From 80c71dc52a6a22efa6c8ea1c77567cce8214d4bc Mon Sep 17 00:00:00 2001
From: "Sergey Semyonov (Serge)" <ion2@mail.ru>
Date: Tue, 27 Jan 2009 00:31:09 +0000
Subject: [PATCH] blit from system memory in PIO mode

git-svn-id: svn://kolibrios.org@1002 a494cfbc-eb01-0410-851d-a64ba20cac60
---
 programs/develop/libraries/pixlib/clip.inc   | 170 +++----
 programs/develop/libraries/pixlib/draw.inc   | 263 +++++++---
 programs/develop/libraries/pixlib/pixlib.c   |   8 +-
 programs/develop/libraries/pixlib/pixlib.h   |   1 +
 programs/develop/libraries/pixlib/pixlib.inc |  65 ++-
 programs/develop/libraries/pixlib/pixmap.inc |  70 ++-
 programs/develop/libraries/pixlib/system.h   |   6 +
 programs/system/drivers/ati2d/accel_2d.h     |   8 +-
 programs/system/drivers/ati2d/accel_2d.inc   | 181 +++++--
 programs/system/drivers/ati2d/ati2d.c        |  15 +-
 programs/system/drivers/ati2d/ati_mem.c      | 140 +++---
 programs/system/drivers/ati2d/init.c         | 482 ++++++++++++-------
 programs/system/drivers/ati2d/init_3d.inc    |   1 -
 programs/system/drivers/ati2d/init_cp.c      | 122 +----
 programs/system/drivers/ati2d/makefile       |   3 +-
 programs/system/drivers/ati2d/pci.c          | 128 +++--
 programs/system/drivers/ati2d/pixmap.inc     |   2 +-
 programs/system/drivers/ati2d/r500.inc       |  27 +-
 18 files changed, 986 insertions(+), 706 deletions(-)

diff --git a/programs/develop/libraries/pixlib/clip.inc b/programs/develop/libraries/pixlib/clip.inc
index cc51a4072b..d8a991adc4 100644
--- a/programs/develop/libraries/pixlib/clip.inc
+++ b/programs/develop/libraries/pixlib/clip.inc
@@ -10,20 +10,20 @@ static int _L1OutCode( clip_t *clip, int x, int y )
 
     Verify that a point is inside or outside the active viewport.   */
 {
-     int   flag;
+    int   flag;
 
-     flag = 0;
-     if( x < clip->xmin ) {
+    flag = 0;
+    if( x < clip->xmin ) {
         flag |= CLIP_LEFT;
-     } else if( x > clip->xmax ) {
+    } else if( x > clip->xmax ) {
         flag |= CLIP_RIGHT;
-     }
-     if( y < clip->ymin ) {
+    }
+    if( y < clip->ymin ) {
         flag |= CLIP_TOP;
-     } else if( y > clip->ymax ) {
+    } else if( y > clip->ymax ) {
         flag |= CLIP_BOTTOM;
-     }
-     return( flag );
+    }
+    return( flag );
 }
 
 
@@ -35,18 +35,18 @@ static void line_inter( int * x1, int* y1, int x2, int y2, int x )
     NOTE : the signs of denom and ( x - *x1 ) cancel out during division
            so make both of them positive before rounding.   */
 {
-     int numer;
-     int denom;
+    int            numer;
+    int            denom;
 
-     denom = abs( x2 - *x1 );
-     numer = 2L * (long)( y2 - *y1 ) * abs( x - *x1 );
-     if( numer > 0 ) {
+    denom = abs( x2 - *x1 );
+    numer = 2L * (long)( y2 - *y1 ) * abs( x - *x1 );
+    if( numer > 0 ) {
         numer += denom;                     /* round to closest pixel   */
-     } else {
+    } else {
         numer -= denom;
-     }
-     *y1 += numer / ( denom << 1 );
-     *x1 = x;
+    }
+    *y1 += numer / ( denom << 1 );
+    *x1 = x;
 }
 
 
@@ -57,39 +57,39 @@ int LineClip( clip_t *clip, int *x1, int *y1, int *x2, int *y2 )
     viewport using the Cohen-Sutherland clipping algorithm. Return the
     clipped coordinates and a decision drawing flag.    */
 {
-     int flag1;
-     int flag2;
+    int    flag1;
+    int    flag2;
 
-     flag1 = _L1OutCode( clip, *x1, *y1 );
-     flag2 = _L1OutCode( clip, *x2, *y2 );
-     for( ;; ) {
+    flag1 = _L1OutCode( clip, *x1, *y1 );
+    flag2 = _L1OutCode( clip, *x2, *y2 );
+    for( ;; ) {
         if( flag1 & flag2 ) break;                  /* trivially outside    */
         if( flag1 == flag2 ) break;                 /* completely inside    */
         if( flag1 == 0 ) {                          /* first point inside   */
-           if( flag2 & CLIP_TOP ) {
-              line_inter( y2, x2, *y1, *x1, clip->ymin );
-           } else if( flag2 & CLIP_BOTTOM ) {
-              line_inter( y2, x2, *y1, *x1, clip->ymax );
-           } else if( flag2 & CLIP_RIGHT ) {
-              line_inter( x2, y2, *x1, *y1, clip->xmax );
-           } else if( flag2 & CLIP_LEFT ) {
-              line_inter( x2, y2, *x1, *y1, clip->xmin );
-           }
-           flag2 = _L1OutCode( clip, *x2, *y2 );
+            if( flag2 & CLIP_TOP ) {
+                line_inter( y2, x2, *y1, *x1, clip->ymin );
+            } else if( flag2 & CLIP_BOTTOM ) {
+                line_inter( y2, x2, *y1, *x1, clip->ymax );
+            } else if( flag2 & CLIP_RIGHT ) {
+                line_inter( x2, y2, *x1, *y1, clip->xmax );
+            } else if( flag2 & CLIP_LEFT ) {
+                line_inter( x2, y2, *x1, *y1, clip->xmin );
+            }
+            flag2 = _L1OutCode( clip, *x2, *y2 );
         } else {                                    /* second point inside  */
-           if( flag1 & CLIP_TOP ) {
-              line_inter( y1, x1, *y2, *x2, clip->ymin );
-           } else if( flag1 & CLIP_BOTTOM ) {
-              line_inter( y1, x1, *y2, *x2, clip->ymax );
-           } else if( flag1 & CLIP_RIGHT ) {
-              line_inter( x1, y1, *x2, *y2, clip->xmax );
-           } else if( flag1 & CLIP_LEFT ) {
-              line_inter( x1, y1, *x2, *y2, clip->xmin );
-           }
-           flag1 = _L1OutCode( clip, *x1, *y1 );
+            if( flag1 & CLIP_TOP ) {
+                line_inter( y1, x1, *y2, *x2, clip->ymin );
+            } else if( flag1 & CLIP_BOTTOM ) {
+                line_inter( y1, x1, *y2, *x2, clip->ymax );
+            } else if( flag1 & CLIP_RIGHT ) {
+                line_inter( x1, y1, *x2, *y2, clip->xmax );
+            } else if( flag1 & CLIP_LEFT ) {
+                line_inter( x1, y1, *x2, *y2, clip->xmin );
+            }
+            flag1 = _L1OutCode( clip, *x1, *y1 );
         }
-     }
-     return( flag1 & flag2 );
+    }
+    return( flag1 & flag2 );
 }
 
 
@@ -98,15 +98,15 @@ static void block_inter( clip_t *clip, int *x, int *y, int flag )
 
     Find the intersection of a block with a boundary of the viewport.   */
 {
-     if( flag & CLIP_TOP ) {
+    if( flag & CLIP_TOP ) {
         *y = clip->ymin;
-     } else if( flag & CLIP_BOTTOM ) {
+    } else if( flag & CLIP_BOTTOM ) {
         *y = clip->ymax;
-     } else if( flag & CLIP_RIGHT ) {
+    } else if( flag & CLIP_RIGHT ) {
         *x = clip->xmax;
-     } else if( flag & CLIP_LEFT ) {
+    } else if( flag & CLIP_LEFT ) {
         *x = clip->xmin;
-     }
+    }
 }
 
 
@@ -118,23 +118,23 @@ int BlockClip(clip_t *clip, int *x1, int *y1, int *x2, int* y2 )
     clipping. Return the clipped coordinates and a decision drawing
     flag ( 0 draw : 1 don't draw ). */
 {
-     int  flag1;
-     int  flag2;
+    int  flag1;
+    int  flag2;
 
-     flag1 = _L1OutCode( clip, *x1, *y1 );
-     flag2 = _L1OutCode( clip, *x2, *y2 );
-     for( ;; ) {
+    flag1 = _L1OutCode( clip, *x1, *y1 );
+    flag2 = _L1OutCode( clip, *x2, *y2 );
+    for( ;; ) {
         if( flag1 & flag2 ) break;                  /* trivially outside    */
         if( flag1 == flag2 ) break;                 /* completely inside    */
         if( flag1 == 0 ) {
-           block_inter( clip, x2, y2, flag2 );
-           flag2 = _L1OutCode( clip,  *x2, *y2 );
+            block_inter( clip, x2, y2, flag2 );
+            flag2 = _L1OutCode( clip,  *x2, *y2 );
         } else {
-           block_inter( clip, x1, y1, flag1 );
-           flag1 = _L1OutCode( clip, *x1, *y1 );
+            block_inter( clip, x1, y1, flag1 );
+            flag1 = _L1OutCode( clip, *x1, *y1 );
         }
-     }
-     return( flag1 & flag2 );
+    }
+    return( flag1 & flag2 );
 }
 
 
@@ -142,39 +142,39 @@ int blit_clip(clip_t *dst_clip,int *dst_x,int *dst_y,
               clip_t *src_clip,int *src_x, int *src_y,
               int *w, int *h)
 {
-     int sx0, sy0, sx1, sy1;
+    int sx0, sy0, sx1, sy1;
 
-     sx0 = *src_x;
-     sy0 = *src_y;
+    sx0 = *src_x;
+    sy0 = *src_y;
 
-     sx1 = sx0 + *w - 1;
-     sy1 = sy0 + *h - 1;
+    sx1 = sx0 + *w - 1;
+    sy1 = sy0 + *h - 1;
 
 
-     if( ! BlockClip( src_clip, &sx0, &sy0, &sx1, &sy1))
-     {
-        int dx0, dy0, dx1, dy1;
+    if( ! BlockClip( src_clip, &sx0, &sy0, &sx1, &sy1))
+    {
+      int dx0, dy0, dx1, dy1;
 
-        dx0 = *dst_x + sx0 - *src_x;
-        dy0 = *dst_y + sy0 - *src_y;
+      dx0 = *dst_x + sx0 - *src_x;
+      dy0 = *dst_y + sy0 - *src_y;
 
-        dx1 = dx0 + sx1 - sx0;
-        dy1 = dy0 + sy1 - sy0;
+      dx1 = dx0 + sx1 - sx0;
+      dy1 = dy0 + sy1 - sy0;
 
-        if( ! BlockClip( dst_clip, &dx0, &dy0, &dx1, &dy1))
-        {
-           *w = dx1 - dx0 + 1;
-           *h = dy1 - dy0 + 1;
+      if( ! BlockClip( dst_clip, &dx0, &dy0, &dx1, &dy1))
+      {
+         *w = dx1 - dx0 + 1;
+         *h = dy1 - dy0 + 1;
 
-           *src_x += dx0 - *dst_x;
-           *src_y += dy0 - *dst_y;
+         *src_x += dx0 - *dst_x;
+         *src_y += dy0 - *dst_y;
 
-           *dst_x = dx0;
-           *dst_y = dx0;
+         *dst_x = dx0;
+         *dst_y = dy0;
 
-           return 0;
-        };
-     };
-     return 1;
+         return 0;
+      };
+    }
+    return 1;
 };
 
diff --git a/programs/develop/libraries/pixlib/draw.inc b/programs/develop/libraries/pixlib/draw.inc
index 56b9948102..7a22fae07e 100644
--- a/programs/develop/libraries/pixlib/draw.inc
+++ b/programs/develop/libraries/pixlib/draw.inc
@@ -558,29 +558,29 @@ int Blit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
          int width, int height)
 {
 
