][ x + 1 ];
+
+      /* are all pixels completely opaque? */
+      if (o00 >= 0xf && o01 >= 0xf && o10 >= 0xf && o11 >= 0xf) {
+        /* set the output chroma to the average of the four pixels */
+        *dst_cr = 128 + (cr00 + cr01 + cr10 + cr11) / 4;
+        *dst_cb = 128 + (cb00 + cb01 + cb10 + cb11) / 4;
+      } else {
+        int t4, cr, cb;
+        
+        /* blending required, so clamp opacity values to allowed range */
+        if (o00 > 0xf) o00 = 0xf;
+        if (o01 > 0xf) o01 = 0xf;
+        if (o10 > 0xf) o10 = 0xf;
+        if (o11 > 0xf) o11 = 0xf;
+
+        /* calculate transparency of background over the four pixels */
+        t4 = (0xf - o00) + (0xf - o01) + (0xf - o10) + (0xf - o11);
+
+        /* get background chroma */
+        cr = -128 + *dst_cr;
+        cb = -128 + *dst_cb;
+
+        /* blend the output chroma to the average of the four pixels */
+        *dst_cr = 128 + (cr * t4 + cr00 * o00 + cr01 * o01 + cr10 * o10 + cr11 * o11) / (4 * 0xf);
+        *dst_cb = 128 + (cb * t4 + cb00 * o00 + cb01 * o01 + cb10 * o10 + cb11 * o11) / (4 * 0xf);
+      }
+    }
+
+    /* next chroma sample */
+    dst_cr++;
+    dst_cb++;
+  }
+}
+
+static uint8_t *(*blend_yuv_grow_extra_data(alphablend_t *extra_data, int osd_width))[ 3 ][ 2 ]
+{
+  struct __attribute__((packed)) header_s {
+    int id;
+    int max_width;
+    uint8_t *data[ 3 ][ 2 ];
+  } **header = (struct header_s **)&extra_data->buffer;
+  
+  int needed_buffer_size = sizeof (**header) + sizeof (uint8_t[ 3 ][ 2 ][ osd_width ]);
+  
+  if (extra_data->buffer_size < needed_buffer_size) {
+    
+    free(extra_data->buffer);
+    extra_data->buffer = xine_xmalloc(needed_buffer_size);
+
+    if (!extra_data->buffer) {
+      extra_data->buffer_size = 0;
+      return 0;
+    }
+    
+    extra_data->buffer_size = needed_buffer_size;
+    (*header)->max_width = 0;
+  }
+
+  if ((*header)->id != ME_FOURCC('y', 'u', 'v', 0) || (*header)->max_width < osd_width) {
+    (*header)->id = ME_FOURCC('y', 'u', 'v', 0);
+    (*header)->max_width = osd_width;
+
+    (*header)->data[ 0 ][ 0 ] = ((uint8_t *)extra_data->buffer) + sizeof (**header);
+    (*header)->data[ 0 ][ 1 ] = (*header)->data[ 0 ][ 0 ] + osd_width;
+    (*header)->data[ 1 ][ 0 ] = (*header)->data[ 0 ][ 1 ] + osd_width;
+    (*header)->data[ 1 ][ 1 ] = (*header)->data[ 1 ][ 0 ] + osd_width;
+    (*header)->data[ 2 ][ 0 ] = (*header)->data[ 1 ][ 1 ] + osd_width;
+    (*header)->data[ 2 ][ 1 ] = (*header)->data[ 2 ][ 0 ] + osd_width;
+  }
+
+  return &(*header)->data;
+}
+  
 void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
-                int dst_width, int dst_height)
+                int dst_width, int dst_height, int dst_pitches[3],
+                alphablend_t *extra_data)
 {
+  int enable_exact_blending = !extra_data->disable_exact_blending;
   clut_t *my_clut;
   uint8_t *my_trans;
 
   int src_width = img_overl->width;
@@ -441,53 +907,91 @@ void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
   rle_elem_t *rle = img_overl->rle;
   rle_elem_t *rle_limit = rle + img_overl->num_rle;
   int x_off = img_overl->x;
   int y_off = img_overl->y;
+  int x_odd = x_off & 1;
+  int y_odd = y_off & 1;
   int ymask,xmask;
   int rle_this_bite;
   int rle_remainder;
   int rlelen;
   int x, y;
   int clip_right;
   uint8_t clr=0;
-
-  uint8_t *dst_y = dst_base[0] + dst_width * y_off + x_off;
-  uint8_t *dst_cr = dst_base[2] +
-    (y_off / 2) * (dst_width / 2) + (x_off / 2) + 1;
-  uint8_t *dst_cb = dst_base[1] +
-    (y_off / 2) * (dst_width / 2) + (x_off / 2) + 1;
+  
+  int any_line_buffered = 0;
+  int exact_blend_width = ((src_width <= (dst_width - x_off)) ? src_width : (dst_width - x_off));
+  int exact_blend_width_m2 = (x_odd + exact_blend_width + 1) & ~1; /* make it a (larger) multiple of 2 */
+  uint8_t *(*blend_yuv_data)[ 3 ][ 2 ] = 0;
+  
+  uint8_t *dst_y = dst_base[0] + dst_pitches[0] * y_off + x_off;
+  uint8_t *dst_cr = dst_base[2] + (y_off / 2) * dst_pitches[1] + (x_off / 2);
+  uint8_t *dst_cb = dst_base[1] + (y_off / 2) * dst_pitches[2] + (x_off / 2);
 #ifdef LOG_BLEND_YUV
   printf("overlay_blend started x=%d, y=%d, w=%d h=%d\n",img_overl->x,img_overl->y,img_overl->width,img_overl->height);
 #endif
   my_clut = (clut_t*) img_overl->clip_color;
   my_trans = img_overl->clip_trans;
 
   /* avoid wraping overlay if drawing to small image */
-  if( (x_off + img_overl->clip_right) < dst_width )
+  if( (x_off + img_overl->clip_right) <= dst_width )
     clip_right = img_overl->clip_right;
   else
-    clip_right = dst_width - 1 - x_off;
+    clip_right = dst_width - x_off;
 
   /* avoid buffer overflow */
-  if( (src_height + y_off) >= dst_height )
-    src_height = dst_height - 1 - y_off;
+  if( (src_height + y_off) > dst_height )
+    src_height = dst_height - y_off;
 
+  if (src_height <= 0)
+    return;
+
+  if (enable_exact_blending) {
+    if (exact_blend_width <= 0)
+      return;
+  
+    blend_yuv_data = blend_yuv_grow_extra_data(extra_data, exact_blend_width_m2);
+    if (!blend_yuv_data)
+      return;
+    
+    /* make linebuffer transparent */
+    memset(&(*blend_yuv_data)[ 0 ][ 0 ][ 0 ], 0, exact_blend_width_m2);
+    memset(&(*blend_yuv_data)[ 0 ][ 1 ][ 0 ], 0, exact_blend_width_m2);
+  }
+  
   rlelen=rle_remainder=0;
   for (y = 0; y < src_height; y++) {
-    ymask = ((img_overl->clip_top > y) || (img_overl->clip_bottom < y));
+    if (rle >= rle_limit) {
+#ifdef LOG_BLEND_YUV
+      printf("y-rle_limit\n");
+#endif
+      break;
+    }
+    
+    ymask = ((y < img_overl->clip_top) || (y >= img_overl->clip_bottom));
     xmask = 0;
 #ifdef LOG_BLEND_YUV
     printf("X started ymask=%d y=%d src_height=%d\n",ymask, y, src_height);
 #endif
 
     for (x = 0; x < src_width;) {
       uint16_t o;
+      
+      if (rle >= rle_limit) {
+#ifdef LOG_BLEND_YUV
+        printf("x-rle_limit\n");
+#endif
+        break;
+      }
+      
 #ifdef LOG_BLEND_YUV
       printf("1:rle_len=%d, remainder=%d, x=%d\n",rlelen, rle_remainder, x);
 #endif
 
       if ((rlelen < 0) || (rle_remainder < 0)) {
+#ifdef LOG_BLEND_YUV
         printf("alphablend: major bug in blend_yuv < 0\n");
+#endif
       } 
       if (rlelen == 0) {
         rle_remainder = rlelen = rle->len;
         clr = rle->color;
@@ -497,24 +1001,23 @@ void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
         rle_remainder = rlelen;
       }
       if ((rle_remainder + x) > src_width) {
         /* Do something for long rlelengths */
-        rle_remainder = src_width - x; 
-        ;
+        rle_remainder = src_width - x;
       }
 #ifdef LOG_BLEND_YUV
       printf("2:rle_len=%d, remainder=%d, x=%d\n",rlelen, rle_remainder, x);
 #endif
 
       if (ymask == 0) {
-        if (x <= img_overl->clip_left) { 
+        if (x < img_overl->clip_left) { 
           /* Starts outside clip area */
-          if ((x + rle_remainder - 1) > img_overl->clip_left ) {
+          if ((x + rle_remainder) > img_overl->clip_left ) {
 #ifdef LOG_BLEND_YUV
             printf("Outside clip left %d, ending inside\n", img_overl->clip_left); 
 #endif
             /* Cutting needed, starts outside, ends inside */
-            rle_this_bite = (img_overl->clip_left - x + 1);
+            rle_this_bite = (img_overl->clip_left - x);
             rle_remainder -= rle_this_bite;
             rlelen -= rle_this_bite;
             my_clut = (clut_t*) img_overl->color;
             my_trans = img_overl->trans;
@@ -556,9 +1059,9 @@ void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
             my_trans = img_overl->clip_trans;
             xmask++;
           }
         } else if (x >= clip_right) {
-          /* Starts outside clip area, ends outsite clip area */
+          /* Starts outside clip area, ends outside clip area */
           if ((x + rle_remainder ) > src_width ) { 
 #ifdef LOG_BLEND_YUV
             printf("Outside clip right %d, ending eol\n", clip_right);
 #endif
@@ -596,58 +1099,189 @@ void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
       o   = my_trans[clr];
 #ifdef LOG_BLEND_YUV
       printf("Trans=%d clr=%d xmask=%d my_clut[clr]=%d\n",o, clr, xmask, my_clut[clr].y);
 #endif
-      if (o) {
-        if(o >= 15) {
-          memset(dst_y + x, my_clut[clr].y, rle_this_bite);
-          if (y & 1) {
-            memset(dst_cr + (x >> 1), my_clut[clr].cr, (rle_this_bite+1) >> 1);
-            memset(dst_cb + (x >> 1), my_clut[clr].cb, (rle_this_bite+1) >> 1);
-          }
-        } else {
-          mem_blend8(dst_y + x, my_clut[clr].y, o, rle_this_bite);
-          if (y & 1) {
-            /* Blending cr and cb should use a different function, with pre -128 to each sample */
-            mem_blend8(dst_cr + (x >> 1), my_clut[clr].cr, o, (rle_this_bite+1) >> 1);
-            mem_blend8(dst_cb + (x >> 1), my_clut[clr].cb, o, (rle_this_bite+1) >> 1);
-          }
+
+      if (x < (dst_width - x_off)) {
+        /* clip against right edge of destination area */
+        if ((x + rle_this_bite) > (dst_width - x_off)) {
+          int toClip = (x + rle_this_bite) - (dst_width - x_off);
+          
+          rle_this_bite -= toClip;
+          rle_remainder += toClip;
+          rlelen += toClip;
         }
 
+        if (enable_exact_blending) {
+          /* remember opacity of current line */
+          memset(&(*blend_yuv_data)[ 0 ][ (y + y_odd) & 1 ][ x + x_odd ], o, rle_this_bite);
+          any_line_buffered |= ((y + y_odd) & 1) ? 2 : 1;
+        }
+      
+        if (o) {
+          if(o >= 15) {
+            memset(dst_y + x, my_clut[clr].y, rle_this_bite);
+            if (!enable_exact_blending) {
+              if ((y + y_odd) & 1) {
+                memset(dst_cr + ((x + x_odd) >> 1), my_clut[clr].cr, (rle_this_bite+1) >> 1);
+                memset(dst_cb + ((x + x_odd) >> 1), my_clut[clr].cb, (rle_this_bite+1) >> 1);
+              }
+            }
+          } else {
+            mem_blend8(dst_y + x, my_clut[clr].y, o, rle_this_bite);
+            if (!enable_exact_blending) {
+              if ((y + y_odd) & 1) {
+                /* Blending cr and cb should use a different function, with pre -128 to each sample */
+                mem_blend8(dst_cr + ((x + x_odd) >> 1), my_clut[clr].cr, o, (rle_this_bite+1) >> 1);
+                mem_blend8(dst_cb + ((x + x_odd) >> 1), my_clut[clr].cb, o, (rle_this_bite+1) >> 1);
+              }
+            }
+          }
+
+          if (enable_exact_blending) {
+            /* remember chroma of current line */
+            memset(&(*blend_yuv_data)[ 1 ][ (y + y_odd) & 1 ][ x + x_odd ], my_clut[ clr ].cr, rle_this_bite);
+            memset(&(*blend_yuv_data)[ 2 ][ (y + y_odd) & 1 ][ x + x_odd ], my_clut[ clr ].cb, rle_this_bite);
+          }
+        }
       }
 #ifdef LOG_BLEND_YUV
       printf("rle_this_bite=%d, remainder=%d, x=%d\n",rle_this_bite, rle_remainder, x);
 #endif
       x += rle_this_bite;
-      if (rle >= rle_limit) {
-#ifdef LOG_BLEND_YUV
-        printf("x-rle_limit\n");
-#endif
-        break;
-      }
-    }
-    if (rle >= rle_limit) {
-#ifdef LOG_BLEND_YUV
-        printf("x-rle_limit\n");
-#endif
-        break;
     }
 
-    dst_y += dst_width;
+    if ((y + y_odd) & 1) {
+      if (enable_exact_blending) {
+        /* blend buffered lines */
+        if (any_line_buffered) {
+          if (!(any_line_buffered & 2)) {
+            /* make second line transparent */
+            memset(&(*blend_yuv_data)[ 0 ][ 1 ][ 0 ], 0, exact_blend_width_m2);
+          }
+          
+          blend_yuv_exact(dst_cr, dst_cb, exact_blend_width, x_odd, blend_yuv_data);
+          
+          any_line_buffered = 0;
+        }
+      }
+      
+      dst_cr += dst_pitches[2];
+      dst_cb += dst_pitches[1];
+    }
+    
+    dst_y += dst_pitches[0];
+  }
 
