Merged the tdfx-2-0 branch onto the trunktdfx-2-0-20000612-merge

27 files changed, 3901 insertions, 1183 deletions

diff --git a/xc/extras/Mesa/src/FX/fxapi.c b/xc/extras/Mesa/src/FX/fxapi.c
index da2b905fa..6d680bcb9 100644
--- a/xc/extras/Mesa/src/FX/fxapi.c
+++ b/xc/extras/Mesa/src/FX/fxapi.c
@@ -940,10 +940,8 @@ fxMesaContext GLAPIENTRY fxMesaCreateContext(GLuint win,
    else
       fxMesa->haveTwoTMUs=GL_FALSE;
 
-   fxMesa->haveDoubleBuffer=doubleBuffer;
    fxMesa->haveAlphaBuffer=alphaBuffer;
    fxMesa->haveGlobalPaletteTexture=GL_FALSE;
-   fxMesa->haveZBuffer=depthSize ? 1 : 0;
    fxMesa->verbose=verbose;
    fxMesa->board=glbCurrentBoard;
 
@@ -1297,7 +1295,7 @@ void GLAPIENTRY fxMesaSwapBuffers(void)
   if(fxMesaCurrentCtx) {
    FLUSH_VB( fxMesaCurrentCtx->glCtx, "swap buffers" );
 
-    if(fxMesaCurrentCtx->haveDoubleBuffer) {
+    if (fxMesaCurrentCtx->glVis->DBflag) {
 
       grBufferSwap(fxMesaCurrentCtx->swapInterval);
 
diff --git a/xc/extras/Mesa/src/FX/fxdd.c b/xc/extras/Mesa/src/FX/fxdd.c
index 2705d6bc1..59a4bec93 100644
--- a/xc/extras/Mesa/src/FX/fxdd.c
+++ b/xc/extras/Mesa/src/FX/fxdd.c
@@ -52,6 +52,7 @@
 
 #if defined(FX)
 
+#include <dlfcn.h>
 #include "image.h"
 #include "types.h"
 #include "fxdrv.h"
@@ -102,34 +103,20 @@ void fxInitPixelTables(fxMesaContext fxMesa, GLboolean bgrOrder)
 /*****                 Miscellaneous functions                    *****/
 /**********************************************************************/
 
-/* Enalbe/Disable dithering */
-static void fxDDDither(GLcontext *ctx, GLboolean enable)
-{
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
-    fprintf(stderr,"fxmesa: fxDDDither()\n");
-  }
-
-  if (enable) {
-    FX_grDitherMode(GR_DITHER_4x4);
-  } else {
-    FX_grDitherMode(GR_DITHER_DISABLE);
-  }
-}
-
 
 /* Return buffer size information */
 static void fxDDBufferSize(GLcontext *ctx, GLuint *width, GLuint *height)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
 
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr,"fxmesa: fxDDBufferSize(...) Start\n");
   }
 
-  *width=fxMesa->width;
-  *height=fxMesa->height;
+  *width = fxMesa->width;
+  *height = fxMesa->height;
 
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr,"fxmesa: fxDDBufferSize(...) End\n");
   }
 }
@@ -139,15 +126,13 @@ static void fxDDBufferSize(GLcontext *ctx, GLuint *width, GLuint *height)
 static void fxDDSetColor(GLcontext *ctx, GLubyte red, GLubyte green,
                          GLubyte blue, GLubyte alpha )
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
   GLubyte col[4];
   ASSIGN_4V( col, red, green, blue, alpha );
-  
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr,"fxmesa: fxDDSetColor(%d,%d,%d,%d)\n",red,green,blue,alpha);
   }
-
-  fxMesa->color=FXCOLOR4(col);
+  fxMesa->color = FXCOLOR4(col);
 }
 
 
@@ -155,19 +140,14 @@ static void fxDDSetColor(GLcontext *ctx, GLubyte red, GLubyte green,
 static void fxDDClearColor(GLcontext *ctx, GLubyte red, GLubyte green,
                            GLubyte blue, GLubyte alpha )
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
   GLubyte col[4];
-
-
-
   ASSIGN_4V( col, red, green, blue, 255 );
-
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr,"fxmesa: fxDDClearColor(%d,%d,%d,%d)\n",red,green,blue,alpha);
   }
- 
-  fxMesa->clearC=FXCOLOR4( col );
-  fxMesa->clearA=alpha;
+  fxMesa->clearC = FXCOLOR4( col );
+  fxMesa->clearA = alpha;
 }
 
 
@@ -177,11 +157,18 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   const GLuint colorMask = *((GLuint *) &ctx->Color.ColorMask);
-  const FxU16 clearD = (FxU16) (ctx->Depth.Clear * 0xffff);
-  GLbitfield softwareMask = mask & (DD_STENCIL_BIT | DD_ACCUM_BIT);
+  const FxU32 clearD = (FxU32) (ctx->Depth.Clear * fxMesa->depthClear);
+  const FxU32 clearS = (FxU32) (ctx->Stencil.Clear);
+  GLbitfield softwareMask = mask & (DD_ACCUM_BIT);
 
-  /* we can't clear stencil or accum buffers */
-  mask &= ~(DD_STENCIL_BIT | DD_ACCUM_BIT);
+  /* we can't clear accum buffers */
+  mask &= ~(DD_ACCUM_BIT);
+
+  if ((mask & DD_STENCIL_BIT) && !fxMesa->haveHwStencil) {
+    /* software stencil buffer */
+    mask &= ~(DD_STENCIL_BIT);
+    softwareMask |= DD_STENCIL_BIT;
+  }
 
   if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr,"fxmesa: fxDDClear(%d,%d,%d,%d)\n", (int) x, (int) y,
@@ -194,16 +181,35 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
     mask &= ~(DD_FRONT_LEFT_BIT | DD_BACK_LEFT_BIT);
   }
 
+
+  /* disable stencil ops if enabled (it screws up clearing) */
+  if (ctx->Stencil.Enabled)
+    FX_grDisable(GR_STENCIL_MODE_EXT);
+
+  if (fxMesa->haveHwStencil) {
+    if (mask & DD_STENCIL_BIT) {
+      FX_grStencilMask(0xff);
+      /* set stencil ref value = desired clear value */
+      FX_grStencilFunc(GR_CMP_ALWAYS, ctx->Stencil.Clear, 0xff);
+    }
+    else {
+      FX_grStencilMask(0x00);
+    }
+  }
+
   /*
    * This could probably be done fancier but doing each possible case
    * explicitly is less error prone.
    */
-  switch (mask) {
+  switch (mask & ~DD_STENCIL_BIT) {
     case DD_BACK_LEFT_BIT | DD_DEPTH_BIT:
       /* back buffer & depth */
       FX_grDepthMask(FXTRUE);
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       if (!ctx->Depth.Mask) {
         FX_grDepthMask(FXFALSE);
       }
@@ -217,17 +223,26 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       FX_grDepthMask(FXTRUE);
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
       FX_grColorMask(FXFALSE,FXFALSE);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       /* clear front */
       FX_grColorMask(FXTRUE, ctx->Color.ColorMask[ACOMP] && fxMesa->haveAlphaBuffer);
       FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       break;
     case DD_BACK_LEFT_BIT:
       /* back buffer only */
       FX_grDepthMask(FXFALSE);
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       if (ctx->Depth.Mask) {
         FX_grDepthMask(FXTRUE);
       }
@@ -236,7 +251,10 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       /* front buffer only */
       FX_grDepthMask(FXFALSE);
       FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       if (ctx->Depth.Mask) {
         FX_grDepthMask(FXTRUE);
       }
@@ -245,9 +263,15 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       /* front and back */
       FX_grDepthMask(FXFALSE);
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       if (ctx->Depth.Mask) {
         FX_grDepthMask(FXTRUE);
       }
@@ -256,11 +280,17 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       /* clear front */
       FX_grDepthMask(FXFALSE);
       FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       /* clear back and depth */
       FX_grDepthMask(FXTRUE);
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       if (!ctx->Depth.Mask) {
         FX_grDepthMask(FXFALSE);
       }
@@ -270,7 +300,10 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
       FX_grColorMask(FXFALSE,FXFALSE);
       FX_grDepthMask(FXTRUE);
-      FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
+      if (mask & DD_STENCIL_BIT)
+        FX_grBufferClearExt(fxMesa->clearC, fxMesa->clearA, clearD, clearS);
+      else
+        FX_grBufferClear(fxMesa->clearC, fxMesa->clearA, clearD);
       FX_grColorMask(FXTRUE, ctx->Color.ColorMask[ACOMP] && fxMesa->haveAlphaBuffer);
       if (ctx->Color.DrawDestMask & FRONT_LEFT_BIT)
         FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
@@ -280,6 +313,14 @@ static GLbitfield fxDDClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
       ;
   }
 
+  if (ctx->Stencil.Enabled) {
+    /* restore stencil state to as it was before the clear */
+    FX_grEnable(GR_STENCIL_MODE_EXT);
+    FX_grStencilMask(ctx->Stencil.WriteMask);
+    FX_grStencilFunc(ctx->Stencil.Function - GL_NEVER,
+                     ctx->Stencil.Ref, ctx->Stencil.ValueMask);
+  }
+
   return softwareMask;
 }
 
@@ -337,6 +378,58 @@ static void fxDDSetReadBuffer(GLcontext *ctx, GLframebuffer *buffer,
 }
 
 
+/*
+ * These functions just set new-state flags.  The exact state
+ * values will be evaluated later.
+ */
+static void
+fxDDStencilFunc(GLcontext *ctx, GLenum func, GLint ref, GLuint mask)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) func;  (void) ref;  (void) mask;
+  fxMesa->new_state |= FX_NEW_STENCIL;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+static void
+fxDDStencilMask(GLcontext *ctx, GLuint mask)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) mask;
+  fxMesa->new_state |= FX_NEW_STENCIL;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+static void
+fxDDStencilOp(GLcontext *ctx, GLenum sfail, GLenum zfail, GLenum zpass)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) sfail;  (void) zfail;  (void) zpass;
+  fxMesa->new_state |= FX_NEW_STENCIL;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+static void
+fxDDDepthFunc(GLcontext *ctx, GLenum func)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) func;
+  fxMesa->new_state |= FX_NEW_DEPTH;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+static void
+fxDDDepthMask(GLcontext *ctx, GLboolean mask)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) mask;
+  fxMesa->new_state |= FX_NEW_DEPTH;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+
+
+
 #ifdef XF86DRI
 /* test if window coord (px,py) is visible */
 static GLboolean inClipRects(fxMesaContext fxMesa, int px, int py)
@@ -353,10 +446,12 @@ static GLboolean inClipRects(fxMesaContext fxMesa, int px, int py)
 #endif
 
 
-static GLboolean fxDDDrawBitmap(GLcontext *ctx, GLint px, GLint py,
-                                GLsizei width, GLsizei height,
-                                const struct gl_pixelstore_attrib *unpack,
-                                const GLubyte *bitmap)
+
+static GLboolean
+bitmap_R5G6B5(GLcontext *ctx, GLint px, GLint py,
+              GLsizei width, GLsizei height,
+              const struct gl_pixelstore_attrib *unpack,
+              const GLubyte *bitmap)
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GrLfbInfo_t info;
@@ -527,11 +622,186 @@ static GLboolean fxDDDrawBitmap(GLcontext *ctx, GLint px, GLint py,
 }
 
 
-static GLboolean fxDDReadPixels( GLcontext *ctx, GLint x, GLint y,
-                                 GLsizei width, GLsizei height,
-                                 GLenum format, GLenum type,
-                                 const struct gl_pixelstore_attrib *packing,
-                                 GLvoid *dstImage )
+static GLboolean
+bitmap_R8G8B8A8(GLcontext *ctx, GLint px, GLint py,
+                GLsizei width, GLsizei height,
+                const struct gl_pixelstore_attrib *unpack,
+                const GLubyte *bitmap)
+{
+  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  GrLfbInfo_t info;
+  GrLfbWriteMode_t mode;
+  /*FxU16 color;*/
+  GLuint color;
+  const struct gl_pixelstore_attrib *finalUnpack;
+  struct gl_pixelstore_attrib scissoredUnpack;
+
+  /* check if there's any raster operations enabled which we can't handle */
+  if (ctx->RasterMask & (ALPHATEST_BIT |
+                         BLEND_BIT |
+                         DEPTH_BIT |
+                         FOG_BIT |
+                         LOGIC_OP_BIT |
+                         SCISSOR_BIT |
+                         STENCIL_BIT |
+                         MASKING_BIT |
+                         ALPHABUF_BIT |
+                         MULTI_DRAW_BIT))
+    return GL_FALSE;
+
+  if (ctx->Scissor.Enabled) {
+    /* This is a bit tricky, but by carefully adjusting the px, py,
+     * width, height, skipPixels and skipRows values we can do
+     * scissoring without special code in the rendering loop.
+     */
+
+    /* we'll construct a new pixelstore struct */
+    finalUnpack = &scissoredUnpack;
+    scissoredUnpack = *unpack;
+    if (scissoredUnpack.RowLength == 0)
+      scissoredUnpack.RowLength = width;
+
+    /* clip left */
+    if (px < ctx->Scissor.X) {
+      scissoredUnpack.SkipPixels += (ctx->Scissor.X - px);
+      width -= (ctx->Scissor.X - px);
+      px = ctx->Scissor.X;
+    }
+    /* clip right */
+    if (px + width >= ctx->Scissor.X + ctx->Scissor.Width) {
+      width -= (px + width - (ctx->Scissor.X + ctx->Scissor.Width));
+    }
+    /* clip bottom */
+    if (py < ctx->Scissor.Y) {
+      scissoredUnpack.SkipRows += (ctx->Scissor.Y - py);
+      height -= (ctx->Scissor.Y - py);
+      py = ctx->Scissor.Y;
+    }
+    /* clip top */
+    if (py + height >= ctx->Scissor.Y + ctx->Scissor.Height) {
+      height -= (py + height - (ctx->Scissor.Y + ctx->Scissor.Height));
+    }
+
+    if (width <= 0 || height <= 0)
+      return GL_TRUE;  /* totally scissored away */
+  }  
+  else {
+    finalUnpack = unpack;
+  }
+
+  /* compute pixel value */
+  {
+    GLint r = (GLint) (ctx->Current.RasterColor[0] * 255.0f);
+    GLint g = (GLint) (ctx->Current.RasterColor[1] * 255.0f);
+    GLint b = (GLint) (ctx->Current.RasterColor[2] * 255.0f);
+    GLint a = (GLint) (ctx->Current.RasterColor[3] * 255.0f);
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+      color = PACK_BGRA32(r, g, b, a);
+    else
+      color = PACK_RGBA32(r, g, b, a);
+  }
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888;
+
+  info.size = sizeof(info);
+  if (!FX_grLfbLock(GR_LFB_WRITE_ONLY,
+                    fxMesa->currentFB,
+                    mode,
+                    GR_ORIGIN_UPPER_LEFT,
+                    FXFALSE,
+                    &info)) {
+#ifndef FX_SILENT
+    fprintf(stderr,"fx Driver: error locking the linear frame buffer\n");
+#endif
+    return GL_TRUE;
+  }
+
+#ifdef XF86DRI
+#define INSIDE(c, x, y) inClipRects((c), (x), (y))
+#else
+#define INSIDE(c, x, y) (1)
+#endif
+
+  {
+    const GLint winX = fxMesa->x_offset;
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    GLint dstStride;
+    GLuint *dst;
+    GLint row;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      dstStride = fxMesa->screen_width;
+      dst = (GLuint *) info.lfbPtr + (winY - py) * dstStride + (winX + px);
+    }
+    else {
+      dstStride = info.strideInBytes / 4;
+      dst = (GLuint *) info.lfbPtr + (winY - py) * dstStride + (winX + px);
+    }
+
+    /* compute dest address of bottom-left pixel in bitmap */
+    for (row = 0; row < height; row++) {
+      const GLubyte *src = (const GLubyte *) _mesa_image_address( finalUnpack,
+                 bitmap, width, height, GL_COLOR_INDEX, GL_BITMAP, 0, row, 0 );
+      if (finalUnpack->LsbFirst) {
+        /* least significan bit first */
+        GLubyte mask = 1U << (finalUnpack->SkipPixels & 0x7);
+        GLint col;
+        for (col=0; col<width; col++) {
+          if (*src & mask) {
+            if (INSIDE(fxMesa, winX + px + col, winY - py - row))
+              dst[col] = color;
+          }
+          if (mask == 128U) {
+            src++;
+            mask = 1U;
+          }
+          else {
+            mask = mask << 1;
+          }
+        }
+        if (mask != 1)
+          src++;
+      }
+      else {
+        /* most significan bit first */
+        GLubyte mask = 128U >> (finalUnpack->SkipPixels & 0x7);
+        GLint col;
+        for (col=0; col<width; col++) {
+          if (*src & mask) {
+            if (INSIDE(fxMesa, winX + px + col, winY - py - row))
+              dst[col] = color;
+          }
+          if (mask == 1U) {
+            src++;
+            mask = 128U;
+          }
+          else {
+            mask = mask >> 1;
+          }
+        }
+        if (mask != 128)
+          src++;
+      }
+      dst -= dstStride;
+    }
+  }
+
+#undef INSIDE
+
+  FX_grLfbUnlock(GR_LFB_WRITE_ONLY,fxMesa->currentFB);
+  return GL_TRUE;
+}
+
+
+static GLboolean
+readpixels_R5G6B5( GLcontext *ctx, GLint x, GLint y,
+                   GLsizei width, GLsizei height,
+                   GLenum format, GLenum type,
+                   const struct gl_pixelstore_attrib *packing,
+                   GLvoid *dstImage )
 {
   if (ctx->Pixel.ScaleOrBiasRGBA || ctx->Pixel.MapColorFlag) {
     return GL_FALSE;  /* can't do this */
@@ -542,6 +812,7 @@ static GLboolean fxDDReadPixels( GLcontext *ctx, GLint x, GLint y,
     GLboolean result = GL_FALSE;
 
     BEGIN_BOARD_LOCK();
+    info.size=sizeof(info);
     if (grLfbLock(GR_LFB_READ_ONLY,
                   fxMesa->currentFB,
                   GR_LFBWRITEMODE_ANY,
@@ -700,7 +971,7 @@ static const GLubyte *fxDDGetString(GLcontext *ctx, GLenum name)
           }
         }
         /* now make the GL_RENDERER string */
-        sprintf(buffer, "Mesa DRI %s 20000510", hardware);
+        sprintf(buffer, "Mesa DRI %s 20000608", hardware);
         return buffer;
       }
     case GL_VENDOR:
@@ -757,6 +1028,19 @@ static const GLubyte *fxDDGetString(GLcontext *ctx, GLenum name)
 
 int fxDDInitFxMesaContext( fxMesaContext fxMesa )
 {
+  /* Get Glide3vn function pointers */
+  {
+    void *handle;
+    handle = dlopen(NULL, RTLD_NOW | RTLD_GLOBAL);
+    if (!handle)
+      return 0;
+    grStencilFuncPtr = dlsym(handle, "grStencilFunc");
+    grStencilMaskPtr = dlsym(handle, "grStencilMask");
+    grStencilOpPtr = dlsym(handle, "grStencilOp");
+    grBufferClearExtPtr = dlsym(handle, "grBufferClearExt");
+    /* call dlclose()? */
+  }
+
   
    FX_setupGrVertexLayout();
    
@@ -783,6 +1067,8 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
    else
       fxMesa->verbose=GL_FALSE;
 
+   fxMesa->depthClear = FX_grGetInteger(FX_ZDEPTH_MAX);
+
    fxMesa->color=0xffffffff;
    fxMesa->clearC=0;
    fxMesa->clearA=0;
@@ -793,7 +1079,6 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
    fxMesa->stats.memTexUpload=0;
 
    fxMesa->tmuSrc=FX_TMU_NONE;
-   fxMesa->lastUnitsMode=FX_UM_NONE;
    fxTMInit(fxMesa);
 
    /* FX units setup */
@@ -808,16 +1093,19 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
    fxMesa->unitsState.blendSrcFuncAlpha=GR_BLEND_ONE;
    fxMesa->unitsState.blendDstFuncAlpha=GR_BLEND_ZERO;
 
+   /*
    fxMesa->unitsState.depthTestEnabled	=GL_FALSE;
    fxMesa->unitsState.depthMask		=GL_TRUE;
    fxMesa->unitsState.depthTestFunc	=GR_CMP_LESS;
+   */
 
    FX_grColorMask(FXTRUE, fxMesa->haveAlphaBuffer ? FXTRUE : FXFALSE);
-   if(fxMesa->haveDoubleBuffer) {
-      fxMesa->currentFB=GR_BUFFER_BACKBUFFER;
+   if (fxMesa->glVis->DBflag) {
+      fxMesa->currentFB = GR_BUFFER_BACKBUFFER;
       FX_grRenderBuffer(GR_BUFFER_BACKBUFFER);
-   } else {
-      fxMesa->currentFB=GR_BUFFER_FRONTBUFFER;
+   }
+   else {
+      fxMesa->currentFB = GR_BUFFER_FRONTBUFFER;
       FX_grRenderBuffer(GR_BUFFER_FRONTBUFFER);
    }
   