-     clip_t src_clip, dst_clip;
+    clip_t src_clip, dst_clip;
 
-     if( ( width <= 0 ) || ( height<=0 ) )
+    if( ( width <= 0 ) || ( height<=0 ) )
         return ERR_PARAM;
 
-/*   if "hardware acceleration present" and
-     "destinastion is primary screen or local videomemory"  and
-     "source is primary screen or local videomemory"
+/*  if "hardware acceleration present" and
+    "destinastion is primary screen or local videomemory"
 */
-     if( (srv_hw2d != 0) &&
-         ( (dst_pixmap == (void*)-1) ||
-           ( (dst_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) )  &&
+    if( (srv_hw2d != 0) &&
+        ( (dst_pixmap == (void*)-1) ||
+           ( (dst_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) ) /* &&
          ( (src_pixmap == (void*)-1) ||
-           ( (src_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) ) )
-     {
+           ( (src_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) ) */ )
+    {
         ioctl_t io;
         pxblit_t *blit = (pxblit_t*)&dst_pixmap;
 
         if((int)dst_pixmap != -1)
-        blit->dst_pixmap = (pixmap_t*)dst_pixmap->handle;
+            blit->dst_pixmap = (pixmap_t*)dst_pixmap->handle;
 
-        if((int)src_pixmap != -1)
-           blit->src_pixmap = (pixmap_t*)src_pixmap->handle;
+        if( (int)src_pixmap != -1 &&
+            (src_pixmap->flags & PX_MEM_MASK) != PX_MEM_SYSTEM)
+            blit->src_pixmap = (pixmap_t*)src_pixmap->handle;
 
         io.handle   = srv_hw2d;
         io.io_code  = PX_BLIT;
@@ -590,83 +590,83 @@ int Blit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
         io.out_size = 0;
 
         return call_service(&io);
-     }
+    }
 
-     dst_pixmap = (dst_pixmap == (void*)-1) ? &scrn_pixmap : dst_pixmap ;
-     src_pixmap = (src_pixmap == (void*)-1) ? &scrn_pixmap : src_pixmap ;
+    dst_pixmap = (dst_pixmap == (void*)-1) ? &scrn_pixmap : dst_pixmap ;
+    src_pixmap = (src_pixmap == (void*)-1) ? &scrn_pixmap : src_pixmap ;
 
-     src_clip.xmin = 0;
-     src_clip.ymin = 0;
-     src_clip.xmax = src_pixmap->width-1;
-     src_clip.ymax = src_pixmap->height-1;
+    src_clip.xmin = 0;
+    src_clip.ymin = 0;
+    src_clip.xmax = src_pixmap->width-1;
+    src_clip.ymax = src_pixmap->height-1;
 
-     dst_clip.xmin = 0;
-     dst_clip.ymin = 0;
-     dst_clip.xmax = dst_pixmap->width-1;
-     dst_clip.ymax = dst_pixmap->height-1;
+    dst_clip.xmin = 0;
+    dst_clip.ymin = 0;
+    dst_clip.xmax = dst_pixmap->width-1;
+    dst_clip.ymax = dst_pixmap->height-1;
 
-     if( !blit_clip(&dst_clip, &dst_x, &dst_y,
-                    &src_clip, &src_x, &src_y,
-                    &width, &height) )
-     {
+    if( !blit_clip(&dst_clip, &dst_x, &dst_y,
+                   &src_clip, &src_x, &src_y,
+                   &width, &height) )
+    {
         color_t *src_addr = &((color_t*)(src_pixmap->mapped))[src_pixmap->pitch*src_y/4+src_x];
         color_t *dst_addr = &((color_t*)(dst_pixmap->mapped))[dst_pixmap->pitch*dst_y/4+dst_x];
 
         while( height-- )
         {
-           int w = width;
-           color_t *tmp_src = src_addr;
-           color_t *tmp_dst = dst_addr;
+            int w = width;
+            color_t *tmp_src = src_addr;
+            color_t *tmp_dst = dst_addr;
 
-           src_addr += src_pixmap->pitch/4;
-           dst_addr += dst_pixmap->pitch/4;
+            src_addr += src_pixmap->pitch/4;
+            dst_addr += dst_pixmap->pitch/4;
 
-           while( w >= 8)
-           {
-              __asm__ __volatile__ (
-              "movq     (%0),   %%mm0\n"
-              "movq    8(%0),   %%mm1\n"
-              "movq   16(%0),   %%mm2\n"
-              "movq   24(%0),   %%mm3\n"
-              "movq   %%mm0,    (%1)\n"
-              "movq   %%mm1,   8(%1)\n"
-              "movq   %%mm2,  16(%1)\n"
-              "movq   %%mm3,  24(%1)\n"
-               :: "r" (tmp_src), "r" (tmp_dst)
-               : "memory", "%mm0", "%mm1", "%mm2", "%mm3");
-              w -= 8;
-              tmp_src += 8;
-              tmp_dst += 8;
-           };
-           if( w >= 4 )
-           {
-              __asm__ __volatile__ (
-              "movq     (%0),   %%mm0\n"
-              "movq    8(%0),   %%mm1\n"
-              "movq   %%mm0,    (%1)\n"
-              "movq   %%mm1,   8(%1)\n"
-              :: "r" (tmp_src), "r" (tmp_dst)
-              : "memory", "%mm0", "%mm1");
-              w -= 4;
-              tmp_src += 4;
-              tmp_dst += 4;
-           };
-           if( w >= 2 )
-           {
-              __asm__ __volatile__ (
-              "movq     (%0),   %%mm0\n"
-              "movq   %%mm0,    (%1)\n"
-              :: "r" (tmp_src), "r" (tmp_dst)
-              : "memory", "%mm0");
-              w -= 2;
-              tmp_src += 2;
-              tmp_dst += 2;
-           };
-           if( w )
-              *tmp_dst = *tmp_src;
+            while( w >= 8)
+            {
+                __asm__ __volatile__ (
+                "movq     (%0),   %%mm0\n"
+                "movq    8(%0),   %%mm1\n"
+                "movq   16(%0),   %%mm2\n"
+                "movq   24(%0),   %%mm3\n"
+                "movq   %%mm0,    (%1)\n"
+                "movq   %%mm1,   8(%1)\n"
+                "movq   %%mm2,  16(%1)\n"
+                "movq   %%mm3,  24(%1)\n"
+                :: "r" (tmp_src), "r" (tmp_dst)
+                : "memory", "%mm0", "%mm1", "%mm2", "%mm3");
+                w -= 8;
+                tmp_src += 8;
+                tmp_dst += 8;
+            };
+            if( w >= 4 )
+            {
+                __asm__ __volatile__ (
+                "movq     (%0),   %%mm0\n"
+                "movq    8(%0),   %%mm1\n"
+                "movq   %%mm0,    (%1)\n"
+                "movq   %%mm1,   8(%1)\n"
+                :: "r" (tmp_src), "r" (tmp_dst)
+                : "memory", "%mm0", "%mm1");
+                w -= 4;
+                tmp_src += 4;
+                tmp_dst += 4;
+            };
+            if( w >= 2 )
+            {
+                __asm__ __volatile__ (
+                "movq     (%0),   %%mm0\n"
+                "movq   %%mm0,    (%1)\n"
+                :: "r" (tmp_src), "r" (tmp_dst)
+                : "memory", "%mm0");
+                w -= 2;
+                tmp_src += 2;
+                tmp_dst += 2;
+            };
+            if( w )
+                *tmp_dst = *tmp_src;
         };
-     };
-     return ERR_OK;
+    };
+    return ERR_OK;
 };
 
 int TransparentBlit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
@@ -705,6 +705,9 @@ int TransparentBlit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
         return call_service(&io);
      };
 
+     dst_pixmap = (dst_pixmap == (void*)-1) ? &scrn_pixmap : dst_pixmap ;
+     src_pixmap = (src_pixmap == (void*)-1) ? &scrn_pixmap : src_pixmap ;
+
      src_clip.xmin = 0;
      src_clip.ymin = 0;
      src_clip.xmax = src_pixmap->width-1;
@@ -722,8 +725,6 @@ int TransparentBlit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
      {
         __m64 clr_key;
 
-        dst_pixmap = (dst_pixmap == (void*)-1) ? &scrn_pixmap : dst_pixmap ;
-        src_pixmap = (src_pixmap == (void*)-1) ? &scrn_pixmap : src_pixmap ;
 
         color_t *src_addr = &((color_t*)(src_pixmap->mapped))[src_pixmap->pitch*src_y/4+src_x];
         color_t *dst_addr = &((color_t*)(dst_pixmap->mapped))[dst_pixmap->pitch*dst_y/4+dst_x];
@@ -766,4 +767,106 @@ int TransparentBlit(pixmap_t *dst_pixmap, int dst_x, int dst_y,
      return ERR_OK;
 }
 
+unsigned long long m_0080 = 0x0080008000800080ULL;
 
+int BlitAlpha(pixmap_t *dst_pixmap, int dst_x, int dst_y,
+              pixmap_t *src_pixmap, int src_x, int src_y,
+              int width, int height, u32_t alpha)
+{
+     clip_t src_clip, dst_clip;
+
+     if( (srv_hw2d != 0) &&
+         ( (dst_pixmap == (void*)-1) ||
+           ( (dst_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) )  &&
+         ( (src_pixmap == (void*)-1) ||
+           ( (src_pixmap->flags & PX_MEM_MASK)==PX_MEM_LOCAL) ) )
+     {
+        ioctl_t io;
+        pxblit_t *blit = (pxblit_t*)&dst_pixmap;
+
+        if((int)dst_pixmap != -1)
+           blit->dst_pixmap = (pixmap_t*)dst_pixmap->handle;
+
+        if((int)src_pixmap != -1)
+           blit->src_pixmap = (pixmap_t*)src_pixmap->handle;
+
+        io.handle   = srv_hw2d;
+        io.io_code  = PX_BLIT_ALPHA;
+        io.input    = blit;
+        io.inp_size = 9;
+        io.output   = NULL;
+        io.out_size = 0;
+
+        return call_service(&io);
+     };
+
+     dst_pixmap = (dst_pixmap == (void*)-1) ? &scrn_pixmap : dst_pixmap ;
+     src_pixmap = (src_pixmap == (void*)-1) ? &scrn_pixmap : src_pixmap ;
+
+     src_clip.xmin = 0;
+     src_clip.ymin = 0;
+     src_clip.xmax = src_pixmap->width-1;
+     src_clip.ymax = src_pixmap->height-1;
+
+     dst_clip.xmin = 0;
+     dst_clip.ymin = 0;
+     dst_clip.xmax = dst_pixmap->width-1;
+     dst_clip.ymax = dst_pixmap->height-1;
+
+
+     if( !blit_clip(&dst_clip, &dst_x, &dst_y,
+                    &src_clip, &src_x, &src_y,
+                    &width, &height) )
+     {
+        __m64 m_alpha;
+        __m64 m_one_alpha;
+
+        color_t *src_addr = &((color_t*)(src_pixmap->mapped))[src_pixmap->pitch*src_y/4+src_x];
+        color_t *dst_addr = &((color_t*)(dst_pixmap->mapped))[dst_pixmap->pitch*dst_y/4+dst_x];
+
+        m_alpha = _mm_cvtsi32_si64((alpha << 16) | alpha);
+        m_alpha = _mm_unpacklo_pi32(m_alpha, m_alpha);
+        m_one_alpha = _mm_subs_pu16((__m64)0x00FF00FF00FF00FFULL, m_alpha);
+
+        while( height-- )
+        {
+           int w = width;
+           color_t *tmp_src = src_addr;
+           color_t *tmp_dst = dst_addr;
+
+           src_addr += src_pixmap->pitch/4;
+           dst_addr += dst_pixmap->pitch/4;
+
+           while( w-- )
+           {
+              __asm__ __volatile__ (
+              "movd (%0), %%mm0               \n\t"
+              "pxor    %%mm1, %%mm1           \n\t"
+              "punpcklbw  %%mm1, %%mm0        \n\t"
+              "pmullw  %[clr_key], %%mm0       \n\t"
+       //       "paddw   %[m_0080], %%mm0       \n\t"
+
+              "movd (%1), %%mm2               \n\t"
+              "punpcklbw  %%mm1, %%mm2        \n\t"
+              "pmullw  %[m_one_alpha], %%mm2  \n\t"
+       //       "paddw   %[m_0080], %%mm2       \n\t"
+
+              "paddw   %%mm2, %%mm0           \n\t"
+              "psrlw   $8, %%mm0              \n\t"
+              "packuswb %%mm0, %%mm0          \n\t"
+              "movd    %%mm0, (%1)"
+              :: "r" (tmp_src),
+                 "r" (tmp_dst),
+              [clr_key] "y" (m_alpha),
+              [m_one_alpha] "y" (m_one_alpha)
+              :"memory","mm0", "mm1", "mm2");
+
+              tmp_src++;
+              tmp_dst++;
+           };
+//           if( w && (*tmp_src != alpha) )
+//              *tmp_dst = *tmp_src;
+        };
+     };
+     return ERR_OK;
+}
diff --git a/programs/develop/libraries/pixlib/pixlib.c b/programs/develop/libraries/pixlib/pixlib.c
index 3f01c1a877..dd46021986 100644
--- a/programs/develop/libraries/pixlib/pixlib.c
+++ b/programs/develop/libraries/pixlib/pixlib.c
@@ -1,4 +1,6 @@
 