-    if (y & 1) {
-      dst_cr += (dst_width + 1) / 2;
-      dst_cb += (dst_width + 1) / 2;
+  if (enable_exact_blending) {
+    /* blend buffered lines */
+    if (any_line_buffered) {
+      if (!(any_line_buffered & 2)) {
+        /* make second line transparent */
+        memset(&(*blend_yuv_data)[ 0 ][ 1 ][ 0 ], 0, exact_blend_width_m2);
+      }
+      
+      blend_yuv_exact(dst_cr, dst_cb, exact_blend_width, x_odd, blend_yuv_data);
     }
   }
+      
 #ifdef LOG_BLEND_YUV
   printf("overlay_blend ended\n");
 #endif
 }
             
+static void blend_yuy2_exact(uint8_t *dst_cr, uint8_t *dst_cb,
+                             int src_width, int x_odd,
+                             uint8_t *(*blend_yuy2_data)[ 3 ])
+{
+  int x;
+  
+  for (x = 0; x < src_width; x += 2) {
+    /* get opacity of the 2 pixels that share chroma */
+    int o0 = (*blend_yuy2_data)[ 0 ][ x + 0 ];
+    int o1 = (*blend_yuy2_data)[ 0 ][ x + 1 ];
+
+    /* are there any pixels a little bit opaque? */
+    if (o0 || o1) {
+      /* get the chroma components of the 2 pixels */
+      int cr0 = -128 + (*blend_yuy2_data)[ 1 ][ x + 0 ];
+      int cr1 = -128 + (*blend_yuy2_data)[ 1 ][ x + 1 ];
+          
+      int cb0 = -128 + (*blend_yuy2_data)[ 2 ][ x + 0 ];
+      int cb1 = -128 + (*blend_yuy2_data)[ 2 ][ x + 1 ];
+
+      /* are all pixels completely opaque? */
+      if (o0 >= 0xf && o1 >= 0xf) {
+        /* set the output chroma to the average of the two pixels */
+        *dst_cr = 128 + (cr0 + cr1) / 2;
+        *dst_cb = 128 + (cb0 + cb1) / 2;
+      } else {
+        int t2, cr, cb;
+        
+        /* blending required, so clamp opacity values to allowed range */
+        if (o0 > 0xf) o0 = 0xf;
+        if (o1 > 0xf) o1 = 0xf;
+
+        /* calculate transparency of background over the two pixels */
+        t2 = (0xf - o0) + (0xf - o1);
+
+        /* get background chroma */
+        cr = -128 + *dst_cr;
+        cb = -128 + *dst_cb;
+
+        /* blend the output chroma to the average of the two pixels */
+        *dst_cr = 128 + (cr * t2 + cr0 * o0 + cr1 * o1) / (2 * 0xf);
+        *dst_cb = 128 + (cb * t2 + cb0 * o0 + cb1 * o1) / (2 * 0xf);
+      }
+    }
+
+    /* next chroma sample */
+    dst_cr += 4;
+    dst_cb += 4;
+  }
+}
+
+static uint8_t *(*blend_yuy2_grow_extra_data(alphablend_t *extra_data, int osd_width))[ 3 ]
+{
+  struct __attribute__((packed)) header_s {
+    int id;
+    int max_width;
+    uint8_t *data[ 3 ];
+  } **header = (struct header_s **)&extra_data->buffer;
+  
+  int needed_buffer_size = sizeof (**header) + sizeof (uint8_t[ 3 ][ osd_width ]);
+  
+  if (extra_data->buffer_size < needed_buffer_size) {
+    
+    free(extra_data->buffer);
+    extra_data->buffer = xine_xmalloc(needed_buffer_size);
+
+    if (!extra_data->buffer) {
+      extra_data->buffer_size = 0;
+      return 0;
+    }
+    
+    extra_data->buffer_size = needed_buffer_size;
+    (*header)->max_width = 0;
+  }
+
+  if ((*header)->id != ME_FOURCC('y', 'u', 'y', '2') || (*header)->max_width < osd_width) {
+    (*header)->id = ME_FOURCC('y', 'u', 'y', '2');
+    (*header)->max_width = osd_width;
+
+    (*header)->data[ 0 ] = ((uint8_t *)extra_data->buffer) + sizeof (**header);
+    (*header)->data[ 1 ] = (*header)->data[ 0 ] + osd_width;
+    (*header)->data[ 2 ] = (*header)->data[ 1 ] + osd_width;
+  }
+
+  return &(*header)->data;
+}
+
 void blend_yuy2 (uint8_t * dst_img, vo_overlay_t * img_overl,
-                int dst_width, int dst_height)
+                 int dst_width, int dst_height, int dst_pitch,
+                 alphablend_t *extra_data)
 {
+  int enable_exact_blending = !extra_data->disable_exact_blending;
   clut_t *my_clut;
   uint8_t *my_trans;
 
   int src_width = img_overl->width;
@@ -655,101 +1289,496 @@ void blend_yuy2 (uint8_t * dst_img, vo_overlay_t * img_overl,
   rle_elem_t *rle = img_overl->rle;
   rle_elem_t *rle_limit = rle + img_overl->num_rle;
   int x_off = img_overl->x;
   int y_off = img_overl->y;
-  int mask;
+  int x_odd = x_off & 1;
+  int ymask;
+  int rle_this_bite;
+  int rle_remainder;
+  int rlelen;
   int x, y;
-  int l;
+  int l = 0;
   int clip_right;
-  uint32_t yuy2;
+
+  union {
+    uint32_t value;
+    uint8_t  b[4];
+    uint16_t h[2];
+  } yuy2;
+
+  uint8_t clr = 0;
+
+  int any_line_buffered = 0;
+  int exact_blend_width = ((src_width <= (dst_width - x_off)) ? src_width : (dst_width - x_off));
+  int exact_blend_width_m2 = (x_odd + exact_blend_width + 1) & ~1; /* make it a (larger) multiple of 2 */
+  uint8_t *(*blend_yuy2_data)[ 3 ] = 0;
   
-  uint8_t *dst_y = dst_img + 2 * (dst_width * y_off + x_off);
+  uint8_t *dst_y = dst_img + dst_pitch * y_off + 2 * x_off;
   uint8_t *dst;
 
   my_clut = (clut_t*) img_overl->clip_color;
   my_trans = img_overl->clip_trans;
 
   /* avoid wraping overlay if drawing to small image */
-  if( (x_off + img_overl->clip_right) < dst_width )
+  if( (x_off + img_overl->clip_right) <= dst_width )
     clip_right = img_overl->clip_right;
   else
-    clip_right = dst_width - 1 - x_off;
+    clip_right = dst_width - x_off;
 
   /* avoid buffer overflow */
-  if( (src_height + y_off) >= dst_height )
-    src_height = dst_height - 1 - y_off;
+  if( (src_height + y_off) > dst_height )
+    src_height = dst_height - y_off;
+
+  if (src_height <= 0)
+    return;
+
+  if (enable_exact_blending) {
+    if (exact_blend_width <= 0)
+      return;
+    
+    blend_yuy2_data = blend_yuy2_grow_extra_data(extra_data, exact_blend_width_m2);
+    if (!blend_yuy2_data)
+      return;
+    
+    /* make linebuffer transparent */
+    memset(&(*blend_yuy2_data)[ 0 ][ 0 ], 0, exact_blend_width_m2);
+  }
 
+  rlelen=rle_remainder=0;
   for (y = 0; y < src_height; y++) {
-    mask = !(img_overl->clip_top > y || img_overl->clip_bottom < y);
+    if (rle >= rle_limit)
+      break;
+    
+    ymask = ((y < img_overl->clip_top) || (y >= img_overl->clip_bottom));
 
     dst = dst_y;
     for (x = 0; x < src_width;) {
-      uint8_t clr;
       uint16_t o;
-      int rlelen;
 
-      clr = rle->color;
-      o   = my_trans[clr];
-      rlelen = rle->len;
-
-      if (o && mask) {
-        /* threat cases where clipping border is inside rle->len pixels */
-        if ( img_overl->clip_left > x ) {
-          if( img_overl->clip_left < x + rlelen ) {
-            rlelen -= img_overl->clip_left - x;
-            x += img_overl->clip_left - x;
+      if (rle >= rle_limit)
+        break;
+    
+      if ((rlelen < 0) || (rle_remainder < 0)) {
+#ifdef LOG_BLEND_YUV
+        printf("alphablend: major bug in blend_yuv < 0\n");
+#endif
+      } 
+      if (rlelen == 0) {
+        rle_remainder = rlelen = rle->len;
+        clr = rle->color;
+        rle++;
+      }
+      if (rle_remainder == 0) {
+        rle_remainder = rlelen;
+      }
+      if ((rle_remainder + x) > src_width) {
+        /* Do something for long rlelengths */
+        rle_remainder = src_width - x;
+      }
+#ifdef LOG_BLEND_YUV
+      printf("2:rle_len=%d, remainder=%d, x=%d\n",rlelen, rle_remainder, x);
+#endif
+
+      if (ymask == 0) {
+        if (x < img_overl->clip_left) { 
+          /* Starts outside clip area */
+          if ((x + rle_remainder) > img_overl->clip_left ) {
+#ifdef LOG_BLEND_YUV
+            printf("Outside clip left %d, ending inside\n", img_overl->clip_left); 
+#endif
+            /* Cutting needed, starts outside, ends inside */
+            rle_this_bite = (img_overl->clip_left - x);
+            rle_remainder -= rle_this_bite;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->color;
+            my_trans = img_overl->trans;
           } else {
-            o = 0;
+#ifdef LOG_BLEND_YUV
+            printf("Outside clip left %d, ending outside\n", img_overl->clip_left); 
+#endif
+          /* no cutting needed, starts outside, ends outside */
+            rle_this_bite = rle_remainder;
+            rle_remainder = 0;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->color;
+            my_trans = img_overl->trans;
           }
-        } else if( clip_right < x + rlelen ) {
-          if( clip_right > x ) {
-            /* fixme: case not implemented */
-            o = 0;            
+        } else if (x < clip_right) {
+          /* Starts inside clip area */
+          if ((x + rle_remainder) > clip_right ) {
+#ifdef LOG_BLEND_YUV
+            printf("Inside clip right %d, ending outside\n", clip_right);
+#endif
+            /* Cutting needed, starts inside, ends outside */
+            rle_this_bite = (clip_right - x);
+            rle_remainder -= rle_this_bite;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->clip_color;
+            my_trans = img_overl->clip_trans;
           } else {
-            o = 0;
+#ifdef LOG_BLEND_YUV
+            printf("Inside clip right %d, ending inside\n", clip_right);
+#endif
+          /* no cutting needed, starts inside, ends inside */
+            rle_this_bite = rle_remainder;
+            rle_remainder = 0;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->clip_color;
+            my_trans = img_overl->clip_trans;
           }
-        } 
+        } else if (x >= clip_right) {
+          /* Starts outside clip area, ends outsite clip area */
+          if ((x + rle_remainder ) > src_width ) { 
+#ifdef LOG_BLEND_YUV
+            printf("Outside clip right %d, ending eol\n", clip_right);
+#endif
+            /* Cutting needed, starts outside, ends at right edge */
+            /* It should never reach here due to the earlier test of src_width */
+            rle_this_bite = (src_width - x );
+            rle_remainder -= rle_this_bite;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->color;
+            my_trans = img_overl->trans;
+          } else {
+          /* no cutting needed, starts outside, ends outside */
+#ifdef LOG_BLEND_YUV
+            printf("Outside clip right %d, ending outside\n", clip_right);
+#endif
+            rle_this_bite = rle_remainder;
+            rle_remainder = 0;
+            rlelen -= rle_this_bite;
+            my_clut = (clut_t*) img_overl->color;
+            my_trans = img_overl->trans;
+          }
+        }
+      } else {
+        /* Outside clip are due to y */
+        /* no cutting needed, starts outside, ends outside */
+        rle_this_bite = rle_remainder;
+        rle_remainder = 0;
+        rlelen -= rle_this_bite;
+        my_clut = (clut_t*) img_overl->color;
+        my_trans = img_overl->trans;
       }
-      
+      o   = my_trans[clr];
+
+      if (x < (dst_width - x_off)) {
+        /* clip against right edge of destination area */
+        if ((x + rle_this_bite) > (dst_width - x_off)) {
+          int toClip = (x + rle_this_bite) - (dst_width - x_off);
+          
+          rle_this_bite -= toClip;
+          rle_remainder += toClip;
+          rlelen += toClip;
+        }
+
+        if (enable_exact_blending) {
+          /* remember opacity of current line */
+          memset(&(*blend_yuy2_data)[ 0 ][ x + x_odd ], o, rle_this_bite);
+          any_line_buffered = 1;
+        }
+
+        if (o) {
+          if (!enable_exact_blending) {
+            l = rle_this_bite>>1;
+            if( !((x_odd+x) & 1) ) {
+              yuy2.b[0] = my_clut[clr].y;
+              yuy2.b[1] = my_clut[clr].cb;
+              yuy2.b[2] = my_clut[clr].y;
+              yuy2.b[3] = my_clut[clr].cr;
+            } else {
+              yuy2.b[0] = my_clut[clr].y;
+              yuy2.b[1] = my_clut[clr].cr;
+              yuy2.b[2] = my_clut[clr].y;
+              yuy2.b[3] = my_clut[clr].cb;
+            }
+          }
+          if (o >= 15) {
+            if (!enable_exact_blending) {
+              while(l--) {
+                *(uint16_t *)dst = yuy2.h[0];
+                dst += 2;
+                *(uint16_t *)dst = yuy2.h[1];
+                dst += 2;
+              }
+              if(rle_this_bite & 1) {
+                *(uint16_t *)dst = yuy2.h[0];
+                dst += 2;
+              }
+            } else {
+              l = rle_this_bite;
+              while (l--) {
+                *dst = my_clut[clr].y;
+                dst += 2;
+              }
+            }
+          } else {
+            if (!enable_exact_blending) {
+              if( l ) {
+                mem_blend32(dst, &yuy2.b[0], o, l);
+                dst += 4*l;
+              }
+              
+              if(rle_this_bite & 1) {
+                *dst = BLEND_BYTE(*dst, yuy2.b[0], o);
+                dst++;
+                *dst = BLEND_BYTE(*dst, yuy2.b[1], o);
+                dst++;
+              }
+            } else {
+              l = rle_this_bite;
+              while (l--) {
+                *dst = BLEND_BYTE(*dst, my_clut[clr].y, o);
+                dst += 2;
+              }
+            }
+          }
 