@@ -833,7 +1121,7 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
       return 0;
    }
 
-   if(fxMesa->haveZBuffer)
+   if (fxMesa->glVis->DepthBits > 0)
       FX_grDepthBufferMode(GR_DEPTHBUFFER_ZBUFFER);
     
 #if (!FXMESA_USE_ARGB)
@@ -846,7 +1134,7 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
    fxMesa->glCtx->Const.MaxTextureUnits=fxMesa->emulateTwoTMUs ? 2 : 1;
    fxMesa->glCtx->NewState|=NEW_DRVSTATE1;
    fxMesa->new_state = NEW_ALL;
-  
+
    fxDDSetupInit();
    fxDDCvaInit();
    fxDDClipInit();
@@ -876,6 +1164,10 @@ int fxDDInitFxMesaContext( fxMesaContext fxMesa )
 	 fxMesa->glCtx->PipelineStage,
 	 fxMesa->glCtx->NrPipelineStages);
 
+   /* this little bit ensures that all Glide state gets initialized */
+   fxMesa->new_state = NEW_ALL;
+   fxMesa->glCtx->Driver.RenderStart = fxSetupFXUnits;
+
    /* Run the config file */
    gl_context_initialize( fxMesa->glCtx );
 
@@ -946,8 +1238,7 @@ static GLboolean fxIsInHardware(GLcontext *ctx)
   if (!ctx->Hint.AllowDrawMem)
      return GL_TRUE;		/* you'll take it and like it */
 
-  if((ctx->RasterMask & (STENCIL_BIT | MULTI_DRAW_BIT)) ||
-     ((ctx->Color.BlendEnabled) && (ctx->Color.BlendEquation!=GL_FUNC_ADD_EXT)) ||
+  if(((ctx->Color.BlendEnabled) && (ctx->Color.BlendEquation!=GL_FUNC_ADD_EXT)) ||
      ((ctx->Color.ColorLogicOpEnabled) && (ctx->Color.LogicOp!=GL_COPY)) ||
      (ctx->Light.Model.ColorControl==GL_SEPARATE_SPECULAR_COLOR) ||
      (!((ctx->Color.ColorMask[RCOMP]==ctx->Color.ColorMask[GCOMP]) &&
@@ -1031,6 +1322,9 @@ static GLboolean fxIsInHardware(GLcontext *ctx)
     }
   }
 
+  if (ctx->Stencil.Enabled && !fxMesa->haveHwStencil)
+    return GL_FALSE;
+
   return GL_TRUE;
 }
 
@@ -1082,44 +1376,47 @@ static void fxDDReducedPrimitiveChange(GLcontext *ctx, GLenum prim)
   }
 }
 
+
 void fxSetupDDPointers(GLcontext *ctx)
 {
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
-    fprintf(stderr,"fxmesa: fxSetupDDPointers()\n");
-  }
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
 
-  ctx->Driver.UpdateState=fxDDUpdateDDPointers;
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
+    fprintf(stderr, "fxmesa: fxSetupDDPointers()\n");
+  }
+  ctx->Driver.UpdateState = fxDDUpdateDDPointers;
+  ctx->Driver.ClearIndex = NULL;
+  ctx->Driver.ClearColor = fxDDClearColor;
+  ctx->Driver.Clear = fxDDClear;
+  ctx->Driver.Index = NULL;
+  ctx->Driver.Color = fxDDSetColor;
+  ctx->Driver.SetDrawBuffer = fxDDSetDrawBuffer;
+  ctx->Driver.SetReadBuffer = fxDDSetReadBuffer;
+  ctx->Driver.GetBufferSize = fxDDBufferSize;
+  ctx->Driver.Finish = fxDDFinish;
+  ctx->Driver.Flush = NULL;
+  ctx->Driver.GetString = fxDDGetString;
+  ctx->Driver.NearFar = fxDDSetNearFar;
+  ctx->Driver.GetParameteri = fxDDGetParameteri;
 
   ctx->Driver.WriteDepthSpan=fxDDWriteDepthSpan;
   ctx->Driver.WriteDepthPixels=fxDDWriteDepthPixels;
   ctx->Driver.ReadDepthSpan=fxDDReadDepthSpan;
   ctx->Driver.ReadDepthPixels=fxDDReadDepthPixels;
          
-  ctx->Driver.GetString=fxDDGetString;
-
-  ctx->Driver.Dither=fxDDDither;
-
-  ctx->Driver.NearFar=fxDDSetNearFar;
-
-  ctx->Driver.GetParameteri=fxDDGetParameteri;
-
-  ctx->Driver.ClearIndex=NULL;
-  ctx->Driver.ClearColor=fxDDClearColor;
-  ctx->Driver.Clear=fxDDClear;
-
-  ctx->Driver.Index=NULL;
-  ctx->Driver.Color=fxDDSetColor;
-
-  ctx->Driver.SetDrawBuffer=fxDDSetDrawBuffer;
-  ctx->Driver.SetReadBuffer=fxDDSetReadBuffer;
-  ctx->Driver.GetBufferSize=fxDDBufferSize;
-
-  ctx->Driver.Bitmap=fxDDDrawBitmap;
-  ctx->Driver.DrawPixels=NULL;
-  ctx->Driver.ReadPixels=fxDDReadPixels;
-
-  ctx->Driver.Finish=fxDDFinish;
-  ctx->Driver.Flush=NULL;
+  if (ctx->Visual->RedBits == 8 &&
+      ctx->Visual->GreenBits == 8 &&
+      ctx->Visual->BlueBits == 8 &&
+      ctx->Visual->AlphaBits == 8) {
+    ctx->Driver.Bitmap = bitmap_R8G8B8A8;
+    ctx->Driver.DrawPixels = NULL;
+    ctx->Driver.ReadPixels = NULL;
+  }
+  else {
+    ctx->Driver.Bitmap = bitmap_R5G6B5;
+    ctx->Driver.DrawPixels = NULL;
+    ctx->Driver.ReadPixels = readpixels_R5G6B5;
+  }
 
   ctx->Driver.RenderStart=NULL;
   ctx->Driver.RenderFinish=NULL;
@@ -1135,6 +1432,12 @@ void fxSetupDDPointers(GLcontext *ctx)
 
   ctx->Driver.RectFunc=NULL;
 
+  if (fxMesa->haveHwStencil) {
+    ctx->Driver.StencilFunc = fxDDStencilFunc;
+    ctx->Driver.StencilMask = fxDDStencilMask;
+    ctx->Driver.StencilOp   = fxDDStencilOp;
+  }
+
   ctx->Driver.AlphaFunc=fxDDAlphaFunc;
   ctx->Driver.BlendFunc=fxDDBlendFunc;
   ctx->Driver.DepthFunc=fxDDDepthFunc;
@@ -1171,6 +1474,7 @@ void fxSetupDDPointers(GLcontext *ctx)
 }
 
 
+
 #else
 
 
diff --git a/xc/extras/Mesa/src/FX/fxddspan.c b/xc/extras/Mesa/src/FX/fxddspan.c
index c620cc3c6..2c53da8d3 100644
--- a/xc/extras/Mesa/src/FX/fxddspan.c
+++ b/xc/extras/Mesa/src/FX/fxddspan.c
@@ -46,33 +46,193 @@
 /* fxdd.c - 3Dfx VooDoo Mesa span and pixel functions */
 
 
-#ifdef HAVE_CONFIG_H
-#include "conf.h"
-#endif
+#include "fxdrv.h"
 
-#if defined(FX)
 
-#include "fxdrv.h"
 
-#ifdef _MSC_VER
-#ifdef _WIN32
-#pragma warning( disable : 4090 4022 )
-/* 4101 : "different 'const' qualifier"
- * 4022 : "pointer mistmatch for actual parameter 'n'
+/*
+ * Examine the cliprects to generate an array of flags to indicate
+ * which pixels in a span are visible.  Note: (x,y) is a screen
+ * coordinate.
  */
-#endif
-#endif
+static void
+generate_vismask(const fxMesaContext fxMesa, GLint x, GLint y, GLint n,
+                 GLubyte vismask[])
+{
+   GLboolean initialized = GL_FALSE;
+   GLint i, j;
+
+   /* turn on flags for all visible pixels */
+   for (i = 0; i < fxMesa->numClipRects; i++) {
+      const XF86DRIClipRectPtr rect = &fxMesa->pClipRects[i];
+
+      if (y >= rect->y1 && y < rect->y2) {
+         if (x >= rect->x1 && x + n <= rect->x2) {
+            /* common case, whole span inside cliprect */
+            MEMSET(vismask, 1, n);
+            return;
+         }
+         if (x < rect->x2 && x + n >= rect->x1) {
+            /* some of the span is inside the rect */
+            GLint start, end;
+            if (!initialized) {
+               MEMSET(vismask, 0, n);
+               initialized = GL_TRUE;
+            }
+            if (x < rect->x1)
+               start = rect->x1 - x;
+            else
+               start = 0;
+            if (x + n > rect->x2)
+               end = rect->x2 - x;
+            else
+               end = n;
+            assert(start >= 0);
+            assert(end <= n);
+            for (j = start; j < end; j++)
+               vismask[j] = 1;
+         }
+      }
+   }
+}
 
 
-#if !defined(FXMESA_USE_ARGB) 
+/*
+ * Examine cliprects and determine if the given screen pixel is visible.
+ */
+static GLboolean
+visible_pixel(const fxMesaContext fxMesa, int scrX, int scrY)
+{
+   int i;
+   for (i = 0; i < fxMesa->numClipRects; i++) {
+      const XF86DRIClipRectPtr rect = &fxMesa->pClipRects[i];
+      if (scrX >= rect->x1 &&
+          scrX <  rect->x2 &&
+          scrY >= rect->y1 &&
+          scrY <  rect->y2)
+        return GL_TRUE;
+   }
+   return GL_FALSE;
+}
 
 
-#if defined(FX_GLIDE3) && defined(XF86DRI)
+
+typedef enum { FBS_READ, FBS_WRITE } FBS_DIRECTION;
+/*
+ *     Read or write a single span from the frame buffer.
+ *     Input Parameters:
+ *         fxMesa:                The context
+ *         target_buffer:         Which buffer to read from.
+ *         xpos, ypos:            Starting Coordinates.
+ *         xlength:               Length of the span.
+ *         data_size:             Size of data elements: 1, 2, or 4.
+ *         buffer:                Pointer to the buffer to read
+ *                                to or write from.  This needs
+ *                                to be turned into a pointer.
+ *         direction:             Read or Write.
+ */
+static void
+rw_fb_span(fxMesaContext fxMesa,
+           GrBuffer_t    target_buffer,
+           FxU32         xpos,
+           FxU32         ypos,
+           FxU32         xlength,
+           FxU32         data_size,
+           void         *buffer,
+           FBS_DIRECTION direction)
+{
+  GrLfbInfo_t info;
+  BEGIN_BOARD_LOCK();
+  info.size=sizeof(info);
+
+  if (grLfbLock(direction == FBS_READ ? GR_LFB_READ_ONLY : GR_LFB_WRITE_ONLY,
+                target_buffer,
+                GR_LFBWRITEMODE_ANY,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winX = fxMesa->x_offset;
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+#ifdef XF86DRI
+    /* stride in data elements */
+    const GLint srcStride =
+      (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+        ? (fxMesa->screen_width)
+        : (info.strideInBytes / data_size);
+#else
+    /* stride in data elements */
+    const GLint srcStride = info.strideInBytes / data_size;
+#endif
+    GLushort *data16 = (GLushort *) info.lfbPtr
+      + (winY - ypos) * srcStride
+      + (winX + xpos);
+    GLubyte *target16 = (GLubyte *)  buffer;
+    GLubyte *data8    = (GLubyte *)  data16;
+    GLubyte *target8  = (GLubyte *)  buffer;
+    GLuint  *data32   = (GLuint *)   data16;
+    GLuint  *target32 = (GLuint *)  buffer;
+    GLuint i, j;
+
+    for (i = j = 0; i < xlength; i += 1, j += 1) {
+      switch (data_size) {
+      case 1:
+        switch (direction) {
+        case FBS_READ:
+          *target8++ = *data8++;
+          break;
+        case FBS_WRITE:
+          *data8++ = *target8++;
+          break;
+        }
+        break;
+      case 2:
+        switch (direction) {
+        case FBS_READ:
+          *target16++ = *data16++;
+          break;
+        case FBS_WRITE:
+          *data16++ = *target16++;
+          break;
+        }
+        break;
+      case 4:
+        switch (direction) {
+        case FBS_READ:
+          *target32++ = *data32++;
+          break;
+        case FBS_WRITE:
+          *data32++ = *target32++;
+          break;
+        }
+        break;
+      }
+    }
+    grLfbUnlock(direction == FBS_READ ? GR_LFB_READ_ONLY : GR_LFB_WRITE_ONLY,
+                target_buffer);
+  }
+  END_BOARD_LOCK();
+}
+           
+static FxBool
+fb_point_is_clipped(fxMesaContext fxMesa,
+                    FxU32 dst_x, FxU32 dst_y)
+{
+  int i;
+  for (i=0; i<fxMesa->numClipRects; i++) {
+    if ((dst_x>=fxMesa->pClipRects[i].x1) && 
+        (dst_x<fxMesa->pClipRects[i].x2) &&
+        (dst_y>=fxMesa->pClipRects[i].y1) && 
+        (dst_y<fxMesa->pClipRects[i].y2)) {
+      return GL_FALSE;
+    }
+  }
+  return GL_TRUE;
+}
 
 static FxBool writeRegionClipped(fxMesaContext fxMesa, GrBuffer_t dst_buffer,
-			  FxU32 dst_x, FxU32 dst_y, GrLfbSrcFmt_t src_format,
-			  FxU32 src_width, FxU32 src_height, FxI32 src_stride,
-			  void *src_data)
+				 FxU32 dst_x, FxU32 dst_y, GrLfbSrcFmt_t src_format,
+				 FxU32 src_width, FxU32 src_height, FxI32 src_stride,
+				 void *src_data)
 {
   int i, x, w, srcElt;
   void *data;
@@ -80,12 +240,12 @@ static FxBool writeRegionClipped(fxMesaContext fxMesa, GrBuffer_t dst_buffer,
   if (src_width==1 && src_height==1) { /* Easy case writing a point */
     for (i=0; i<fxMesa->numClipRects; i++) {
       if ((dst_x>=fxMesa->pClipRects[i].x1) && 
-	  (dst_x<fxMesa->pClipRects[i].x2) &&
-	  (dst_y>=fxMesa->pClipRects[i].y1) && 
-	  (dst_y<fxMesa->pClipRects[i].y2)) {
-	FX_grLfbWriteRegion(dst_buffer, dst_x, dst_y, src_format,
-			    1, 1, src_stride, src_data);
-	return GL_TRUE;
+          (dst_x<fxMesa->pClipRects[i].x2) &&
+          (dst_y>=fxMesa->pClipRects[i].y1) && 
+          (dst_y<fxMesa->pClipRects[i].y2)) {
+        FX_grLfbWriteRegion(dst_buffer, dst_x, dst_y, src_format,
+                            1, 1, src_stride, src_data);
+        return GL_TRUE;
       }
     }
   } else if (src_height==1) { /* Writing a span */
@@ -97,20 +257,20 @@ static FxBool writeRegionClipped(fxMesaContext fxMesa, GrBuffer_t dst_buffer,
     }
     for (i=0; i<fxMesa->numClipRects; i++) {
       if (dst_y>=fxMesa->pClipRects[i].y1 && dst_y<fxMesa->pClipRects[i].y2) {
-	if (dst_x<fxMesa->pClipRects[i].x1) {
-	  x=fxMesa->pClipRects[i].x1;
-	  data=((char*)src_data)+srcElt*(x - dst_x);
-	  w=src_width-(x-dst_x);
-	} else {
-	  x=dst_x;
-	  data=src_data;
-	  w=src_width;
-	}
-	if (x+w>fxMesa->pClipRects[i].x2) {
-	  w=fxMesa->pClipRects[i].x2-x;
-	}
-	FX_grLfbWriteRegion(dst_buffer, x, dst_y, src_format, w, 1,
-			    src_stride, data);
+        if (dst_x<fxMesa->pClipRects[i].x1) {
+          x=fxMesa->pClipRects[i].x1;
+          data=((char*)src_data)+srcElt*(x - dst_x);
+          w=src_width-(x-dst_x);
+        } else {
+          x=dst_x;
+          data=src_data;
+          w=src_width;
+        }
+        if (x+w>fxMesa->pClipRects[i].x2) {
+          w=fxMesa->pClipRects[i].x2-x;
+        }
+        FX_grLfbWriteRegion(dst_buffer, x, dst_y, src_format, w, 1,
+                            src_stride, data);
       }
     }
   } else { /* Punt on the case of arbitrary rectangles */
@@ -119,88 +279,42 @@ static FxBool writeRegionClipped(fxMesaContext fxMesa, GrBuffer_t dst_buffer,
   return GL_TRUE;
 }
 
-#else
-
-#define writeRegionClipped(fxm,dst_buffer,dst_x,dst_y,src_format,src_width,src_height,src_stride,src_data)		\
-  FX_grLfbWriteRegion(dst_buffer,dst_x,dst_y,src_format,src_width,src_height,src_stride,src_data)
-
-#endif
 
 
 /* KW: Rearranged the args in the call to grLfbWriteRegion().
  */
-#define LFB_WRITE_SPAN_MESA(dst_buffer,		\
-			    dst_x,		\
-			    dst_y,		\
-			    src_width,		\
-			    src_stride,		\
-			    src_data)		\
-  writeRegionClipped(fxMesa, dst_buffer,	\
-		   dst_x,			\
-		   dst_y,			\
-		   GR_LFB_SRC_FMT_8888,		\
-		   src_width,			\
-		   1,				\
-		   src_stride,			\
-		   src_data)			\
-
-
-#else /* !defined(FXMESA_USE_RGBA) */
-
-#define writeRegionClipped(fxm,dst_buffer,dst_x,dst_y,src_format,src_width,src_height,src_stride,src_data)		\
-  FX_grLfbWriteRegion(dst_buffer,dst_x,dst_y,src_format,src_width,src_height,src_stride,src_data)
-
-
-#define MESACOLOR_TO_ARGB(c) (				\
-             ( ((unsigned int)(c[ACOMP]))<<24 ) |	\
-             ( ((unsigned int)(c[RCOMP]))<<16 ) |	\
-             ( ((unsigned int)(c[GCOMP]))<<8 )  |	\
-             (  (unsigned int)(c[BCOMP])) )
-  
-inline void LFB_WRITE_SPAN_MESA(GrBuffer_t dst_buffer, 
-			 FxU32 dst_x, 
-			 FxU32 dst_y, 
-			 FxU32 src_width,
-			 FxI32 src_stride, 
-			 void *src_data )
-{
-   /* Covert to ARGB */
-   GLubyte (*rgba)[4] = src_data;
-   GLuint argb[MAX_WIDTH];
-   int i;
-   
-   for (i = 0; i < src_width; i++)
-   {
-      argb[i] = MESACOLOR_TO_ARGB(rgba[i]);
-   }
-   writeRegionClipped( /*fxMesa,*/ NULL, dst_buffer,
-		       dst_x,
-		       dst_y,
-		       GR_LFB_SRC_FMT_8888,
-		       src_width,
-		       1,
-		       src_stride,
-		       (void*)argb);
-}
- 
-#endif /* !defined(FXMESA_USE_RGBA) */
+#define LFB_WRITE_SPAN_MESA(dst_buffer,         \
+                            dst_x,              \
+                            dst_y,              \
+                            src_width,          \
+                            src_stride,         \
+                            src_data)           \
+  writeRegionClipped(fxMesa, dst_buffer,        \
+                   dst_x,                       \
+                   dst_y,                       \
+                   GR_LFB_SRC_FMT_8888,         \
+                   src_width,                   \
+                   1,                           \
+                   src_stride,                  \
+                   src_data)                    \
+
 
 
-/************************************************************************/
-/*****                    Span functions                            *****/
-/************************************************************************/
 
+/*
+ * 16bpp span/pixel functions
+ */
 
-static void fxDDWriteRGBASpan(const GLcontext *ctx, 
-                              GLuint n, GLint x, GLint y,
-                              const GLubyte rgba[][4], const GLubyte mask[])
+static void
+write_R5G6B5_rgba_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                       const GLubyte rgba[][4], const GLubyte mask[])
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GLuint i;
   GLint bottom=fxMesa->height+fxMesa->y_offset-1; 
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDWriteRGBASpan(...)\n");
+    fprintf(stderr,"fxmesa: fxDDWriteRGBASpan(...)\n");
   }
 
   x+=fxMesa->x_offset;
@@ -213,7 +327,7 @@ static void fxDDWriteRGBASpan(const GLcontext *ctx,
       } else {
         if (span > 0) {
           LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x+i-span, bottom-y,
-                           /* GR_LFB_SRC_FMT_8888,*/ span, /*1,*/ 0, (void *) rgba[i-span] );
+			       /* GR_LFB_SRC_FMT_8888,*/ span, /*1,*/ 0, (void *) rgba[i-span] );
           span = 0;
         }
       }
@@ -221,16 +335,16 @@ static void fxDDWriteRGBASpan(const GLcontext *ctx,
 
     if (span > 0)
       LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x+n-span, bottom-y,
-                        /* GR_LFB_SRC_FMT_8888, */ span, /*1,*/ 0, (void *) rgba[n-span] );
+			   /* GR_LFB_SRC_FMT_8888, */ span, /*1,*/ 0, (void *) rgba[n-span] );
   } else
     LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x, bottom-y,/* GR_LFB_SRC_FMT_8888,*/
-                      n,/* 1,*/ 0, (void *) rgba );
+			 n,/* 1,*/ 0, (void *) rgba );
 }
 