+//#define KOLIBRI_PE
+
 #include "types.h"
 #include "system.h"
 
@@ -64,9 +66,9 @@ int __stdcall start(int state)
      scrn_pixmap.width   = scrnsize >> 16;
      scrn_pixmap.height  = scrnsize & 0xFFFF;
      scrn_pixmap.format  = PICT_a8r8g8b8;
-     scrn_pixmap.flags   = 0;
+     scrn_pixmap.flags   = PX_MEM_LOCAL;
      scrn_pixmap.pitch   = scrnpitch;
-     scrn_pixmap.mapped  = (void*)0xFE000000;
+     scrn_pixmap.mapped  = (void*)LFB_BASE;
 
      br_slab.available = 256;
      br_slab.start = brushes;
@@ -129,6 +131,7 @@ char szDrawRect[]        = "DrawRect";
 char szFillRect[]        = "FillRect";
 char szBlit[]            = "Blit";
 char szTransparentBlit[] = "TransparentBlit";
+char szBlitAlpha[]       = "BlitAlpha";
 
 
 export_t EXPORTS[] __asm__("EXPORTS") =
@@ -155,6 +158,7 @@ export_t EXPORTS[] __asm__("EXPORTS") =
            { szFillRect,        FillRect        },
            { szBlit,            Blit            },
            { szTransparentBlit, TransparentBlit },
+           { szBlitAlpha,       BlitAlpha       },
 
            { NULL, NULL },
          };
diff --git a/programs/develop/libraries/pixlib/pixlib.h b/programs/develop/libraries/pixlib/pixlib.h
index c7d0b47c37..2df0353274 100644
--- a/programs/develop/libraries/pixlib/pixlib.h
+++ b/programs/develop/libraries/pixlib/pixlib.h
@@ -151,6 +151,7 @@ typedef struct
   int        src_y;
   int        w;
   int        h;
+  color_t    key;
 }pxblit_t;
 
 #define PX_CREATE              1
diff --git a/programs/develop/libraries/pixlib/pixlib.inc b/programs/develop/libraries/pixlib/pixlib.inc
index 30e0634754..f11c309fc0 100644
--- a/programs/develop/libraries/pixlib/pixlib.inc
+++ b/programs/develop/libraries/pixlib/pixlib.inc
@@ -126,6 +126,21 @@ macro Blit dstpix, dstx, dsty, srcpix, srcx, srcy, w, h
            add esp, 8*4
 }
 
+macro BlitAlpha dstpix, dstx, dsty, srcpix, srcx, srcy, w, h, alpha
+{
+           pushd alpha
+           pushd h
+           pushd w
+           pushd srcy
+           pushd srcx
+           pushd srcpix
+           pushd dsty
+           pushd dstx
+           pushd dstpix
+           call [imp_BlitAlpha]
+           add esp, 9*4
+}
+
 macro TransparentBlit dstpix, dstx, dsty, srcpix, srcx, srcy, w, h, key
 {
            pushd key
@@ -141,6 +156,25 @@ macro TransparentBlit dstpix, dstx, dsty, srcpix, srcx, srcy, w, h, key
            add esp, 9*4
 }
 
+
+macro fix_cwd  path
+{
+          pushd path
+
+          push '/'
+          push path
+          call _strrchr
+
+          mov byte [eax], 0
+
+          mov eax, 30
+          mov ebx, 1
+          mov ecx, [esp+8]
+          int 0x40
+
+          add esp, 12
+}
+
 szPxlib              db  '/rd/1/lib/pixlib.obj',0
 
 szStart              db  'START',0
@@ -162,6 +196,7 @@ szDrawRect           db  'DrawRect',0
 szFillRect           db  'FillRect',0
 szBlit               db  'Blit',0
 szTransparentBlit    db  'TransparentBlit',0
+szBlitAlpha          db  'BlitAlpha',0
 
 align 4
 
@@ -186,6 +221,7 @@ imp_DrawRect         dd  szDrawRect
 imp_FillRect         dd  szFillRect
 imp_Blit             dd  szBlit
 imp_TransparentBlit  dd  szTransparentBlit
+imp_BlitAlpha        dd  szBlitAlpha
 
                      dd  0
 
@@ -257,4 +293,31 @@ load_pxlib:
            xor eax, eax
            ret
 
-
+align 4
+_strrchr:
+	push	ebp
+	mov	ebp, esp
+	push	edi
+        mov     edi, [8+ebp]
+        mov     ecx, -1
+        xor     al, al
+        cld
+        repne
+        scasb
+        not     ecx
+        dec     edi
+        mov     al, [12+ebp]
+        std
+        repne
+        scasb
+        cld
+        jne     .failure
+        lea     eax, [edi+1]
+        pop     edi
+        pop     ebp
+        ret
+.failure:
+        xor     eax, eax
+        pop     edi
+        pop     ebp
+        ret
diff --git a/programs/develop/libraries/pixlib/pixmap.inc b/programs/develop/libraries/pixlib/pixmap.inc
index 0e84963035..09c30a8dc4 100644
--- a/programs/develop/libraries/pixlib/pixmap.inc
+++ b/programs/develop/libraries/pixlib/pixmap.inc
@@ -23,17 +23,17 @@ static void  free_pixmap(pixmap_t *pixmap)
 
 pixmap_t* CreatePixmap(unsigned width, unsigned height, u32_t format, u32_t flags)
 {
-     pixmap_t *pixmap;
+    pixmap_t *pixmap;
 
-     if( (width  == 0) || ( width > 2048)||
-         (height == 0) || (height > 2048)||
-         (format != PICT_a8r8g8b8))
+    if( (width  == 0) || ( width > 2048)||
+        (height == 0) || (height > 2048)||
+        (format != PICT_a8r8g8b8))
         return NULL;
 
-     pixmap = alloc_pixmap();
+    pixmap = alloc_pixmap();
 
-     if( pixmap )
-     {
+    if( pixmap )
+    {
         void *raw;
         int  pitch;
 
@@ -43,45 +43,43 @@ pixmap_t* CreatePixmap(unsigned width, unsigned height, u32_t format, u32_t flag
         pixmap->flags   = flags;
 
         if( (srv_hw2d != 0) &&
-            ( (flags & PX_MEM_MASK)==PX_MEM_LOCAL) )
+           ( (flags & PX_MEM_MASK)==PX_MEM_LOCAL) )
         {
-           ioctl_t io;
-           io.handle   = srv_hw2d;
-           io.io_code  = PX_CREATE;
-           io.input    = pixmap;
-           io.inp_size = 7;
-           io.output   = NULL;
-           io.out_size = 0;
+            ioctl_t io;
+            io.handle   = srv_hw2d;
+            io.io_code  = PX_CREATE;
+            io.input    = pixmap;
+            io.inp_size = 7;
+            io.output   = NULL;
+            io.out_size = 0;
 
-           if (call_service(&io)==ERR_OK)
-              return pixmap;
-           else
-           {
-              free_pixmap(pixmap) ;
-              return NULL;
-           }
+            if (call_service(&io)==ERR_OK)
+                return pixmap;
+            else{
+                free_pixmap(pixmap) ;
+                return NULL;
+            }
         }
         else
-     {
+        {
 /*
    Only system memory
 */
-        pixmap->flags &= ~PX_MEM_MASK;
+            pixmap->flags &= ~PX_MEM_MASK;
 
-        pitch = ((width+8)&~8)*4;
-        raw = UserAlloc(pitch * height);
+            pitch = ((width+8)&~8)*4;
+            raw = UserAlloc(pitch * height);
 
-        if (! raw)
-        {
-           free_pixmap(pixmap);
-           return NULL;
+            if ( !raw ){
+                free_pixmap(pixmap);
+                return NULL;
+            };
+            pixmap->pitch   = pitch;
+            pixmap->mapped  = raw;
         };
-        pixmap->pitch   = pitch;
-        pixmap->mapped  = raw;
-     };
-     return pixmap;
-  };
-  return NULL;
+        return pixmap;
+    };
+    return NULL;
 };
 
 int DestroyPixmap( pixmap_t *pixmap)
diff --git a/programs/develop/libraries/pixlib/system.h b/programs/develop/libraries/pixlib/system.h
index 2025535fe1..bff94cad07 100644
--- a/programs/develop/libraries/pixlib/system.h
+++ b/programs/develop/libraries/pixlib/system.h
@@ -18,6 +18,12 @@ typedef struct
 
 ///////////////////////////////////////////////////////////////////////////////
 
+#ifdef KOLIBRI_PE
+  #define  LFB_BASE   0xDF000000
+#else
+  #define  LFB_BASE   0xFE000000
+#endif
+
 
 void usleep(u32_t delay);
 
diff --git a/programs/system/drivers/ati2d/accel_2d.h b/programs/system/drivers/ati2d/accel_2d.h
index 7e22621baa..effa95c808 100644
--- a/programs/system/drivers/ati2d/accel_2d.h
+++ b/programs/system/drivers/ati2d/accel_2d.h
@@ -64,8 +64,14 @@ typedef struct
   void      *local;
 }local_pixmap_t;
 
+#define  PX_MEM_SYSTEM    0
+#define  PX_MEM_LOCAL     1
+#define  PX_MEM_GART      2
 
-#define   PX_LOCK            1
+#define  PX_MEM_MASK      3
+
+
+#define   PX_LOCK         1
 
 typedef struct
 {
diff --git a/programs/system/drivers/ati2d/accel_2d.inc b/programs/system/drivers/ati2d/accel_2d.inc
index 0f5fa9a0ea..16a4e7b669 100644
--- a/programs/system/drivers/ati2d/accel_2d.inc
+++ b/programs/system/drivers/ati2d/accel_2d.inc
@@ -567,43 +567,140 @@ int FillRect(io_fill_t *fill)
 };
 