-      if (o && mask) {
-        l = rlelen>>1;
-        if( !(x & 1) ) {
-          yuy2 =  my_clut[clr].y + (my_clut[clr].cb << 8) +
-                 (my_clut[clr].y << 16) + (my_clut[clr].cr << 24);
+          if (enable_exact_blending) {
+            /* remember chroma of current line */
+            memset(&(*blend_yuy2_data)[ 1 ][ x + x_odd ], my_clut[ clr ].cr, rle_this_bite);
+            memset(&(*blend_yuy2_data)[ 2 ][ x + x_odd ], my_clut[ clr ].cb, rle_this_bite);
+          }
         } else {
-          yuy2 =  my_clut[clr].y + (my_clut[clr].cr << 8) +
-                 (my_clut[clr].y << 16) + (my_clut[clr].cb << 24);
+          dst += rle_this_bite*2;
         }
+      }
+      
+      x += rle_this_bite;
+    }
+    
+    if (enable_exact_blending) {
+      /* blend buffered line */
+      if (any_line_buffered) {
+        blend_yuy2_exact(dst_y - x_odd * 2 + 3, dst_y - x_odd * 2 + 1, exact_blend_width, x_odd, blend_yuy2_data);
         
-        if (o >= 15) {
-          while(l--) {
-            *((uint32_t *)dst)++ = yuy2;
+        any_line_buffered = 0;
+      }
+    }
+
+    dst_y += dst_pitch;
+  }
+}
+
+void clear_xx44_palette(xx44_palette_t *p) 
+{
+  register int i;
+  register uint32_t *cluts = p->cluts;
+  register int *ids = p->lookup_cache;
+
+  i= p->size;
+  while(i--)
+    *cluts++ = 0;
+  i = 2*OVL_PALETTE_SIZE;
+  while(i--)
+    *ids++ = -1;
+  p->max_used=1; 
+}
+
+void init_xx44_palette(xx44_palette_t *p, unsigned num_entries) 
+{
+  p->size = (num_entries > XX44_PALETTE_SIZE) ? XX44_PALETTE_SIZE : num_entries; 
+}
+
+void dispose_xx44_palette(xx44_palette_t *p) 
+{
+}
+
+static void colorToPalette(const uint32_t *icolor, unsigned char *palette_p,
+                           unsigned num_xvmc_components, char *xvmc_components) 
+{
+  const clut_t *color = (const clut_t *) icolor;
+  int i;
+  for (i=0; i<num_xvmc_components; ++i) {
+    switch(xvmc_components[i]) {
+    case 'V': *palette_p = color->cr; break;
+    case 'U': *palette_p = color->cb; break;
+    case 'Y': 
+    default:  *palette_p = color->y; break;
+    }
+    *palette_p++;
+  }
+}
+
+
+void xx44_to_xvmc_palette(const xx44_palette_t *p,unsigned char *xvmc_palette,
+                          unsigned first_xx44_entry, unsigned num_xx44_entries,
+                          unsigned num_xvmc_components, char *xvmc_components) 
+{
+  register int i;
+  register const uint32_t *cluts = p->cluts + first_xx44_entry;
+
+  for (i=0; i<num_xx44_entries; ++i) {
+    if ((cluts - p->cluts) < p->size) {
+      colorToPalette(cluts++, xvmc_palette, num_xvmc_components, xvmc_components);
+      xvmc_palette += num_xvmc_components;
+    }
+  }
+} 
+
+static int xx44_paletteIndex(xx44_palette_t *p, int color, uint32_t clut) 
+{
+
+  register int i;
+  register uint32_t *cluts = p->cluts;
+  register int tmp;
+
+  if ((tmp = p->lookup_cache[color]) >= 0) 
+    if (cluts[tmp] == clut) return tmp;
+
+  for (i=0; i<p->max_used; ++i) {
+    if (*cluts++ == clut) return p->lookup_cache[color] = i;
+  }
+
+  if (p->max_used == p->size -1) {
+    printf("video_out: Warning! Out of xx44 palette colors!\n");
+    return 1;
+  }
+  p->cluts[p->max_used] = clut;
+  return p->lookup_cache[color] = p->max_used++;
+}
+
+static void memblend_xx44(uint8_t *mem,uint8_t val, register size_t size, uint8_t mask)
+{
+  register uint8_t masked_val = val & mask;
+
+  while(size--) {
+    if ((*mem & mask) <= masked_val ) *mem = val;
+    mem++;
+  }
+}
+
+void blend_xx44 (uint8_t *dst_img, vo_overlay_t *img_overl,
+                int dst_width, int dst_height, int dst_pitch, 
+                alphablend_t *extra_data,
+                xx44_palette_t *palette,int ia44) 
+{
+  int  src_width  = img_overl->width;
+  int  src_height = img_overl->height;
+  rle_elem_t *rle        = img_overl->rle;
+  rle_elem_t *rle_limit  = rle + img_overl->num_rle;
+  int mask;
+  int x_off = img_overl->x;
+  int y_off = img_overl->y;
+  int x, y;
+  uint8_t norm_pixel,clip_pixel;  
+  uint8_t *dst_y;
+  uint8_t *dst;
+  uint8_t alphamask = (ia44) ? 0x0F : 0xF0;
+  int clip_right;
+
+  if (!img_overl)
+    return;
+
+  dst_y = dst_img + dst_pitch*y_off + x_off;
+
+  if( (x_off + img_overl->width) <= dst_width )
+    clip_right = img_overl->width;
+  else
+    clip_right = dst_width - x_off;
+
+  if ((src_height + y_off) > dst_height)
+    src_height = dst_height - y_off;
+
+  for (y = 0; y < src_height; y++) {
+
+    mask = !(y < img_overl->clip_top || y >= img_overl->clip_bottom);
+    dst = dst_y;
+
+    for (x = 0; x < src_width;) {
+      int len = (x + rle->len > clip_right) ? clip_right - x : rle->len;
+
+      if (len > 0) {
+        norm_pixel = (uint8_t)((xx44_paletteIndex(palette,rle->color,
+                                             img_overl->color[rle->color]) << 4) | 
+                               (img_overl->trans[rle->color] & 0x0F));
+        clip_pixel = (uint8_t)((xx44_paletteIndex(palette,rle->color+OVL_PALETTE_SIZE,
+                                             img_overl->clip_color[rle->color]) << 4) | 
+                               (img_overl->clip_trans[rle->color] & 0x0F));
+        if (!ia44) {
+          norm_pixel = ((norm_pixel & 0x0F) << 4) | ((norm_pixel & 0xF0) >> 4);
+          clip_pixel = ((clip_pixel & 0x0F) << 4) | ((clip_pixel & 0xF0) >> 4);
+        }
+        if (mask) {
+          if (x < img_overl->clip_left) {
+            if (x + len <= img_overl->clip_left) {
+              memblend_xx44(dst,norm_pixel,len, alphamask);
+              dst += len;
+            } else {
+              memblend_xx44(dst,norm_pixel,img_overl->clip_left -x, alphamask);
+              dst += img_overl->clip_left - x;
+              len -= img_overl->clip_left - x;
+              if (len <= img_overl->clip_right - img_overl->clip_left) {
+                memblend_xx44(dst,clip_pixel,len, alphamask);
+                dst += len;
+              } else {
+                memblend_xx44(dst,clip_pixel,img_overl->clip_right - img_overl->clip_left,
+                              alphamask);
+                dst += img_overl->clip_right - img_overl->clip_left;
+                len -= img_overl->clip_right - img_overl->clip_left;
+                memblend_xx44(dst,norm_pixel,len, alphamask);
+                dst += len;
+              }
+            }
+          } else if (x < img_overl->clip_right) {  
+            if (len <= img_overl->clip_right - x) {
+              memblend_xx44(dst,clip_pixel,len, alphamask);
+              dst += len;
+            } else {
+              memblend_xx44(dst,clip_pixel,img_overl->clip_right - x,alphamask);
+              dst += img_overl->clip_right - x;
+              len -= img_overl->clip_right - x;
+              memblend_xx44(dst,norm_pixel,len, alphamask);
+              dst += len;
+            }
+          } else {
+            memblend_xx44(dst,norm_pixel,len, alphamask);
+            dst += len;
           }
-          if(rlelen & 1)
-            *((uint16_t *)dst)++ = yuy2 & 0xffff;
         } else {
-          if( l ) {
-            mem_blend32(dst, (uint8_t *)&yuy2, o, l);
-            dst += 4*l;
-          }
-          
-          if(rlelen & 1) {
-            *dst = BLEND_BYTE(*dst, *((uint8_t *)&yuy2), o);
-            dst++;
-            *dst = BLEND_BYTE(*dst, *((uint8_t *)&yuy2+1), o);
-            dst++;
-          }
+          memblend_xx44(dst,norm_pixel,len, alphamask);
+          dst += len;
         }
-      } else {
-        dst += rlelen*2;
       }
-
-      x += rlelen;
+      x += rle->len;
       rle++;
       if (rle >= rle_limit) break;
-    }
+    }  
     if (rle >= rle_limit) break;
+    dst_y += dst_pitch;
+  }
+}
+
+static void alphablend_disable_exact_osd_alpha_blending_changed(void *user_data, xine_cfg_entry_t *entry)
+{
+  alphablend_t *extra_data = (alphablend_t *)user_data;
+
+  extra_data->disable_exact_blending = entry->num_value;
+}
 
-    dst_y += dst_width*2;
+void _x_alphablend_init(alphablend_t *extra_data, xine_t *xine)
+{
+  config_values_t *config = xine->config;
+  
+  extra_data->buffer = 0;
+  extra_data->buffer_size = 0;
+
+  extra_data->disable_exact_blending =
+    config->register_bool(config, "video.output.disable_exact_alphablend", 0,
+      _("disable exact alpha blending of overlays"),
+      _("If you experience a performance impact when an On Screen Display or other "
+        "overlays like DVD subtitles are active, then you might want to enable this option.\n"
+        "The result is that alpha blending of overlays is less accurate than before, "
+        "but the CPU usage will be decreased as well."),
+      10, alphablend_disable_exact_osd_alpha_blending_changed, extra_data);
+}
+
+void _x_alphablend_free(alphablend_t *extra_data)
+{
+  if (extra_data->buffer) {
+    free(extra_data->buffer);
+    extra_data->buffer = NULL;
   }
+
+  extra_data->buffer_size = 0;
 }
+
diff --git a/noncore/multimedia/opieplayer2/alphablend.h b/noncore/multimedia/opieplayer2/alphablend.h
index 7230f41..1edb5ba 100644
--- a/noncore/multimedia/opieplayer2/alphablend.h
+++ b/noncore/multimedia/opieplayer2/alphablend.h
@@ -26,32 +26,109 @@
 #define __ALPHABLEND_H__
 
 #include <xine/video_out.h>
 
+typedef struct {
+  void *buffer;
+  int buffer_size;
+
+  int disable_exact_blending;
+} alphablend_t;
+
+void _x_alphablend_init(alphablend_t *extra_data, xine_t *xine);
+void _x_alphablend_free(alphablend_t *extra_data);
+
+/* _MSC_VER port changes */
+#undef ATTRIBUTE_PACKED
+#undef PRAGMA_PACK_BEGIN 
+#undef PRAGMA_PACK_END
+
+#if __GNUC__ > 2 || (__GNUC__ == 2 && __GNUC_MINOR__ >= 95) || defined(__ICC)
+#define ATTRIBUTE_PACKED __attribute__ ((packed))
+#define PRAGMA_PACK 0
+#endif
+
+#if !defined(ATTRIBUTE_PACKED)
+#define ATTRIBUTE_PACKED
+#define PRAGMA_PACK 1
+#endif
+
+#if PRAGMA_PACK
+#pragma pack(8)
+#endif
+
 typedef struct {         /* CLUT == Color LookUp Table */
   uint8_t cb    : 8;
   uint8_t cr    : 8;
   uint8_t y     : 8;
   uint8_t foo   : 8;
-} __attribute__ ((packed)) clut_t;
+} ATTRIBUTE_PACKED clut_t;
+
+
+#if PRAGMA_PACK
+#pragma pack()
+#endif
+
+#define XX44_PALETTE_SIZE 32
+
+typedef struct {
+  unsigned size;
+  unsigned max_used;
+  uint32_t cluts[XX44_PALETTE_SIZE];
+  /* cache palette entries for both colors and clip_colors */
+  int lookup_cache[OVL_PALETTE_SIZE*2];
+} xx44_palette_t;
+
 
 void blend_rgb16 (uint8_t * img, vo_overlay_t * img_overl,
                   int img_width, int img_height,
-                  int dst_width, int dst_height);
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data);
 
 void blend_rgb24 (uint8_t * img, vo_overlay_t * img_overl,
                   int img_width, int img_height,
-                  int dst_width, int dst_height);
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data);
 
 void blend_rgb32 (uint8_t * img, vo_overlay_t * img_overl,
                   int img_width, int img_height,
-                  int dst_width, int dst_height);
+                  int dst_width, int dst_height,
+                  alphablend_t *extra_data);
 
 void blend_yuv (uint8_t *dst_base[3], vo_overlay_t * img_overl,
-                int dst_width, int dst_height);
+                int dst_width, int dst_height, int dst_pitches[3],
+                alphablend_t *extra_data);
 
 void blend_yuy2 (uint8_t * dst_img, vo_overlay_t * img_overl,
-                int dst_width, int dst_height);
+                 int dst_width, int dst_height, int dst_pitch,
+                 alphablend_t *extra_data);
+
+/*
+ * This function isn't too smart about blending. We want to avoid creating new
+ * colors in the palette as a result from two non-zero colors needed to be
+ * blended. Instead we choose the color with the highest alpha value to be
+ * visible. Some parts of the code taken from the "VeXP" project.
+ */
+
+void blend_xx44 (uint8_t *dst_img, vo_overlay_t *img_overl,
+                int dst_width, int dst_height, int dst_pitch, 
+                alphablend_t *extra_data,
+                xx44_palette_t *palette,int ia44);
+
+/*
+ * Functions to handle the xine-specific palette.
+ */
+
+void clear_xx44_palette(xx44_palette_t *p);
+void init_xx44_palette(xx44_palette_t *p, unsigned num_entries);
+void dispose_xx44_palette(xx44_palette_t *p);
+
+/*
+ * Convert the xine-specific palette to something useful.
+ */
+
+void xx44_to_xvmc_palette(const xx44_palette_t *p,unsigned char *xvmc_palette,
+                          unsigned first_xx44_entry, unsigned num_xx44_entries,
+                          unsigned num_xvmc_components, char *xvmc_components);
 