 
-static void fxDDWriteRGBSpan(const GLcontext *ctx, 
-                             GLuint n, GLint x, GLint y,
-                             const GLubyte rgb[][3], const GLubyte mask[])
+static void
+write_R5G6B5_rgb_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                      const GLubyte rgb[][3], const GLubyte mask[])
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GLuint i;
@@ -238,7 +352,7 @@ static void fxDDWriteRGBSpan(const GLcontext *ctx,
   GLubyte rgba[MAX_WIDTH][4];
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDWriteRGBSpan()\n");
+    fprintf(stderr,"fxmesa: fxDDWriteRGBSpan()\n");
   }
 
   x+=fxMesa->x_offset;
@@ -255,7 +369,7 @@ static void fxDDWriteRGBSpan(const GLcontext *ctx,
       } else {
         if (span > 0) {
           LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x+i-span, bottom-y,
-                            /*GR_LFB_SRC_FMT_8888,*/ span,/* 1,*/ 0, (void *) rgba );
+			       /*GR_LFB_SRC_FMT_8888,*/ span,/* 1,*/ 0, (void *) rgba );
           span = 0;
         }
       }
@@ -263,7 +377,7 @@ static void fxDDWriteRGBSpan(const GLcontext *ctx,
 
     if (span > 0)
       LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x+n-span, bottom-y,
-                        /*GR_LFB_SRC_FMT_8888,*/ span,/* 1,*/ 0, (void *) rgba );
+			   /*GR_LFB_SRC_FMT_8888,*/ span,/* 1,*/ 0, (void *) rgba );
   } else {
     for (i=0;i<n;i++) {
       rgba[i][RCOMP]=rgb[i][0];
@@ -273,14 +387,14 @@ static void fxDDWriteRGBSpan(const GLcontext *ctx,
     }
 
     LFB_WRITE_SPAN_MESA( fxMesa->currentFB, x, bottom-y,/* GR_LFB_SRC_FMT_8888,*/
-                      n,/* 1,*/ 0, (void *) rgba );
+			 n,/* 1,*/ 0, (void *) rgba );
   }
 }
 
 
-static void fxDDWriteMonoRGBASpan(const GLcontext *ctx, 
-                                  GLuint n, GLint x, GLint y,
-                                  const GLubyte mask[])
+static void
+write_R5G6B5_mono_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                       const GLubyte mask[])
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GLuint i;
@@ -288,7 +402,7 @@ static void fxDDWriteMonoRGBASpan(const GLcontext *ctx,
   GLuint data[MAX_WIDTH];
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDWriteMonoRGBASpan(...)\n");
+    fprintf(stderr,"fxmesa: fxDDWriteMonoRGBASpan(...)\n");
   }
 
   x+=fxMesa->x_offset;
@@ -302,8 +416,8 @@ static void fxDDWriteMonoRGBASpan(const GLcontext *ctx,
       } else {
         if (span > 0) {
           writeRegionClipped(fxMesa,  fxMesa->currentFB, x+i-span, bottom-y,
-                            GR_LFB_SRC_FMT_8888, span, 1, 0,
-                            (void *) data );
+			     GR_LFB_SRC_FMT_8888, span, 1, 0,
+			     (void *) data );
           span = 0;
         }
       }
@@ -311,59 +425,31 @@ static void fxDDWriteMonoRGBASpan(const GLcontext *ctx,
 
     if (span > 0)
       writeRegionClipped(fxMesa,  fxMesa->currentFB, x+n-span, bottom-y,
-                        GR_LFB_SRC_FMT_8888, span, 1, 0,
-                        (void *) data );
+			 GR_LFB_SRC_FMT_8888, span, 1, 0,
+			 (void *) data );
   } else {
     for (i=0;i<n;i++) {
       data[i]=(GLuint) fxMesa->color;
     }
 
     writeRegionClipped(fxMesa,  fxMesa->currentFB, x, bottom-y, GR_LFB_SRC_FMT_8888,
-                      n, 1, 0, (void *) data );
+		       n, 1, 0, (void *) data );
   }
 }
 
 
-#if 0
-static void fxDDReadRGBASpan(const GLcontext *ctx, 
-                             GLuint n, GLint x, GLint y, GLubyte rgba[][4])
-{
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  GLushort data[MAX_WIDTH];
-  GLuint i;
-  GLint bottom=fxMesa->height+fxMesa->y_offset-1;
-
-  printf("read span %d, %d, %d\n", x,y,n);
-  if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDReadRGBASpan(...)\n");
-  }
-
-  assert(n < MAX_WIDTH);
-
-  x+=fxMesa->x_offset;
-  FX_grLfbReadRegion( fxMesa->currentFB, x, bottom-y, n, 1, 0, data);
-
-  for (i=0;i<n;i++) {
-    GLushort pixel = data[i];
-    rgba[i][RCOMP] = FX_PixelToR[pixel];
-    rgba[i][GCOMP] = FX_PixelToG[pixel];
-    rgba[i][BCOMP] = FX_PixelToB[pixel];
-    rgba[i][ACOMP] = 255;
-  }
-}
-#endif
-
-
 /*
  * Read a span of 16-bit RGB pixels.  Note, we don't worry about cliprects
  * since OpenGL says obscured pixels have undefined values.
  */
-static void read_R5G6B5_span(const GLcontext *ctx, 
-                             GLuint n, GLint x, GLint y, GLubyte rgba[][4])
+static void
+read_R5G6B5_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                 GLubyte rgba[][4])
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
   GrLfbInfo_t info;
   BEGIN_BOARD_LOCK();
+  info.size=sizeof(info);
   if (grLfbLock(GR_LFB_READ_ONLY,
                 fxMesa->currentFB,
                 GR_LFBWRITEMODE_ANY,
@@ -374,13 +460,13 @@ static void read_R5G6B5_span(const GLcontext *ctx,
     const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
 #ifdef XF86DRI
     const GLint srcStride = (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
-                          ? (fxMesa->screen_width) : (info.strideInBytes / 2);
+      ? (fxMesa->screen_width) : (info.strideInBytes / 2);
 #else
     const GLint srcStride = info.strideInBytes / 2; /* stride in GLushorts */
 #endif
     const GLushort *data16 = (const GLushort *) info.lfbPtr
-                           + (winY - y) * srcStride
-                           + (winX + x);
+      + (winY - y) * srcStride
+      + (winX + x);
     const GLuint *data32 = (const GLuint *) data16;
     GLuint i, j;
     GLuint extraPixel = (n & 1);
@@ -412,91 +498,716 @@ static void read_R5G6B5_span(const GLcontext *ctx,
 }
 
 
-/************************************************************************/
-/*****                    Pixel functions                           *****/
-/************************************************************************/
-
-static void fxDDWriteRGBAPixels(const GLcontext *ctx,
-                                GLuint n, const GLint x[], const GLint y[],
-                                CONST GLubyte rgba[][4], const GLubyte mask[])
+static void
+write_R5G6B5_pixels(const GLcontext *ctx,
+                    GLuint n, const GLint x[], const GLint y[],
+                    CONST GLubyte rgba[][4], const GLubyte mask[])
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GLuint i;
   GLint bottom=fxMesa->height+fxMesa->y_offset-1;
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDWriteRGBAPixels(...)\n");
+    fprintf(stderr,"fxmesa: fxDDWriteRGBAPixels(...)\n");
   }
 
   for(i=0;i<n;i++)
     if(mask[i])
-       LFB_WRITE_SPAN_MESA(fxMesa->currentFB, x[i]+fxMesa->x_offset, bottom-y[i],
-                       1, 1, (void *)rgba[i]);
+      LFB_WRITE_SPAN_MESA(fxMesa->currentFB, x[i]+fxMesa->x_offset, bottom-y[i],
+			  1, 1, (void *)rgba[i]);
 }
 
-static void fxDDWriteMonoRGBAPixels(const GLcontext *ctx,
-                                    GLuint n, const GLint x[], const GLint y[],
-                                    const GLubyte mask[])
+
+static void
+write_R5G6B5_mono_pixels(const GLcontext *ctx,
+                         GLuint n, const GLint x[], const GLint y[],
+                         const GLubyte mask[])
 {
   fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
   GLuint i;
   GLint bottom=fxMesa->height+fxMesa->y_offset-1;
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDWriteMonoRGBAPixels(...)\n");
+    fprintf(stderr,"fxmesa: fxDDWriteMonoRGBAPixels(...)\n");
   }
 
   for(i=0;i<n;i++)
     if(mask[i])
       writeRegionClipped(fxMesa, fxMesa->currentFB,x[i]+fxMesa->x_offset,bottom-y[i],
-                       GR_LFB_SRC_FMT_8888,1,1,0,(void *) &fxMesa->color);
+			 GR_LFB_SRC_FMT_8888,1,1,0,(void *) &fxMesa->color);
 }
 
+static void
+read_R5G6B5_pixels(const GLcontext *ctx,
+                   GLuint n, const GLint x[], const GLint y[],
+                   GLubyte rgba[][4], const GLubyte mask[])
+{
+  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
+  GLuint i;
+  GLint bottom=fxMesa->height+fxMesa->y_offset-1;
 
-static void read_R5G6B5_pixels(const GLcontext *ctx,
-                               GLuint n, const GLint x[], const GLint y[],
-                               GLubyte rgba[][4], const GLubyte mask[])
+  if (MESA_VERBOSE&VERBOSE_DRIVER) {
+    fprintf(stderr,"fxmesa: fxDDReadRGBAPixels(...)\n");
+  }
+
+  for(i=0;i<n;i++) {
+    if(mask[i]) {
+      GLushort pixel;
+      FX_grLfbReadRegion(fxMesa->currentFB,x[i],bottom-y[i],1,1,0,&pixel);
+      rgba[i][RCOMP] = FX_PixelToR[pixel];
+      rgba[i][GCOMP] = FX_PixelToG[pixel];
+      rgba[i][BCOMP] = FX_PixelToB[pixel];
+      rgba[i][ACOMP] = 255;
+    }
+  }
+}
+
+
+/*
+ * 24bpp span/pixel functions
+ */
+
+static void
+write_R8G8B8_rgb_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                      const GLubyte rgb[][3], const GLubyte mask[])
 {
-  fxMesaContext fxMesa = (fxMesaContext) ctx->DriverCtx;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
   GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_888;
+  else
+    mode = GR_LFBWRITEMODE_888;
+
   BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+    const GLint scrX = winX + x;
+    const GLint scrY = winY - y;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      /*GLint dstStride = fxMesa->screen_width * 3;*/
+      GLint dstStride = info.strideInBytes / 1;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 1;
+      GLuint *dst32 = (GLuint *) dst;
+      GLubyte visMask[MAX_WIDTH];
+      GLuint i;
+      generate_vismask(fxMesa, scrX, scrY, n, visMask);
+      for (i = 0; i < n; i++) {
+        if (visMask[i] && (!mask || mask[i])) {
+          dst32[i] = PACK_BGRA32(rgb[i][0], rgb[i][1], rgb[i][2], 255);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLint dstStride = info.strideInBytes;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      GLuint *dst32 = (GLuint *) dst;
+      if (mask) {
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          if (mask[i]) {
+            dst32[i] = PACK_RGBA32(rgb[i][0], rgb[i][1], rgb[i][2], 255);
+          }
+        }
+      }
+      else {
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          dst32[i] = PACK_RGBA32(rgb[i][2], rgb[i][1], rgb[i][0], 255);
+        }
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
+  END_BOARD_LOCK();
+}
+
+
+
+static void
+write_R8G8B8_rgba_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                       const GLubyte rgba[][4], const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888 /*565*/;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+    const GLint scrX = winX + x;
+    const GLint scrY = winY - y;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      /* XXX have to do cliprect clipping! */
+      GLint dstStride = fxMesa->screen_width * 4;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      GLuint *dst32 = (GLuint *) dst;
+      GLubyte visMask[MAX_WIDTH];
+      GLuint i;
+      generate_vismask(fxMesa, scrX, scrY, n, visMask);
+      for (i = 0; i < n; i++) {
+        if (visMask[i] && (!mask || mask[i])) {
+          dst32[i] = PACK_BGRA32(rgba[i][0], rgba[i][1], rgba[i][2], rgba[i][3]);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLint dstStride = info.strideInBytes;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      if (mask) {
+        const GLuint *src32 = (const GLuint *) rgba;
+        GLuint *dst32 = (GLuint *) dst;
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          if (mask[i]) {
+            dst32[i] = src32[i];
+          }
+        }
+      }
+      else {
+        /* no mask, write all pixels */
+        MEMCPY(dst, rgba, 4 * n);
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
+  END_BOARD_LOCK();
+}
+
+
+static void
+write_R8G8B8_mono_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                       const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GLubyte rgba[MAX_WIDTH][4];
+  GLuint *data = (GLuint *) rgba;
+  GLuint i;
+
+  /* XXX this is a simple-minded implementation but good enough for now */
+  for (i = 0; i < n; i++) {
+    data[i] = (GLuint) fxMesa->color;
+  }
+  write_R8G8B8_rgba_span(ctx, n, x, y, (const GLubyte (*)[4]) rgba, mask);
+}
+
+
+static void
+read_R8G8B8_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                 GLubyte rgba[][4])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (1 || fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+    mode = GR_LFBWRITEMODE_8888;
+  }
+  else {
+    mode = GR_LFBWRITEMODE_565; /*888*/ /*565*/;
+  }
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
   if (grLfbLock(GR_LFB_READ_ONLY,
                 fxMesa->currentFB,
-                GR_LFBWRITEMODE_ANY,
+                mode, /*GR_LFBWRITEMODE_ANY,*/
                 GR_ORIGIN_UPPER_LEFT,
                 FXFALSE,
                 &info)) {
-#ifdef XF86DRI
-    const GLint srcStride = (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
-                          ? (fxMesa->screen_width) : (info.strideInBytes / 2);
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLint srcStride = fxMesa->screen_width * 4;
+      const GLubyte *src = (const GLubyte *) info.lfbPtr
+                         + (winY - y) * srcStride + (winX + x) * 4;
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        rgba[i][0] = src[i * 4 + 2];
+        rgba[i][1] = src[i * 4 + 1];
+        rgba[i][2] = src[i * 4 + 0];
+        rgba[i][3] = src[i * 4 + 3];
+      }
+    }
+    else {
+      /* back buffer */
+      GLint srcStride = /*info.strideInBytes;*/ 8192 / 2;  /* XXX a hack! */
+      const GLubyte *src = (const GLubyte *) info.lfbPtr
+                         + (winY - y) * srcStride + (winX + x) * 4;
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        rgba[i][0] = src[i * 4 + 2];
+        rgba[i][1] = src[i * 4 + 1];
+        rgba[i][2] = src[i * 4 + 0];
+        rgba[i][3] = src[i * 4 + 3];
+      }
+    }
+    grLfbUnlock(GR_LFB_READ_ONLY, fxMesa->currentFB);
+  }
+  END_BOARD_LOCK();
+}
+
+
+static void
+write_R8G8B8_pixels(const GLcontext *ctx,
+                    GLuint n, const GLint x[], const GLint y[],
+                    CONST GLubyte rgba[][4], const GLubyte mask[])
+{
+#if 00
+  GLuint i;
+  for (i = 0; i < n; i++) {
+    write_R8G8B8_rgba_span(ctx, 1, x[i], y[i], rgba + i, mask + i);
+  }
+
 #else
-    const GLint srcStride = info.strideInBytes / 2; /* stride in GLushorts */
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888 /*565*/;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+        const GLint winX = fxMesa->x_offset;
+        const GLint scrX = winX + x[i];
+        const GLint scrY = winY - y[i];
+        if (mask[i] && visible_pixel(fxMesa, scrX, scrY)) {
+          GLint dstStride = fxMesa->screen_width * 4;
+          GLubyte *dst = (GLubyte *) info.lfbPtr + scrY * dstStride + scrX * 4;
+          GLuint *dst32 = (GLuint *) dst;
+          *dst32 = PACK_BGRA32(rgba[i][0], rgba[i][1],
+                               rgba[i][2], rgba[i][3]);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+        const GLint winX = fxMesa->x_offset;
+        const GLint scrX = winX + x[i];
+        const GLint scrY = winY - y[i];
+        if (mask[i] && visible_pixel(fxMesa, scrX, scrY)) {
+          GLint dstStride = info.strideInBytes;
+          GLubyte *dst = (GLubyte *) info.lfbPtr + scrY * dstStride + scrX * 4;
+          GLuint *dst32 = (GLuint *) dst;
+          const GLuint *src32 = (const GLuint *) rgba;
+          *dst32 = src32[i];
+        }
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
+  END_BOARD_LOCK();
 #endif
+}
+
+
+static void
+write_R8G8B8_mono_pixels(const GLcontext *ctx,
+                         GLuint n, const GLint x[], const GLint y[],
+                         const GLubyte mask[])
+{
+}
+
+
+static void
+read_R8G8B8_pixels(const GLcontext *ctx,
+                   GLuint n, const GLint x[], const GLint y[],
+                   GLubyte rgba[][4], const GLubyte mask[])
+{
+  printf("read_R8G8B8_pixels %d\n", n);
+}
+
+
+
+/*
+ * 32bpp span/pixel functions
+ */
+
+static void
+write_R8G8B8A8_rgb_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                        const GLubyte rgb[][3], const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888 /*565*/;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
     const GLint winX = fxMesa->x_offset;
+    const GLint scrX = winX + x;
+    const GLint scrY = winY - y;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLint dstStride = fxMesa->screen_width * 4;
+      GLubyte * dst = (GLubyte *) info.lfbPtr
+                    + (winY - y) * dstStride + (winX + x) * 4;
+      GLuint *dst32 = (GLuint *) dst;
+      GLubyte visMask[MAX_WIDTH];
+      GLuint i;
+      generate_vismask(fxMesa, scrX, scrY, n, visMask);
+      for (i = 0; i < n; i++) {
+        if (visMask[i] && (!mask || mask[i])) {
+          dst32[i] = PACK_BGRA32(rgb[i][0], rgb[i][1], rgb[i][2], 255);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLint dstStride = info.strideInBytes;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      if (mask) {
+        GLuint *dst32 = (GLuint *) dst;
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          if (mask[i]) {
+            dst32[i] = PACK_RGBA32(rgb[i][0], rgb[i][1], rgb[i][2], 255);
+          }
+        }
+      }
+      else {
+        GLuint *dst32 = (GLuint *) dst;
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          dst32[i] = PACK_RGBA32(rgb[i][0], rgb[i][1], rgb[i][2], 255);
+        }
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
+  END_BOARD_LOCK();
+}
+
+
+/*
+ *XXX test of grLfbWriteRegion in 32bpp mode.  Doesn't seem to work!
+ */
+#if 0
+static void
+write_R8G8B8A8_rgb_span2(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                        const GLubyte rgb[][3], const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GLint bottom = fxMesa->height + fxMesa->y_offset - 1; 
+  GLint x2 = fxMesa->x_offset +x;
+  GLint y2 = bottom - y;
+
+  FX_grLfbWriteRegion(fxMesa->currentFB, x2, y2, GR_LFB_SRC_FMT_888,
+                      n, 1, 0, rgb);
+}
+#endif
+
+
+static void
+write_R8G8B8A8_rgba_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                         const GLubyte rgba[][4], const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888 /*565*/;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
     const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
-    GLuint i;
-    for(i=0;i<n;i++) {
-      if(mask[i]) {
-        const GLushort *data16 = (const GLushort *) info.lfbPtr
-                               + (winY - y[i]) * srcStride
-                               + (winX + x[i]);
-        const GLushort pixel = *data16;
-        rgba[i][RCOMP] = FX_PixelToR[pixel];
-        rgba[i][GCOMP] = FX_PixelToG[pixel];
-        rgba[i][BCOMP] = FX_PixelToB[pixel];
-        rgba[i][ACOMP] = 255;
+    const GLint winX = fxMesa->x_offset;
+    const GLint scrX = winX + x;
+    const GLint scrY = winY - y;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLint dstStride = fxMesa->screen_width * 4;
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      GLuint *dst32 = (GLuint *) dst;
+      GLubyte visMask[MAX_WIDTH];
+      GLuint i;
+      generate_vismask(fxMesa, scrX, scrY, n, visMask);
+      for (i = 0; i < n; i++) {
+        if (visMask[i] && (!mask || mask[i])) {
+          dst32[i] = PACK_BGRA32(rgba[i][0], rgba[i][1], rgba[i][2], rgba[i][3]);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLint dstStride = 8192; /* XXX a hack  info.strideInBytes; */
+      GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+      if (mask) {
+        const GLuint *src32 = (const GLuint *) rgba;
+        GLuint *dst32 = (GLuint *) dst;
+        GLuint i;
+        for (i = 0; i < n; i++) {
+          if (mask[i]) {
+            dst32[i] = src32[i];
+          }
+        }
+      }
+      else {
+        /* no mask, write all pixels */
+        MEMCPY(dst, rgba, 4 * n);
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
+  else {
+    info.strideInBytes = -1;
+  }
+  END_BOARD_LOCK();
+}
+
+
+static void
+write_R8G8B8A8_mono_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                         const GLubyte mask[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GLubyte rgba[MAX_WIDTH][4];
+  GLuint *data = (GLuint *) rgba;
+  GLuint i;
+
+  /* XXX this is a simple-minded implementation but good enough for now */
+  for (i = 0; i < n; i++) {
+    data[i] = (GLuint) fxMesa->color;
+  }
+  write_R8G8B8A8_rgba_span(ctx, n, x, y, (const GLubyte (*)[4]) rgba, mask);
+}
+
+
+static void
+read_R8G8B8A8_span(const GLcontext *ctx, GLuint n, GLint x, GLint y,
+                   GLubyte rgba[][4])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+    mode = GR_LFBWRITEMODE_8888;
+  }
+  else {
+    mode = GR_LFBWRITEMODE_8888; /*565;*/
+  }
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_READ_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLint srcStride = fxMesa->screen_width;
+      const GLuint *src32 = (const GLuint *) info.lfbPtr
+                          + (winY - y) * srcStride + (winX + x);
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        const GLuint p = src32[i];
+        rgba[i][0] = (p >> 16) & 0xff;
+        rgba[i][1] = (p >>  8) & 0xff;
+        rgba[i][2] = (p >>  0) & 0xff;
+        rgba[i][3] = (p >> 24) & 0xff;
+      }
+    }
+    else {
+      /* back buffer */
+      GLint srcStride = 1024;  /* XXX a hack */
+      const GLuint *src32 = (const GLuint *) info.lfbPtr
+                          + (winY - y) * srcStride + (winX + x);
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        GLuint p = src32[i];
+        rgba[i][0] = (p >> 16) & 0xff;
+        rgba[i][1] = (p >>  8) & 0xff;
+        rgba[i][2] = (p >>  0) & 0xff;
+        rgba[i][3] = (p >> 24) & 0xff;
       }
     }
     grLfbUnlock(GR_LFB_READ_ONLY, fxMesa->currentFB);
   }
+  else
+    info.strideInBytes = -1;
+  END_BOARD_LOCK();
+}
+
+
+static void
+write_R8G8B8A8_pixels(const GLcontext *ctx,
+                      GLuint n, const GLint x[], const GLint y[],
+                      CONST GLubyte rgba[][4], const GLubyte mask[])
+{
+#if 00
+  GLuint i;
+  for (i = 0; i < n; i++) {
+    write_R8G8B8A8_rgba_span(ctx, 1, x[i], y[i], rgba + i, mask + i);
+  }
+
+#else
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbWriteMode_t mode;
+  GrLfbInfo_t info;
+
+  if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT)
+    mode = GR_LFBWRITEMODE_8888;
+  else
+    mode = GR_LFBWRITEMODE_888 /*565*/;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                fxMesa->currentFB,
+                mode,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    if (fxMesa->glCtx->Color.DrawBuffer == GL_FRONT) {
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+        const GLint winX = fxMesa->x_offset;
+        const GLint scrX = winX + x[i];
+        const GLint scrY = winY - y[i];
+        if (mask[i] && visible_pixel(fxMesa, scrX, scrY)) {
+          GLint dstStride = fxMesa->screen_width * 4;
+          GLubyte *dst = (GLubyte *) info.lfbPtr + scrY * dstStride + scrX * 4;
+          GLuint *dst32 = (GLuint *) dst;
+          *dst32 = PACK_BGRA32(rgba[i][0], rgba[i][1],
+                               rgba[i][2], rgba[i][3]);
+        }
+      }
+    }
+    else {
+      /* back buffer */
+      GLuint i;
+      for (i = 0; i < n; i++) {
+        const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+        const GLint winX = fxMesa->x_offset;
+        const GLint scrX = winX + x[i];
+        const GLint scrY = winY - y[i];
+        if (mask[i] && visible_pixel(fxMesa, scrX, scrY)) {
+          GLint dstStride = info.strideInBytes;
+          GLubyte *dst = (GLubyte *) info.lfbPtr + scrY * dstStride + scrX * 4;
+          GLuint *dst32 = (GLuint *) dst;
+          const GLuint *src32 = (const GLuint *) rgba;
+          *dst32 = src32[i];
+        }
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, fxMesa->currentFB);
+  }
   END_BOARD_LOCK();
+#endif
+}
+
+
+static void
+write_R8G8B8A8_mono_pixels(const GLcontext *ctx,
+                           GLuint n, const GLint x[], const GLint y[],
+                           const GLubyte mask[])
+{
+  fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx; 
+  GLuint i;
+  GLuint color = fxMesa->color;
+  for (i = 0; i < n; i++) {
+    if (mask[i]) {
+      write_R8G8B8A8_rgba_span(ctx, 1, x[i], y[i],
+                               (const GLubyte (*)[4]) &color, mask + i);
+    }
+  }
+}
+
+
+
+static void
+read_R8G8B8A8_pixels(const GLcontext *ctx,
+                     GLuint n, const GLint x[], const GLint y[],
+                     GLubyte rgba[][4], const GLubyte mask[])
+{
+  GLuint i;
+  for (i = 0; i < n; i++) {
+    if (mask[i]) {
+      read_R8G8B8A8_span(ctx, 1, x[i], y[i], rgba + i);
+    }
+  }
+  printf("read_R8G8B8A8_pixels %d\n", n);
 }
 