 
+#define ADDRREG(addr)  ((volatile u32_t *)(rhd.MMIOBase + (addr)))
+
+
+static int blit_host(u32_t dstpitch, int dstx, int dsty,
+                     u32_t  src, int srcx, int srcy,
+                     int w, int h, int srcpitch)
+{
+    u32_t     ifl;
+    color_t  *src_addr;
+
+    ifl = safe_cli();
+
+#if R300_PIO
+
+    R5xxFIFOWait(5);
+
+    OUTREG(R5XX_DP_GUI_MASTER_CNTL,
+               RADEON_GMC_DST_PITCH_OFFSET_CNTL  |
+//               RADEON_GMC_DST_CLIPPING           |
+               RADEON_GMC_BRUSH_NONE             |
+               RADEON_GMC_DST_32BPP              |
+               RADEON_GMC_SRC_DATATYPE_COLOR     |
+               RADEON_DP_SRC_SOURCE_HOST_DATA    |
+          //     RADEON_GMC_BYTE_MSB_TO_LSB        |
+               R5XX_GMC_CLR_CMP_CNTL_DIS         |
+               R5XX_GMC_WR_MSK_DIS               |
+               R5XX_ROP3_S
+              );
+
+    OUTREG(R5XX_DP_CNTL, R5XX_DST_X_LEFT_TO_RIGHT |
+                         R5XX_DST_Y_TOP_TO_BOTTOM);
+
+    OUTREG(R5XX_DST_PITCH_OFFSET, dstpitch);
+
+//        OUTREG(RADEON_SC_TOP_LEFT, (y << 16) | ((x+skipleft) & 0xffff));
+//        OUTREG(RADEON_SC_BOTTOM_RIGHT,  ((y+h) << 16) | ((x+w) & 0xffff));
+
+    OUTREG(RADEON_DST_Y_X, (dsty << 16) | (dstx & 0xffff));
+    OUTREG(RADEON_DST_HEIGHT_WIDTH, (h << 16) | w);
+
+
+    src_addr = &((color_t*)src)[srcpitch*srcy/4+srcx];
+
+    while ( h-- )
+    {
+        color_t *tmp_src = src_addr;
+        src_addr += srcpitch/4;
+
+        int left = w;
+
+        while( left )
+        {
+            volatile u32_t *d;
+
+            if( left > 8 )
+            {
+                int i;
+
+                R5xxFIFOWait(8);
+                d = ADDRREG(RADEON_HOST_DATA0);
+
+				/* Unrolling doesn't improve performance */
+                for ( i = 0; i < 8; i++)
+                    *d++ = *tmp_src++;
+                left -= 8;
+            }
+            else
+            {
+                R5xxFIFOWait(left);
+
+                if( h )
+                    d = ADDRREG(RADEON_HOST_DATA7) - (left - 1);
+                else
+                    d = ADDRREG(RADEON_HOST_DATA_LAST) - (left - 1);
+
+                for ( ; left; --left)
+                    *d++ = *tmp_src++;
+                left = 0;
+            };
+        };
+    };
+
+#endif
+
+    safe_sti(ifl);
+
+    return ERR_OK;
+}
+
 
 int Blit(io_blit_t *blit)
 {
-     clip_t src_clip, dst_clip;
+    clip_t src_clip, dst_clip;
 
-     local_pixmap_t *srcpixmap;
-     local_pixmap_t *dstpixmap;
+    local_pixmap_t *srcpixmap;
+    local_pixmap_t *dstpixmap;
 
-     //dbgprintf("Pixblit src: %x dst: %x\n",blit->srcpix, blit->dstpix);
+    //dbgprintf("Pixblit src: %x dst: %x\n",blit->srcpix, blit->dstpix);
 
-     dstpixmap = (blit->dstpix == (void*)-1) ? &scr_pixmap : blit->dstpix ;
-     srcpixmap = (blit->srcpix == (void*)-1) ? &scr_pixmap : blit->srcpix ;
+    dstpixmap = (blit->dstpix == (void*)-1) ? &scr_pixmap : blit->dstpix ;
+    srcpixmap = (blit->srcpix == (void*)-1) ? &scr_pixmap : blit->srcpix ;
 
-     //dbgprintf("srcpixmap: %x dstpixmap: %x\n",srcpixmap, dstpixmap);
+    //dbgprintf("srcpixmap: %x dstpixmap: %x\n",srcpixmap, dstpixmap);
 
-     //dbgprintf("dst.width: %d dst.height: %d\n", dstpixmap->width,dstpixmap->height);
-     //dbgprintf("src.width: %d src.height: %d\n", srcpixmap->width,srcpixmap->height);
-     //dbgprintf("srcpitch: %x dstpitch: %x\n",
-     //           srcpixmap->pitch_offset,dstpixmap->pitch_offset);
+    //dbgprintf("dst.width: %d dst.height: %d\n", dstpixmap->width,dstpixmap->height);
+    //dbgprintf("src.width: %d src.height: %d\n", srcpixmap->width,srcpixmap->height);
+    //dbgprintf("srcpitch: %x dstpitch: %x\n",
+    //           srcpixmap->pitch_offset,dstpixmap->pitch_offset);
 
-     src_clip.xmin = 0;
-     src_clip.ymin = 0;
-     src_clip.xmax = srcpixmap->width-1;
-     src_clip.ymax = srcpixmap->height-1;
+    src_clip.xmin = 0;
+    src_clip.ymin = 0;
+    src_clip.xmax = srcpixmap->width-1;
+    src_clip.ymax = srcpixmap->height-1;
 
-     dst_clip.xmin = 0;
-     dst_clip.ymin = 0;
-     dst_clip.xmax = dstpixmap->width-1;
-     dst_clip.ymax = dstpixmap->height-1;
+    dst_clip.xmin = 0;
+    dst_clip.ymin = 0;
+    dst_clip.xmax = dstpixmap->width-1;
+    dst_clip.ymax = dstpixmap->height-1;
 
-     if( !blit_clip(&dst_clip, &blit->dst_x, &blit->dst_y,
-                    &src_clip, &blit->src_x, &blit->src_y,
-                    &blit->w, &blit->h) )
-     {
+    if( !blit_clip(&dst_clip, &blit->dst_x, &blit->dst_y,
+                   &src_clip, &blit->src_x, &blit->src_y,
+                   &blit->w, &blit->h) )
+    {
         u32_t *ring, write;
         u32_t ifl;
 
+        if( (srcpixmap->flags & PX_MEM_MASK)==PX_MEM_SYSTEM)
+            return blit_host(dstpixmap->pitch_offset,
+                             blit->dst_x, blit->dst_y,
+                             srcpixmap->mapped,
+                             blit->src_x, blit->src_y,
+                             blit->w, blit->h,
+                             srcpixmap->pitch);
+
         ifl = safe_cli();
 
 #if R300_PIO
@@ -632,31 +729,31 @@ int Blit(io_blit_t *blit)
         OUTREG(R5XX_DST_HEIGHT_WIDTH,(blit->h<<16)|blit->w);
 
 #else
-        BEGIN_RING(7);
+    BEGIN_RING(7);
 
-          OUT_RING(CP_PACKET3(RADEON_CNTL_BITBLT, 5));
+        OUT_RING(CP_PACKET3(RADEON_CNTL_BITBLT, 5));
 
-          OUT_RING(RADEON_GMC_SRC_PITCH_OFFSET_CNTL  |
-                   RADEON_GMC_DST_PITCH_OFFSET_CNTL  |
-                   RADEON_GMC_BRUSH_NONE             |
-                   RADEON_GMC_DST_32BPP              |
-                   RADEON_GMC_SRC_DATATYPE_COLOR     |
-                   RADEON_DP_SRC_SOURCE_MEMORY       |
-                   R5XX_GMC_CLR_CMP_CNTL_DIS         |
-                   R5XX_GMC_WR_MSK_DIS               |
-                   R5XX_ROP3_S
-                  );
+        OUT_RING(RADEON_GMC_SRC_PITCH_OFFSET_CNTL  |
+                 RADEON_GMC_DST_PITCH_OFFSET_CNTL  |
+                 RADEON_GMC_BRUSH_NONE             |
+                 RADEON_GMC_DST_32BPP              |
+                 RADEON_GMC_SRC_DATATYPE_COLOR     |
+                 RADEON_DP_SRC_SOURCE_MEMORY       |
+                 R5XX_GMC_CLR_CMP_CNTL_DIS         |
+                 R5XX_GMC_WR_MSK_DIS               |
+                 R5XX_ROP3_S
+                );
 
-          OUT_RING(srcpixmap->pitch_offset);
-          OUT_RING(dstpixmap->pitch_offset);
+        OUT_RING(srcpixmap->pitch_offset);
+        OUT_RING(dstpixmap->pitch_offset);
 
-          OUT_RING((blit->src_x<<16)|blit->src_y);
-          OUT_RING((blit->dst_x<<16)|blit->dst_y);
-          OUT_RING((blit->w<<16)|blit->h);
-        COMMIT_RING();
+        OUT_RING((blit->src_x<<16)|blit->src_y);
+        OUT_RING((blit->dst_x<<16)|blit->dst_y);
+        OUT_RING((blit->w<<16)|blit->h);
+    COMMIT_RING();
 
 #endif
-       safe_sti(ifl);
+        safe_sti(ifl);
      };
      return ERR_OK;
 };
diff --git a/programs/system/drivers/ati2d/ati2d.c b/programs/system/drivers/ati2d/ati2d.c
index 001748cadb..8633806558 100644
--- a/programs/system/drivers/ati2d/ati2d.c
+++ b/programs/system/drivers/ati2d/ati2d.c
@@ -1,6 +1,11 @@
 
 
-#define R300_PIO     0       /* now we have cp */
+#define R300_PIO        1
+
+
+#define API_VERSION     0x01000100
+
+#define SRV_GETVERSION  0
 
 
 #include "types.h"
@@ -72,8 +77,11 @@ u32_t __stdcall drvEntry(int action)
 
     R5xx2DInit();
 
+#if !R300_PIO
+
     Init3DEngine(&rhd);
 
+#endif
 
     retval = RegService("HDRAW", srv_2d);
     dbgprintf("reg service %s as: %x\n", "HDRAW", retval);
@@ -82,11 +90,6 @@ u32_t __stdcall drvEntry(int action)
 };
 
 
-#define API_VERSION     0x01000100
-
-#define SRV_GETVERSION  0
-
-
 int __stdcall srv_2d(ioctl_t *io)
 {
   u32_t *inp;
diff --git a/programs/system/drivers/ati2d/ati_mem.c b/programs/system/drivers/ati2d/ati_mem.c
index 3052d55b8f..a98bb7115b 100644
--- a/programs/system/drivers/ati2d/ati_mem.c
+++ b/programs/system/drivers/ati2d/ati_mem.c
@@ -39,49 +39,50 @@
  * already mapped into each client's address space.
  */
 
-struct mem_block {
-	struct mem_block *next;
-	struct mem_block *prev;
-  u32_t  start;
-  size_t size;
+struct mem_block
+{
+  struct mem_block *next;
+  struct mem_block *prev;
+  u32_t             start;
+  size_t            size;
 };
 
 /* Initialize.  How to check for an uninitialized heap?
  */
 static int init_heap(struct mem_block **heap, int start, int size)
 {
-  struct mem_block *blocks = malloc(sizeof(*blocks));
+    struct mem_block *blocks = malloc(sizeof(*blocks));
 
-	if (!blocks)
-    return -1; //-ENOMEM;
+    if (!blocks)
+        return -1; //-ENOMEM;
 
-  *heap = malloc(sizeof(**heap));
-  if (!*heap)
-  {
-    free(blocks);
-    return -1; //-ENOMEM;
+    *heap = malloc(sizeof(**heap));
+    if (!*heap)
+    {
+        free(blocks);
+        return -1; //-ENOMEM;
 	}
 
-  blocks->start = start;
-  blocks->size  = size;
-  blocks->next  = blocks->prev = *heap;
+    blocks->start = start;
+    blocks->size  = size;
+    blocks->next  = blocks->prev = *heap;
 
-  __clear(*heap,sizeof(**heap));
+    __clear(*heap,sizeof(**heap));
 	(*heap)->next = (*heap)->prev = blocks;
-  (*heap)->start |= USED_BLOCK;
+    (*heap)->start |= USED_BLOCK;
 	return 0;
 }
 
 static struct mem_block **get_heap(RHDPtr rhdPtr, int region)
 {
-  switch (region)
-  {
-    case RHD_MEM_GART:
-      return &rhdPtr->gart_heap;
-    case RHD_MEM_FB:
-      return &rhdPtr->fb_heap;
-    default:
-      return NULL;
+    switch (region)
+    {
+        case RHD_MEM_GART:
+            return &rhdPtr->gart_heap;
+        case RHD_MEM_FB:
+            return &rhdPtr->fb_heap;
+        default:
+        return NULL;
 	}
 }
 
@@ -91,17 +92,17 @@ static struct mem_block *split_block(struct mem_block *p, int size)
 	/* Maybe cut off the end of an existing block */
     if (size < p->size)
     {
-      struct mem_block *newblock = malloc(sizeof(*newblock));
-      if (!newblock)
-        goto out;
-      newblock->start = p->start + size;
-      newblock->size = p->size - size;
-      newblock->next = p->next;
-      newblock->prev = p;
-      p->next->prev = newblock;
-      p->next = newblock;
-      p->size = size;
-      p->start|=USED_BLOCK;
+        struct mem_block *newblock = malloc(sizeof(*newblock));
+        if (!newblock)
+            goto out;
+        newblock->start = p->start + size;
+        newblock->size = p->size - size;
+        newblock->next = p->next;
+        newblock->prev = p;
+        p->next->prev = newblock;
+        p->next = newblock;
+        p->size = size;
+        p->start|=USED_BLOCK;
     }
 
 out:
@@ -112,11 +113,10 @@ static struct mem_block *alloc_block(struct mem_block *heap, int size)
 {
 	struct mem_block *p;
 