-void crop_overlay (vo_overlay_t * overlay);
 
 #endif
diff --git a/noncore/multimedia/opieplayer2/lib.cpp b/noncore/multimedia/opieplayer2/lib.cpp
index 4ae8490..84d28ce 100644
--- a/noncore/multimedia/opieplayer2/lib.cpp
+++ b/noncore/multimedia/opieplayer2/lib.cpp
@@ -48,24 +48,19 @@ _;:,   .>  :=|.         This program is free software; you can
 #include <unistd.h>
 
 typedef void (*display_xine_frame_t) (void *user_data, uint8_t* frame,
                       int  width, int height,int bytes );
+typedef void (*vo_scale_cb) (void*, int, int, double,
+                             int*, int*, int*, int*, double*, int*, int* );
+typedef void (*dest_size_cb) (void*, int, int, double, int*, int*, double*);
+
 
 extern "C" {
-    xine_vo_driver_t* init_video_out_plugin( xine_t *xine, void* video, display_xine_frame_t, void * );
+    xine_vo_driver_t* init_video_out_plugin( xine_t *xine, void* video, display_xine_frame_t, void *, vo_scale_cb,  dest_size_cb );
     int null_is_showing_video( const xine_vo_driver_t* self );
     void null_set_show_video( const xine_vo_driver_t* self, int show );
-    int null_is_fullscreen( const xine_vo_driver_t* self );
-    void null_set_fullscreen( const xine_vo_driver_t* self, int screen );
-    int null_is_scaling( const xine_vo_driver_t* self );
-    void null_set_scaling( const xine_vo_driver_t* self, int scale );
-    void null_set_gui_width( const xine_vo_driver_t* self, int width );
-    void null_set_gui_height( const xine_vo_driver_t* self, int height );
     void null_set_mode( const xine_vo_driver_t* self, int depth,  int rgb  );
-    void null_set_videoGamma( const  xine_vo_driver_t* self , int value );
     void null_display_handler( const xine_vo_driver_t* self, display_xine_frame_t t, void* user_data );
-
-    void null_preload_decoders( xine_stream_t *stream );
 }
 
 using namespace XINE;
 
@@ -126,9 +121,12 @@ void Lib::initialize()
 
     // allocate oss for sound
     // and fb for framebuffer
     m_audioOutput = xine_open_audio_driver( m_xine,  "oss", NULL );
-    m_videoOutput = ::init_video_out_plugin( m_xine, NULL, xine_display_frame, this );
+    m_videoOutput = ::init_video_out_plugin( m_xine, NULL,
+                                             xine_display_frame, this,
+                                             xine_vo_scale_cb,
+                                             xine_dest_cb );
 
     m_stream = xine_stream_new (m_xine,  m_audioOutput,  m_videoOutput );
     xine_set_param( m_stream, XINE_PARAM_AUDIO_CLOSE_DEVICE, 1);
 
@@ -138,10 +136,8 @@ void Lib::initialize()
 
     m_queue = xine_event_new_queue (m_stream);
     xine_event_create_listener_thread (m_queue, xine_event_handler, this);
 
-    ::null_preload_decoders( m_stream );
-
     m_duringInitialization = false;
 }
 
 Lib::~Lib() {
@@ -164,12 +160,10 @@ Lib::~Lib() {
 
 void Lib::resize ( const QSize &s ) {
     assert( m_initialized || m_duringInitialization );
 
-    if ( s. width ( ) && s. height ( ) ) {
-        ::null_set_gui_width( m_videoOutput,  s. width() );
-        ::null_set_gui_height( m_videoOutput,  s. height() );
-    }
+    if ( s. width ( ) && s. height ( ) )
+        m_videoSize = s;
 }
 
 int Lib::majorVersion() {
      int major, minor, sub;
@@ -421,33 +415,35 @@ bool Lib::isShowingVideo() const {
 
 void Lib::showVideoFullScreen( bool fullScreen ) {
     assert( m_initialized );
 
-    ::null_set_fullscreen( m_videoOutput, fullScreen );
+    #warning use xine
 }
 
 bool Lib::isVideoFullScreen() const {
     assert( m_initialized );
 
-    return ::null_is_fullscreen( m_videoOutput );
+    #warning use xine
+    return false;
 }
 
 void Lib::setScaling( bool scale ) {
     assert( m_initialized );
 
-    ::null_set_scaling( m_videoOutput, scale );
+    xine_set_param( m_stream, XINE_PARAM_VO_ASPECT_RATIO,
+                    scale ? XINE_VO_ASPECT_AUTO : XINE_VO_ASPECT_SQUARE );
 }
 
 void Lib::setGamma( int value ) {
     assert( m_initialized );
-
-   ::null_set_videoGamma( m_videoOutput, value );
+    xine_set_param( m_stream, XINE_PARAM_VO_BRIGHTNESS, value );
 }
 
 bool Lib::isScaling() const {
     assert( m_initialized );
 
-    return ::null_is_scaling( m_videoOutput );
+    int aratio = xine_get_param( m_stream, XINE_PARAM_VO_ASPECT_RATIO );
+    return aratio == XINE_VO_ASPECT_AUTO;
 }
 
 void Lib::xine_event_handler( void* user_data, const xine_event_t* t ) {
     ( (Lib*)user_data)->handleXineEvent( t );
@@ -457,15 +453,39 @@ void Lib::xine_display_frame( void* user_data, uint8_t *frame,
                               int width,  int height,  int bytes ) {
     ( (Lib*)user_data)->drawFrame( frame, width, height, bytes );
 }
 
+void Lib::xine_vo_scale_cb( void *user_data, int video_with, int video_height,
+                            double video_pixel_aspect,
+                            int *dest_x, int *dest_y, int *dest_width,
+                            int *dest_height, double *dest_pixel_aspect,
+                            int *win_x, int *win_y ) {
+    QSize size = ((Lib*)user_data)->m_videoSize;
+    if (!size.isValid())
+        return;
+
+    *dest_x      = 0;
+    *dest_y      = 0;
+    *dest_width  = size.width();
+    *dest_height = size.height();
+    *win_x       = 0;
+    *win_y       = 0;
+}
+
+void Lib::xine_dest_cb( void* user_data, int, int, double,
+                        int *dest_width, int* dest_height, double* ) {
+    QSize size = ((Lib*)user_data)->m_videoSize;
+    if ( !size.isValid() )
+        return;
+
+    *dest_width  = size.width();
+    *dest_height = size.height();
+}
+
 void Lib::drawFrame( uint8_t* frame,  int width,  int height,  int bytes ) {
     assert( m_initialized );
 
-    if ( !m_video ) {
+    if ( !m_video )
         return;
-    }
-
-//    assert( m_wid );
 
     if (m_wid) m_wid-> setVideoFrame ( frame, width, height, bytes );
 }
diff --git a/noncore/multimedia/opieplayer2/lib.h b/noncore/multimedia/opieplayer2/lib.h
index 2607193..a6ac033 100644
--- a/noncore/multimedia/opieplayer2/lib.h
+++ b/noncore/multimedia/opieplayer2/lib.h
@@ -186,29 +186,33 @@ namespace XINE {
 
         virtual void run();
 
     private:
-        void initialize();
-
-        int m_bytes_per_pixel;
-        bool m_initialized:1;
-        bool m_duringInitialization:1;
-        bool m_video:1;
-        XineVideoWidget *m_wid;
-        xine_t *m_xine;
-    xine_stream_t *m_stream;
-        xine_cfg_entry_t *m_config;
-        xine_vo_driver_t *m_videoOutput;
-        xine_ao_driver_t* m_audioOutput;
-    xine_event_queue_t *m_queue;
-
-        void handleXineEvent( const xine_event_t* t );
-    void handleXineEvent( int type, int data, const char* name );
-        void drawFrame( uint8_t* frame, int width, int height, int bytes );
+        void initialize();
+
+        int m_bytes_per_pixel;
+        bool m_initialized:1;
+        bool m_duringInitialization:1;
+        bool m_video:1;
+        XineVideoWidget *m_wid;
+        QSize m_videoSize;
+        xine_t *m_xine;
+        xine_stream_t *m_stream;
+        xine_cfg_entry_t *m_config;
+        xine_vo_driver_t *m_videoOutput;
+        xine_ao_driver_t* m_audioOutput;
+        xine_event_queue_t *m_queue;
+
+        void handleXineEvent( const xine_event_t* t );
+        void handleXineEvent( int type, int data, const char* name );
+        void drawFrame( uint8_t* frame, int width, int height, int bytes );
         // C -> C++ bridge for the event system
         static void xine_event_handler( void* user_data, const xine_event_t* t);
         static void xine_display_frame( void* user_data, uint8_t* frame ,
-                                        int width, int height, int bytes );
+                                        int width, int height, int bytes );
+        static void xine_vo_scale_cb(void *, int, int, double,
+                                     int*,int*,int*,int*,double*,int*,int*);
+        static void xine_dest_cb(void*,int,int,double,int*,int*,double*);
     };
 };
 
 
diff --git a/noncore/multimedia/opieplayer2/mainTest.cpp b/noncore/multimedia/opieplayer2/mainTest.cpp
deleted file mode 100644
index e374c89..0000000
--- a/noncore/multimedia/opieplayer2/mainTest.cpp
+++ b/dev/null
@@ -1,16 +0,0 @@
-
-#include <stdlib.h>
-#include <stdio.h>
-#include <unistd.h>
-
-#include "lib.h"
-
-int main( int argc,  char *argv[] ) {
-    printf("FixME\n");
-    //return 0;
-    XINE::Lib lib;
-    QString str = QString::fromLatin1( argv[1] );
-    lib.play( str );
-    for (;;) sleep( 60 );
-    return 0;
-}
diff --git a/noncore/multimedia/opieplayer2/nullvideo.c b/noncore/multimedia/opieplayer2/nullvideo.c
index 6769a37..99fa329 100644
--- a/noncore/multimedia/opieplayer2/nullvideo.c
+++ b/noncore/multimedia/opieplayer2/nullvideo.c
@@ -1,9 +1,8 @@
 /*
                             This file is part of the Opie Project
 
                              Copyright (c)  2002 Max Reiss <harlekin@handhelds.org>
-                             Copyright (c)  2002 LJP <>
                              Copyright (c)  2002 Holger Freyther <zecke@handhelds.org>
                              Copyright (c)  2002-2003 Miguel Freitas of xine
               =.
             .=l.
@@ -39,9 +38,8 @@
 
 #include <xine.h>
 #include <xine/video_out.h>
 #include <xine/xine_internal.h>
-//#include <xine/xineutils.h>
 #include <xine/vo_scale.h>
 #include <xine/buffer.h>
 
 #include <pthread.h>
@@ -50,42 +48,45 @@
 
 #define printf(x,...)
 
 /*
-#define LOG
-*/
+ * #define LOG
+ */
 
 /* the caller for our event draw handler */
 typedef void (*display_xine_frame_t) (void *user_data, uint8_t* frame,
-              int  width, int height,int bytes );
+                                      int  width, int height,int bytes );
+typedef void (*vo_scale_cb) (void*, int, int, double,
+                             int*, int*, int*, double*, int*, int* );
+typedef void (*dest_size_cb) (void*, int, int, double, int*, int*, double*);
+
 
 typedef struct null_driver_s null_driver_t;
 
 struct null_driver_s {
   vo_driver_t     vo_driver;
 
-  uint32_t             m_capabilities;
   int                  m_show_video;
   int                  m_video_fullscreen;
-  int                  m_is_scaling;
 
   int                  depth, bpp, bytes_per_pixel;
   int                  yuv2rgb_mode;
   int                  yuv2rgb_swap;
-  int                  yuv2rgb_gamma;
+  int                  yuv2rgb_brightness;
   uint8_t             *yuv2rgb_cmap;
   yuv2rgb_factory_t   *yuv2rgb_factory;
 
   vo_overlay_t        *overlay;
-  vo_scale_t           sc;  
-
-  int                  gui_width;
-  int                  gui_height;
-  int                  gui_changed;
+  vo_scale_t           sc;
 
-  double               display_ratio;
-  void*                caller;
+  /*
+   * DISPLAY to widget...
+   */
   display_xine_frame_t frameDis;
+  void *userData;
+
+  xine_t              *xine;
+  alphablend_t         alphablend_extra_data;
 };
 
 typedef struct opie_frame_s opie_frame_t;
 struct opie_frame_s {
@@ -104,24 +105,34 @@ struct opie_frame_s {
   yuv2rgb_t         *yuv2rgb;
   uint8_t           *rgb_dst;
   int                yuv_stride;
   int                stripe_height, stripe_inc;
- 
+
   null_driver_t     *output;
 };
 
 static uint32_t null_get_capabilities( vo_driver_t *self ){
-  null_driver_t* this = (null_driver_t*)self;
-  return this->m_capabilities;
+    self = self;
+    return VO_CAP_YUY2 | VO_CAP_YV12;
 }
 
 static void null_frame_proc_slice (vo_frame_t *vo_img, uint8_t **src) {
   opie_frame_t  *frame = (opie_frame_t *) vo_img ;
 
   vo_img->proc_called = 1;
 