 
 
-/************************************************************************/
-/*****                    Depth functions                           *****/
-/************************************************************************/
+/*
+ * Depth buffer read/write functions.
+ */
 
 void fxDDWriteDepthSpan(GLcontext *ctx,
                         GLuint n, GLint x, GLint y, const GLdepth depth[],
@@ -504,9 +1215,11 @@ void fxDDWriteDepthSpan(GLcontext *ctx,
 {
   fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
   GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
+  GLuint depth_size = fxMesa->glVis->DepthBits;
+  GLuint stencil_size = fxMesa->glVis->StencilBits;
 
   if (MESA_VERBOSE & VERBOSE_DRIVER) {
-     fprintf(stderr, "fxmesa: fxDDWriteDepthSpan(...)\n");
+    fprintf(stderr, "fxmesa: fxDDWriteDepthSpan(...)\n");
   }
 
   x += fxMesa->x_offset;
@@ -515,20 +1228,77 @@ void fxDDWriteDepthSpan(GLcontext *ctx,
     GLint i;
     for (i = 0; i < n; i++) {
       if (mask[i]) {
-        GLshort d = depth[i];
-        writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, x + i, bottom - y,
-                           GR_LFB_SRC_FMT_ZA16, 1, 1, 0, (void *) &d);
+	GLshort d16;
+	GLuint d32;
+	switch (depth_size) {
+	case 16:
+	  d16 = depth[i];
+	  writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, x + i, bottom - y,
+			     GR_LFB_SRC_FMT_ZA16, 1, 1, 0, (void *) &d16);
+	  break;
+	case 32:
+	  if (!fb_point_is_clipped(fxMesa, x+i, bottom-y)) {
+	    if (stencil_size > 0) {
+	      rw_fb_span(fxMesa,
+			 GR_BUFFER_AUXBUFFER,
+			 x + i, bottom - y,
+			 1,
+			 sizeof(GLdepth),
+			 &d32,
+			 FBS_READ);
+	      d32 = depth[i] & 0xFF000000;
+	      break;
+	    } else {
+	      d32 = depth[i];
+	    }
+	    rw_fb_span(fxMesa,
+		       GR_BUFFER_AUXBUFFER,
+		       x + i, bottom - y,
+		       1,
+		       sizeof(GLdepth),
+		       &d32,
+		       FBS_WRITE);
+	  }
+	}
       }
     }
-  }
-  else {
+  } else {
     GLushort depth16[MAX_WIDTH];
     GLint i;
-    for (i = 0; i < n; i++) {
-      depth16[i] = depth[i];
+    GLuint d32;
+    switch (depth_size) {
+    case 16:
+      for (i = 0; i < n; i++) {
+	depth16[i] = depth[i];
+      }
+      writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, x, bottom - y,
+			 GR_LFB_SRC_FMT_ZA16, n, 1, 0, (void *) depth16);
+      break;
+    case 32:
+      for (i = 0; i < n; i++) {
+	if (fb_point_is_clipped(fxMesa, x+i, bottom-y)) {
+	  if (stencil_size > 0) {
+	    rw_fb_span(fxMesa,
+		       GR_BUFFER_AUXBUFFER,
+		       x + i, bottom - y,
+		       1,
+		       sizeof(GLdepth),
+		       &d32,
+		       FBS_READ);
+	    d32 = (d32 & 0xFF0000) | (depth[i] & 0x00FFFFFF);
+	  } else {
+	    d32 = depth[i];
+	  }
+	  rw_fb_span(fxMesa,
+		     GR_BUFFER_AUXBUFFER,
+		     x + i, bottom - y,
+		     1,
+		     sizeof(GLdepth),
+		     &d32,
+		     FBS_WRITE);
+	}
+      }	   
     }
-    writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, x, bottom - y,
-                       GR_LFB_SRC_FMT_ZA16, n, 1, 0, (void *) depth16);
   }
 }
 
@@ -540,15 +1310,34 @@ void fxDDReadDepthSpan(GLcontext *ctx,
   GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
   GLushort depth16[MAX_WIDTH];
   GLuint i;
+  GLuint depth_size = fxMesa->glVis->DepthBits;
+  GLuint stencil_size = fxMesa->glVis->StencilBits;
 
   if (MESA_VERBOSE & VERBOSE_DRIVER) {
-     fprintf(stderr, "fxmesa: fxDDReadDepthSpan(...)\n");
+    fprintf(stderr, "fxmesa: fxDDReadDepthSpan(...)\n");
   }
 
   x += fxMesa->x_offset;
-  FX_grLfbReadRegion(GR_BUFFER_AUXBUFFER, x, bottom - y, n, 1, 0, depth16);
-  for (i = 0; i < n; i++) {
-    depth[i] = depth16[i];
+  switch (depth_size) {
+  case 16:
+    FX_grLfbReadRegion(GR_BUFFER_AUXBUFFER, x, bottom - y, n, 1, 0, depth16);
+    for (i = 0; i < n; i++) {
+      depth[i] = depth16[i];
+    }
+    break;
+  case 32:
+    rw_fb_span(fxMesa,
+	       GR_BUFFER_AUXBUFFER,
+	       x, bottom - y,
+	       n,
+	       sizeof(GLdepth),
+	       depth,
+	       FBS_READ);
+    if (stencil_size > 0) {
+      for (i = 0; i < n; i++) {
+	depth[i] &= 0xFFFFFF;
+      }
+    }
   }
 }
 
@@ -561,6 +1350,9 @@ void fxDDWriteDepthPixels(GLcontext *ctx,
   fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
   GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
   GLuint i;
+  GLuint d32;
+  GLuint depth_size = fxMesa->glVis->DepthBits;
+  GLuint stencil_size = fxMesa->glVis->StencilBits;
 
   if (MESA_VERBOSE & VERBOSE_DRIVER) {
     fprintf(stderr, "fxmesa: fxDDWriteDepthPixels(...)\n");
@@ -570,9 +1362,41 @@ void fxDDWriteDepthPixels(GLcontext *ctx,
     if (mask[i]) {
       int xpos = x[i] + fxMesa->x_offset;
       int ypos = bottom - y[i];
-      GLushort d = depth[i];
-      writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, xpos, ypos,
-                         GR_LFB_SRC_FMT_ZA16, 1, 1, 0, (void *) &d);
+      GLushort d16 = depth[i];
+      switch (depth_size) {
+      case 16:
+	writeRegionClipped(fxMesa, GR_BUFFER_AUXBUFFER, xpos, ypos,
+			   GR_LFB_SRC_FMT_ZA16, 1, 1, 0,
+			   (void *) &d16);
+	break;
+      case 32:
+        if (!fb_point_is_clipped(fxMesa, xpos, ypos)) {
+          if (stencil_size > 0) {
+	    /*
+	     * Should I sign extend this?  I don't
+	     * believe so, since GLdepth is unsigned.
+	     */
+            rw_fb_span(fxMesa,
+                       GR_BUFFER_AUXBUFFER,
+                       xpos, ypos,
+                       1,
+                       sizeof(GLdepth),
+                       &d32,
+                       FBS_READ);
+            d32 = (d32 &~ 0xFF000000) | (depth[i] & 0x00FFFFFF);
+          } else {
+            d32 = depth[i];
+          }
+          rw_fb_span(fxMesa,
+                     GR_BUFFER_AUXBUFFER,
+                     xpos, ypos,
+                     1,
+                     sizeof(GLdepth),
+                     &d32,
+                     FBS_WRITE);
+        }
+        break;
+      }
     }
   }
 }
@@ -592,51 +1416,409 @@ void fxDDReadDepthPixels(GLcontext *ctx, GLuint n,
   for (i = 0; i < n; i++) {
     int xpos = x[i] + fxMesa->x_offset;
     int ypos = bottom - y[i];
-    GLushort d;
-    FX_grLfbReadRegion(GR_BUFFER_AUXBUFFER, xpos, ypos, 1, 1, 0, &d);
-    depth[i] = d;
+    GLushort d16;
+    GLuint d32;
+    GLuint depth_size = fxMesa->glVis->DepthBits;
+    GLuint stencil_size = fxMesa->glVis->StencilBits;
+    
+    assert((depth_size == 16) || (depth_size == 32));
+    switch (depth_size) {
+    case 16:
+      FX_grLfbReadRegion(GR_BUFFER_AUXBUFFER, xpos, ypos, 1, 1, 0, &d16);
+      depth[i] = d16;
+      break;
+    case 32:
+      rw_fb_span(fxMesa,
+		 GR_BUFFER_AUXBUFFER,
+		 xpos, ypos,
+		 1,
+		 depth_size/8,
+		 &d32,
+		 FBS_READ);
+      /*
+       * Get rid of the stencil bits.  Should I sign
+       * extend this?  I don't believe so, since GLdepth
+       * is unsigned.
+       */
+      if (stencil_size > 0) {
+	d32 &= 0xFFFFFF;
+      }
+      depth[i] = d32;
+      break;
+    default:
+      assert(0);
+    }
   }
 }
 
 
+/*
+ * Stencil buffer read/write functions.
+ */
 
+#define DEPTH_VALUE_TO_STENCIL_VALUE(d)         (((d) >> 24) & 0xFF)
+#define STENCIL_VALUE_TO_DEPTH_VALUE(s,d) \
+        ((((s) & 0xFF) << 24) | ((d) & 0xFFFFFF))
+#define ASSEMBLE_STENCIL_VALUE(os, ns, mask) \
+        (((os) &~ (mask)) | ((ns) & (mask)))
 
-/************************************************************************/
+#if 00
+/*
+ * Read a horizontal span of stencil values from the stencil buffer.
+ */
+void
+fxDDReadStencilSpan( GLcontext *ctx,
+                     GLuint n,
+                     GLint x,
+                     GLint y,
+                     GLstencil stencil[] )
+{
+  fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
+  GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
+  GLuint i;
+  int xpos = x + fxMesa->x_offset;
+  int ypos = bottom - y;
 
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
+    fprintf(stderr, "fxmesa: fxDDReadStencilSpan(...)\n");
+  }
 
-void fxSetupDDSpanPointers(GLcontext *ctx)
-{
-  ctx->Driver.WriteRGBASpan       =fxDDWriteRGBASpan;
-  ctx->Driver.WriteRGBSpan        =fxDDWriteRGBSpan;
-  ctx->Driver.WriteMonoRGBASpan   =fxDDWriteMonoRGBASpan;
-  ctx->Driver.WriteRGBAPixels     =fxDDWriteRGBAPixels;
-  ctx->Driver.WriteMonoRGBAPixels =fxDDWriteMonoRGBAPixels;
+  for (i = 0; i < n; i++, ypos += 1) {
+    GLuint d;
+    rw_fb_span(fxMesa,
+	       GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_READ);
+    stencil[i] = DEPTH_VALUE_TO_STENCIL_VALUE(d);
+  }
+}
 
-  ctx->Driver.WriteCI8Span        =NULL;
-  ctx->Driver.WriteCI32Span       =NULL;
-  ctx->Driver.WriteMonoCISpan     =NULL;
-  ctx->Driver.WriteCI32Pixels     =NULL;
-  ctx->Driver.WriteMonoCIPixels   =NULL;
+/*
+ * Write a horizontal span of stencil values into the stencil buffer.
+ * If mask is NULL, write all stencil values.
+ * Else, only write stencil[i] if mask[i] is non-zero.
+ */
+void
+fxDDWriteStencilSpan( GLcontext *ctx,
+                      GLuint n,
+                      GLint x,
+                      GLint y,
+                      const GLstencil stencil[],
+                      const GLubyte mask[] )
+{
+  fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
+  GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
+  GLuint i;
+  int xpos = x + fxMesa->x_offset;
+  int ypos = bottom - y;
 
-  /*  ctx->Driver.ReadRGBASpan        =fxDDReadRGBASpan;*/
-  ctx->Driver.ReadRGBASpan = read_R5G6B5_span;
-  ctx->Driver.ReadRGBAPixels      = read_R5G6B5_pixels;
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
+    fprintf(stderr, "fxmesa: fxDDWriteStencilSpan(...)\n");
+  }
 
-  ctx->Driver.ReadCI32Span        =NULL;
-  ctx->Driver.ReadCI32Pixels      =NULL;
+  for (i = 0; i < n; i++, ypos += 1) {
+    GLdepth d;
+    GLstencil ns;
+    rw_fb_span(fxMesa,
+	       GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_READ);
+   /*
+    * Find the old stencil value, and strip off the bits
+    * we are going to write.
+    */
+    ns = ASSEMBLE_STENCIL_VALUE(DEPTH_VALUE_TO_STENCIL_VALUE(d),
+                                stencil[i],
+                                mask[i]);
+   /*
+    * Now, assemble the new StenDepth value, with the old
+    * depth value, the old stencil value in the bits
+    * where mask is 0, and the new stencil value in the bits
+    * where mask is 1.
+    */
+    d = STENCIL_VALUE_TO_DEPTH_VALUE(ns, d);
+    rw_fb_span(fxMesa,
+	       GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_WRITE);
+  }
 }
 
+/* Write an array of stencil values into the stencil buffer.
+ * If mask is NULL, write all stencil values.
+ * Else, only write stencil[i] if mask[i] is non-zero.
+ */
+void
+fxDDReadStencilPixels( GLcontext *ctx,
+                       GLuint n,
+                       const GLint x[],
+                       const GLint y[],
+                       GLstencil stencil[])
+{
+  fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
+  GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
+  GLuint i;
 
-#else
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
+    fprintf(stderr, "fxmesa: fxDDReadStencilPixels(...)\n");
+  }
 
+  for (i = 0; i < n; i++) {
+    int xpos = x[i] + fxMesa->x_offset;
+    int ypos = bottom - y[i];
+    GLuint d;
+    rw_fb_span(fxMesa,
+               GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_READ);
+    stencil[i] = DEPTH_VALUE_TO_STENCIL_VALUE(d);
+  }
+}
 
 /*
- * Need this to provide at least one external definition.
+ * Read an array of stencil values from the stencil buffer.
  */
+void
+fxDDWriteStencilPixels( GLcontext *ctx,
+                        GLuint n,
+                        const GLint x[],
+                        const GLint y[],
+                        const GLstencil stencil[],
+			const GLstencil mask[])
+{
+  fxMesaContext fxMesa = (fxMesaContext)ctx->DriverCtx;
+  GLint bottom = fxMesa->height + fxMesa->y_offset - 1;
+  GLuint i;
+
+  if (MESA_VERBOSE & VERBOSE_DRIVER) {
+    fprintf(stderr, "fxmesa: fxDDWriteStencilPixels(...)\n");
+  }
 