-  list_for_each(p, heap)
-  {
-
-    if ( !(p->start & USED_BLOCK) && size <= p->size)
-      return split_block(p, size);
+    list_for_each(p, heap)
+    {
+        if ( !(p->start & USED_BLOCK) && size <= p->size)
+            return split_block(p, size);
 	}
 
 	return NULL;
@@ -140,70 +140,70 @@ static void free_block(struct mem_block *p)
 	 * 'heap' to stop it being subsumed.
 	 */
 
-  p->start &= ~USED_BLOCK;
+    p->start &= ~USED_BLOCK;
 
-  if ( !(p->next->start & USED_BLOCK))
-  {
+    if ( !(p->next->start & USED_BLOCK))
+    {
 		struct mem_block *q = p->next;
 		p->size += q->size;
 		p->next = q->next;
 		p->next->prev = p;
-    free(q);
+        free(q);
 	}
 
-  if ( !(p->prev->start & USED_BLOCK))
-  {
+    if ( !(p->prev->start & USED_BLOCK))
+    {
 		struct mem_block *q = p->prev;
 		q->size += p->size;
 		q->next = p->next;
 		q->next->prev = q;
-    free(p);
+        free(p);
 	}
 }
 
 int rhdInitHeap(RHDPtr rhdPtr)
 {
-  int base = rhdPtr->FbFreeStart;
+    int base = rhdPtr->FbFreeStart;
 
-  return init_heap(&rhdPtr->fb_heap, base, rhdPtr->FbFreeSize);
+    return init_heap(&rhdPtr->fb_heap, base, rhdPtr->FbFreeSize);
 };
 
 void *rhd_mem_alloc(RHDPtr rhdPtr,int region, int size)
 {
 	struct mem_block *block, **heap;
 
-  heap = get_heap(rhdPtr, region);
-  if (!heap || !*heap)
-    return NULL;
+    heap = get_heap(rhdPtr, region);
+    if (!heap || !*heap)
+        return NULL;
 
 	/* Make things easier on ourselves: all allocations at least
 	 * 4k aligned.
 	 */
 
-  size = (size+4095) & ~4095;
+    size = (size+4095) & ~4095;
 
-  block = alloc_block(*heap, size);
+    block = alloc_block(*heap, size);
 
 	if (!block)
-    return NULL;
+        return NULL;
 
-  return (void*)(block->start & ~USED_BLOCK);
+    return (void*)(block->start & ~USED_BLOCK);
 }
 
 int rhd_mem_free(RHDPtr rhdPtr, int region, void *offset)
 {
 	struct mem_block *block, **heap;
 
-  heap = get_heap(rhdPtr, region);
-	if (!heap || !*heap)
-    return -1;
+    heap = get_heap(rhdPtr, region);
+    if (!heap || !*heap)
+        return -1;
 
-  block = find_block(*heap, (int)offset);
+    block = find_block(*heap, (int)offset);
 	if (!block)
-    return -1;
+        return -1;
 
-  if ( !(block->start & USED_BLOCK))
-    return -1;
+    if ( !(block->start & USED_BLOCK))
+        return -1;
 
 	free_block(block);
 	return 0;
@@ -212,13 +212,13 @@ int rhd_mem_free(RHDPtr rhdPtr, int region, void *offset)
 void dump_mem()
 {
 	struct mem_block *p;
-  struct mem_block **heap;
+    struct mem_block **heap;
 
-  heap = &rhd.fb_heap;
+    heap = &rhd.fb_heap;
 
-  list_for_each(p, *heap)
-  {
-    dbgprintf("block: %x  next: %x  prev: %x  start: %x  size:%x\n",
+    list_for_each(p, *heap)
+    {
+        dbgprintf("block: %x  next: %x  prev: %x  start: %x  size:%x\n",
               p,p->next,p->prev,p->start,p->size);
 	}
 }
diff --git a/programs/system/drivers/ati2d/init.c b/programs/system/drivers/ati2d/init.c
index 020decbf31..ea6658f348 100644
--- a/programs/system/drivers/ati2d/init.c
+++ b/programs/system/drivers/ati2d/init.c
@@ -19,25 +19,25 @@ unsigned INMC(RHDPtr info, int addr)
 
     if ((info->ChipFamily == CHIP_FAMILY_RS690) ||
        (info->ChipFamily == CHIP_FAMILY_RS740)) {
-       OUTREG(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
-       data = INREG(RS690_MC_DATA);
+        OUTREG(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
+        data = INREG(RS690_MC_DATA);
     } else if (info->ChipFamily == CHIP_FAMILY_RS600) {
-       OUTREG(RS600_MC_INDEX, (addr & RS600_MC_INDEX_MASK));
-       data = INREG(RS600_MC_DATA);
+        OUTREG(RS600_MC_INDEX, (addr & RS600_MC_INDEX_MASK));
+        data = INREG(RS600_MC_DATA);
     } else if (IS_AVIVO_VARIANT) {
-       OUTREG(AVIVO_MC_INDEX, (addr & 0xff) | 0x7f0000);
-       (void)INREG(AVIVO_MC_INDEX);
-       data = INREG(AVIVO_MC_DATA);
+        OUTREG(AVIVO_MC_INDEX, (addr & 0xff) | 0x7f0000);
+        (void)INREG(AVIVO_MC_INDEX);
+        data = INREG(AVIVO_MC_DATA);
 
-       OUTREG(AVIVO_MC_INDEX, 0);
-       (void)INREG(AVIVO_MC_INDEX);
+        OUTREG(AVIVO_MC_INDEX, 0);
+        (void)INREG(AVIVO_MC_INDEX);
     } else {
-       OUTREG(R300_MC_IND_INDEX, addr & 0x3f);
-       (void)INREG(R300_MC_IND_INDEX);
-       data = INREG(R300_MC_IND_DATA);
+        OUTREG(R300_MC_IND_INDEX, addr & 0x3f);
+        (void)INREG(R300_MC_IND_INDEX);
+        data = INREG(R300_MC_IND_DATA);
 
-       OUTREG(R300_MC_IND_INDEX, 0);
-       (void)INREG(R300_MC_IND_INDEX);
+        OUTREG(R300_MC_IND_INDEX, 0);
+        (void)INREG(R300_MC_IND_INDEX);
     }
 
     return data;
@@ -46,65 +46,65 @@ unsigned INMC(RHDPtr info, int addr)
      /* Write MC information */
 void OUTMC(RHDPtr info, int addr, u32_t data)
 {
-     if ((info->ChipFamily == CHIP_FAMILY_RS690) ||
-         (info->ChipFamily == CHIP_FAMILY_RS740)) {
+    if ((info->ChipFamily == CHIP_FAMILY_RS690) ||
+        (info->ChipFamily == CHIP_FAMILY_RS740)) {
         OUTREG(RS690_MC_INDEX, ((addr & RS690_MC_INDEX_MASK) | RS690_MC_INDEX_WR_EN));
-        OUTREG(RS690_MC_DATA, data);
+         OUTREG(RS690_MC_DATA, data);
         OUTREG(RS690_MC_INDEX, RS690_MC_INDEX_WR_ACK);
-     }
-     else if (info->ChipFamily == CHIP_FAMILY_RS600) {
+    }
+    else if (info->ChipFamily == CHIP_FAMILY_RS600) {
         OUTREG(RS600_MC_INDEX, ((addr & RS600_MC_INDEX_MASK) | RS600_MC_INDEX_WR_EN));
         OUTREG(RS600_MC_DATA, data);
         OUTREG(RS600_MC_INDEX, RS600_MC_INDEX_WR_ACK);
-     }
-     else if (IS_AVIVO_VARIANT) {
+    }
+    else if (IS_AVIVO_VARIANT) {
         OUTREG(AVIVO_MC_INDEX, (addr & 0xff) | 0xff0000);
         (void)INREG(AVIVO_MC_INDEX);
         OUTREG(AVIVO_MC_DATA, data);
         OUTREG(AVIVO_MC_INDEX, 0);
         (void)INREG(AVIVO_MC_INDEX);
-     }
-     else {
+    }
+    else {
         OUTREG(R300_MC_IND_INDEX, (((addr) & 0x3f) | R300_MC_IND_WR_EN));
         (void)INREG(R300_MC_IND_INDEX);
         OUTREG(R300_MC_IND_DATA, data);
         OUTREG(R300_MC_IND_INDEX, 0);
         (void)INREG(R300_MC_IND_INDEX);
-     }
+    }
 }
 
 static Bool avivo_get_mc_idle(RHDPtr info)
 {
 
-     if (info->ChipFamily >= CHIP_FAMILY_R600) {
+    if (info->ChipFamily >= CHIP_FAMILY_R600) {
 	/* no idea where this is on r600 yet */
         return TRUE;
-     }
-     else if (info->ChipFamily == CHIP_FAMILY_RV515) {
+    }
+    else if (info->ChipFamily == CHIP_FAMILY_RV515) {
         if (INMC(info, RV515_MC_STATUS) & RV515_MC_STATUS_IDLE)
-           return TRUE;
+            return TRUE;
         else
-           return FALSE;
-     }
-     else if (info->ChipFamily == CHIP_FAMILY_RS600)
-     {
+            return FALSE;
+    }
+    else if (info->ChipFamily == CHIP_FAMILY_RS600)
+    {
         if (INMC(info, RS600_MC_STATUS) & RS600_MC_STATUS_IDLE)
-           return TRUE;
+            return TRUE;
         else
-           return FALSE;
-     }
-     else if ((info->ChipFamily == CHIP_FAMILY_RS690) ||
-	       (info->ChipFamily == CHIP_FAMILY_RS740)) {
+            return FALSE;
+    }
+    else if ((info->ChipFamily == CHIP_FAMILY_RS690) ||
+          (info->ChipFamily == CHIP_FAMILY_RS740)) {
         if (INMC(info, RS690_MC_STATUS) & RS690_MC_STATUS_IDLE)
-           return TRUE;
+            return TRUE;
         else
-           return FALSE;
+            return FALSE;
     }
     else {
-       if (INMC(info, R520_MC_STATUS) & R520_MC_STATUS_IDLE)
-           return TRUE;
-       else
-           return FALSE;
+        if (INMC(info, R520_MC_STATUS) & R520_MC_STATUS_IDLE)
+            return TRUE;
+        else
+            return FALSE;
     }
 }
 
@@ -448,115 +448,115 @@ static void RADEONUpdateMemMapRegisters(RHDPtr info)
 #endif
 
 
-
 static void RADEONInitMemoryMap(RHDPtr info)
 {
-     u32_t       mem_size;
-     u32_t       aper_size;
+    u32_t       mem_size;
+    u32_t       aper_size;
 
-     radeon_read_mc_fb_agp_location(info, LOC_FB | LOC_AGP, &info->mc_fb_location,
-                    &info->mc_agp_location, &info->mc_agp_location_hi);
+    radeon_read_mc_fb_agp_location(info, LOC_FB | LOC_AGP, &info->mc_fb_location,
+                   &info->mc_agp_location, &info->mc_agp_location_hi);
 
-      dbgprintf("  MC_FB_LOCATION   : 0x%08x\n", (unsigned)info->mc_fb_location);
-      dbgprintf("  MC_AGP_LOCATION  : 0x%08x\n", (unsigned)info->mc_agp_location);
+    dbgprintf("  MC_FB_LOCATION   : 0x%08x\n", (unsigned)info->mc_fb_location);
+    dbgprintf("  MC_AGP_LOCATION  : 0x%08x\n", (unsigned)info->mc_agp_location);
 