-  if (!frame->output->m_show_video) {
-    /* printf("nullvideo: no video\n");  */
+  /*
+   * drop hard if we don't show the video
+   */
+  if (!frame->output->m_show_video)
+    return;
+
+  if( frame->frame.crop_left  || frame->frame.crop_top ||
+      frame->frame.crop_right || frame->frame.crop_bottom )
+  {
+    /* we don't support crop, so don't even waste cpu cycles.
+     * cropping will be performed by video_out.c
+     */
     return;
   }
 
   if (frame->format == XINE_IMGFMT_YV12) {
@@ -141,17 +152,17 @@ static void null_frame_field (vo_frame_t *vo_img, int which_field) {
   opie_frame_t  *frame = (opie_frame_t *) vo_img ;
 
   switch (which_field) {
   case VO_TOP_FIELD:
-    frame->rgb_dst    = (uint8_t *)frame->data;
+    frame->rgb_dst    = frame->data;
     frame->stripe_inc = 2*frame->stripe_height * frame->bytes_per_line;
     break;
   case VO_BOTTOM_FIELD:
-    frame->rgb_dst    = (uint8_t *)frame->data + frame->bytes_per_line ;
+    frame->rgb_dst    = frame->data + frame->bytes_per_line ;
     frame->stripe_inc = 2*frame->stripe_height * frame->bytes_per_line;
     break;
   case VO_BOTH_FIELDS:
-    frame->rgb_dst    = (uint8_t *)frame->data;
+    frame->rgb_dst    = frame->data;
     break;
   }
 }
 
@@ -176,8 +187,10 @@ static vo_frame_t* null_alloc_frame( vo_driver_t* self ){
   fprintf (stderr, "nullvideo: alloc_frame\n");
 #endif
 
   frame = (opie_frame_t*)xine_xmalloc ( sizeof(opie_frame_t) );
+  if(!frame)
+      return NULL;
 
   memcpy (&frame->sc, &this->sc, sizeof(vo_scale_t));
 
   pthread_mutex_init (&frame->frame.mutex, NULL);
@@ -186,8 +199,9 @@ static vo_frame_t* null_alloc_frame( vo_driver_t* self ){
 
   /* initialize the frame*/
   frame->frame.driver = self;
   frame->frame.proc_slice = null_frame_proc_slice;
+  frame->frame.proc_frame = NULL;
   frame->frame.field = null_frame_field;
   frame->frame.dispose = null_frame_dispose;
 
   /*
@@ -198,9 +212,9 @@ static vo_frame_t* null_alloc_frame( vo_driver_t* self ){
   return (vo_frame_t*) frame;
 }
 
 
-static void null_frame_compute_ideal_size( null_driver_t *this, 
+static void null_frame_compute_ideal_size( null_driver_t *this,
                                            opie_frame_t *frame ) {
     this = this;
 
     _x_vo_scale_compute_ideal_size(&frame->sc);
@@ -228,25 +242,24 @@ static void null_frame_reallocate( null_driver_t *this, opie_frame_t *frame,
                                    uint32_t width, uint32_t height, int format){
     /*
      * (re-) allocate
      */
-    if( frame->data ) {
-      if( frame->chunk[0] ){
-          free( frame->chunk[0] );
-          frame->chunk[0] = NULL;
-      }
-      if( frame->chunk[1] ){
-          free ( frame->chunk[1] );
-          frame->chunk[1] = NULL;
-      }
-      if( frame->chunk[2] ){
-          free ( frame->chunk[2] );
-          frame->chunk[2] = NULL;
-      }
-      free ( frame->data );
+    if( frame->chunk[0] ){
+        free( frame->chunk[0] );
+        frame->chunk[0] = NULL;
+    }
+    if( frame->chunk[1] ){
+        free ( frame->chunk[1] );
+        frame->chunk[1] = NULL;
     }
+    if( frame->chunk[2] ){
+        free ( frame->chunk[2] );
+        frame->chunk[2] = NULL;
+    }
+    if(frame->data)
+        free ( frame->data );
 
-      frame->data = xine_xmalloc (frame->sc.output_width
+    frame->data = xine_xmalloc (frame->sc.output_width
                                 * frame->sc.output_height
                                 * this->bytes_per_pixel );
 
     if( format == XINE_IMGFMT_YV12 ) {
@@ -300,39 +313,44 @@ static void null_setup_colorspace_converter(opie_frame_t *frame, int flags ) {
 }
 
 static void null_update_frame_format( vo_driver_t* self, vo_frame_t* img,
                                       uint32_t width, uint32_t height,
-                                      double ratio_code,  int format, 
+                                      double ratio_code,  int format,
                                       int flags ){
   null_driver_t* this = (null_driver_t*) self;
   opie_frame_t*  frame = (opie_frame_t*)img;
 
 #ifdef LOG
   fprintf (stderr, "nullvideo: update_frame_format\n");
 #endif
 
+  int gui_width  = 0;
+  int gui_height = 0;
+  double gui_pixel_aspect = 0.0;
+
   flags &= VO_BOTH_FIELDS;
 
-  /* find out if we need to adapt this frame */
+  frame->sc.dest_size_cb(frame->sc.user_data, width, height,
+                         frame->sc.video_pixel_aspect,
+                         &gui_width, &gui_height, &gui_pixel_aspect);
 
+  /* find out if we need to adapt this frame */
   if ((width != frame->sc.delivered_width)
       || (height != frame->sc.delivered_height)
       || (ratio_code != frame->sc.delivered_ratio)
       || (flags != frame->flags)
       || (format != frame->format)
       || (this->sc.user_ratio != frame->sc.user_ratio)
-      || (this->gui_width != frame->sc.gui_width)
-      || (this->gui_height != frame->sc.gui_height)) {
-
+      || (gui_width  != frame->sc.gui_width)
+      || (gui_height != frame->sc.gui_height)) {
     frame->sc.delivered_width      = width;
     frame->sc.delivered_height     = height;
     frame->sc.delivered_ratio      = ratio_code;
     frame->flags                   = flags;
     frame->format                  = format;
     frame->sc.user_ratio           = this->sc.user_ratio;
-    frame->sc.gui_width            = this->gui_width;
-    frame->sc.gui_height           = this->gui_height;
-    frame->sc.gui_pixel_aspect     = 1.0;
+    frame->sc.gui_width  = gui_width;
+    frame->sc.gui_height = gui_height;
 
 
     null_frame_compute_ideal_size(this, frame);
     null_frame_compute_rgb_size(this, frame);
@@ -362,17 +380,17 @@ static void null_update_frame_format( vo_driver_t* self, vo_frame_t* img,
 
     if (frame->data) {
         switch (flags) {
             case VO_TOP_FIELD:
-                frame->rgb_dst    = (uint8_t *)frame->data;
+                frame->rgb_dst    = frame->data;
                 frame->stripe_inc = 2 * frame->stripe_height * frame->bytes_per_line;
                 break;
             case VO_BOTTOM_FIELD:
-                frame->rgb_dst    = (uint8_t *)frame->data + frame->bytes_per_line ;
+                frame->rgb_dst    = frame->data + frame->bytes_per_line ;
                 frame->stripe_inc = 2 * frame->stripe_height * frame->bytes_per_line;
                 break;
             case VO_BOTH_FIELDS:
-                frame->rgb_dst    = (uint8_t *)frame->data;
+                frame->rgb_dst    = frame->data;
                 frame->stripe_inc = frame->stripe_height * frame->bytes_per_line;
                 break;
         }
     }
@@ -386,9 +404,9 @@ static void null_display_frame( vo_driver_t* self, vo_frame_t *frame_gen ){
   if (!this->m_show_video)
     return;
 
   if( display != NULL ) {
-    (*display)(this->caller, frame->data,
+    (*display)(this->userData, frame->data,
                frame->sc.output_width, frame->sc.output_height,
                frame->bytes_per_line );
   }
 
@@ -441,30 +459,33 @@ static void null_overlay_blend ( vo_driver_t *this_gen, vo_frame_t *frame_gen,
       null_overlay_clut_yuv2rgb(this,overlay,frame);
 
     switch(this->bpp) {
         case 16:
-            blend_rgb16((uint8_t *)frame->data,
+            blend_rgb16(frame->data,
                         overlay,
                         frame->sc.output_width,
                         frame->sc.output_height,
                         frame->sc.delivered_width,
-                        frame->sc.delivered_height);
+                        frame->sc.delivered_height,
+                        &this->alphablend_extra_data);
         break;
         case 24:
-            blend_rgb24((uint8_t *)frame->data,
+            blend_rgb24(frame->data,
                         overlay,
                         frame->sc.output_width,
                         frame->sc.output_height,
                         frame->sc.delivered_width,
-                        frame->sc.delivered_height);
+                        frame->sc.delivered_height,
+                        &this->alphablend_extra_data);
         break;
         case 32:
-            blend_rgb32((uint8_t *)frame->data,
+            blend_rgb32(frame->data,
                         overlay,
                         frame->sc.output_width,
                         frame->sc.output_height,
                         frame->sc.delivered_width,
-                        frame->sc.delivered_height);
+                        frame->sc.delivered_height,
+                        &this->alphablend_extra_data);
         break;
         default:
         /* It should never get here */
             break;
@@ -473,9 +494,8 @@ static void null_overlay_blend ( vo_driver_t *this_gen, vo_frame_t *frame_gen,
 }
 
 
 static int null_get_property( vo_driver_t* self, int property ){
-#if 0
   null_driver_t *this = (null_driver_t *)self;
 
   switch(property)
   {
@@ -489,18 +509,13 @@ static int null_get_property( vo_driver_t* self, int property ){
           return this->sc.gui_height;
     default:
         break;
   }
-#else
-  property = property;
-  self = self;
-#endif
 
   return 0;
 }
 static int null_set_property( vo_driver_t* self, int property,
                               int value ){
-#if 0
   null_driver_t *this = (null_driver_t *)self;
 
   switch(property)
   {
@@ -516,90 +531,86 @@ static int null_set_property( vo_driver_t* self, int property,
           break;
       default:
           break;
   }
-#else
-  self = self;
-  property = property;
-#endif
 
   return value;
 }
 static void null_get_property_min_max( vo_driver_t* self,
                                        int property, int *min,
                                        int *max ){
-  self = self;
-  property = property;
-
-  *max = 0;
-  *min = 0;
+ if(property == VO_PROP_BRIGHTNESS) {
+     *min = -100;
+     *max = +100;
+ }else {
+     *min = 0;
+     *max = 0;
+ }
 }
 static int null_gui_data_exchange( vo_driver_t* self,
                                    int data_type,
                                    void *data ){
-  self = self;
-  data_type = data_type;
-  data = data;
-
   return 0;
 }
 
 static void null_dispose ( vo_driver_t* self ){
   null_driver_t* this = (null_driver_t*)self;
+  _x_alphablend_free(&this->alphablend_extra_data);
   free ( this );
 }
 static int null_redraw_needed( vo_driver_t* self ){
-    self = self;
-
-    return 0;
+    null_driver_t *this = (null_driver_t*)self;
+    return _x_vo_scale_redraw_needed(&this->sc);
 }
 
 
 xine_video_port_t* init_video_out_plugin( xine_t *xine,
-                                         void* video,
-                                         display_xine_frame_t frameDisplayFunc,
-                                         void *userData ){
-  video = video;
-
-
+                                          void* video,
+                                          display_xine_frame_t frameDisplayFunc,
+                                          void *userData, vo_scale_cb frame_cb,
+                                          dest_size_cb dest_cb) {
   null_driver_t *vo;
-  vo = (null_driver_t*)malloc( sizeof(null_driver_t ) );
-
-  /* memset? */
-  memset(vo,0, sizeof(null_driver_t ) );
+  vo = (null_driver_t*)xine_xmalloc(sizeof(null_driver_t ));
+  vo->xine = xine;
 
+  _x_alphablend_init(&vo->alphablend_extra_data, xine);
   _x_vo_scale_init (&vo->sc, 0, 0, xine->config);
+  vo->sc.gui_width  = 18;
+  vo->sc.gui_height = 6;
+  vo->sc.user_ratio = XINE_VO_ASPECT_AUTO;
+  vo->sc.user_data  = userData;
+  vo->sc.frame_output_cb = frame_cb;
+  vo->sc.dest_size_cb    = dest_cb;
 
-  vo->sc.gui_pixel_aspect = 1.0;
 
   vo->m_show_video = 0; // false
-  vo->m_video_fullscreen = 0;
-  vo->m_is_scaling = 0;
-  vo->display_ratio = 1.0;
-  vo->gui_width = 16;
-  vo->gui_height = 8;
-  vo->frameDis = NULL;
 
   /* install callback handlers*/
-  vo->vo_driver.get_capabilities      = null_get_capabilities;
-  vo->vo_driver.alloc_frame           = null_alloc_frame;
-  vo->vo_driver.update_frame_format   = null_update_frame_format;
-  vo->vo_driver.display_frame         = null_display_frame;
-  vo->vo_driver.overlay_blend         = null_overlay_blend;
-  vo->vo_driver.get_property          = null_get_property;
-  vo->vo_driver.set_property          = null_set_property;
-  vo->vo_driver.get_property_min_max  = null_get_property_min_max;
-  vo->vo_driver.gui_data_exchange     = null_gui_data_exchange;
-  vo->vo_driver.dispose               = null_dispose;
-  vo->vo_driver.redraw_needed         = null_redraw_needed;
-    
+  vo->vo_driver.get_capabilities    = null_get_capabilities;
+  vo->vo_driver.alloc_frame         = null_alloc_frame;
+  vo->vo_driver.update_frame_format = null_update_frame_format;
+  vo->vo_driver.display_frame       = null_display_frame;
+  vo->vo_driver.overlay_begin       = 0;
+  vo->vo_driver.overlay_blend       = null_overlay_blend;
+  vo->vo_driver.overlay_end         = 0;
+  vo->vo_driver.get_property        = null_get_property;
+  vo->vo_driver.set_property        = null_set_property;
+  vo->vo_driver.get_property_min_max= null_get_property_min_max;
+  vo->vo_driver.gui_data_exchange   = null_gui_data_exchange;
+  vo->vo_driver.dispose             = null_dispose;
+  vo->vo_driver.redraw_needed       = null_redraw_needed;
 