-int gl_fx_dummy_function_span(void)
+  for (i = 0; i < n; i++) {
+    int xpos = x[i] + fxMesa->x_offset;
+    int ypos = bottom - y[i];
+    GLuint d;
+    rw_fb_span(fxMesa,
+               GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_READ);
+    d = STENCIL_VALUE_TO_DEPTH_VALUE(stencil[i], d);
+    rw_fb_span(fxMesa,
+               GR_BUFFER_AUXBUFFER,
+	       xpos, ypos,
+	       1,
+	       sizeof(GLdepth),
+	       &d,
+	       FBS_WRITE);
+  }
+}
+#endif /* disable fxddStencil* funcs */
+
+
+#define EXTRACT_S_FROM_ZS(zs) (((zs) >> 24) & 0xFF)
+#define EXTRACT_Z_FROM_ZS(zs) ((zs) & 0xffffff)
+#define BUILD_ZS(z, s)  (((s) << 24) | (z))
+
+static void write_stencil_span(GLcontext *ctx, GLuint n, GLint x, GLint y,
+                               const GLstencil stencil[],
+                               const GLubyte mask[] )
 {
-  return 0;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbInfo_t info;
+  int s;
+  void *d;
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_WRITE_ONLY,
+                GR_BUFFER_AUXBUFFER,
+                GR_LFBWRITEMODE_Z32,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+    const GLint scrX = winX + x;
+    const GLint scrY = winY - y;
+
+    GLint dstStride = info.strideInBytes;
+    GLubyte *dst = (GLubyte *) info.lfbPtr
+                   + (winY - y) * dstStride + (winX + x) * 4;
+    GLuint *dst32 = (GLuint *) dst;
+    GLubyte visMask[MAX_WIDTH];
+    GLuint i;
+    generate_vismask(fxMesa, scrX, scrY, n, visMask);
+    s = dstStride;
+    d = dst32;
+    for (i = 0; i < n; i++) {
+      if (visMask[i] && (!mask || mask[i])) {
+        GLuint zs = dst32[i];
+        GLuint z = EXTRACT_Z_FROM_ZS(zs);
+        zs = BUILD_ZS(z, stencil[i]);
+        dst32[i] = /*zs;*/ stencil[i];
+      }
+    }
+    grLfbUnlock(GR_LFB_WRITE_ONLY, GR_BUFFER_AUXBUFFER);
+  }
+  else {
+    s = -1;
+    d = 0;
+  }
+  END_BOARD_LOCK();
+  /*
+  printf("write stencil span %d %p\n", s, d);
+  printf("info: size=%d lfbPtr=%p stride=%x writeMode=%x origin=%x\n",
+    info.size, info.lfbPtr, info.strideInBytes, info.writeMode, info.origin);
+  */
+}
+
+
+static void
+read_stencil_span(GLcontext *ctx, GLuint n, GLint x, GLint y,
+                  GLstencil stencil[])
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  GrLfbInfo_t info;
+  /*
+  int s;
+  void *d;
+  */
+
+  BEGIN_BOARD_LOCK();
+  info.size = sizeof(info);
+  if (grLfbLock(GR_LFB_READ_ONLY,
+                GR_BUFFER_AUXBUFFER,
+                GR_LFBWRITEMODE_Z32,
+                GR_ORIGIN_UPPER_LEFT,
+                FXFALSE,
+                &info)) {
+    const GLint winY = fxMesa->y_offset + fxMesa->height - 1;
+    const GLint winX = fxMesa->x_offset;
+    GLint srcStride = /*info.strideInBytes;*/ 8192 ;  /* XXX a hack! */
+    const GLubyte *src = (const GLubyte *) info.lfbPtr
+                       + (winY - y) * srcStride + (winX + x) * 4;
+    const GLuint *src32 = (const GLuint *) src;
+    GLuint i;
+    /*
+    s = srcStride;
+    d = src32;
+    */
+    for (i = 0; i < n; i++) {
+      stencil[i] = EXTRACT_S_FROM_ZS(src32[i]);
+    }
+    grLfbUnlock(GR_LFB_READ_ONLY, GR_BUFFER_AUXBUFFER);
+  }
+  END_BOARD_LOCK();
+  /*
+  printf("read stencil span %d %p\n", s, d);
+  printf("info: size=%d lfbPtr=%p stride=%x writeMode=%x origin=%x\n",
+    info.size, info.lfbPtr, info.strideInBytes, info.writeMode, info.origin);
+  */
+}
+
+
+static void
+write_stencil_pixels( GLcontext *ctx, GLuint n,
+                      const GLint x[], const GLint y[],
+                      const GLstencil stencil[], const GLubyte mask[])
+{
+  /* XXX optimize this */
+  GLuint i;
+  for (i = 0; i < n; i++) {
+    if (mask[i]) {
+      write_stencil_span(ctx, 1, x[i], y[i], stencil + i, mask + i);
+    }
+  }
+}
+
+
+static void
+read_stencil_pixels(GLcontext *ctx, GLuint n, const GLint x[], const GLint y[],
+                    GLstencil stencil[])
+{
+  /* XXX optimize this */
+  GLuint i;
+  for (i = 0; i < n; i++) {
+    read_stencil_span(ctx, 1, x[i], y[i], stencil + i);
+  }
 }
 
-#endif  /* FX */
+
+
+void fxSetupDDSpanPointers(GLcontext *ctx)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+
+  if (ctx->Visual->RedBits   == 5 &&
+      ctx->Visual->GreenBits == 6 &&
+      ctx->Visual->BlueBits  == 5 &&
+      ctx->Visual->AlphaBits == 0) {
+    /* 16bpp mode */
+    ctx->Driver.WriteRGBASpan       = write_R5G6B5_rgba_span;
+    ctx->Driver.WriteRGBSpan        = write_R5G6B5_rgb_span;
+    ctx->Driver.WriteMonoRGBASpan   = write_R5G6B5_mono_span;
+    ctx->Driver.WriteRGBAPixels     = write_R5G6B5_pixels;
+    ctx->Driver.WriteMonoRGBAPixels = write_R5G6B5_mono_pixels;
+    ctx->Driver.ReadRGBASpan        = read_R5G6B5_span;
+    ctx->Driver.ReadRGBAPixels      = read_R5G6B5_pixels;
+  }
+  else if (ctx->Visual->RedBits   == 8 &&
+           ctx->Visual->GreenBits == 8 &&
+           ctx->Visual->BlueBits  == 8 &&
+           ctx->Visual->AlphaBits == 0) {
+    /* 24bpp mode */
+    ctx->Driver.WriteRGBASpan       = write_R8G8B8_rgba_span;
+    ctx->Driver.WriteRGBSpan        = write_R8G8B8_rgb_span;
+    ctx->Driver.WriteMonoRGBASpan   = write_R8G8B8_mono_span;
+    ctx->Driver.WriteRGBAPixels     = write_R8G8B8_pixels;
+    ctx->Driver.WriteMonoRGBAPixels = write_R8G8B8_mono_pixels;
+    ctx->Driver.ReadRGBASpan        = read_R8G8B8_span;
+    ctx->Driver.ReadRGBAPixels      = read_R8G8B8_pixels;
+  }
+  else if (ctx->Visual->RedBits   == 8 &&
+           ctx->Visual->GreenBits == 8 &&
+           ctx->Visual->BlueBits  == 8 &&
+           ctx->Visual->AlphaBits == 8) {
+    /* 32bpp mode */
+    ctx->Driver.WriteRGBASpan       = write_R8G8B8A8_rgba_span;
+    ctx->Driver.WriteRGBSpan        = write_R8G8B8A8_rgb_span;
+    ctx->Driver.WriteMonoRGBASpan   = write_R8G8B8A8_mono_span;
+    ctx->Driver.WriteRGBAPixels     = write_R8G8B8A8_pixels;
+    ctx->Driver.WriteMonoRGBAPixels = write_R8G8B8A8_mono_pixels;
+    ctx->Driver.ReadRGBASpan        = read_R8G8B8A8_span;
+    ctx->Driver.ReadRGBAPixels      = read_R8G8B8A8_pixels;
+  }
+  else {
+    abort();
+  }
+
+  if (fxMesa->haveHwStencil) {
+    ctx->Driver.WriteStencilSpan    = write_stencil_span;
+    ctx->Driver.ReadStencilSpan     = read_stencil_span;
+    ctx->Driver.WriteStencilPixels  = write_stencil_pixels;
+    ctx->Driver.ReadStencilPixels   = read_stencil_pixels;
+  }
+
+  ctx->Driver.WriteCI8Span        = NULL;
+  ctx->Driver.WriteCI32Span       = NULL;
+  ctx->Driver.WriteMonoCISpan     = NULL;
+  ctx->Driver.WriteCI32Pixels     = NULL;
+  ctx->Driver.WriteMonoCIPixels   = NULL;
+  ctx->Driver.ReadCI32Span        = NULL;
+  ctx->Driver.ReadCI32Pixels      = NULL;
+}
diff --git a/xc/extras/Mesa/src/FX/fxddtex.c b/xc/extras/Mesa/src/FX/fxddtex.c
index db0acb6b4..69ceebaa9 100644
--- a/xc/extras/Mesa/src/FX/fxddtex.c
+++ b/xc/extras/Mesa/src/FX/fxddtex.c
@@ -496,209 +496,68 @@ static int logbase2(int n)
 
 /* Need different versions for different cpus.
  */
-#define INT_TRICK(l2) (0x800000 * l2)
-
-
-int fxTexGetInfo(int w, int h, GrLOD_t *lodlevel, GrAspectRatio_t *ar,
+#define INT_TRICK(l2) (0x800000 * (l2))
+int fxTexGetInfo(int w, int h, GrLOD_t *lodlevel, 
+		 GrAspectRatio_t *aspectratio,
                  float *sscale, float *tscale,
                  int *i_sscale, int *i_tscale,
                  int *wscale, int *hscale)
 {
-
-  static GrLOD_t lod[9]={GR_LOD_256,GR_LOD_128,GR_LOD_64,GR_LOD_32,
-                         GR_LOD_16,GR_LOD_8,GR_LOD_4,GR_LOD_2,GR_LOD_1};
-
-  int logw,logh,ws,hs;
-  GrLOD_t l;
-  GrAspectRatio_t aspectratio;
-  float s,t;
-  int is,it;
+  int logw, logh, ar, l, is, it, ws, hs;
+  float s, t;
 
   logw=logbase2(w);
   logh=logbase2(h);
-
-  switch(logw-logh) {
-  case 0:
-    aspectratio=GR_ASPECT_1x1;
-    l=lod[8-logw];
-    s=t=256.0f;
-    is=it=INT_TRICK(8);
-    ws=hs=1;
-    break;
-  case 1:
-    aspectratio=GR_ASPECT_2x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=128.0f;
-    is=INT_TRICK(8);it=INT_TRICK(7);
-    ws=1;
-    hs=1;
-    break;
-  case 2:
-    aspectratio=GR_ASPECT_4x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=64.0f;
-    is=INT_TRICK(8);it=INT_TRICK(6);
-    ws=1;
-    hs=1;
-    break;
-  case 3:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=1;
-    break;
-  case 4:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=2;
-    break;
-  case 5:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=4;
-    break;
-  case 6:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=8;
-    break;
-  case 7:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=16;
-    break;
-  case 8:
-    aspectratio=GR_ASPECT_8x1;
-    l=lod[8-logw];
-    s=256.0f;
-    t=32.0f;
-    is=INT_TRICK(8);it=INT_TRICK(5);
-    ws=1;
-    hs=32;
-    break;
-  case -1:
-    aspectratio=GR_ASPECT_1x2;
-    l=lod[8-logh];
-    s=128.0f;
-    t=256.0f;
-    is=INT_TRICK(7);it=INT_TRICK(8);
-    ws=1;
-    hs=1;
-    break;
-  case -2:
-    aspectratio=GR_ASPECT_1x4;
-    l=lod[8-logh];
-    s=64.0f;
-    t=256.0f;
-    is=INT_TRICK(6);it=INT_TRICK(8);
+  ar=logw-logh;
+  /* Hardware only allows a maximum aspect ratio of 8x1, so handle
+     |ar|>3 by scaling the image and using an 8x1 aspect ratio */
+  if (ar>=0) {
+    l=logw;
+    s=256.0;
+    is=INT_TRICK(8);
     ws=1;
+    if (ar<3) {
+      t=256>>ar;
+      it=INT_TRICK(8-ar);
+      hs=1;
+    } else {
+      t=32.0;
+      it=INT_TRICK(5);
+      hs=1<<(ar-3);
+    }
+  } else {
+    l=logh;
+    t=256.0;
+    it=INT_TRICK(8);
     hs=1;
-    break;
-  case -3:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=1;
-    hs=1;
-    break;
-  case -4:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=2;
-    hs=1;
-    break;
-  case -5:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=4;
-    hs=1;
-    break;
-  case -6:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=8;
-    hs=1;
-    break;
-  case -7:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=16;
-    hs=1;
-    break;
-  case -8:
-    aspectratio=GR_ASPECT_1x8;
-    l=lod[8-logh];
-    s=32.0f;
-    t=256.0f;
-    is=INT_TRICK(5);it=INT_TRICK(8);
-    ws=32;
-    hs=1;
-    break;
-  default:
-    return 0;
-    break;
+    if (-ar<3) {
+      s=256>>-ar;
+      is=INT_TRICK(8+ar);
+      ws=1;
+    } else {
+      s=32.0;
+      is=INT_TRICK(5);
+      ws=1<<(-ar-3);
+    }
   }
-
-  if(lodlevel)
-    (*lodlevel)=l;
-
-  if(ar)
-    (*ar)=aspectratio;
-
-  if(sscale)
-    (*sscale)=s;
-
-  if(tscale)
-    (*tscale)=t;
-
-  if(wscale)
-    (*wscale)=ws;
-
-  if(hscale)
-    (*hscale)=hs;
-
-  if (i_sscale)
-     *i_sscale = is;
-
-  if (i_tscale)
-     *i_tscale = it;
-
-
+  if (ar<-3) ar=-3;
+  if (ar>3) ar=3;
+
+  /* The above numbers are calculated sensibly and work for Glide3, but
+     we change them to the whacky glide2 values if needed. */
+#ifdef FX_GLIDE3
+  if (lodlevel) *lodlevel=l;
+  if (aspectratio) *aspectratio=ar;
+#else
+  if (lodlevel) *lodlevel=8-l;
+  if (aspectratio) *aspectratio=3-ar;
+#endif
+  if (sscale) *sscale=s;
+  if (tscale) *tscale=t;
+  if (wscale) *wscale=ws;
+  if (hscale) *hscale=hs;
+  if (i_sscale) *i_sscale = is;
+  if (i_tscale) *i_tscale = it;
   return 1;
 }
 
@@ -758,12 +617,17 @@ void fxTexGetFormat(GLenum glformat, GrTextureFormat_t *tfmt, GLint *ifmt)
     case GL_R3_G3_B2:
     case GL_RGB4:
     case GL_RGB5:
+      if(tfmt)
+        (*tfmt)=GR_TEXFMT_RGB_565;
+      if(ifmt)
+        (*ifmt)=GL_RGB;
+      break;
     case GL_RGB8:
     case GL_RGB10:
     case GL_RGB12:
     case GL_RGB16:
       if(tfmt)
-        (*tfmt)=GR_TEXFMT_RGB_565;
+        (*tfmt)=GR_TEXFMT_ARGB_8888;
       if(ifmt)
         (*ifmt)=GL_RGB;
       break;
@@ -771,12 +635,17 @@ void fxTexGetFormat(GLenum glformat, GrTextureFormat_t *tfmt, GLint *ifmt)
     case GL_RGBA:
     case GL_RGBA2:
     case GL_RGBA4:
+      if(tfmt)
+        (*tfmt)=GR_TEXFMT_ARGB_4444;
+      if(ifmt)
+        (*ifmt)=GL_RGBA;
+      break;
     case GL_RGBA8:
     case GL_RGB10_A2:
     case GL_RGBA12:
     case GL_RGBA16:
       if(tfmt)
-        (*tfmt)=GR_TEXFMT_ARGB_4444;
+        (*tfmt)=GR_TEXFMT_ARGB_8888;
       if(ifmt)
         (*ifmt)=GL_RGBA;
       break;
@@ -910,23 +779,29 @@ GLboolean fxDDTexImage2D(GLcontext *ctx, GLenum target, GLint level,
       case GL_R3_G3_B2:
       case GL_RGB4:
       case GL_RGB5:
+        texelSize = 2;
+        intFormat = MESA_R5_G6_B5;
+        break;
       case GL_RGB8:
       case GL_RGB10:
       case GL_RGB12:
       case GL_RGB16:
-        texelSize = 2;
-        intFormat = MESA_R5_G6_B5;
+        texelSize = 4;
+        intFormat = MESA_A8_R8_G8_B8;
         break;
       case 4:
       case GL_RGBA:
       case GL_RGBA2:
       case GL_RGBA4:
+        texelSize = 2;
+        intFormat = MESA_A4_R4_G4_B4;
+        break;
       case GL_RGBA8:
       case GL_RGB10_A2:
       case GL_RGBA12:
       case GL_RGBA16:
-        texelSize = 2;
-        intFormat = MESA_A4_R4_G4_B4;
+        texelSize = 4;
+        intFormat = MESA_A8_R8_G8_B8;
         break;
       case GL_RGB5_A1:
         texelSize = 2;
@@ -948,6 +823,7 @@ GLboolean fxDDTexImage2D(GLcontext *ctx, GLenum target, GLint level,
       mml->data = MALLOC(dstWidth * dstHeight * texelSize);
       if (!mml->data)
         return GL_FALSE;
+      mml->texelSize = texelSize;
       mml->glideFormat = gldformat;
       mml->width = dstWidth;
       mml->height = dstHeight;
@@ -1060,6 +936,14 @@ GLboolean fxDDTexSubImage2D(GLcontext *ctx, GLenum target, GLint level,
                                          texImage->Width, texImage->Height,
                                          format, type, pixels, packing);
       break;
+    case GR_TEXFMT_ARGB_8888:
+      dstStride = mml->width * 4;
+      result = _mesa_convert_texsubimage(MESA_A8_R8_G8_B8, xoffset, yoffset,
+                                         mml->width, mml->height, mml->data,
+                                         dstStride, width, height,
+                                         texImage->Width, texImage->Height,
+                                         format, type, pixels, packing);
+      break;
     case GR_TEXFMT_ARGB_1555:
       dstStride = mml->width * 2;
       result = _mesa_convert_texsubimage(MESA_A1_R5_G5_B5, xoffset, yoffset,
@@ -1149,6 +1033,11 @@ GLvoid *fxDDGetTexImage(GLcontext *ctx, GLenum target, GLint level,
         glFormat = GL_RGB;
         srcStride = mml->width * 2;
         break;
+      case GR_TEXFMT_ARGB_8888:
+        mesaFormat = MESA_A8_R8_G8_B8;
+        glFormat = GL_RGBA;
+        srcStride = mml->width * 4;
+        break;
       case GR_TEXFMT_ARGB_4444:
         mesaFormat = MESA_A4_R4_G4_B4;
         glFormat = GL_RGBA;
diff --git a/xc/extras/Mesa/src/FX/fxdrv.h b/xc/extras/Mesa/src/FX/fxdrv.h
index c85c73d8e..eea75fc56 100644
--- a/xc/extras/Mesa/src/FX/fxdrv.h
+++ b/xc/extras/Mesa/src/FX/fxdrv.h
@@ -162,6 +162,7 @@ typedef struct {
 #endif
 #endif
 
+
 #define FX_VB_COLOR(fxm, color)				\
   do {							\
     if (sizeof(GLint) == 4*sizeof(GLubyte)) {		\
@@ -257,6 +258,20 @@ typedef struct {
 #define FX_UM_ALPHA_ITERATED        0x04000000
 #define FX_UM_ALPHA_CONSTANT        0x08000000
 
+
+#define PACK_BGRA32(R, G, B, A)  \
+    ( (((GLuint) (R)) << 16) | \
+      (((GLuint) (G)) <<  8) | \
+      (((GLuint) (B))      ) | \
+      (((GLuint) (A)) << 24) )
+
+#define PACK_RGBA32(R, G, B, A)  \
+    ( (((GLuint) (R))      ) | \
+      (((GLuint) (G)) <<  8) | \
+      (((GLuint) (B)) << 16) | \
+      (((GLuint) (A)) << 24) )
+
+
 typedef void (*tfxRenderVBFunc)(GLcontext *);
 
 /*
@@ -269,6 +284,7 @@ typedef struct MemRange_t {
 
 typedef struct {
   GLsizei width, height;              /* image size */
+  GLint texelSize;		      /* How many bytes to a texel */
   GrTextureFormat_t glideFormat;      /* Glide image format */
   unsigned short *data;               /* Glide-formated texture image */
 } tfxMipMapLevel;
@@ -339,27 +355,20 @@ extern tfxLineClipFunc fxLineClipTab[0x8];
 
 typedef struct {
   /* Alpha test */
-
   GLboolean alphaTestEnabled;
   GrCmpFnc_t alphaTestFunc;
   GrAlpha_t alphaTestRefValue;
 
   /* Blend function */
-
   GLboolean blendEnabled;
   GrAlphaBlendFnc_t blendSrcFuncRGB;
   GrAlphaBlendFnc_t blendDstFuncRGB;
   GrAlphaBlendFnc_t blendSrcFuncAlpha;
   GrAlphaBlendFnc_t blendDstFuncAlpha;
-
-  /* Depth test */
-
-  GLboolean depthTestEnabled;
-  GLboolean depthMask;
-  GrCmpFnc_t depthTestFunc;
 } tfxUnitsState;
 
 
+
 /* Flags for render_index.
  */
 #define FX_OFFSET             0x1
@@ -380,6 +389,7 @@ typedef struct {
 #define FX_NEW_SCISSOR        0x20
 #define FX_NEW_COLOR_MASK     0x40
 #define FX_NEW_CULL           0x80
+#define FX_NEW_STENCIL        0x100
 