 
-     /* We shouldn't use info->videoRam here which might have been clipped
-      * but the real video RAM instead
-      */
-     if (info->ChipFamily >= CHIP_FAMILY_R600){
+    /* We shouldn't use info->videoRam here which might have been clipped
+     * but the real video RAM instead
+     */
+    if (info->ChipFamily >= CHIP_FAMILY_R600){
         mem_size = INREG(R600_CONFIG_MEMSIZE);
         aper_size = INREG(R600_CONFIG_APER_SIZE);
-     }
-     else {
+    }
+    else {
         mem_size = INREG(RADEON_CONFIG_MEMSIZE);
         aper_size = INREG(RADEON_CONFIG_APER_SIZE);
-     }
+    }
 
-     if (mem_size == 0)
+    if (mem_size == 0)
         mem_size = 0x800000;
 
-     /* Fix for RN50, M6, M7 with 8/16/32(??) MBs of VRAM -
-        Novell bug 204882 + along with lots of ubuntu ones */
-     if (aper_size > mem_size)
+    /* Fix for RN50, M6, M7 with 8/16/32(??) MBs of VRAM -
+       Novell bug 204882 + along with lots of ubuntu ones */
+    if (aper_size > mem_size)
         mem_size = aper_size;
 
 
-     if ( (info->ChipFamily != CHIP_FAMILY_RS600) &&
-          (info->ChipFamily != CHIP_FAMILY_RS690) &&
-          (info->ChipFamily != CHIP_FAMILY_RS740)) {
+    if ( (info->ChipFamily != CHIP_FAMILY_RS600) &&
+         (info->ChipFamily != CHIP_FAMILY_RS690) &&
+         (info->ChipFamily != CHIP_FAMILY_RS740))
+    {
         if (info->IsIGP)
-           info->mc_fb_location = INREG(RADEON_NB_TOM);
+          info->mc_fb_location = INREG(RADEON_NB_TOM);
         else
         {
-           u32_t aper0_base;
+            u32_t aper0_base;
 
-           if (info->ChipFamily >= CHIP_FAMILY_R600) {
-              aper0_base = INREG(R600_CONFIG_F0_BASE);
-           }
-           else {
-              aper0_base = INREG(RADEON_CONFIG_APER_0_BASE);
-           }
-           dbgprintf("aper0 base %x\n", aper0_base );
+            if (info->ChipFamily >= CHIP_FAMILY_R600) {
+                aper0_base = INREG(R600_CONFIG_F0_BASE);
+            }
+            else {
+                aper0_base = INREG(RADEON_CONFIG_APER_0_BASE);
+            }
+            dbgprintf("aper0 base %x\n", aper0_base );
 
-         /* Recent chips have an "issue" with the memory controller, the
-          * location must be aligned to the size. We just align it down,
-          * too bad if we walk over the top of system memory, we don't
-          * use DMA without a remapped anyway.
-          * Affected chips are rv280, all r3xx, and all r4xx, but not IGP
-          */
-           if ( info->ChipFamily == CHIP_FAMILY_RV280 ||
-                info->ChipFamily == CHIP_FAMILY_R300 ||
-                info->ChipFamily == CHIP_FAMILY_R350 ||
-                info->ChipFamily == CHIP_FAMILY_RV350 ||
-                info->ChipFamily == CHIP_FAMILY_RV380 ||
-                info->ChipFamily == CHIP_FAMILY_R420 ||
-                info->ChipFamily == CHIP_FAMILY_RV410)
-              aper0_base &= ~(mem_size - 1);
+        /* Recent chips have an "issue" with the memory controller, the
+         * location must be aligned to the size. We just align it down,
+         * too bad if we walk over the top of system memory, we don't
+         * use DMA without a remapped anyway.
+         * Affected chips are rv280, all r3xx, and all r4xx, but not IGP
+         */
+            if ( info->ChipFamily == CHIP_FAMILY_RV280 ||
+                 info->ChipFamily == CHIP_FAMILY_R300 ||
+                 info->ChipFamily == CHIP_FAMILY_R350 ||
+                 info->ChipFamily == CHIP_FAMILY_RV350 ||
+                 info->ChipFamily == CHIP_FAMILY_RV380 ||
+                 info->ChipFamily == CHIP_FAMILY_R420 ||
+                 info->ChipFamily == CHIP_FAMILY_RV410)
+                aper0_base &= ~(mem_size - 1);
 
-           if ( info->ChipFamily >= CHIP_FAMILY_R600) {
-              info->mc_fb_location = (aper0_base >> 24) |
-              (((aper0_base + mem_size - 1) & 0xff000000U) >> 8);
-              dbgprintf("mc fb loc is %08x\n", (unsigned int)info->mc_fb_location);
-           }
-           else {
-              info->mc_fb_location = (aper0_base >> 16) |
-              ((aper0_base + mem_size - 1) & 0xffff0000U);
-              dbgprintf("mc fb loc is %08x\n", (unsigned int)info->mc_fb_location);
-           }
+            if ( info->ChipFamily >= CHIP_FAMILY_R600) {
+                 info->mc_fb_location = (aper0_base >> 24) |
+                 (((aper0_base + mem_size - 1) & 0xff000000U) >> 8);
+                dbgprintf("mc fb loc is %08x\n", (unsigned int)info->mc_fb_location);
+            }
+            else {
+                info->mc_fb_location = (aper0_base >> 16) |
+                ((aper0_base + mem_size - 1) & 0xffff0000U);
+                dbgprintf("mc fb loc is %08x\n", (unsigned int)info->mc_fb_location);
+            }
         }
-     }
-     if (info->ChipFamily >= CHIP_FAMILY_R600) {
+    }
+    if (info->ChipFamily >= CHIP_FAMILY_R600) {
         info->fbLocation = (info->mc_fb_location & 0xffff) << 24;
-     }
-     else {
+    }
+    else {
         info->fbLocation = (info->mc_fb_location & 0xffff) << 16;
-     }
-     /* Just disable the damn AGP apertures for now, it may be
-      * re-enabled later by the DRM
-      */
+    }
+    /* Just disable the damn AGP apertures for now, it may be
+     * re-enabled later by the DRM
+     */
 
-//     if (IS_AVIVO_VARIANT) {
-//        if (info->ChipFamily >= CHIP_FAMILY_R600) {
-//           OUTREG(R600_HDP_NONSURFACE_BASE, (info->mc_fb_location << 16) & 0xff0000);
-//        }
-//        else {
-//           OUTREG(AVIVO_HDP_FB_LOCATION, info->mc_fb_location);
-//        }
-//        info->mc_agp_location = 0x003f0000;
-//     }
-//     else
-//         info->mc_agp_location = 0xffffffc0;
+//    if (IS_AVIVO_VARIANT) {
+//       if (info->ChipFamily >= CHIP_FAMILY_R600) {
+//          OUTREG(R600_HDP_NONSURFACE_BASE, (info->mc_fb_location << 16) & 0xff0000);
+//       }
+//       else {
+//          OUTREG(AVIVO_HDP_FB_LOCATION, info->mc_fb_location);
+//       }
+//       info->mc_agp_location = 0x003f0000;
+//    }
+//    else
+//        info->mc_agp_location = 0xffffffc0;
 
-      dbgprintf("RADEONInitMemoryMap() : \n");
-      dbgprintf("  mem_size         : 0x%08x\n", (u32_t)mem_size);
-      dbgprintf("  MC_FB_LOCATION   : 0x%08x\n", (unsigned)info->mc_fb_location);
-      dbgprintf("  MC_AGP_LOCATION  : 0x%08x\n", (unsigned)info->mc_agp_location);
-      dbgprintf("  FB_LOCATION   : 0x%08x\n", (unsigned)info->fbLocation);
+    dbgprintf("RADEONInitMemoryMap() : \n");
+    dbgprintf("  mem_size         : 0x%08x\n", (u32_t)mem_size);
+    dbgprintf("  MC_FB_LOCATION   : 0x%08x\n", (unsigned)info->mc_fb_location);
+    dbgprintf("  MC_AGP_LOCATION  : 0x%08x\n", (unsigned)info->mc_agp_location);
+    dbgprintf("  FB_LOCATION   : 0x%08x\n", (unsigned)info->fbLocation);
 
 #if !R300_PIO
 
-      RADEONUpdateMemMapRegisters(info);
+    RADEONUpdateMemMapRegisters(info);
 
 #endif
 
@@ -813,31 +813,31 @@ static Bool RADEONPreInitVRAM(RHDPtr info)
 
 static Bool RADEONPreInitChipType(RHDPtr rhdPtr)
 {
-     u32_t cmd_stat;
+    u32_t cmd_stat;
 
-     rhdPtr->ChipErrata = 0;
+    rhdPtr->ChipErrata = 0;
 
-     if ( (rhdPtr->ChipFamily == CHIP_FAMILY_R300) &&
-          ((_RHDRegRead(rhdPtr,RADEON_CONFIG_CNTL) & RADEON_CFG_ATI_REV_ID_MASK)
-           == RADEON_CFG_ATI_REV_A11))
+    if ( (rhdPtr->ChipFamily == CHIP_FAMILY_R300) &&
+         ((_RHDRegRead(rhdPtr,RADEON_CONFIG_CNTL) & RADEON_CFG_ATI_REV_ID_MASK)
+          == RADEON_CFG_ATI_REV_A11))
         rhdPtr->ChipErrata |= CHIP_ERRATA_R300_CG;
 
-     if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RV200) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) )
+    if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RV200) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) )
         rhdPtr->ChipErrata |= CHIP_ERRATA_PLL_DUMMYREADS;
 
-     if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RV100) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS100) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) )
+    if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RV100) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS100) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) )
         rhdPtr->ChipErrata |= CHIP_ERRATA_PLL_DELAY;
 
-     rhdPtr->cardType = CARD_PCI;
+    rhdPtr->cardType = CARD_PCI;
 
 
-     cmd_stat = pciReadLong(rhdPtr->PciTag, PCI_CMD_STAT_REG);
+    cmd_stat = pciReadLong(rhdPtr->PciTag, PCI_CMD_STAT_REG);
 
-     if (cmd_stat & RADEON_CAP_LIST)
-     {
+    if (cmd_stat & RADEON_CAP_LIST)
+    {
         u32_t cap_ptr, cap_id;
 
         cap_ptr = pciReadLong(rhdPtr->PciTag, RADEON_CAPABILITIES_PTR_PCI_CONFIG);
@@ -845,42 +845,42 @@ static Bool RADEONPreInitChipType(RHDPtr rhdPtr)
 
         while(cap_ptr != RADEON_CAP_ID_NULL)
         {
-           cap_id = pciReadLong(rhdPtr->PciTag, cap_ptr);
-           if ((cap_id & 0xff)== RADEON_CAP_ID_AGP) {
-              rhdPtr->cardType = CARD_AGP;
-              break;
-           }
-           if ((cap_id & 0xff)== RADEON_CAP_ID_EXP) {
-              rhdPtr->cardType = CARD_PCIE;
-              break;
-           }
-           cap_ptr = (cap_id >> 8) & RADEON_CAP_PTR_MASK;
+            cap_id = pciReadLong(rhdPtr->PciTag, cap_ptr);
+            if ((cap_id & 0xff)== RADEON_CAP_ID_AGP) {
+                 rhdPtr->cardType = CARD_AGP;
+                break;
+            }
+            if ((cap_id & 0xff)== RADEON_CAP_ID_EXP) {
+                rhdPtr->cardType = CARD_PCIE;
+                break;
+            }
+            cap_ptr = (cap_id >> 8) & RADEON_CAP_PTR_MASK;
         }
-     }
+    }
 
-     dbgprintf("%s card detected\n",(rhdPtr->cardType==CARD_PCI) ? "PCI" :
-               (rhdPtr->cardType==CARD_PCIE) ? "PCIE" : "AGP");
+    dbgprintf("%s card detected\n",(rhdPtr->cardType==CARD_PCI) ? "PCI" :
+              (rhdPtr->cardType==CARD_PCIE) ? "PCIE" : "AGP");
 
     /* treat PCIE IGP cards as PCI  */
-     if (rhdPtr->cardType == CARD_PCIE && rhdPtr->IsIGP)
-         rhdPtr->cardType = CARD_PCI;
+    if (rhdPtr->cardType == CARD_PCIE && rhdPtr->IsIGP)
+        rhdPtr->cardType = CARD_PCI;
 