   /* capabilities */
-  vo->m_capabilities = VO_CAP_YUY2 | VO_CAP_YV12;
-  vo->yuv2rgb_factory = yuv2rgb_factory_init (MODE_16_RGB, vo->yuv2rgb_swap,
+  vo->yuv2rgb_mode       = MODE_16_RGB;
+  vo->yuv2rgb_brightness = 0;
+  vo->yuv2rgb_factory = yuv2rgb_factory_init (vo->yuv2rgb_mode,
+                                              vo->yuv2rgb_swap,
                                               vo->yuv2rgb_cmap);
+  vo->yuv2rgb_factory->set_csc_levels(vo->yuv2rgb_factory,
+                                      vo->yuv2rgb_brightness,
+                                      128,128);
 
-  vo->caller = userData;
+  vo->userData = userData;
   vo->frameDis = frameDisplayFunc;
 
   return _x_vo_new_port(xine, &vo->vo_driver, 0);
 }
@@ -620,42 +631,17 @@ int null_is_showing_video( xine_vo_driver_t* self ){
 void null_set_show_video( xine_vo_driver_t* self, int show ) {
   ((null_driver_t*)self->driver)->m_show_video = show;
 }
 
-int null_is_fullscreen( xine_vo_driver_t* self ){
-  return ((null_driver_t*)self->driver)->m_video_fullscreen;
-}
-void null_set_fullscreen( xine_vo_driver_t* self, int screen ){
-  ((null_driver_t*)self->driver)->m_video_fullscreen = screen;
-}
-int null_is_scaling( xine_vo_driver_t* self ){
-  return ((null_driver_t*)self->driver)->m_is_scaling;
-}
-
-void null_set_videoGamma( xine_vo_driver_t* self , int value ) {
-  ((null_driver_t*) self->driver) ->yuv2rgb_gamma = value;
-  ((null_driver_t*) self->driver) ->yuv2rgb_factory->set_gamma( ((null_driver_t*) self->driver) ->yuv2rgb_factory, value );
-}
-
-void null_set_scaling( xine_vo_driver_t* self, int scale ) {
-  ((null_driver_t*)self->driver)->m_is_scaling = scale;
-}
-
-void null_set_gui_width( xine_vo_driver_t* self, int width ) {
-  ((null_driver_t*)self->driver)->gui_width = width;
-}
-void null_set_gui_height( xine_vo_driver_t* self, int height ) {
-  ((null_driver_t*)self->driver)->gui_height = height;
-}
 
 
 void null_set_mode( xine_vo_driver_t* self, int depth,  int rgb  ) {
   null_driver_t* this = (null_driver_t*)self->driver;
 
   this->bytes_per_pixel = (depth + 7 ) / 8;
   this->bpp = this->bytes_per_pixel * 8;
   this->depth = depth;
- 
+
   switch ( this->depth ) {
   case 32:
     if( rgb == 0 )
       this->yuv2rgb_mode = MODE_32_RGB;
@@ -696,40 +682,14 @@ void null_set_mode( xine_vo_driver_t* self, int depth,  int rgb  ) {
       this->yuv2rgb_mode = MODE_8_BGR;
     }
     break;
   };
-  //free(this->yuv2rgb_factory );
-  // this->yuv2rgb_factory = yuv2rgb_factory_init (this->yuv2rgb_mode, this->yuv2rgb_swap, 
-  //            this->yuv2rgb_cmap);
 };
 
-void null_display_handler( xine_vo_driver_t* self, display_xine_frame_t t, 
-         void* user_data ) {
+void null_display_handler( xine_vo_driver_t* self, display_xine_frame_t t,
+                           void* user_data ) {
   null_driver_t* this = (null_driver_t*) self->driver;
-  this->caller = user_data;
+  this->userData = user_data;
   this->frameDis = t;
 }
 
-void null_preload_decoders( xine_stream_t *stream )
-{
-    static const uint32_t preloadedAudioDecoders[] = { BUF_AUDIO_MPEG, BUF_AUDIO_VORBIS };
-    static const uint8_t preloadedAudioDecoderCount = sizeof( preloadedAudioDecoders ) / sizeof( preloadedAudioDecoders[ 0 ] );
-    static const uint32_t preloadedVideoDecoders[] = { BUF_VIDEO_MPEG, BUF_VIDEO_MPEG4, BUF_VIDEO_DIVX5 };
-    static const uint8_t preloadedVideoDecoderCount = sizeof( preloadedVideoDecoders ) / sizeof( preloadedVideoDecoders[ 0 ] );
-
-    uint8_t i;
-#if 0
-
-    for ( i = 0; i < preloadedAudioDecoderCount; ++i ) {
-        audio_decoder_t *decoder = get_audio_decoder( stream, ( preloadedAudioDecoders[ i ] >> 16 ) & 0xff );
-        decoder = decoder;
-/*        free_audio_decoder( stream, decoder ); */
-    }
-
-    for ( i = 0; i < preloadedVideoDecoderCount; ++i ) {
-        video_decoder_t *decoder = get_video_decoder( stream, ( preloadedVideoDecoders[ i ] >> 16 ) & 0xff );
-         decoder = decoder;
-/*         free_video_decoder( stream, decoder ); */
-    }
-#endif
-}
 
diff --git a/noncore/multimedia/opieplayer2/yuv2rgb.c b/noncore/multimedia/opieplayer2/yuv2rgb.c
index 8e34052..487ed20 100644
--- a/noncore/multimedia/opieplayer2/yuv2rgb.c
+++ b/noncore/multimedia/opieplayer2/yuv2rgb.c
@@ -1,8 +1,9 @@
 /*
  * yuv2rgb.c
  *
- * This file is part of xine, a unix video player.
+ * Copyright (C) 2003-2004 the xine project
+ * This file is part of xine, a free video player.
  *
  * based on work from mpeg2dec:
  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
  *
@@ -30,10 +31,16 @@
 #include <string.h>
 #include <inttypes.h>
 
 #include "yuv2rgb.h"
-#include <xine/xineutils.h>
 
+#define LOG_MODULE "yuv2rgb"
+#define LOG_VERBOSE
+/*
+#define LOG
+*/
+
+#include <xine/xineutils.h>
 
 static int prof_scale_line = -1;
 
 static scale_line_func_t find_scale_line_func(int step);
@@ -50,32 +57,69 @@ const int32_t Inverse_Table_6_9[8][4] = {
   {117579, 136230, 16907, 35559}  /* SMPTE 240M (1987) */
 };
 
 
-static void *my_malloc_aligned (size_t alignment, size_t size, void **chunk) {
-
+static void *my_malloc_aligned (size_t alignment, size_t size, void **chunk)
+{
   char *pMem;
 
   pMem = xine_xmalloc (size+alignment);
 
   *chunk = pMem;
 
-  while ((int) pMem % alignment)
+  while ((uintptr_t) pMem % alignment)
     pMem++;
 
   return pMem;
 }
 
 
+static int yuv2rgb_next_slice (yuv2rgb_t *this, uint8_t **dest)
+{
+  int y0, y1;
+
+  if (dest == NULL) {
+    this->slice_offset = 0;
+    this->slice_height = 16;
+    return 0;
+  }
+  if (this->slice_height == this->source_height) {
+    return this->dest_height;
+  }
+
+  y0 = (this->slice_offset * this->dest_height) / this->source_height;
+  y1 = ((this->slice_offset + this->slice_height) * this->dest_height) / this->source_height;
+  *dest += (this->rgb_stride * y0);
+
+  if ((this->slice_offset + this->slice_height) >= this->source_height) {
+    this->slice_offset = 0;
+    return (this->dest_height - y0);
+  } else {
+    this->slice_offset += this->slice_height;
+    return (y1 - y0);
+  }
+}
+
+static void yuv2rgb_dispose (yuv2rgb_t *this)
+{
+  free (this->y_chunk);
+  free (this->u_chunk);
+  free (this->v_chunk);
+#ifdef HAVE_MLIB
+  free (this->mlib_chunk);
+#endif
+  free (this);
+}
+
 static int yuv2rgb_configure (yuv2rgb_t *this, 
                               int source_width, int source_height,
                               int y_stride, int uv_stride,
                               int dest_width, int dest_height,
                               int rgb_stride) {
-  /*
+/*  
   printf ("yuv2rgb setup (%d x %d => %d x %d)\n", source_width, source_height,
           dest_width, dest_height);
-          */
+        */  
   if (prof_scale_line == -1)
     prof_scale_line = xine_profiler_allocate_slot("xshm scale line");
 
   this->source_width  = source_width;
@@ -84,9 +128,11 @@ static int yuv2rgb_configure (yuv2rgb_t *this,
   this->uv_stride     = uv_stride;
   this->dest_width    = dest_width;
   this->dest_height   = dest_height;
   this->rgb_stride    = rgb_stride;
-  
+  this->slice_height  = source_height;
+  this->slice_offset  = 0;
+
   if (this->y_chunk) {
     free (this->y_chunk);
     this->y_buffer = this->y_chunk = NULL;
   }
@@ -98,12 +144,25 @@ static int yuv2rgb_configure (yuv2rgb_t *this,
     free (this->v_chunk);
     this->v_buffer = this->v_chunk = NULL;
   }
 
-  
+#ifdef HAVE_MLIB
+  if (this->mlib_chunk) {
+    free (this->mlib_chunk);
+    this->mlib_buffer = this->mlib_chunk = NULL;
+  }
+  if (this->mlib_resize_chunk) {
+    free (this->mlib_resize_chunk);
+    this->mlib_resize_buffer = this->mlib_resize_chunk = NULL;
+  }
+#endif
+
   this->step_dx = source_width  * 32768 / dest_width;
   this->step_dy = source_height * 32768 / dest_height;
-    
+/*
+  printf("yuv2rgb config: src_ht=%i, dst_ht=%i\n",source_height, dest_height);
+  printf("yuv2rgb config: step_dy=%i %f\n",this->step_dy, (float)this->step_dy / 32768.0);
+*/    
   this->scale_line = find_scale_line_func(this->step_dx);
 
   if ((source_width == dest_width) && (source_height == dest_height)) {
     this->do_scale = 0;
@@ -137,13 +196,25 @@ static int yuv2rgb_configure (yuv2rgb_t *this,
       return 0;
     this->v_buffer = my_malloc_aligned (16, (dest_width+1)/2, &this->v_chunk);
     if (!this->v_buffer)
       return 0;
+
+#if HAVE_MLIB
+    /* Only need these if we are resizing and in mlib code */
+    this->mlib_buffer = my_malloc_aligned (16, source_width*source_height*4, &this->mlib_chunk);
+    if (!this->mlib_buffer)
+      return 0;
+    /* Only need this one if we are 24 bit */
+    if((rgb_stride / dest_width) == 3) {
+      this->mlib_resize_buffer = my_malloc_aligned (16, dest_width*dest_height*4, &this->mlib_resize_chunk);
+      if (!this->mlib_resize_buffer)
+      return 0;
+    }
+        #endif
   }
   return 1;
 }
 
-
 static void scale_line_gen (uint8_t *source, uint8_t *dest,
                             int width, int step) {
 
   /*
@@ -1203,18 +1274,31 @@ static scale_line_func_t find_scale_line_func(int step) {
     {  1,  2, scale_line_1_2,   "2*zoom" },
     {  1,  1, scale_line_1_1,   "non-scaled" },
   };
   int i;
+        #ifdefLOG
+  /* to filter out multiple messages about the scale_line variant we're using */
+  static int reported_for_step;
+#endif
 
   for (i = 0; i < sizeof(scale_line)/sizeof(scale_line[0]); i++) {
     if (step == scale_line[i].src_step*32768/scale_line[i].dest_step) {
-        //printf("yuv2rgb: using %s optimized scale_line\n", scale_line[i].desc);
+        #ifdefLOG
+      if (step != reported_for_step)
+        printf("yuv2rgb: using %s optimized scale_line\n", scale_line[i].desc);
+      reported_for_step = step;
+#endif
       return scale_line[i].func;
     }
   }
-  //printf("yuv2rgb: using generic scale_line with interpolation\n");
-  return scale_line_gen;
 
+        #ifdefLOG
+  if (step != reported_for_step)
+    printf("yuv2rgb: using generic scale_line with interpolation\n");
+  reported_for_step = step;
+#endif
+
+  return scale_line_gen;
 }
 
 
 static void scale_line_2 (uint8_t *source, uint8_t *dest,
@@ -1271,9 +1355,9 @@ static void scale_line_4 (uint8_t *source, uint8_t *dest,
   }
 }
 
 
-        #define RGB(i)                                          \
+        #define X_RGB(i)                                                \
                 U = pu[i];                                      \
                 V = pv[i];                                      \
                 r = this->table_rV[V];                          \
                 g = (void *) (((uint8_t *)this->table_gU[U]) + this->table_gV[V]);\
@@ -1347,9 +1431,9 @@ static void yuv2rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = (uint32_t*)_dst;
       py_1  = this->y_buffer;
@@ -1358,18 +1442,18 @@ static void yuv2rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1(3);
 
           pu += 4;
           pv += 4;
@@ -1411,9 +1495,9 @@ static void yuv2rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst,
           height++;
       } while( dy>=32768);
     }
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = (uint32_t*)_dst;
       dst_2 = (void*)( (uint8_t *)_dst + this->rgb_stride );
       py_1 = _py;
@@ -1422,21 +1506,21 @@ static void yuv2rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst,
       pv   = _pv;
 
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1(0);
         DST2(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2(1);
         DST1(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1(2);
         DST2(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2(3);
         DST1(3);
       
         pu += 4;
@@ -1478,9 +1562,9 @@ static void yuv2rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = _dst;
       py_1  = this->y_buffer;
@@ -1489,18 +1573,18 @@ static void yuv2rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1RGB(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1RGB(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1RGB(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1RGB(3);
 
           pu += 4;
           pv += 4;
@@ -1542,9 +1626,9 @@ static void yuv2rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst,
           height++;
       } while (dy>=32768);
     }
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = _dst;
       dst_2 = (void*)( (uint8_t *)_dst + this->rgb_stride );
       py_1  = _py;
@@ -1553,21 +1637,21 @@ static void yuv2rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst,
       pv    = _pv;
 
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1RGB(0);
         DST2RGB(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2RGB(1);
         DST1RGB(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1RGB(2);
         DST2RGB(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2RGB(3);
         DST1RGB(3);
 
         pu += 4;
@@ -1609,9 +1693,9 @@ static void yuv2rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = _dst;
       py_1  = this->y_buffer;
@@ -1620,18 +1704,18 @@ static void yuv2rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1BGR(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1BGR(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1BGR(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1BGR(3);
 
           pu += 4;
           pv += 4;
@@ -1674,9 +1758,9 @@ static void yuv2rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst,
       } while( dy>=32768 );
     }
 