 /* FX struct stored in VB->driver_data.
  */
@@ -421,13 +431,9 @@ extern GLubyte FX_PixelToB[0x10000];
 
 
 struct tfxMesaContext {
-  GuTexPalette glbPalette;
-
   GLcontext *glCtx;              /* the core Mesa context */
-#if !defined(XFree86Server) && !defined(GLX_DIRECT_RENDERING)
   GLvisual *glVis;               /* describes the color buffer */
   GLframebuffer *glBuffer;       /* the ancillary buffers */
-#endif
 
   GLint board;                   /* the board used for this context */
   GLint width, height;           /* size of color buffer */
@@ -435,6 +441,7 @@ struct tfxMesaContext {
   GrBuffer_t currentFB;
 
   GLboolean bgrOrder;
+  GLuint depthClear;
   GrColor_t color;
   GrColor_t clearC;
   GrAlpha_t clearA;
@@ -444,6 +451,8 @@ struct tfxMesaContext {
   tfxUnitsState unitsState;
   tfxUnitsState restoreUnitsState; /* saved during multipass */
 
+  GuTexPalette glbPalette;
+
   GLuint tmu_source[FX_NUM_TMU];
   GLuint tex_dest[MAX_TEXTURE_UNITS];
   GLuint setupindex;
@@ -467,7 +476,6 @@ struct tfxMesaContext {
 
   GLuint texBindNumber;
   GLint tmuSrc;
-  GLuint lastUnitsMode;
   GLuint freeTexMem[FX_NUM_TMU];
   MemRange *tmPool;
   MemRange *tmFree[FX_NUM_TMU];
@@ -502,8 +510,7 @@ struct tfxMesaContext {
   GLboolean haveTwoTMUs;	/* True if we really have 2 tmu's  */
   GLboolean emulateTwoTMUs;	/* True if we present 2 tmu's to mesa.  */
   GLboolean haveAlphaBuffer;
-  GLboolean haveZBuffer;
-  GLboolean haveDoubleBuffer;
+  GLboolean haveHwStencil;
   GLboolean haveGlobalPaletteTexture;
   GLint swapInterval;
   GLint maxPendingSwapBuffers;
@@ -592,8 +599,6 @@ extern void fxDDTexUseGlbPalette(GLcontext *, GLboolean);
 extern void fxDDEnable(GLcontext *, GLenum, GLboolean);
 extern void fxDDAlphaFunc(GLcontext *, GLenum, GLclampf);
 extern void fxDDBlendFunc(GLcontext *, GLenum, GLenum);
-extern void fxDDDepthMask(GLcontext *, GLboolean);
-extern void fxDDDepthFunc(GLcontext *, GLenum);
 
 extern void fxDDRegisterVB( struct vertex_buffer *VB );
 extern void fxDDUnregisterVB( struct vertex_buffer *VB );
diff --git a/xc/extras/Mesa/src/FX/fxglidew.c b/xc/extras/Mesa/src/FX/fxglidew.c
index 5155c825a..f01644a60 100644
--- a/xc/extras/Mesa/src/FX/fxglidew.c
+++ b/xc/extras/Mesa/src/FX/fxglidew.c
@@ -55,6 +55,13 @@
 #include <stdlib.h>
 #include <string.h>
 
+
+grStencilFunc_t grStencilFuncPtr = NULL;
+grStencilMask_t grStencilMaskPtr = NULL;
+grStencilOp_t grStencilOpPtr = NULL;
+grBufferClearExt_t grBufferClearExtPtr = NULL;
+
+
 FxI32 FX_grGetInteger_NoLock(FxU32 pname)
 {
 #if !defined(FX_GLIDE3)
@@ -71,6 +78,8 @@ FxI32 FX_grGetInteger_NoLock(FxU32 pname)
     case FX_TEXTURE_ALIGN:
         /* This is a guess from reading the glide3 docs */
         return 8;
+    case FX_ZDEPTH_MAX:
+      return 0xFFFF;
     default:
        if (MESA_VERBOSE&VERBOSE_DRIVER) {
           fprintf(stderr,"Wrong parameter in FX_grGetInteger!\n");
@@ -90,6 +99,12 @@ FxI32 FX_grGetInteger_NoLock(FxU32 pname)
      case FX_TEXTURE_ALIGN:
        grname = pname;
        break;
+     case FX_ZDEPTH_MAX: {
+       int zvals[2];
+
+       grGet(GR_ZDEPTH_MIN_MAX, 8, zvals);
+       return zvals[0];
+     }
      default:
        if (MESA_VERBOSE&VERBOSE_DRIVER) {
           fprintf(stderr,"Wrong parameter in FX_grGetInteger!\n");
diff --git a/xc/extras/Mesa/src/FX/fxglidew.h b/xc/extras/Mesa/src/FX/fxglidew.h
index 6da1ea5f3..4ad2942af 100644
--- a/xc/extras/Mesa/src/FX/fxglidew.h
+++ b/xc/extras/Mesa/src/FX/fxglidew.h
@@ -47,7 +47,51 @@
 #define __FX_GLIDE_WARPER__
 
 #include <glide.h>
+#include <g3ext.h>
 
+/*
+ * These are glide extension definitions.  These are not
+ * defined in glide.h.  They should really be defined in
+ * g3ext.h, but they are not.
+ */
+#if 0
+FX_ENTRY void FX_CALL 
+grStencilFunc(GrCmpFnc_t fnc, GrStencil_t ref, GrStencil_t mask);
+
+FX_ENTRY void FX_CALL 
+grStencilMask(GrStencil_t write_mask);
+
+FX_ENTRY void FX_CALL 
+grStencilOp(
+            GrStencilOp_t stencil_fail,
+            GrStencilOp_t depth_fail,
+            GrStencilOp_t depth_pass);
+
+FX_ENTRY void FX_CALL 
+grBufferClearExt(
+              GrColor_t color,
+              GrAlpha_t alpha,
+              FxU32     depth,
+              GrStencil_t stencil);
+#endif
+
+
+typedef void (*grStencilFunc_t)(GrCmpFnc_t fnc, GrStencil_t ref, GrStencil_t mask);
+typedef void (*grStencilMask_t)(GrStencil_t write_mask);
+typedef void (*grStencilOp_t)(GrStencilOp_t stencil_fail, GrStencilOp_t depth_fail, GrStencilOp_t depth_pass);
+typedef void (*grBufferClearExt_t)(GrColor_t color, GrAlpha_t alpha, FxU32 depth, GrStencil_t stencil);
+
+extern grStencilFunc_t grStencilFuncPtr;
+extern grStencilMask_t grStencilMaskPtr;
+extern grStencilOp_t grStencilOpPtr;
+extern grBufferClearExt_t grBufferClearExtPtr;
+
+
+FX_ENTRY void FX_CALL
+grEnable(GrEnableMode_t mode);
+
+FX_ENTRY void FX_CALL
+grEnable(GrEnableMode_t mode);
 /* 
  * General context: 
  */
@@ -74,12 +118,13 @@
 	#define FX_PENDING_BUFFERSWAPS          GR_PENDING_BUFFERSWAPS
 	#define FX_TEXTURE_ALIGN		GR_TEXTURE_ALIGN
 #endif
+#define FX_ZDEPTH_MAX 0x100
 
 /*
  * Genral warper functions for Glide2/Glide3:
  */ 
-extern FxI32 FX_grGetInteger(FxU32 pname);
 extern FxI32 FX_grGetInteger_NoLock(FxU32 pname);
+extern FxI32 FX_grGetInteger(FxU32 pname);
 
 /*
  * Glide2 emulation on Glide3:
@@ -253,7 +298,6 @@ typedef struct
 
 #endif
 
-
 /*
  * Glide2 functions for Glide3
  */
@@ -567,6 +611,55 @@ extern void FX_grDrawPolygonVertexList(int n, GrVertex *v);
     END_CLIP_LOOP();			\
   } while (0)
 
+#define FX_grBufferClearExt(c, a, d, s)	\
+  do {					\
+    BEGIN_CLIP_LOOP();			\
+    (*grBufferClearExtPtr)(c, a, d, s);	\
+    END_CLIP_LOOP();			\
+  } while (0)
+
+/*
+ * Enable/Disable
+ */
+#define FX_grEnable(m)                   \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    grEnable(m);                         \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+#define FX_grDisable(m)                  \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    grDisable(m);                        \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+/*
+ * Stencil operations.
+ */
+#define FX_grStencilFunc(fnc, ref, mask) \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    (*grStencilFuncPtr)((fnc), (ref), (mask)); \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+#define FX_grStencilMask(write_mask)     \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    (*grStencilMaskPtr)(write_mask);           \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+
+#define FX_grStencilOp(stencil_fail, depth_fail, depth_pass) \
+  do {				                             \
+    BEGIN_BOARD_LOCK();		                             \
+    (*grStencilOpPtr)((stencil_fail), (depth_fail), (depth_pass)); \
+    END_BOARD_LOCK();		                             \
+  } while (0)
+
 #define FX_grDepthMask(m)	\
   do {				\
     BEGIN_BOARD_LOCK();		\
diff --git a/xc/extras/Mesa/src/FX/fxsetup.c b/xc/extras/Mesa/src/FX/fxsetup.c
index f1a56650b..f07a2ecec 100644
--- a/xc/extras/Mesa/src/FX/fxsetup.c
+++ b/xc/extras/Mesa/src/FX/fxsetup.c
@@ -62,16 +62,11 @@ static void fxSetupTextureSingleTMU_NoLock(GLcontext *ctx, GLuint textureset);
 static void fxSetupDoubleTMU_NoLock(fxMesaContext fxMesa, 
 			     struct gl_texture_object *tObj0,
 			     struct gl_texture_object *tObj1);
-static void fxSetupTexture_NoLock(GLcontext *ctx);
-static void fxSetupTexture(GLcontext *ctx);
-static void fxSetupBlend(GLcontext *ctx);
-static void fxSetupDepthTest(GLcontext *ctx);
-static void fxSetupScissor(GLcontext *ctx);
-static void fxSetupCull(GLcontext *ctx);
 static void gl_print_fx_state_flags( const char *msg, GLuint flags);
-/*static GLboolean fxMultipassBlend(struct vertex_buffer *, GLuint);*/
 static GLboolean fxMultipassTexture( struct vertex_buffer *, GLuint );
 
+
+
 static void fxTexValidate(GLcontext *ctx, struct gl_texture_object *tObj)
 {
   tfxTexInfo *ti=fxTMGetTexInfo(tObj);
@@ -478,11 +473,6 @@ static void fxSetupTextureSingleTMU_NoLock(GLcontext *ctx, GLuint textureset)
   else
     unitsmode=fxGetTexSetConfiguration(ctx,NULL,tObj);
 
-/*    if(fxMesa->lastUnitsMode==unitsmode) */
-/*      return; */
-
-  fxMesa->lastUnitsMode=unitsmode;
-
   fxMesa->stw_hint_state = 0;
   FX_grHints_NoLock(GR_HINT_STWHINT,0);
 
@@ -600,7 +590,8 @@ static void fxSetupTextureSingleTMU_NoLock(GLcontext *ctx, GLuint textureset)
   }
 }
 
-static void fxSetupTextureSingleTMU(GLcontext *ctx, GLuint textureset) {
+static void fxSetupTextureSingleTMU(GLcontext *ctx, GLuint textureset)
+{
   BEGIN_BOARD_LOCK();
   fxSetupTextureSingleTMU_NoLock(ctx, textureset);
   END_BOARD_LOCK();
@@ -765,11 +756,6 @@ static void fxSetupTextureDoubleTMU_NoLock(GLcontext *ctx)
 
   unitsmode=fxGetTexSetConfiguration(ctx,tObj0,tObj1);
 
-/*    if(fxMesa->lastUnitsMode==unitsmode) */
-/*      return; */
-
-  fxMesa->lastUnitsMode=unitsmode;
-
   fxMesa->stw_hint_state |= GR_STWHINT_ST_DIFF_TMU1;
   FX_grHints_NoLock(GR_HINT_STWHINT, fxMesa->stw_hint_state);
 
@@ -1055,8 +1041,6 @@ static void fxSetupTextureNone_NoLock(GLcontext *ctx)
 			   localc,
 			   GR_COMBINE_OTHER_NONE,
 			   FXFALSE);
-
-  fxMesa->lastUnitsMode=FX_UM_NONE;
 }
 
 /************************************************************************/
@@ -1107,7 +1091,8 @@ static void fxSetupTexture_NoLock(GLcontext *ctx)
   }
 }
 
-static void fxSetupTexture(GLcontext *ctx) {
+static void fxSetupTexture(GLcontext *ctx)
+{
   BEGIN_BOARD_LOCK();
   fxSetupTexture_NoLock(ctx);
   END_BOARD_LOCK();
@@ -1240,15 +1225,19 @@ void fxDDBlendFunc(GLcontext *ctx, GLenum sfactor, GLenum dfactor)
 
 static void fxSetupBlend(GLcontext *ctx)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  tfxUnitsState *us = &fxMesa->unitsState;
+
+  assert(us->blendEnabled == ctx->Color.BlendEnabled);
 
-  if(us->blendEnabled)
-     FX_grAlphaBlendFunction(us->blendSrcFuncRGB,us->blendDstFuncRGB,
-			  us->blendSrcFuncAlpha,us->blendDstFuncAlpha);
+  if (us->blendEnabled)
+     FX_grAlphaBlendFunction(us->blendSrcFuncRGB, us->blendDstFuncRGB,
+                             us->blendSrcFuncAlpha, us->blendDstFuncAlpha);
   else
-     FX_grAlphaBlendFunction(GR_BLEND_ONE,GR_BLEND_ZERO,GR_BLEND_ONE,GR_BLEND_ZERO);
+     FX_grAlphaBlendFunction(GR_BLEND_ONE, GR_BLEND_ZERO, 
+                             GR_BLEND_ONE, GR_BLEND_ZERO);
 }
+
   
 /************************************************************************/
 /************************** Alpha Test SetUp ****************************/
@@ -1256,8 +1245,8 @@ static void fxSetupBlend(GLcontext *ctx)
 
 void fxDDAlphaFunc(GLcontext *ctx, GLenum func, GLclampf ref)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  tfxUnitsState *us = &fxMesa->unitsState;
   GrCmpFnc_t newfunc;
 
   switch(func) {
@@ -1317,83 +1306,118 @@ static void fxSetupAlphaTest(GLcontext *ctx)
      FX_grAlphaTestFunction(GR_CMP_ALWAYS);
 }
 
-/************************************************************************/
-/************************** Depth Test SetUp ****************************/
-/************************************************************************/
 
-void fxDDDepthFunc(GLcontext *ctx, GLenum func)
+/*
+ * Evaluate all depth-test state and make the Glide calls.
+ */
+static void
+fxSetupDepthTest(GLcontext *ctx)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
-  GrCmpFnc_t dfunc;
-
-  switch(func) {
-  case GL_NEVER:
-    dfunc=GR_CMP_NEVER;
-    break;
-  case GL_LESS:
-    dfunc=GR_CMP_LESS;
-    break;
-  case GL_GEQUAL:
-    dfunc=GR_CMP_GEQUAL;
-    break;
-  case GL_LEQUAL:
-    dfunc=GR_CMP_LEQUAL;
-    break;
-  case GL_GREATER:
-    dfunc=GR_CMP_GREATER;
-    break;
-  case GL_NOTEQUAL:
-    dfunc=GR_CMP_NOTEQUAL;
-    break;
-  case GL_EQUAL:
-    dfunc=GR_CMP_EQUAL;
-    break;
-  case GL_ALWAYS:
-    dfunc=GR_CMP_ALWAYS;
-    break;
-  default:
-    fprintf(stderr,"fx Driver: internal error in fxDDDepthFunc()\n");
-    fxCloseHardware();
-    exit(-1);
-    break;
+  if (ctx->Depth.Test) {
+    GrCmpFnc_t dfunc;
+    switch (ctx->Depth.Func) {
+      case GL_NEVER:
+        dfunc = GR_CMP_NEVER;
+        break;
+      case GL_LESS:
+        dfunc = GR_CMP_LESS;
+        break;
+      case GL_GEQUAL:
+        dfunc = GR_CMP_GEQUAL;
+        break;
+      case GL_LEQUAL:
+        dfunc = GR_CMP_LEQUAL;
+        break;
+      case GL_GREATER:
+        dfunc = GR_CMP_GREATER;
+        break;
+      case GL_NOTEQUAL:
+        dfunc = GR_CMP_NOTEQUAL;
+        break;
+      case GL_EQUAL:
+        dfunc = GR_CMP_EQUAL;
+        break;
+      case GL_ALWAYS:
+        dfunc = GR_CMP_ALWAYS;
+        break;
+      default:
+        gl_problem(ctx, "bad depth mode in fxSetupDepthTest");
+        dfunc = GR_CMP_ALWAYS;
+    }
+    FX_grDepthBufferFunction(dfunc);
+    FX_grDepthMask(ctx->Depth.Mask);
   }
-
-  if(dfunc!=us->depthTestFunc) {
-    us->depthTestFunc=dfunc;
-    fxMesa->new_state |= FX_NEW_DEPTH;
-    ctx->Driver.RenderStart = fxSetupFXUnits;
+  else {
+    /* depth test always passes, don't update Z buffer */
+    FX_grDepthBufferFunction(GR_CMP_ALWAYS);
+    FX_grDepthMask(FXFALSE);
   }
-
 }
 
-void fxDDDepthMask(GLcontext *ctx, GLboolean flag)
-{
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
 
-  if(flag!=us->depthMask) {
-    us->depthMask=flag;
-    fxMesa->new_state |= FX_NEW_DEPTH;
-    ctx->Driver.RenderStart = fxSetupFXUnits;
+/*
+ * Evaluate all stencil state and make the Glide calls.
+ */
+static GrStencil_t
+fxConvertGLStencilOp(GLenum op)
+{
+  switch (op) {
+    case GL_KEEP:
+      return GR_STENCILOP_KEEP;
+    case GL_ZERO:
+      return GR_STENCILOP_ZERO;
+    case GL_REPLACE:
+      return GR_STENCILOP_REPLACE;
+    case GL_INCR:
+      return GR_STENCILOP_INCR_CLAMP;
+    case GL_DECR:
+      return GR_STENCILOP_DECR_CLAMP;
+    case GL_INVERT:
+      return GR_STENCILOP_INVERT;
+    default:
+      gl_problem(NULL, "bad stencil op in fxConvertGLStencilOp");
   }
+  return GR_STENCILOP_KEEP;  /* never get, silence compiler warning */
 }
 
-static void fxSetupDepthTest(GLcontext *ctx)
+/*
+ * This function is called just before any rendering is done.
+ * It will validate the stencil parameters.
+ */
+static void
+fxSetupStencilTest(GLcontext *ctx)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
-
-  if (us->depthTestEnabled) {
-     FX_grDepthBufferFunction(us->depthTestFunc);
-     FX_grDepthMask(us->depthMask);
+  if (ctx->Stencil.Enabled) {
+    GrStencil_t sfail = fxConvertGLStencilOp(ctx->Stencil.FailFunc);
+    GrStencil_t zfail = fxConvertGLStencilOp(ctx->Stencil.ZFailFunc);
+    GrStencil_t zpass = fxConvertGLStencilOp(ctx->Stencil.ZPassFunc);
+    FX_grStencilOp(sfail, zfail, zpass);
+    FX_grStencilFunc(ctx->Stencil.Function - GL_NEVER,
+                     ctx->Stencil.Ref, ctx->Stencil.ValueMask);
+    FX_grStencilMask(ctx->Stencil.WriteMask);
+    FX_grEnable(GR_STENCIL_MODE_EXT);
   }
   else {
-     FX_grDepthBufferFunction(GR_CMP_ALWAYS);
-     FX_grDepthMask(FXFALSE);
+    FX_grDisable(GR_STENCIL_MODE_EXT);
   }
 }
 