-     if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RS100) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS300) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS400) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS480) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS600) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS690) ||
-          (rhdPtr->ChipFamily == CHIP_FAMILY_RS740))
+    if ( (rhdPtr->ChipFamily == CHIP_FAMILY_RS100) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS200) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS300) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS400) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS480) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS600) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS690) ||
+         (rhdPtr->ChipFamily == CHIP_FAMILY_RS740))
         rhdPtr->has_tcl = FALSE;
-     else {
+    else {
         rhdPtr->has_tcl = TRUE;
-     }
+    }
 
-     rhdPtr->LinearAddr = rhdPtr->memBase[RHD_FB_BAR];
+    rhdPtr->LinearAddr = rhdPtr->memBase[RHD_FB_BAR];
 
-     return TRUE;
+    return TRUE;
 }
 
 #if 0
@@ -958,29 +958,29 @@ Bool RHDPreInit()
     RHDPtr info;
 
     /* We need access to IO space already */
-     if ( !rhdMapMMIO(&rhd) ) {
+    if ( !rhdMapMMIO(&rhd) ) {
         dbgprintf("Failed to map MMIO.\n");
         return FALSE;
-     };
+    };
 
 
-     if( !RADEONPreInitChipType(&rhd))
-        return FALSE;
+    if( !RADEONPreInitChipType(&rhd))
+       return FALSE;
 
-     if (!RADEONPreInitVRAM(&rhd))
-        return FALSE;
+    if (!RADEONPreInitVRAM(&rhd))
+       return FALSE;
 
-     RADEONInitMemoryMap(&rhd);
+    RADEONInitMemoryMap(&rhd);
 
-     if (!rhd.videoRam)
-     {
-        dbgprintf("No Video RAM detected.\n");
-        goto error1;
-     }
-     dbgprintf("VideoRAM: %d kByte\n",rhd.videoRam);
+    if (!rhd.videoRam)
+    {
+       dbgprintf("No Video RAM detected.\n");
+       goto error1;
+    }
+    dbgprintf("VideoRAM: %d kByte\n",rhd.videoRam);
 
  //    rhd.FbFreeStart = 0;
-     rhd.FbFreeSize = rhd.videoRam << 10;
+    rhd.FbFreeSize = rhd.videoRam << 10;
 
  // if( !rhdMapFB(&rhd))
 //    return FALSE;
@@ -988,19 +988,131 @@ Bool RHDPreInit()
 //  rhd.FbScanoutStart = 0;
 //  rhd.FbScanoutSize  = 8*1024*1024;
 
-  rhd.FbFreeStart    = 10*1024*1024;
-  rhd.FbFreeSize     = rhd.FbMapSize - rhd.FbFreeStart - rhd.FbSecureSize;
+    rhd.FbFreeStart    = 10*1024*1024;
+    rhd.FbFreeSize     = rhd.FbMapSize - rhd.FbFreeStart - rhd.FbSecureSize;
 
-  rhdInitHeap(&rhd);
+    rhdInitHeap(&rhd);
 
-     info = &rhd;
+    info = &rhd;
 
-
-   return TRUE;
+    return TRUE;
 
 error1:
-  return FALSE;
+
+    return FALSE;
+};
+
+static void RADEONPllErrataAfterIndex()
+{
+    if (!(rhd.ChipErrata & CHIP_ERRATA_PLL_DUMMYREADS))
+       return;
+
+    /* This workaround is necessary on rv200 and RS200 or PLL
+     * reads may return garbage (among others...)
+     */
+    (void)INREG(RADEON_CLOCK_CNTL_DATA);
+    (void)INREG(RADEON_CRTC_GEN_CNTL);
+}
+
+
+static void RADEONPllErrataAfterData()
+{
+
+    /* This function is required to workaround a hardware bug in some (all?)
+     * revisions of the R300.  This workaround should be called after every
+     * CLOCK_CNTL_INDEX register access.  If not, register reads afterward
+     * may not be correct.
+     */
+    if (rhd.ChipFamily <= CHIP_FAMILY_RV380)
+    {
+        u32_t save, tmp;
+
+	save = INREG(RADEON_CLOCK_CNTL_INDEX);
+	tmp = save & ~(0x3f | RADEON_PLL_WR_EN);
+	OUTREG(RADEON_CLOCK_CNTL_INDEX, tmp);
+	tmp = INREG(RADEON_CLOCK_CNTL_DATA);
+	OUTREG(RADEON_CLOCK_CNTL_INDEX, save);
+    }
+}
+
+
+/* Read PLL register */
+static u32_t RADEONINPLL(int addr)
+{
+    u32_t       data;
+
+    OUTREG8(RADEON_CLOCK_CNTL_INDEX, addr & 0x3f);
+    RADEONPllErrataAfterIndex();
+    data = INREG(RADEON_CLOCK_CNTL_DATA);
+    RADEONPllErrataAfterData();
+
+    return data;
+};
+
+/* Write PLL information */
+static void RADEONOUTPLL(int addr, u32_t data)
+{
+    OUTREG8(RADEON_CLOCK_CNTL_INDEX, (((addr) & 0x3f) |
+				      RADEON_PLL_WR_EN));
+    RADEONPllErrataAfterIndex();
+    OUTREG(RADEON_CLOCK_CNTL_DATA, data);
+    RADEONPllErrataAfterData();
+}
+
+static void init_pipes(RHDPtr info)
+{
+    u32_t gb_tile_config = 0;
+
+    if ( (info->ChipFamily == CHIP_FAMILY_RV410) ||
+         (info->ChipFamily == CHIP_FAMILY_R420)  ||
+         (info->ChipFamily == CHIP_FAMILY_RS600) ||
+         (info->ChipFamily == CHIP_FAMILY_RS690) ||
+         (info->ChipFamily == CHIP_FAMILY_RS740) ||
+         (info->ChipFamily == CHIP_FAMILY_RS400) ||
+         (info->ChipFamily == CHIP_FAMILY_RS480) || IS_R500_3D)
+    {
+        u32_t gb_pipe_sel = INREG(R400_GB_PIPE_SELECT);
+
+        info->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
+        if (IS_R500_3D)
+            OUTPLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
+    }
+    else
+    {
+        if ((info->ChipFamily == CHIP_FAMILY_R300) ||
+           (info->ChipFamily == CHIP_FAMILY_R350))
+        {
+        /* R3xx chips */
+            info->num_gb_pipes = 2;
+        }
+        else {
+        /* RV3xx chips */
+            info->num_gb_pipes = 1;
+        }
+    }
+
+    if (IS_R300_3D || IS_R500_3D)
+    {
+
+        dbgprintf("num quad-pipes is %d\n", info->num_gb_pipes);
+
+        switch(info->num_gb_pipes) {
+            case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
+            case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
+            case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
+            default:
+                case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
+        }
+
+        OUTREG(R300_GB_TILE_CONFIG, gb_tile_config);
+        OUTREG(RADEON_WAIT_UNTIL, RADEON_WAIT_2D_IDLECLEAN | RADEON_WAIT_3D_IDLECLEAN);
+        OUTREG(R300_DST_PIPE_CONFIG, INREG(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
+        OUTREG(R300_RB2D_DSTCACHE_MODE, (INREG(R300_RB2D_DSTCACHE_MODE) |
+                                       R300_DC_AUTOFLUSH_ENABLE |
+                                       R300_DC_DC_DISABLE_IGNORE_PE));
+    }
+    else
+        OUTREG(RADEON_RB3D_CNTL, 0);
 };
 
 
-
diff --git a/programs/system/drivers/ati2d/init_3d.inc b/programs/system/drivers/ati2d/init_3d.inc
index 095c5b7d2e..d10ca80462 100644
--- a/programs/system/drivers/ati2d/init_3d.inc
+++ b/programs/system/drivers/ati2d/init_3d.inc
@@ -635,7 +635,6 @@ int Init3DEngine(RHDPtr info)
 	FINISH_ACCEL();
     }
 
-
     safe_sti(ifl);
 
 }
diff --git a/programs/system/drivers/ati2d/init_cp.c b/programs/system/drivers/ati2d/init_cp.c
index 54607f3521..ad6cecade7 100644
--- a/programs/system/drivers/ati2d/init_cp.c
+++ b/programs/system/drivers/ati2d/init_cp.c
@@ -20,65 +20,9 @@
 
 
 #define RADEON_IDLE_RETRY      16 /* Fall out of idle loops after this count */
-#define RADEON_TIMEOUT    2000000 /* Fall out of wait loops after this count */
+#define RADEON_TIMEOUT    4000000 /* Fall out of wait loops after this count */
 
 
-void RADEONPllErrataAfterIndex()
-{
-    if (!(rhd.ChipErrata & CHIP_ERRATA_PLL_DUMMYREADS))
-       return;
-
-    /* This workaround is necessary on rv200 and RS200 or PLL
-     * reads may return garbage (among others...)
-     */
-    (void)INREG(RADEON_CLOCK_CNTL_DATA);
-    (void)INREG(RADEON_CRTC_GEN_CNTL);
-}
-
-
-void RADEONPllErrataAfterData()
-{
-
-    /* This function is required to workaround a hardware bug in some (all?)
-     * revisions of the R300.  This workaround should be called after every
-     * CLOCK_CNTL_INDEX register access.  If not, register reads afterward
-     * may not be correct.
-     */
-    if (rhd.ChipFamily <= CHIP_FAMILY_RV380)
-    {
-        u32_t save, tmp;
-
-	save = INREG(RADEON_CLOCK_CNTL_INDEX);
-	tmp = save & ~(0x3f | RADEON_PLL_WR_EN);
-	OUTREG(RADEON_CLOCK_CNTL_INDEX, tmp);
-	tmp = INREG(RADEON_CLOCK_CNTL_DATA);
-	OUTREG(RADEON_CLOCK_CNTL_INDEX, save);
-    }
-}
-
-
-/* Read PLL register */
-u32_t RADEONINPLL(int addr)
-{
-    u32_t       data;
-
-    OUTREG8(RADEON_CLOCK_CNTL_INDEX, addr & 0x3f);
-    RADEONPllErrataAfterIndex();
-    data = INREG(RADEON_CLOCK_CNTL_DATA);
-    RADEONPllErrataAfterData();
-
-    return data;
-};
-
-/* Write PLL information */
-void RADEONOUTPLL(int addr, u32_t data)
-{
-    OUTREG8(RADEON_CLOCK_CNTL_INDEX, (((addr) & 0x3f) |
-				      RADEON_PLL_WR_EN));
-    RADEONPllErrataAfterIndex();
-    OUTREG(RADEON_CLOCK_CNTL_DATA, data);
-    RADEONPllErrataAfterData();
-}
 
 void RADEONEngineFlush(RHDPtr info)
 {
@@ -148,61 +92,6 @@ static int radeon_do_wait_for_idle()
 }
 
 
-static void init_pipes(RHDPtr info)
-{
-     u32_t gb_tile_config = 0;
-
-     if ( (info->ChipFamily == CHIP_FAMILY_RV410) ||
-          (info->ChipFamily == CHIP_FAMILY_R420)  ||
-          (info->ChipFamily == CHIP_FAMILY_RS600) ||
-          (info->ChipFamily == CHIP_FAMILY_RS690) ||
-          (info->ChipFamily == CHIP_FAMILY_RS740) ||
-          (info->ChipFamily == CHIP_FAMILY_RS400) ||
-          (info->ChipFamily == CHIP_FAMILY_RS480) || IS_R500_3D)
-     {
-         u32_t gb_pipe_sel = INREG(R400_GB_PIPE_SELECT);
-
-         info->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
-         if (IS_R500_3D)
-            OUTPLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
-     }
-     else
-     {
-        if ((info->ChipFamily == CHIP_FAMILY_R300) ||
-            (info->ChipFamily == CHIP_FAMILY_R350))
-        {
-         /* R3xx chips */
-            info->num_gb_pipes = 2;
-        }
-        else {
-         /* RV3xx chips */
-           info->num_gb_pipes = 1;
-        }
-     }
-
-     if (IS_R300_3D || IS_R500_3D)
-     {
-
-        dbgprintf("num quad-pipes is %d\n", info->num_gb_pipes);
-
-        switch(info->num_gb_pipes) {
-           case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
-           case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
-           case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
-           default:
-           case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
-        }
-
-        OUTREG(R300_GB_TILE_CONFIG, gb_tile_config);
-        OUTREG(RADEON_WAIT_UNTIL, RADEON_WAIT_2D_IDLECLEAN | RADEON_WAIT_3D_IDLECLEAN);
-        OUTREG(R300_DST_PIPE_CONFIG, INREG(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
-        OUTREG(R300_RB2D_DSTCACHE_MODE, (INREG(R300_RB2D_DSTCACHE_MODE) |
-                                        R300_DC_AUTOFLUSH_ENABLE |
-                                        R300_DC_DC_DISABLE_IGNORE_PE));
-    }
-    else
-       OUTREG(RADEON_RB3D_CNTL, 0);
-};
 