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = _dst;
       dst_2 = (void*)( (uint8_t *)_dst + this->rgb_stride );
       py_1 = _py;
@@ -1684,21 +1768,21 @@ static void yuv2rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst,
       pu   = _pu;
       pv   = _pv;
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1BGR(0);
         DST2BGR(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2BGR(1);
         DST1BGR(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1BGR(2);
         DST2BGR(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2BGR(3);
         DST1BGR(3);
 
         pu += 4;
@@ -1740,9 +1824,9 @@ static void yuv2rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = (uint16_t*)_dst;
       py_1  = this->y_buffer;
@@ -1751,18 +1835,18 @@ static void yuv2rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1(3);
 
           pu += 4;
           pv += 4;
@@ -1804,9 +1888,9 @@ static void yuv2rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst,
           height++;
       } while( dy>=32768);
     }
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = (uint16_t*)_dst;
       dst_2 = (void*)( (uint8_t *)_dst + this->rgb_stride );
       py_1 = _py;
@@ -1814,21 +1898,21 @@ static void yuv2rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst,
       pu   = _pu;
       pv   = _pv;
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1(0);
         DST2(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2(1);
         DST1(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1(2);
         DST2(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2(3);
         DST1(3);
 
         pu += 4;
@@ -1870,9 +1954,9 @@ static void yuv2rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = (uint8_t*)_dst;
       py_1  = this->y_buffer;
@@ -1881,18 +1965,18 @@ static void yuv2rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1(3);
 
           pu += 4;
           pv += 4;
@@ -1934,9 +2018,9 @@ static void yuv2rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst,
           height++;
       } while( dy>=32768 );
     }
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = (uint8_t*)_dst;
       dst_2 = (void*)( (uint8_t *)_dst + this->rgb_stride );
       py_1 = _py;
@@ -1945,21 +2029,21 @@ static void yuv2rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst,
       pv   = _pv;
 
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1(0);
         DST2(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2(1);
         DST1(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1(2);
         DST2(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2(3);
         DST1(3);
       
         pu += 4;
@@ -1989,9 +2073,9 @@ static void yuv2rgb_c_gray (yuv2rgb_t *this, uint8_t * _dst,
   if (this->do_scale) {
     scale_line_func_t scale_line = this->scale_line;
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (;;) {
       scale_line (_py, _dst, this->dest_width, this->step_dx);
 
@@ -2015,9 +2099,9 @@ static void yuv2rgb_c_gray (yuv2rgb_t *this, uint8_t * _dst,
          _py += this->y_stride;
       */
     }
   } else {
-    for (height = this->source_height; --height >= 0; ) {
+    for (height = this->next_slice (this, &_dst); --height >= 0; ) {
       xine_fast_memcpy(_dst, _py, this->dest_width);
       _dst += this->rgb_stride;
       _py += this->y_stride;
     }
@@ -2045,9 +2129,9 @@ static void yuv2rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst,
     scale_line (_py, this->y_buffer,
                 this->dest_width, this->step_dx);
 
     dy = 0;
-    dst_height = this->dest_height;
+    dst_height = this->next_slice (this, &_dst);
 
     for (height = 0;; ) {
       dst_1 = _dst;
       py_1  = this->y_buffer;
@@ -2056,18 +2140,18 @@ static void yuv2rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst,
 
       width = this->dest_width >> 3;
 
       do {
-          RGB(0);
+          X_RGB(0);
           DST1CMAP(0);
 
-          RGB(1);
+          X_RGB(1);
           DST1CMAP(1);
       
-          RGB(2);
+          X_RGB(2);
           DST1CMAP(2);
 
-          RGB(3);
+          X_RGB(3);
           DST1CMAP(3);
 
           pu += 4;
           pv += 4;
@@ -2109,9 +2193,9 @@ static void yuv2rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst,
           height++;
       } while( dy>=32768 );
     }
   } else {
-    height = this->source_height >> 1;
+    height = this->next_slice (this, &_dst) >> 1;
     do {
       dst_1 = _dst;
       dst_2 = _dst + this->rgb_stride;
       py_1 = _py;
@@ -2119,21 +2203,21 @@ static void yuv2rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst,
       pu   = _pu;
       pv   = _pv;
       width = this->source_width >> 3;
       do {
-        RGB(0);
+        X_RGB(0);
         DST1CMAP(0);
         DST2CMAP(0);
 
-        RGB(1);
+        X_RGB(1);
         DST2CMAP(1);
         DST1CMAP(1);
 
-        RGB(2);
+        X_RGB(2);
         DST1CMAP(2);
         DST2CMAP(2);
 
-        RGB(3);
+        X_RGB(3);
         DST2CMAP(3);
         DST1CMAP(3);
 
         pu += 4;
@@ -2160,9 +2244,10 @@ static int div_round (int dividend, int divisor)
   else
     return -((-dividend + (divisor>>1)) / divisor);
 }
 
-static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped) 
+static void yuv2rgb_set_csc_levels (yuv2rgb_factory_t *this,
+                                    int brightness, int contrast, int saturation)
 {
   int i;
   uint8_t table_Y[1024];
   uint32_t * table_32 = 0;
@@ -2176,20 +2261,26 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
   int cbu = Inverse_Table_6_9[this->matrix_coefficients][1];
   int cgu = -Inverse_Table_6_9[this->matrix_coefficients][2];
   int cgv = -Inverse_Table_6_9[this->matrix_coefficients][3];
 
+  int mode = this->mode;
+  int swapped = this->swapped;
+
   for (i = 0; i < 1024; i++) {
     int j;
 
-    j = (76309 * (i - 384 - 16) + 32768) >> 16;
+    j = (76309 * (i - 384 - 16 + brightness) + 32768) >> 16;
     j = (j < 0) ? 0 : ((j > 255) ? 255 : j);
     table_Y[i] = j;
   }
 
   switch (mode) {
   case MODE_32_RGB:
   case MODE_32_BGR:
-    table_32 = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint32_t));
+    if (this->table_base == NULL) {
+      this->table_base = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint32_t));
+    }
+    table_32 = this->table_base;
 
     entry_size = sizeof (uint32_t);
     table_r = table_32 + 197;
     table_b = table_32 + 197 + 685;
@@ -2216,9 +2307,12 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
     break;
 
   case MODE_24_RGB:
   case MODE_24_BGR:
-    table_8 = malloc ((256 + 2*232) * sizeof (uint8_t));
+    if (this->table_base == NULL) {
+      this->table_base = malloc ((256 + 2*232) * sizeof (uint8_t));
+    }
+    table_8 = this->table_base;
 
     entry_size = sizeof (uint8_t);
     table_r = table_g = table_b = table_8 + 232;
 
@@ -2229,9 +2323,12 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
   case MODE_15_BGR:
   case MODE_16_BGR:
   case MODE_15_RGB:
   case MODE_16_RGB:
-    table_16 = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint16_t));
+    if (this->table_base == NULL) {
+      this->table_base = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint16_t));
+    }
+    table_16 = this->table_base;
 
     entry_size = sizeof (uint16_t);
     table_r = table_16 + 197;
     table_b = table_16 + 197 + 685;
@@ -2269,9 +2366,12 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
     break;
 
   case MODE_8_RGB:
   case MODE_8_BGR:
-    table_8 = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint8_t));
+    if (this->table_base == NULL) {
+      this->table_base = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint8_t));
+    }
+    table_8 = this->table_base;
 
     entry_size = sizeof (uint8_t);
     table_r = table_8 + 197;
     table_b = table_8 + 197 + 685;
@@ -2293,9 +2393,12 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
   case MODE_8_GRAY:
     return;
 
   case MODE_PALETTE:
-    table_16 = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint16_t));
+    if (this->table_base == NULL) {
+      this->table_base = malloc ((197 + 2*682 + 256 + 132) * sizeof (uint16_t));
+    }
+    table_16 = this->table_base;
 
     entry_size = sizeof (uint16_t);
     table_r = table_16 + 197;
     table_b = table_16 + 197 + 685;
@@ -2317,10 +2420,10 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
     break;
 
 
   default:
-    fprintf (stderr, "mode %d not supported by yuv2rgb\n", mode);
-    abort();
+    lprintf ("mode %d not supported by yuv2rgb\n", mode);
+    _x_abort();
   }
   
   for (i = 0; i < 256; i++) {
     this->table_rV[i] = (((uint8_t *) table_r) +
@@ -2330,10 +2433,12 @@ static void yuv2rgb_setup_tables (yuv2rgb_factory_t *this, int mode, int swapped
     this->table_gV[i] = entry_size * div_round (cgv * (i-128), 76309);
     this->table_bU[i] = (((uint8_t *)table_b) +
                          entry_size * div_round (cbu * (i-128), 76309));
   }
-  this->gamma = 0;
-  this->entry_size = entry_size;
+
+#if defined(ARCH_X86) || defined(ARCH_X86_64)
+  mmx_yuv2rgb_set_csc_levels (this, brightness, contrast, saturation);
+#endif  
 }
 
 static uint32_t yuv2rgb_single_pixel_32 (yuv2rgb_t *this, uint8_t y, uint8_t u, uint8_t v)
 {
@@ -2447,10 +2552,10 @@ static void yuv2rgb_c_init (yuv2rgb_factory_t *this)
     this->yuv2rgb_fun = yuv2rgb_c_palette;
     break;
 
   default:
-    printf ("yuv2rgb: mode %d not supported by yuv2rgb\n", this->mode);
-    abort();
+    lprintf ("mode %d not supported by yuv2rgb\n", this->mode);
+    _x_abort();
   }
 
 }
 
@@ -2490,10 +2595,10 @@ static void yuv2rgb_single_pixel_init (yuv2rgb_factory_t *this) {
     this->yuv2rgb_single_pixel_fun = yuv2rgb_single_pixel_palette;
     break;
 
   default:
-    printf ("yuv2rgb: mode %d not supported by yuv2rgb\n", this->mode);
-    abort();
+    lprintf ("mode %d not supported by yuv2rgb\n", this->mode);
+    _x_abort();
   }
 }
 
 
@@ -2519,9 +2624,9 @@ static void yuy22rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer,
                 this->dest_width, this->step_dx);
   
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
   
   for (;;) {
     dst_1 = (uint32_t*)_dst;
     py_1  = this->y_buffer;
@@ -2531,18 +2636,18 @@ static void yuy22rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     width = this->dest_width >> 3;
 
     do {
 
-      RGB(0);
+      X_RGB(0);
       DST1(0);
 
-      RGB(1);
+      X_RGB(1);
       DST1(1);
       
-      RGB(2);
+      X_RGB(2);
       DST1(2);
       
-      RGB(3);
+      X_RGB(3);
       DST1(3);
 
       pu += 4;
       pv += 4;
@@ -2563,9 +2668,9 @@ static void yuy22rgb_c_32 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
     /*
       dy -= 32768;
       _p += this->y_stride*2;
@@ -2598,9 +2703,9 @@ static void yuy22rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
   
   for (;;) {
     dst_1 = _dst;
     py_1  = this->y_buffer;
@@ -2609,18 +2714,18 @@ static void yuy22rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     width = this->dest_width >> 3;
     
     do {
-      RGB(0);
+      X_RGB(0);
       DST1RGB(0);
       
-      RGB(1);
+      X_RGB(1);
       DST1RGB(1);
       
-      RGB(2);
+      X_RGB(2);
       DST1RGB(2);
       
-      RGB(3);
+      X_RGB(3);
       DST1RGB(3);
 
       pu += 4;
       pv += 4;
@@ -2641,9 +2746,9 @@ static void yuy22rgb_c_24_rgb (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
     /*
       dy -= 32768;
       _p += this->y_stride*2;
@@ -2676,9 +2781,9 @@ static void yuy22rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer, 
                 this->dest_width, this->step_dx);
 
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
   
   for (;;) {
     dst_1 = _dst;
     py_1  = this->y_buffer;
@@ -2687,18 +2792,18 @@ static void yuy22rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     width = this->dest_width >> 3;
     
     do {
-      RGB(0);
+      X_RGB(0);
       DST1BGR(0);
       
-      RGB(1);
+      X_RGB(1);
       DST1BGR(1);
       
-      RGB(2);
+      X_RGB(2);
       DST1BGR(2);
 
-      RGB(3);
+      X_RGB(3);
       DST1BGR(3);
       
       pu += 4;
       pv += 4;
@@ -2719,9 +2824,9 @@ static void yuy22rgb_c_24_bgr (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
 
     scale_line_4 (_p+1, this->u_buffer,
                   this->dest_width >> 1, this->step_dx);
@@ -2750,9 +2855,9 @@ static void yuy22rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer, 
                 this->dest_width, this->step_dx);
   
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
 
   for (;;) {
     dst_1 = (uint16_t*)_dst;
     py_1  = this->y_buffer;
@@ -2761,18 +2866,18 @@ static void yuy22rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     width = this->dest_width >> 3;
     
     do {
-      RGB(0);
+      X_RGB(0);
       DST1(0);
 
-      RGB(1);
+      X_RGB(1);
       DST1(1);
       
-      RGB(2);
+      X_RGB(2);
       DST1(2);
       
-      RGB(3);
+      X_RGB(3);
       DST1(3);
 
       pu += 4;
       pv += 4;
@@ -2793,9 +2898,9 @@ static void yuy22rgb_c_16 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
     
     scale_line_4 (_p+1, this->u_buffer,
                   this->dest_width >> 1, this->step_dx);
@@ -2824,9 +2929,9 @@ static void yuy22rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer,
                 this->dest_width, this->step_dx);
   
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
   
   for (;;) {
     dst_1 = _dst;
     py_1  = this->y_buffer;
@@ -2835,18 +2940,18 @@ static void yuy22rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     width = this->dest_width >> 3;
     
     do {
-      RGB(0);
+      X_RGB(0);
       DST1(0);
       
-      RGB(1);
+      X_RGB(1);
       DST1(1);
 
-      RGB(2);
+      X_RGB(2);
       DST1(2);
       