+
+/*
+ * Set the state so that stencil is either enabled or disabled.
+ * This is called from Mesa only.  Glide is invoked at
+ * setup time, not now.
+ */
+static void
+fxDDEnableStencil(GLcontext *ctx, GLboolean state)
+{
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  (void) state;
+  fxMesa->new_state |= FX_NEW_STENCIL;
+  ctx->Driver.RenderStart = fxSetupFXUnits;
+}
+
+
 /************************************************************************/
 /**************************** Color Mask SetUp **************************/
 /************************************************************************/
@@ -1591,63 +1615,71 @@ static void fxSetupCull(GLcontext *ctx)
 
 void fxDDEnable(GLcontext *ctx, GLenum cap, GLboolean state)
 {
-  fxMesaContext fxMesa=(fxMesaContext)ctx->DriverCtx;
-  tfxUnitsState *us=&fxMesa->unitsState;
+  fxMesaContext fxMesa = FX_CONTEXT(ctx);
+  tfxUnitsState *us = &fxMesa->unitsState;
 
   if (MESA_VERBOSE&VERBOSE_DRIVER) {
-     fprintf(stderr,"fxmesa: fxDDEnable(...)\n");
+    fprintf(stderr,"fxmesa: fxDDEnable(...)\n");
   }
 
   switch(cap) {
-  case GL_ALPHA_TEST:
-    if(state!=us->alphaTestEnabled) {
-      us->alphaTestEnabled=state;
-      fxMesa->new_state |= FX_NEW_ALPHA;
+    case GL_ALPHA_TEST:
+      if(state!=us->alphaTestEnabled) {
+        us->alphaTestEnabled=state;
+        fxMesa->new_state |= FX_NEW_ALPHA;
+        ctx->Driver.RenderStart = fxSetupFXUnits;
+      }
+      break;
+    case GL_BLEND:
+      if(state!=us->blendEnabled) {
+        us->blendEnabled=state;
+        fxMesa->new_state |= FX_NEW_BLEND;
+        ctx->Driver.RenderStart = fxSetupFXUnits;
+      }
+      break;
+    case GL_DEPTH_TEST:
+      fxMesa->new_state |= FX_NEW_DEPTH;
       ctx->Driver.RenderStart = fxSetupFXUnits;
-    }
-    break;
-  case GL_BLEND:
-    if(state!=us->blendEnabled) {
-      us->blendEnabled=state;
-      fxMesa->new_state |= FX_NEW_BLEND;
+      break;
+    case GL_DITHER:
+      if (state)
+        FX_grDitherMode(GR_DITHER_4x4);
+      else
+        FX_grDitherMode(GR_DITHER_DISABLE);
+      break;
+    case GL_SCISSOR_TEST:
+      fxMesa->new_state |= FX_NEW_SCISSOR;
       ctx->Driver.RenderStart = fxSetupFXUnits;
-    }
-    break;
-  case GL_DEPTH_TEST:
-    if(state!=us->depthTestEnabled) {
-      us->depthTestEnabled=state;
-      fxMesa->new_state |= FX_NEW_DEPTH;
+      break;
+    case GL_SHARED_TEXTURE_PALETTE_EXT:
+      fxDDTexUseGlbPalette(ctx, state);
+      break;
+    case GL_FOG:
+      fxMesa->new_state |= FX_NEW_FOG;
       ctx->Driver.RenderStart = fxSetupFXUnits;
-    }
-    break;
-  case GL_SCISSOR_TEST:
-     fxMesa->new_state |= FX_NEW_SCISSOR;
-     ctx->Driver.RenderStart = fxSetupFXUnits;
-     break;
-  case GL_SHARED_TEXTURE_PALETTE_EXT:
-     fxDDTexUseGlbPalette(ctx, state);
-     break;
-  case GL_FOG:
-     fxMesa->new_state |= FX_NEW_FOG;
-     ctx->Driver.RenderStart = fxSetupFXUnits;
-     break;
-  case GL_CULL_FACE:
-     fxMesa->new_state |= FX_NEW_CULL;
-     ctx->Driver.RenderStart = fxSetupFXUnits;
-     break;
-  case GL_LINE_SMOOTH:
-  case GL_LINE_STIPPLE:
-  case GL_POINT_SMOOTH:
-  case GL_POLYGON_SMOOTH:
-  case GL_TEXTURE_2D:
+      break;
+    case GL_CULL_FACE:
+      fxMesa->new_state |= FX_NEW_CULL;
+      ctx->Driver.RenderStart = fxSetupFXUnits;
+      break;
+    case GL_LINE_SMOOTH:
+    case GL_LINE_STIPPLE:
+    case GL_POINT_SMOOTH:
+    case GL_POLYGON_SMOOTH:
+    case GL_TEXTURE_2D:
       fxMesa->new_state |= FX_NEW_TEXTURING;
       ctx->Driver.RenderStart = fxSetupFXUnits;
       break;
-  default:
-    ;  /* XXX no-op? */
+    case GL_STENCIL_TEST:
+      fxMesa->new_state |= FX_NEW_STENCIL;
+      ctx->Driver.RenderStart = fxSetupFXUnits;
+      break;
+    default:
+      ;  /* no-op */
   }    
 }
 
+
 #if 0
 /*
   Multipass to do GL_BLEND texture functions
@@ -1676,6 +1708,12 @@ static GLboolean fxMultipassBlend(struct vertex_buffer *VB, GLuint pass)
       }
       fxDDDepthMask(ctx, FALSE);
     }
+   /*
+    * Disable stencil as well.
+    */
+    if (ctx->Stencil.Enabled) {
+      fxDDEnableStencil(ctx, GL_FALSE);
+    } 
     /* Enable Cc*Ct mode */
     /* XXX Set the Constant Color ? */
     fxDDEnable(ctx, GL_BLEND, GL_TRUE);
@@ -1692,6 +1730,7 @@ static GLboolean fxMultipassBlend(struct vertex_buffer *VB, GLuint pass)
     fxSetupTextureSingleTMU(ctx, XXX);
     fxSetupBlend(ctx);
     fxSetupDepthTest(ctx);
+    fxSetupStencilText(ctx);
     break;
   }
 
@@ -1737,13 +1776,15 @@ static GLboolean fxMultipassTexture( struct vertex_buffer *VB, GLuint pass )
 	 case GL_ALWAYS:
 	    break;
 	 default:
-	    fxDDDepthFunc( ctx, GL_EQUAL );
+            /*fxDDDepthFunc( ctx, GL_EQUAL );*/
+            FX_grDepthBufferFunction(GR_CMP_EQUAL);
 	    break;
 	 }
 
-	 fxDDDepthMask( ctx, GL_FALSE ); 
+	 /*fxDDDepthMask( ctx, GL_FALSE ); */
+         FX_grDepthMask(FXFALSE);
       }
-      
+      fxDDEnableStencil(ctx, GL_FALSE);
       if (ctx->Texture.Unit[1].EnvMode == GL_MODULATE) {
 	 fxDDEnable( ctx, GL_BLEND, GL_TRUE );
 	 fxDDBlendFunc( ctx, GL_DST_COLOR, GL_ZERO );
@@ -1763,6 +1804,7 @@ static GLboolean fxMultipassTexture( struct vertex_buffer *VB, GLuint pass )
       fxSetupTextureSingleTMU( ctx, 0 ); 
       fxSetupBlend( ctx );
       fxSetupDepthTest( ctx );
+      fxSetupStencilTest( ctx );
       break;
    }
 
@@ -1820,10 +1862,13 @@ void fxSetupFXUnits( GLcontext *ctx )
 
      if (newstate & FX_NEW_ALPHA)
 	fxSetupAlphaTest(ctx);
-     
+
      if (newstate & FX_NEW_DEPTH)
 	fxSetupDepthTest(ctx);
 
+     if (newstate & FX_NEW_STENCIL)
+       fxSetupStencilTest(ctx);
+
      if (newstate & FX_NEW_FOG)
 	fxSetupFog(ctx);
 
@@ -1835,14 +1880,11 @@ void fxSetupFXUnits( GLcontext *ctx )
 
      if (newstate & FX_NEW_CULL)
 	fxSetupCull(ctx);     
-
      fxMesa->new_state = 0;
 /*       ctx->Driver.RenderStart = 0; */
   }
 }
 
-
-
 #else
 
 
diff --git a/xc/extras/Mesa/src/FX/fxtexman.c b/xc/extras/Mesa/src/FX/fxtexman.c
index 2ffaeadd7..5066efa55 100644
--- a/xc/extras/Mesa/src/FX/fxtexman.c
+++ b/xc/extras/Mesa/src/FX/fxtexman.c
@@ -545,12 +545,8 @@ void fxTMReloadSubMipMapLevel(fxMesaContext fxMesa,
   fxTexGetInfo(ti->mipmapLevel[0].width, ti->mipmapLevel[0].height,
 	       &lodlevel, NULL, NULL, NULL, NULL, NULL, NULL, NULL);
 
-  if((ti->info.format==GR_TEXFMT_INTENSITY_8) ||
-     (ti->info.format==GR_TEXFMT_P_8) ||
-     (ti->info.format==GR_TEXFMT_ALPHA_8))
-    data=ti->mipmapLevel[level].data+((yoffset*ti->mipmapLevel[level].width)>>1);
-  else
-    data=ti->mipmapLevel[level].data+yoffset*ti->mipmapLevel[level].width;
+  data=ti->mipmapLevel[level].data +
+    yoffset*ti->mipmapLevel[level].width*ti->mipmapLevel[level].texelSize;
 
   switch(tmu) {
   case FX_TMU0:
diff --git a/xc/extras/Mesa/src/FX/fxtritmp.h b/xc/extras/Mesa/src/FX/fxtritmp.h
index 9e59b4fc7..6411d167a 100644
--- a/xc/extras/Mesa/src/FX/fxtritmp.h
+++ b/xc/extras/Mesa/src/FX/fxtritmp.h
@@ -307,11 +307,11 @@ static void TAG(fx_line)(GLcontext *ctx, GLuint e1, GLuint e2, GLuint pv)
    GrVertex *v2 = (GrVertex *)gWin[e2].f;
    GLfloat w = ctx->Line.Width*.5;
 
-   if (IND & FX_FLAT) 
-   {
-      /* Flat shading seems broke. Explicitly set vertex colors to same color */
-      GOURAUD2(v1,color[pv]);
-      GOURAUD2(v2,color[pv]);
+   if (IND & FX_FLAT) {
+      v1->r = v2->r = UBYTE_COLOR_TO_FLOAT_255_COLOR(color[pv][0]);
+      v1->g = v2->g = UBYTE_COLOR_TO_FLOAT_255_COLOR(color[pv][1]);
+      v1->b = v2->b = UBYTE_COLOR_TO_FLOAT_255_COLOR(color[pv][2]);
+      v1->a = v2->a = UBYTE_COLOR_TO_FLOAT_255_COLOR(color[pv][3]);
    } 
    else if (IND & FX_TWOSIDE) 
    {
@@ -334,8 +334,7 @@ static void TAG(fx_line)(GLcontext *ctx, GLuint e1, GLuint e2, GLuint pv)
    else
       DRAW_LINE(v1,v2,w);
 
-   if (IND & FX_FRONT_BACK) 
-   {
+   if (IND & FX_FRONT_BACK) {
       FX_grColorMask(ctx->Color.ColorMask[RCOMP] ||
 		  ctx->Color.ColorMask[GCOMP] ||
 		  ctx->Color.ColorMask[BCOMP],
diff --git a/xc/lib/GL/mesa/src/drv/gamma/gamma_gl.c b/xc/lib/GL/mesa/src/drv/gamma/gamma_gl.c
index 2ae1001ee..47d1139a5 100644
--- a/xc/lib/GL/mesa/src/drv/gamma/gamma_gl.c
+++ b/xc/lib/GL/mesa/src/drv/gamma/gamma_gl.c
@@ -2483,6 +2483,8 @@ void _gamma_NewList(GLuint list, GLenum mode)
 
 void _gamma_Normal3b(GLbyte nx, GLbyte ny, GLbyte nz)
 {
+    GLfloat x,y,z;
+
     DEBUG_GLCMDS(("Normal3b: %d %d %d\n", nx, ny, nz));
 
     _gamma_Normal3f((GLfloat)nx,(GLfloat)ny,(GLfloat)nz);
@@ -2490,6 +2492,8 @@ void _gamma_Normal3b(GLbyte nx, GLbyte ny, GLbyte nz)
 
 void _gamma_Normal3bv(const GLbyte *v)
 {
+    GLfloat p[3];
+
     DEBUG_GLCMDS(("Normal3bv: %d %d %d\n", v[0], v[1], v[2]));
 
     _gamma_Normal3f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2]);
@@ -2528,6 +2532,8 @@ void _gamma_Normal3fv(const GLfloat *v)
 
 void _gamma_Normal3i(GLint nx, GLint ny, GLint nz)
 {
+    GLfloat x,y,z;
+
     DEBUG_GLCMDS(("Normal3i: %d %d %d\n", (int)nx, (int)ny, (int)nz));
 
     _gamma_Normal3f((GLfloat)nx,(GLfloat)ny,(GLfloat)nz);
@@ -2535,6 +2541,8 @@ void _gamma_Normal3i(GLint nx, GLint ny, GLint nz)
 
 void _gamma_Normal3iv(const GLint *v)
 {
+    GLfloat p[3];
+
     DEBUG_GLCMDS(("Normal3iv: %d %d %d\n", (int)v[0], (int)v[1], (int)v[2]));
 
     _gamma_Normal3f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2]);
@@ -2542,6 +2550,8 @@ void _gamma_Normal3iv(const GLint *v)
 
 void _gamma_Normal3s(GLshort nx, GLshort ny, GLshort nz)
 {
+    GLfloat x,y,z;
+
     DEBUG_GLCMDS(("Normal3s: %d %d %d\n", nx, ny, nz));
 
     _gamma_Normal3f((GLfloat)nx,(GLfloat)ny,(GLfloat)nz);
@@ -2549,6 +2559,8 @@ void _gamma_Normal3s(GLshort nx, GLshort ny, GLshort nz)
 
 void _gamma_Normal3sv(const GLshort *v)
 {
+    GLfloat p[3];
+
     DEBUG_GLCMDS(("Normal3sv: %d %d %d\n", v[0], v[1], v[2]));
 
     _gamma_Normal3f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2]);
@@ -4146,6 +4158,8 @@ void _gamma_Vertex2fv(const GLfloat *v)
 
 void _gamma_Vertex2i(GLint x, GLint y)
 {
+    GLfloat a,b;
+
     DEBUG_GLCMDS(("Vertex2i: %d %d\n", (int)x, (int)y));
 
     _gamma_Vertex2f((GLfloat)x,(GLfloat)y);
@@ -4153,6 +4167,8 @@ void _gamma_Vertex2i(GLint x, GLint y)
 
 void _gamma_Vertex2iv(const GLint *v)
 {
+    GLfloat p[2];
+
     DEBUG_GLCMDS(("Vertex2iv: %d %d\n", (int)v[0], (int)v[1]));
 
     _gamma_Vertex2f((GLfloat)v[0],(GLfloat)v[1]);
@@ -4160,6 +4176,8 @@ void _gamma_Vertex2iv(const GLint *v)
 
 void _gamma_Vertex2s(GLshort x, GLshort y)
 {
+    GLfloat a,b;
+
     DEBUG_GLCMDS(("Vertex2s: %d %d\n", x, y));
 
     _gamma_Vertex2f((GLfloat)x,(GLfloat)y);
@@ -4167,6 +4185,8 @@ void _gamma_Vertex2s(GLshort x, GLshort y)
 
 void _gamma_Vertex2sv(const GLshort *v)
 {
+    GLfloat p[2];
+
     DEBUG_GLCMDS(("Vertex2sv: %d %d\n", v[0], v[1]));
 
     _gamma_Vertex2f((GLfloat)v[0],(GLfloat)v[1]);
@@ -4205,6 +4225,8 @@ void _gamma_Vertex3fv(const GLfloat *v)
 
 void _gamma_Vertex3i(GLint x, GLint y, GLint z)
 {
+    GLfloat a,b,c;
+
     DEBUG_GLCMDS(("Vertex3i: %d %d %d\n", (int)x, (int)y, (int)z));
 
     _gamma_Vertex3f((GLfloat)x,(GLfloat)y,(GLfloat)z);
@@ -4212,6 +4234,8 @@ void _gamma_Vertex3i(GLint x, GLint y, GLint z)
 
 void _gamma_Vertex3iv(const GLint *v)
 {
+    GLfloat p[3];
+
     DEBUG_GLCMDS(("Vertex3iv: %d %d %d\n", (int)v[0], (int)v[1], (int)v[2]));
 
     _gamma_Vertex3f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2]);
@@ -4219,6 +4243,8 @@ void _gamma_Vertex3iv(const GLint *v)
 
 void _gamma_Vertex3s(GLshort x, GLshort y, GLshort z)
 {
+    GLfloat a,b,c;
+
     DEBUG_GLCMDS(("Vertex3s: %d %d %d\n", x, y, z));
 
     _gamma_Vertex3f((GLfloat)x,(GLfloat)y,(GLfloat)z);
@@ -4226,6 +4252,8 @@ void _gamma_Vertex3s(GLshort x, GLshort y, GLshort z)
 
 void _gamma_Vertex3sv(const GLshort *v)
 {
+    GLfloat p[3];
+
     DEBUG_GLCMDS(("Vertex3sv: %d %d %d\n", v[0], v[1], v[2]));
 
     _gamma_Vertex3f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2]);
@@ -4265,6 +4293,8 @@ void _gamma_Vertex4fv(const GLfloat *v)
 
 void _gamma_Vertex4i(GLint x, GLint y, GLint z, GLint w)
 {
+    GLfloat a,b,c,d;
+
     DEBUG_GLCMDS(("Vertex4i: %d %d %d %d\n", (int)x, (int)y, (int)z, (int)w));
 
     _gamma_Vertex4f((GLfloat)x,(GLfloat)y,(GLfloat)z,(GLfloat)w);
@@ -4272,6 +4302,8 @@ void _gamma_Vertex4i(GLint x, GLint y, GLint z, GLint w)
 
 void _gamma_Vertex4iv(const GLint *v)
 {
+    GLfloat p[4];
+
     DEBUG_GLCMDS(("Vertex4iv: %d %d %d %d\n",
 		  (int)v[0], (int)v[1], (int)v[2], (int)v[3]));
 
@@ -4280,6 +4312,8 @@ void _gamma_Vertex4iv(const GLint *v)
 
 void _gamma_Vertex4s(GLshort x, GLshort y, GLshort z, GLshort w)
 {
+    GLfloat a,b,c,d;
+
     DEBUG_GLCMDS(("Vertex4s: %d %d %d %d\n", x, y, z, w));
 
     _gamma_Vertex4f((GLfloat)x,(GLfloat)y,(GLfloat)z,(GLfloat)w);
@@ -4287,6 +4321,8 @@ void _gamma_Vertex4s(GLshort x, GLshort y, GLshort z, GLshort w)
 
 void _gamma_Vertex4sv(const GLshort *v)
 {
+    GLfloat p[4];
+
     DEBUG_GLCMDS(("Vertex4sv: %d %d %d %d\n", v[0], v[1], v[2], v[3]));
 
     _gamma_Vertex4f((GLfloat)v[0],(GLfloat)v[1],(GLfloat)v[2],(GLfloat)v[3]);
diff --git a/xc/lib/GL/mesa/src/drv/tdfx/tdfx_inithw.c b/xc/lib/GL/mesa/src/drv/tdfx/tdfx_inithw.c
index 7d74318e1..0bafa3190 100644
--- a/xc/lib/GL/mesa/src/drv/tdfx/tdfx_inithw.c
+++ b/xc/lib/GL/mesa/src/drv/tdfx/tdfx_inithw.c
@@ -68,18 +68,12 @@ GLboolean tdfxInitHW(__DRIdrawablePrivate *driDrawPriv,
     fxMesa->haveTwoTMUs=GL_TRUE;
 
   /* !!! We are forcing these !!! */
-  fxMesa->haveDoubleBuffer=GL_TRUE; 
   fxMesa->haveAlphaBuffer=GL_FALSE; 
-  fxMesa->haveZBuffer=GL_TRUE;
   fxMesa->haveGlobalPaletteTexture=GL_FALSE;
 
   fxMesa->glideContext =  FX_grSstWinOpen_NoLock((FxU32)-1, GR_RESOLUTION_NONE,
 						 GR_REFRESH_NONE,
-#if  FXMESA_USE_ARGB
-						 GR_COLORFORMAT_ARGB,
-#else
 						 GR_COLORFORMAT_ABGR,
-#endif
 						 GR_ORIGIN_LOWER_LEFT, 2, 1);
 
   grDRIResetSAREA();
diff --git a/xc/lib/GL/mesa/src/drv/tdfx/tdfx_xmesa.c b/xc/lib/GL/mesa/src/drv/tdfx/tdfx_xmesa.c
index 0a99539e8..b40f0f054 100644
--- a/xc/lib/GL/mesa/src/drv/tdfx/tdfx_xmesa.c
+++ b/xc/lib/GL/mesa/src/drv/tdfx/tdfx_xmesa.c
@@ -107,7 +107,6 @@ void XMesaResetDriver(__DRIscreenPrivate *sPriv)
   sPriv->private = NULL;
 }
 