 /* ================================================================
  * CP control, initialization
@@ -357,10 +246,10 @@ void init_ring_buffer(RHDPtr info)
 
 void radeon_engine_reset(RHDPtr info)
 {
-     u32_t  clock_cntl_index;
-     u32_t  mclk_cntl;
-     u32_t  rbbm_soft_reset;
-     u32_t  host_path_cntl;
+    u32_t  clock_cntl_index;
+    u32_t  mclk_cntl;
+    u32_t  rbbm_soft_reset;
+    u32_t  host_path_cntl;
 
     if (info->ChipFamily <= CHIP_FAMILY_RV410)
     {
@@ -501,6 +390,7 @@ Bool init_cp(RHDPtr info)
 
      radeon_cp_start(&rhd);
 
+    return TRUE;
 };
 
 
diff --git a/programs/system/drivers/ati2d/makefile b/programs/system/drivers/ati2d/makefile
index 85623ffe8a..5cac7babf6 100644
--- a/programs/system/drivers/ati2d/makefile
+++ b/programs/system/drivers/ati2d/makefile
@@ -22,8 +22,7 @@ SRC_DEP:=    init.c    		\
              blend.inc          \
              r500.inc           \
              pixmap.inc		\
-             accel_2d.inc  	\
-             accel_3d.inc
+             accel_2d.inc
 
 ATI_SRC:=    ati2d.c     
 
diff --git a/programs/system/drivers/ati2d/pci.c b/programs/system/drivers/ati2d/pci.c
index 8096087247..db7391acc8 100644
--- a/programs/system/drivers/ati2d/pci.c
+++ b/programs/system/drivers/ati2d/pci.c
@@ -34,83 +34,81 @@ const RADEONCardInfo *RadeonDevMatch(u16_t dev,const RADEONCardInfo *list)
 
 RHDPtr FindPciDevice()
 {
-  const RADEONCardInfo *dev;
-  u32_t bus, last_bus;
+    const RADEONCardInfo *dev;
+    u32_t bus, last_bus;
 
-  if( (last_bus = PciApi(1))==-1)
-    return 0;
+    if( (last_bus = PciApi(1))==-1)
+        return 0;
 
-  for(bus=0;bus<=last_bus;bus++)
-  {
-    u32_t devfn;
-
-    for(devfn=0;devfn<256;devfn++)
+    for(bus=0;bus<=last_bus;bus++)
     {
-      u32_t id;
-      id = PciRead32(bus,devfn, 0);
+        u32_t devfn;
 
-      if( (u16_t)id != VENDOR_ATI)
-        continue;
-
-      rhd.PciDeviceID = (id>>16);
-
-      if( (dev = RadeonDevMatch(rhd.PciDeviceID, RADEONCards))!=NULL)
-      {
-        u32_t reg2C;
-        int i;
-
-        rhd.chipset = (char*)xf86TokenToString(RADEONChipsets, rhd.PciDeviceID);
-        if (!rhd.chipset)
+        for(devfn=0;devfn<256;devfn++)
         {
-           dbgprintf("ChipID 0x%04x is not recognized\n", rhd.PciDeviceID);
-           return FALSE;
-        }
-        dbgprintf("Chipset: \"%s\" (ChipID = 0x%04x)\n",rhd.chipset,rhd.PciDeviceID);
+            u32_t id;
+            id = PciRead32(bus,devfn, 0);
 
-        rhd.bus = bus;
-        rhd.devfn = devfn;
-        rhd.PciTag = pciTag(bus,(devfn>>3)&0x1F,devfn&0x7);
+            if( (u16_t)id != VENDOR_ATI)
+                continue;
 
-        rhd.ChipFamily  = dev->chip_family;
-        rhd.IsMobility  = dev->mobility;
-        rhd.IsIGP       = dev->igp;
-        rhd.HasCRTC2    = !dev->nocrtc2;
+            rhd.PciDeviceID = (id>>16);
 
-        reg2C = PciRead32(bus,devfn, 0x2C);
-
-        rhd.subvendor_id = reg2C & 0xFFFF;;
-        rhd.subdevice_id = reg2C >> 16;
-
-        if (rhd.ChipFamily >= CHIP_FAMILY_R600)
-           dbgprintf("R600 unsupported yet.\nExit\n");
-
-
-        for (i = 0; i < 6; i++)
-        {
-          u32_t base;
-          Bool validSize;
-
-          base = PciRead32(bus,devfn, PCI_MAP_REG_START + (i << 2));
-          if(base)
-          {
-            if (base & PCI_MAP_IO)
+            if( (dev = RadeonDevMatch(rhd.PciDeviceID, RADEONCards))!=NULL)
             {
-              rhd.ioBase[i] = (u32_t)PCIGETIO(base);
-              rhd.memtype[i]   = base & PCI_MAP_IO_ATTR_MASK;
+                u32_t reg2C;
+                int i;
+
+                rhd.chipset = (char*)xf86TokenToString(RADEONChipsets, rhd.PciDeviceID);
+                if (!rhd.chipset){
+                    dbgprintf("ChipID 0x%04x is not recognized\n", rhd.PciDeviceID);
+                    return FALSE;
+                }
+                dbgprintf("Chipset: \"%s\" (ChipID = 0x%04x)\n",
+                           rhd.chipset,rhd.PciDeviceID);
+
+                rhd.bus = bus;
+                rhd.devfn = devfn;
+                rhd.PciTag = pciTag(bus,(devfn>>3)&0x1F,devfn&0x7);
+
+                rhd.ChipFamily  = dev->chip_family;
+                rhd.IsMobility  = dev->mobility;
+                rhd.IsIGP       = dev->igp;
+                rhd.HasCRTC2    = !dev->nocrtc2;
+
+                reg2C = PciRead32(bus,devfn, 0x2C);
+
+                rhd.subvendor_id = reg2C & 0xFFFF;;
+                rhd.subdevice_id = reg2C >> 16;
+
+                if (rhd.ChipFamily >= CHIP_FAMILY_R600)
+                    dbgprintf("R600 unsupported yet.\nExit\n");
+
+
+                for (i = 0; i < 6; i++)
+                {
+                    u32_t base;
+                    Bool validSize;
+
+                    base = PciRead32(bus,devfn, PCI_MAP_REG_START + (i << 2));
+                    if(base)
+                    {
+                        if (base & PCI_MAP_IO){
+                            rhd.ioBase[i] = (u32_t)PCIGETIO(base);
+                            rhd.memtype[i]   = base & PCI_MAP_IO_ATTR_MASK;
+                        }
+                        else{
+                            rhd.memBase[i] = (u32_t)PCIGETMEMORY(base);
+                            rhd.memtype[i] = base & PCI_MAP_MEMORY_ATTR_MASK;
+                        }
+                    }
+                    rhd.memsize[i] = pciGetBaseSize(bus,devfn, i, TRUE, &validSize);
+                }
+                return &rhd;
             }
-            else
-            {
-              rhd.memBase[i] = (u32_t)PCIGETMEMORY(base);
-              rhd.memtype[i] = base & PCI_MAP_MEMORY_ATTR_MASK;
-            }
-          }
-          rhd.memsize[i] = pciGetBaseSize(bus,devfn, i, TRUE, &validSize);
         }
-        return &rhd;
-      }
     };
-  };
-  return NULL;
+    return NULL;
 }
 
 
diff --git a/programs/system/drivers/ati2d/pixmap.inc b/programs/system/drivers/ati2d/pixmap.inc
index a397c367f8..edd9e8958c 100644
--- a/programs/system/drivers/ati2d/pixmap.inc
+++ b/programs/system/drivers/ati2d/pixmap.inc
@@ -50,7 +50,7 @@ int CreatePixmap(pixmap_t *io)
            pixmap->width   = io->width;
            pixmap->height  = io->height;
            pixmap->format  = PICT_a8r8g8b8;
-           pixmap->flags   = io->flags;
+           pixmap->flags   = PX_MEM_LOCAL; //io->flags;
            pixmap->pitch   = pitch;
            pixmap->mapped  = mapped;
            pixmap->pitch_offset =  ((pitch/64)<<22)| (((u32_t)local+rhd.fbLocation)>>10);
diff --git a/programs/system/drivers/ati2d/r500.inc b/programs/system/drivers/ati2d/r500.inc
index 42bab0ca64..fc1506d760 100644
--- a/programs/system/drivers/ati2d/r500.inc
+++ b/programs/system/drivers/ati2d/r500.inc
@@ -39,8 +39,8 @@ R5xx2DFlush()
                 R5XX_DSTCACHE_FLUSH_ALL, R5XX_DSTCACHE_FLUSH_ALL);
 
     for (i = 0; i < R5XX_LOOP_COUNT; i++)
-      if (!(INREG(R5XX_DSTCACHE_CTLSTAT) & R5XX_DSTCACHE_BUSY))
-        return TRUE;
+        if (!(INREG(R5XX_DSTCACHE_CTLSTAT) & R5XX_DSTCACHE_BUSY))
+            return TRUE;
 
     dbgprintf("%s: Timeout 0x%08x.\n", __func__,
          (unsigned int)INREG(R5XX_DSTCACHE_CTLSTAT));
@@ -54,24 +54,23 @@ R5xx2DIdleLocal()                                //R100-R500
 
       /* wait for fifo to clear */
     for (i = 0; i < R5XX_LOOP_COUNT; i++)
-      if (64 == (INREG(R5XX_RBBM_STATUS) & R5XX_RBBM_FIFOCNT_MASK))
-        break;
+        if (64 == (INREG(R5XX_RBBM_STATUS) & R5XX_RBBM_FIFOCNT_MASK))
+            break;
 
     if (i == R5XX_LOOP_COUNT) {
-      dbgprintf("%s: FIFO Timeout 0x%08X.\n", __func__,INREG(R5XX_RBBM_STATUS));
-      return FALSE;
+        dbgprintf("%s: FIFO Timeout 0x%08X.\n", __func__,INREG(R5XX_RBBM_STATUS));
+        return FALSE;
     }
 
       /* wait for engine to go idle */
     for (i = 0; i < R5XX_LOOP_COUNT; i++) {
-      if (!(INREG(R5XX_RBBM_STATUS) & R5XX_RBBM_ACTIVE)) {
-        R5xx2DFlush();
-        return TRUE;
-      }
+        if (!(INREG(R5XX_RBBM_STATUS) & R5XX_RBBM_ACTIVE)) {
+            R5xx2DFlush();
+            return TRUE;
+        }
     }
     dbgprintf("%s: Idle Timeout 0x%08X.\n", __func__,INREG(R5XX_RBBM_STATUS));
     return FALSE;
-
 }
 
 
@@ -174,6 +173,7 @@ void  R5xx2DInit()
     scr_pixmap.width  = rhd.displayWidth;
     scr_pixmap.height = rhd.displayHeight;
     scr_pixmap.format = PICT_a8r8g8b8;
+    scr_pixmap.flags  = PX_MEM_LOCAL;
     scr_pixmap.pitch  = rhd.displayWidth * 4     ;//screenpitch;
     scr_pixmap.local  = (void*)rhd.fbLocation;
     scr_pixmap.pitch_offset =  rhd.dst_pitch_offset;
@@ -188,9 +188,10 @@ void  R5xx2DInit()
 
     OUTREG(R5XX_SURFACE_CNTL, rhd.surface_cntl);
 
-#if R300_PIO
-#else
+#if !R300_PIO
+
     init_cp(&rhd);
+
 #endif
 
     R5xx2DSetup();