-      RGB(3);
+      X_RGB(3);
       DST1(3);
 
       pu += 4;
       pv += 4;
@@ -2867,9 +2972,9 @@ static void yuy22rgb_c_8 (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
     
     scale_line_4 (_p+1, this->u_buffer,
                   this->dest_width >> 1, this->step_dx);
@@ -2888,9 +2993,9 @@ static void yuy22rgb_c_gray (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   uint8_t * y;
 
   if (this->do_scale) {
     dy = 0;
-    height = this->dest_height;
+    height = this->next_slice (this, &_dst);
   
     for (;;) {
       scale_line_2 (_p, _dst, this->dest_width, this->step_dx);
     
@@ -2907,21 +3012,21 @@ static void yuy22rgb_c_gray (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
       if (height <= 0)
         break;
 
-      _p += this->y_stride*2*(dy>>15);
+      _p += this->y_stride*(dy>>15);
       dy &= 32767;
     }
   } else {
-    for (height = this->source_height; --height >= 0; ) { 
+    for (height = this->next_slice (this, &_dst); --height >= 0; ) { 
       dst = _dst;
       y = _p;
       for (width = this->source_width; --width >= 0; ) {
         *dst++ = *y;
         y += 2;
       }
       _dst += this->rgb_stride;
-      _p += this->y_stride*2;
+      _p += this->y_stride;
     }
   }
 }
 
@@ -2941,9 +3046,9 @@ static void yuy22rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
   scale_line_2 (_p, this->y_buffer,
                 this->dest_width, this->step_dx);
     
   dy = 0;
-  height = this->dest_height;
+  height = this->next_slice (this, &_dst);
   
   for (;;) {
     dst_1 = _dst;
     py_1  = this->y_buffer;
@@ -2952,18 +3057,18 @@ static void yuy22rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
     
     width = this->dest_width >> 3;
     
     do {
-      RGB(0);
+      X_RGB(0);
       DST1CMAP(0);
 
-      RGB(1);
+      X_RGB(1);
       DST1CMAP(1);
 
-      RGB(2);
+      X_RGB(2);
       DST1CMAP(2);
 
-      RGB(3);
+      X_RGB(3);
       DST1CMAP(3);
 
       pu += 4;
       pv += 4;
@@ -2984,9 +3089,9 @@ static void yuy22rgb_c_palette (yuv2rgb_t *this, uint8_t * _dst, uint8_t * _p)
 
     if (height <= 0)
       break;
 
-    _p += this->y_stride*2*(dy>>15);
+    _p += this->y_stride*(dy>>15);
     dy &= 32767;
 
     scale_line_4 (_p+1, this->u_buffer,
                   this->dest_width >> 1, this->step_dx);
@@ -3032,121 +3137,127 @@ static void yuy22rgb_c_init (yuv2rgb_factory_t *this)
     this->yuy22rgb_fun = yuy22rgb_c_palette;
     break;
 
   default:
-    printf ("yuv2rgb: mode %d not supported for yuy2\n", this->mode);
+    lprintf ("mode %d not supported for yuy2\n", this->mode);
   }
 }
 
-yuv2rgb_t *yuv2rgb_create_converter (yuv2rgb_factory_t *factory) {
+static yuv2rgb_t *yuv2rgb_create_converter (yuv2rgb_factory_t *factory) {
 
   yuv2rgb_t *this = xine_xmalloc (sizeof (yuv2rgb_t));
-  
+
+          this->swapped                  = factory->swapped;
   this->cmap                     = factory->cmap;
 
   this->y_chunk = this->y_buffer = NULL;
   this->u_chunk = this->u_buffer = NULL;
   this->v_chunk = this->v_buffer = NULL;
 
+#ifdef HAVE_MLIB
+  this->mlib_chunk = this->mlib_buffer = NULL;
+  this->mlib_resize_chunk = this->mlib_resize_buffer = NULL;
+  this->mlib_filter_type = MLIB_BILINEAR;
+#endif
+
   this->table_rV                 = factory->table_rV;
   this->table_gU                 = factory->table_gU;
   this->table_gV                 = factory->table_gV;
   this->table_bU                 = factory->table_bU;
+  this->table_mmx                = factory->table_mmx;
 
   this->yuv2rgb_fun              = factory->yuv2rgb_fun;
   this->yuy22rgb_fun             = factory->yuy22rgb_fun;
   this->yuv2rgb_single_pixel_fun = factory->yuv2rgb_single_pixel_fun;
 
   this->configure                = yuv2rgb_configure;
+  this->next_slice               = yuv2rgb_next_slice;
+  this->dispose                  = yuv2rgb_dispose;
   return this;
 }
 
 /*
  * factory functions 
  */
-void yuv2rgb_set_gamma (yuv2rgb_factory_t *this, int gamma) {
-    
-    int i;
-    
-    for (i = 0; i < 256; i++) {
-        (uint8_t *)this->table_rV[i] += this->entry_size*(gamma - this->gamma);
-        (uint8_t *)this->table_gU[i] += this->entry_size*(gamma - this->gamma);
-        (uint8_t *)this->table_bU[i] += this->entry_size*(gamma - this->gamma);
-    }
-#ifdef ARCH_X86
-  mmx_yuv2rgb_set_gamma(gamma);
-#endif  
-  this->gamma = gamma;
-}
 
-int yuv2rgb_get_gamma (yuv2rgb_factory_t *this) {
+static void yuv2rgb_factory_dispose (yuv2rgb_factory_t *this) {
 
-  return this->gamma;
+  free (this->table_base);
+  free (this->table_mmx_base);
+  free (this);
 }
 
 yuv2rgb_factory_t* yuv2rgb_factory_init (int mode, int swapped, 
                                          uint8_t *cmap) {
 
   yuv2rgb_factory_t *this;
-
-#ifdef ARCH_X86
   uint32_t mm = xine_mm_accel();
-#endif
 
   this = malloc (sizeof (yuv2rgb_factory_t));
 
   this->mode                = mode;
   this->swapped             = swapped;
   this->cmap                = cmap;
   this->create_converter    = yuv2rgb_create_converter;
-  this->set_gamma           = yuv2rgb_set_gamma;
-  this->get_gamma           = yuv2rgb_get_gamma;
+  this->set_csc_levels      = yuv2rgb_set_csc_levels;
+  this->dispose             = yuv2rgb_factory_dispose;
   this->matrix_coefficients = 6;
+  this->table_base          = NULL;
+  this->table_mmx           = NULL;
+  this->table_mmx_base      = NULL;
 
 
-  yuv2rgb_setup_tables (this, mode, swapped);
+  yuv2rgb_set_csc_levels (this, 0, 128, 128);
 
   /*
    * auto-probe for the best yuv2rgb function
    */
 
   this->yuv2rgb_fun = NULL;
-#ifdef ARCH_X86
+#if defined(ARCH_X86) || defined(ARCH_X86_64)
   if ((this->yuv2rgb_fun == NULL) && (mm & MM_ACCEL_X86_MMXEXT)) {
 
     yuv2rgb_init_mmxext (this);
 
+#ifdef LOG
     if (this->yuv2rgb_fun != NULL)
       printf ("yuv2rgb: using MMXEXT for colorspace transform\n");
+#endif
   }
 
   if ((this->yuv2rgb_fun == NULL) && (mm & MM_ACCEL_X86_MMX)) {
 
     yuv2rgb_init_mmx (this);
 
+#ifdef LOG
     if (this->yuv2rgb_fun != NULL)
       printf ("yuv2rgb: using MMX for colorspace transform\n");
+#endif
+  }
+#ifdef __arm__
+  if (this->yuv2rgb_fun == NULL) {
+  yuv2rgb_init_arm ( this );
+
+  if(this->yuv2rgb_fun != NULL)
+       printf("yuv2rgb: using arm4l assembler for colorspace transform\n" );
   }
 #endif
+
+#endif
 #if HAVE_MLIB
-  if (this->yuv2rgb_fun == NULL) {
+  if ((this->yuv2rgb_fun == NULL) && (mm & MM_ACCEL_MLIB)) {
 
     yuv2rgb_init_mlib (this);
 
+#ifdef LOG
     if (this->yuv2rgb_fun != NULL)
       printf ("yuv2rgb: using medialib for colorspace transform\n");
-  }
 #endif
-#ifdef __arm__
-  if (this->yuv2rgb_fun == NULL) {
-  yuv2rgb_init_arm ( this );
-  
-  if(this->yuv2rgb_fun != NULL)
-          printf("yuv2rgb: using arm4l assembler for colorspace transform\n" );
   }
 #endif
   if (this->yuv2rgb_fun == NULL) {
-    printf ("yuv2rgb: no accelerated colorspace conversion found\n");
+    lprintf ("no accelerated colorspace conversion found\n");
+
     yuv2rgb_c_init (this);
   }
 
   /*
@@ -3163,5 +3274,4 @@ yuv2rgb_factory_t* yuv2rgb_factory_init (int mode, int swapped,
   yuv2rgb_single_pixel_init (this);
 
   return this;
 }
-
diff --git a/noncore/multimedia/opieplayer2/yuv2rgb.h b/noncore/multimedia/opieplayer2/yuv2rgb.h
index e453243..1833889 100644
--- a/noncore/multimedia/opieplayer2/yuv2rgb.h
+++ b/noncore/multimedia/opieplayer2/yuv2rgb.h
@@ -1,8 +1,12 @@
 
 #ifndef HAVE_YUV2RGB_H
 #define HAVE_YUV2RGB_h
 
+#ifdef HAVE_MLIB
+#include <mlib_video.h>
+#endif
+
 #include <inttypes.h>
 
 typedef struct yuv2rgb_s yuv2rgb_t;
 
@@ -21,9 +25,8 @@ typedef void (*yuv2rgb_fun_t) (yuv2rgb_t *this, uint8_t * image, uint8_t * py, u
 typedef void (*yuy22rgb_fun_t) (yuv2rgb_t *this, uint8_t * image, uint8_t * p);
 
 typedef uint32_t (*yuv2rgb_single_pixel_fun_t) (yuv2rgb_t *this, uint8_t y, uint8_t u, uint8_t v);
 
-
 /*
  * modes supported - feel free to implement yours
  */
 
@@ -40,9 +43,8 @@ typedef uint32_t (*yuv2rgb_single_pixel_fun_t) (yuv2rgb_t *this, uint8_t y, uint
         #defineMODE_8_GRAY  11
 #define MODE_PALETTE 12
 
 struct yuv2rgb_s {
-
   /*
    * configure converter for scaling factors
    */
   int (*configure) (yuv2rgb_t *this,
@@ -51,8 +53,18 @@ struct yuv2rgb_s {
                     int dest_width, int dest_height,
                     int rgb_stride);
 
   /*
+   * start a new field or frame if dest is NULL
+   */
+  int (*next_slice) (yuv2rgb_t *this, uint8_t **dest);
+
+  /*
+   * free resources
+   */
+  void (*dispose) (yuv2rgb_t *this);
+
+  /*
    * this is the function to call for the yuv2rgb and scaling process
    */
   yuv2rgb_fun_t     yuv2rgb_fun;
 
@@ -73,80 +85,88 @@ struct yuv2rgb_s {
   int               source_width, source_height;
   int               y_stride, uv_stride;
   int               dest_width, dest_height;
   int               rgb_stride;
+  int               slice_height, slice_offset;
   int               step_dx, step_dy;
-  int               do_scale;
+  int               do_scale, swapped;
 
   uint8_t          *y_buffer;
   uint8_t          *u_buffer;
   uint8_t          *v_buffer;
           void           *y_chunk;
           void           *u_chunk;
           void           *v_chunk;
 
+#ifdef HAVE_MLIB
+  uint8_t          *mlib_buffer;
+  uint8_t          *mlib_resize_buffer;
+          void           *mlib_chunk;
+          void           *mlib_resize_chunk;
+  mlib_filter      mlib_filter_type;
+#endif
+
   void            **table_rV;
   void            **table_gU;
   int              *table_gV;
   void            **table_bU;
+  void             *table_mmx;
 
   uint8_t          *cmap;
-  scale_line_func_t scale_line;
-
+  scale_line_func_t scale_line;  
 } ;
 
 /*
  * convenience class to easily create a lot of converters
  */
 
 struct yuv2rgb_factory_s {
-
   yuv2rgb_t* (*create_converter) (yuv2rgb_factory_t *this);
 
-  /*
-   * adjust gamma (-100 to 100 looks fine)
+  /* 
+   * set color space conversion levels
    * for all converters produced by this factory
    */
-  void (*set_gamma) (yuv2rgb_factory_t *this, int gamma);
+  void (*set_csc_levels) (yuv2rgb_factory_t *this,
+                          int brightness, int contrast, int saturation);
 
   /*
-   * get gamma value
+   * free resources
    */
-  int (*get_gamma) (yuv2rgb_factory_t *this);
+  void (*dispose) (yuv2rgb_factory_t *this);
 
   /* private data */
 
   int      mode;
   int      swapped;
   uint8_t *cmap;
 
-  int      gamma;
-  int      entry_size;
-
   uint32_t matrix_coefficients;
 
+  void    *table_base;
   void    *table_rV[256];
   void    *table_gU[256];
   int      table_gV[256];
   void    *table_bU[256];
+  void    *table_mmx_base;
+  void    *table_mmx;
 
   /* preselected functions for mode/swap/hardware */
   yuv2rgb_fun_t               yuv2rgb_fun;
   yuy22rgb_fun_t              yuy22rgb_fun;
   yuv2rgb_single_pixel_fun_t  yuv2rgb_single_pixel_fun;
-
 };
 
 yuv2rgb_factory_t *yuv2rgb_factory_init (int mode, int swapped, uint8_t *colormap);
 
-
+                   
 /*
  * internal stuff below this line
  */
 
-void mmx_yuv2rgb_set_gamma(int gamma);
+void mmx_yuv2rgb_set_csc_levels(yuv2rgb_factory_t *this,
+                                int brightness, int contrast, int saturation);
 void yuv2rgb_init_mmxext (yuv2rgb_factory_t *this);
 void yuv2rgb_init_mmx (yuv2rgb_factory_t *this);
 void yuv2rgb_init_mlib (yuv2rgb_factory_t *this);
-void yuv2rgb_init_arm (yuv2rgb_factory_t *this);
 
 #endif