-
 GLvisual *XMesaCreateVisual(Display *dpy,
                             __DRIscreenPrivate *driScrnPriv,
                             const XVisualInfo *visinfo,
@@ -150,6 +149,18 @@ GLboolean XMesaCreateContext(Display *dpy, GLvisual *mesaVis,
 
   cPriv->hHWContext = driContextPriv->hHWContext;
   cPriv->tdfxScrnPriv = sPriv;
+  /* deviceID = 0x05 = Voodoo3 */
+  /* deviceID = 0x09 = Voodoo5 (and Voodoo5?) */
+  cPriv->haveHwStencil = sPriv->deviceID == 0x9 && sPriv->cpp == 4;
+
+  cPriv->glVis=mesaVis;
+  cPriv->glBuffer=gl_create_framebuffer(mesaVis,
+                        GL_FALSE,  /* software depth buffer? */
+                        mesaVis->StencilBits > 0 && !cPriv->haveHwStencil,
+                        mesaVis->AccumRedBits > 0,
+                        GL_FALSE   /* software alpha channel */
+                      );
+
   cPriv->screen_width=sPriv->width;
   cPriv->screen_height=sPriv->height;
   cPriv->new_state = ~0;
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/glint/dualmx_accel.c b/xc/programs/Xserver/hw/xfree86/drivers/glint/dualmx_accel.c
index 974ee3b4f..c8ec93434 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/glint/dualmx_accel.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/glint/dualmx_accel.c
@@ -749,7 +749,7 @@ DualMXWriteBitmap(ScrnInfoPtr pScrn,
     DualMXLoadCoord(pScrn, x, y, x+w, h, 0, 1);
 
     if(bg == -1) {
-	/* >>>>> set fg <<<<<<<< */
+	/* === set fg === */
 	REPLICATE(fg);
 	GLINT_WAIT(3);
 	if (rop == GXcopy) {
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c b/xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c
index 1e7fad981..b728093d7 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c
@@ -23,14 +23,16 @@
  *           Dirk Hohndel, <hohndel@suse.de>
  *           Stefan Dirsch, <sndirsch@suse.de>
  *           Mark Vojkovich, <mvojkovi@ucsd.edu>
+ *           Michel Dänzer, <michdaen@iiic.ethz.ch>
  *
  * this work is sponsored by S.u.S.E. GmbH, Fuerth, Elsa GmbH, Aachen and
  * Siemens Nixdorf Informationssysteme
  * 
  * Permedia 2 accelerated options.
  */
-/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c,v 1.17 1999/07/11 08:49:26 dawes Exp $ */
+/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/glint/pm2_accel.c,v 1.23 2000/02/25 21:02:50 dawes Exp $ */
 
+#include "Xarch.h"
 #include "xf86.h"
 #include "xf86_OSproc.h"
 #include "xf86_ansic.h"
@@ -39,18 +41,26 @@
 #include "xf86Pci.h"
 
 #include "miline.h"
-#define PSZ 8
-#include "cfb.h"
-#undef PSZ
-#include "cfb16.h"
-#include "cfb24.h"
-#include "cfb32.h"
+
+#include "fb.h"
 
 #include "glint_regs.h"
 #include "glint.h"
 
 #include "xaalocal.h"		/* For replacements */
 
+#define DEBUG 0
+
+#if DEBUG
+# define TRACE_ENTER(str)       ErrorF("pm2_accel: " str " %d\n",pScrn->scrnIndex)
+# define TRACE_EXIT(str)        ErrorF("pm2_accel: " str " done\n")
+# define TRACE(str)             ErrorF("pm2_accel trace: " str "\n")
+#else
+# define TRACE_ENTER(str)
+# define TRACE_EXIT(str)
+# define TRACE(str)
+#endif
+
 static void Permedia2SetupForFillRectSolid(ScrnInfoPtr pScrn, int color,
 				int rop, unsigned int planemask);
 static void Permedia2SubsequentFillRectSolid(ScrnInfoPtr pScrn, int x,
@@ -79,13 +89,7 @@ static void Permedia2SetupForSolidLine(ScrnInfoPtr pScrn, int color,
 				int rop, unsigned int planemask);
 static void Permedia2SubsequentHorVertLine(ScrnInfoPtr pScrn, int x, int y, 
 				int len, int dir);
-static void Permedia2SubsequentSolidBresenhamLine8bpp(ScrnInfoPtr pScrn,
-        			int x, int y, int dmaj, int dmin, int e, 
-				int len, int octant);
-static void Permedia2SubsequentSolidBresenhamLine16bpp(ScrnInfoPtr pScrn,
-        			int x, int y, int dmaj, int dmin, int e, 
-				int len, int octant);
-static void Permedia2SubsequentSolidBresenhamLine32bpp(ScrnInfoPtr pScrn,
+static void Permedia2SubsequentSolidBresenhamLine(ScrnInfoPtr pScrn,
         			int x, int y, int dmaj, int dmin, int e, 
 				int len, int octant);
 static void Permedia2WriteBitmap(ScrnInfoPtr pScrn, int x, int y, int w, int h, 
@@ -139,6 +143,12 @@ static void Permedia2PolySegmentThinSolidWrapper(DrawablePtr pDraw, GCPtr pGC,
 
 #define MAX_FIFO_ENTRIES 256
 
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+# define STIPPLE_SWAP	1<<18		/* Mirror stipple pattern horizontally */
+#else
+# define STIPPLE_SWAP	0
+#endif
+
 void
 Permedia2InitializeEngine(ScrnInfoPtr pScrn)
 {
@@ -146,6 +156,8 @@ Permedia2InitializeEngine(ScrnInfoPtr pScrn)
 
     /* Initialize the Accelerator Engine to defaults */
 
+    TRACE_ENTER("Permedia2InitializeEngine");
+
     GLINT_SLOW_WRITE_REG(UNIT_DISABLE,	ScissorMode);
     GLINT_SLOW_WRITE_REG(UNIT_ENABLE,	FBWriteMode);
     GLINT_SLOW_WRITE_REG(0, 		dXSub);
@@ -183,14 +195,26 @@ Permedia2InitializeEngine(ScrnInfoPtr pScrn)
     GLINT_SLOW_WRITE_REG(UNIT_DISABLE,	FBSourceBase);
     GLINT_SLOW_WRITE_REG(UNIT_DISABLE,	LBWindowBase);
 
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+    pGlint->RasterizerSwap = 1;
+#else
+    pGlint->RasterizerSwap = 0;
+#endif
+
     switch (pScrn->bitsPerPixel) {
 	case 8:
 	    pGlint->PixelWidth = 0x0; /* 8 Bits */
 	    pGlint->TexMapFormat = pGlint->pprod;
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	    pGlint->RasterizerSwap |= 3<<15;	/* Swap host data */
+#endif
 	    break;
 	case 16:
 	    pGlint->PixelWidth = 0x1; /* 16 Bits */
 	    pGlint->TexMapFormat = pGlint->pprod | 1<<19;
+#if X_BYTE_ORDER == X_BIG_ENDIAN
+	    pGlint->RasterizerSwap |= 2<<15;	/* Swap host data */
+#endif
 	    break;
 	case 24:
  	    pGlint->PixelWidth = 0x4; /* 24 Bits */
@@ -223,6 +247,8 @@ Permedia2InitializeEngine(ScrnInfoPtr pScrn)
     GLINT_SLOW_WRITE_REG(0, StartXSub);
     GLINT_SLOW_WRITE_REG(0, StartY);
     GLINT_SLOW_WRITE_REG(0, GLINTCount);
+
+    TRACE_EXIT("Permedia2InitializeEngine");
 }
 
 Bool
@@ -262,17 +288,8 @@ Permedia2AccelInit(ScreenPtr pScreen)
         infoPtr->PolylinesThinSolidFlags = 0;
         infoPtr->SetupForSolidLine = Permedia2SetupForSolidLine;
         infoPtr->SubsequentSolidHorVertLine = Permedia2SubsequentHorVertLine;
-	switch(pScrn->bitsPerPixel) {
-	case 8:		infoPtr->SubsequentSolidBresenhamLine = 
-				Permedia2SubsequentSolidBresenhamLine8bpp;
-			break;
-	case 16:	infoPtr->SubsequentSolidBresenhamLine = 
-				Permedia2SubsequentSolidBresenhamLine16bpp;
-			break;
-	case 32:	infoPtr->SubsequentSolidBresenhamLine = 
-				Permedia2SubsequentSolidBresenhamLine32bpp;
-			break;
-	}
+	infoPtr->SubsequentSolidBresenhamLine = 
+				Permedia2SubsequentSolidBresenhamLine;
 	infoPtr->PolySegmentThinSolid = Permedia2PolySegmentThinSolidWrapper;
 	infoPtr->PolylinesThinSolid = Permedia2PolylinesThinSolidWrapper;
     	infoPtr->SetupForSolidFill = Permedia2SetupForFillRectSolid;
@@ -483,6 +500,7 @@ Permedia2SetupForScreenToScreenCopy(ScrnInfoPtr pScrn,
 				unsigned int planemask, int transparency_color)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
+    TRACE_ENTER("Permedia2SetupForScreenToScreenCopy");
 
     pGlint->BltScanDirection = 0;
     if (xdir == 1) pGlint->BltScanDirection |= XPositive;
@@ -502,6 +520,7 @@ Permedia2SetupForScreenToScreenCopy(ScrnInfoPtr pScrn,
 							FBRM_DstEnable;
     }
     LOADROP(rop);
+    TRACE_EXIT("Permedia2SetupForScreenToScreenCopy");
 }
 
 static void
@@ -511,6 +530,7 @@ Permedia2SubsequentScreenToScreenCopy(ScrnInfoPtr pScrn, int x1, int y1,
     GLINTPtr pGlint = GLINTPTR(pScrn);
     char align;
 
+    TRACE_ENTER("Permedia2SubsequentScreenToScreenCopy");
     /* We can only use GXcopy for Packed modes */
     if (pGlint->ROP != GXcopy) {
 	GLINT_WAIT(5);
@@ -528,6 +548,7 @@ Permedia2SubsequentScreenToScreenCopy(ScrnInfoPtr pScrn, int x1, int y1,
     }
 
     GLINT_WRITE_REG(PrimitiveRectangle | pGlint->BltScanDirection, Render);
+    TRACE_EXIT("Permedia2SubsequentScreenToScreenCopy");
 }
 
 
@@ -543,6 +564,7 @@ Permedia2PolylinesThinSolidWrapper(
     XAAInfoRecPtr infoRec = GET_XAAINFORECPTR_FROM_GC(pGC);
     GLINTPtr pGlint = GLINTPTR(infoRec->pScrn);
     pGlint->CurrentGC = pGC;
+    pGlint->CurrentDrawable = pDraw;
     if(infoRec->NeedToSync) (*infoRec->Sync)(infoRec->pScrn);
     XAAPolyLines(pDraw, pGC, mode, npt, pPts);
 }
@@ -557,6 +579,7 @@ Permedia2PolySegmentThinSolidWrapper(
     XAAInfoRecPtr infoRec = GET_XAAINFORECPTR_FROM_GC(pGC);
     GLINTPtr pGlint = GLINTPTR(infoRec->pScrn);
     pGlint->CurrentGC = pGC;
+    pGlint->CurrentDrawable = pDraw;
     if(infoRec->NeedToSync) (*infoRec->Sync)(infoRec->pScrn);
     XAAPolySegment(pDraw, pGC, nseg, pSeg);
 }
@@ -600,11 +623,10 @@ Permedia2SubsequentHorVertLine(ScrnInfoPtr pScrn,int x,int y,int len,int dir)
 }
 
 static void 
-Permedia2SubsequentSolidBresenhamLine8bpp( ScrnInfoPtr pScrn,
+Permedia2SubsequentSolidBresenhamLine( ScrnInfoPtr pScrn,
         int x, int y, int dmaj, int dmin, int e, int len, int octant)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
 
     if(dmaj == dmin) {
 	GLINT_WAIT(6);
@@ -627,93 +649,13 @@ Permedia2SubsequentSolidBresenhamLine8bpp( ScrnInfoPtr pScrn,
 	return;
     }
 
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfbBresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth >> 2, 
+    fbBres(pGlint->CurrentDrawable, pGlint->CurrentGC, 0,
                 (octant & XDECREASING) ? -1 : 1, 
                 (octant & YDECREASING) ? -1 : 1, 
                 (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
+                x, y, dmin + e, dmin, -dmaj, len);
 }
 
-static void 
-Permedia2SubsequentSolidBresenhamLine16bpp( ScrnInfoPtr pScrn,
-        int x, int y, int dmaj, int dmin, int e, int len, int octant)
-{
-    GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
-
-    if(dmaj == dmin) {
-	GLINT_WAIT(6);
-	if(octant & YDECREASING) {
-	    GLINT_WRITE_REG(-1<<16, dY);
-	} else {
-	    GLINT_WRITE_REG(1<<16, dY);
-	}
-
-	if(octant & XDECREASING) {
-	    GLINT_WRITE_REG(-1<<16, dXDom);
-	} else {
-	    GLINT_WRITE_REG(1<<16, dXDom);
-	}
-
-	GLINT_WRITE_REG(x<<16, StartXDom);
-	GLINT_WRITE_REG(y<<16, StartY);
-	GLINT_WRITE_REG(len,GLINTCount);
-	GLINT_WRITE_REG(PrimitiveLine, Render);
-	return;
-    }
-
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfb16BresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth >> 1, 
-                (octant & XDECREASING) ? -1 : 1, 
-                (octant & YDECREASING) ? -1 : 1, 
-                (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
-}
-
-static void 
-Permedia2SubsequentSolidBresenhamLine32bpp( ScrnInfoPtr pScrn,
-        int x, int y, int dmaj, int dmin, int e, int len, int octant)
-{
-    GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
-
-    if(dmaj == dmin) {
-	GLINT_WAIT(6);
-	if(octant & YDECREASING) {
-	    GLINT_WRITE_REG(-1<<16, dY);
-	} else {
-	    GLINT_WRITE_REG(1<<16, dY);
-	}
-
-	if(octant & XDECREASING) {
-	    GLINT_WRITE_REG(-1<<16, dXDom);
-	} else {
-	    GLINT_WRITE_REG(1<<16, dXDom);
-	}
-
-	GLINT_WRITE_REG(x<<16, StartXDom);
-	GLINT_WRITE_REG(y<<16, StartY);
-	GLINT_WRITE_REG(len,GLINTCount);
-	GLINT_WRITE_REG(PrimitiveLine, Render);
-	return;
-    }
-
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfb32BresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth, 
-                (octant & XDECREASING) ? -1 : 1, 
-                (octant & YDECREASING) ? -1 : 1, 
-                (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
-}
-
-
 static void
 Permedia2SetupForFillRectSolid24bpp(ScrnInfoPtr pScrn, int color,
 					 int rop, unsigned int planemask)
@@ -737,6 +679,7 @@ Permedia2SetupForFillRectSolid(ScrnInfoPtr pScrn, int color,
 				    int rop, unsigned int planemask)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
+    TRACE_ENTER("Permedia2SetupForFillRectSolid");
 
     REPLICATE(color);
 
@@ -753,6 +696,7 @@ Permedia2SetupForFillRectSolid(ScrnInfoPtr pScrn, int color,
 	GLINT_WRITE_REG(pGlint->pprod|FBRM_DstEnable|FBRM_Packed, FBReadMode);
     }
     LOADROP(rop);
+    TRACE_EXIT("Permedia2SetupForFillRectSolid");
 }
 
 static void
@@ -769,6 +713,7 @@ Permedia2SubsequentFillRectSolid(ScrnInfoPtr pScrn, int x, int y, int w, int h)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
     int speed = 0;
+    TRACE_ENTER("Permedia2SubsequentFillRectSolid");
 
     if (pGlint->ROP == GXcopy) {
 	GLINT_WAIT(3);
@@ -782,6 +727,7 @@ Permedia2SubsequentFillRectSolid(ScrnInfoPtr pScrn, int x, int y, int w, int h)
   	speed = 0;
     }
     GLINT_WRITE_REG(PrimitiveRectangle | XPositive | YPositive | speed, Render);
+    TRACE_EXIT("Permedia2SubsequentFillRectSolid");
 }
 
 static void MoveBYTE(
@@ -840,6 +786,7 @@ static void MoveDWORDS(
      *(dest + 2) = *(src + 2);
 }
 
+
 static void 
 Permedia2SetupForMono8x8PatternFill24bpp(ScrnInfoPtr pScrn, 
 					   int patternx, int patterny, 
@@ -882,6 +829,7 @@ Permedia2SetupForMono8x8PatternFill(ScrnInfoPtr pScrn,
 					   unsigned int planemask)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
+    TRACE_ENTER("Permedia2SetupForMono8x8PatternFill");
 
     if (bg == -1) pGlint->FrameBufferReadMode = -1;
 	else    pGlint->FrameBufferReadMode = 0;
@@ -891,6 +839,10 @@ Permedia2SetupForMono8x8PatternFill(ScrnInfoPtr pScrn,
     REPLICATE(pGlint->ForeGroundColor);
     REPLICATE(pGlint->BackGroundColor);
   
+#if DEBUG
+    ErrorF("patternx: %x patterny: %x\n", patternx, patterny);
+#endif
+
     GLINT_WAIT(13);
     DO_PLANEMASK(planemask);
     GLINT_WRITE_REG((patternx & 0xFF), AreaStipplePattern0);
@@ -910,6 +862,7 @@ Permedia2SetupForMono8x8PatternFill(ScrnInfoPtr pScrn,
 	GLINT_WRITE_REG(pGlint->pprod | FBRM_DstEnable, FBReadMode);
     }
     LOADROP(rop);
+    TRACE_EXIT("Permedia2SetupForMono8x8PatternFill");
 }
 
 static void 
@@ -926,7 +879,7 @@ Permedia2SubsequentMono8x8PatternFillRect24bpp(ScrnInfoPtr pScrn,
     if (pGlint->FrameBufferReadMode != -1) {
 	GLINT_WRITE_REG(pGlint->BackGroundColor, ConstantColor);
 	GLINT_WRITE_REG(patternx<<7|patterny<<12| ASM_InvertPattern |
-						UNIT_ENABLE, AreaStippleMode);
+				STIPPLE_SWAP | UNIT_ENABLE, AreaStippleMode);
 	GLINT_WRITE_REG(AreaStippleEnable | XPositive | 
 					YPositive | PrimitiveRectangle, Render);
     }
@@ -945,6 +898,8 @@ Permedia2SubsequentMono8x8PatternFillRect(ScrnInfoPtr pScrn,
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
   
+    TRACE_ENTER("Permedia2SubsequentMono8x8PatternFillRect()");
+	
     GLINT_WAIT(9);
     Permedia2LoadCoord(pScrn, x, y, w, h);
 
@@ -952,14 +907,14 @@ Permedia2SubsequentMono8x8PatternFillRect(ScrnInfoPtr pScrn,
 	if (pGlint->ROP == GXcopy) {
       	    GLINT_WRITE_REG(pGlint->BackGroundColor, FBBlockColor);
     	    GLINT_WRITE_REG(ASM_InvertPattern|patternx<<7|patterny<<12|
-			UNIT_ENABLE, AreaStippleMode);
+			STIPPLE_SWAP | UNIT_ENABLE, AreaStippleMode);
     	    GLINT_WRITE_REG(AreaStippleEnable | FastFillEnable | 
 			XPositive | YPositive | PrimitiveRectangle, Render);
 	} else { 
   	    GLINT_WRITE_REG(pGlint->ForeGroundColor, ConstantColor);
       	    GLINT_WRITE_REG(pGlint->BackGroundColor, Texel0);
 	    GLINT_WRITE_REG(patternx<<7|patterny<<12| 
-			UNIT_ENABLE, AreaStippleMode);
+			STIPPLE_SWAP | UNIT_ENABLE, AreaStippleMode);
   	    GLINT_WRITE_REG(AreaStippleEnable | XPositive | TextureEnable |
 				YPositive | PrimitiveRectangle, Render);
 	    return;
@@ -973,9 +928,11 @@ Permedia2SubsequentMono8x8PatternFillRect(ScrnInfoPtr pScrn,
   	GLINT_WRITE_REG(pGlint->ForeGroundColor, ConstantColor);
 	pGlint->FrameBufferReadMode = 0;
     }
-    GLINT_WRITE_REG(patternx<<7|patterny<<12|UNIT_ENABLE, AreaStippleMode);
+    GLINT_WRITE_REG(patternx<<7|patterny<<12|STIPPLE_SWAP|UNIT_ENABLE, AreaStippleMode);
     GLINT_WRITE_REG(AreaStippleEnable | pGlint->FrameBufferReadMode | 
 			XPositive | YPositive | PrimitiveRectangle, Render);
+
+    TRACE_EXIT("Permedia2SubsequentMono8x8PatternFillRect()");
 }
 
 static void
@@ -988,6 +945,7 @@ Permedia2SetupForCPUToScreenColorExpandFill(
     GLINTPtr pGlint = GLINTPTR(pScrn);
     int dobackground = 0;
 
+    TRACE_ENTER("Permedia2SetupForCPUToScreenColorExpandFill");
     if (bg != -1) dobackground |= ForceBackgroundColor;
 
     GLINT_WAIT(7);
@@ -1010,11 +968,12 @@ Permedia2SetupForCPUToScreenColorExpandFill(
 	} else {
             GLINT_WRITE_REG(fg, FBBlockColor);
 	}
-	GLINT_WRITE_REG(0, RasterizerMode);
+	GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
         pGlint->FrameBufferReadMode = FastFillEnable;
     } else {
 	GLINT_WRITE_REG(UNIT_ENABLE, ColorDDAMode);
-	GLINT_WRITE_REG(BitMaskPackingEachScanline|dobackground,RasterizerMode);
+	GLINT_WRITE_REG(BitMaskPackingEachScanline|dobackground|
+				pGlint->RasterizerSwap,RasterizerMode);
         GLINT_WRITE_REG(fg, ConstantColor);
 	if (dobackground) {
             pGlint->FrameBufferReadMode = TextureEnable;
@@ -1023,6 +982,7 @@ Permedia2SetupForCPUToScreenColorExpandFill(
             pGlint->FrameBufferReadMode = 0;
     }
     LOADROP(rop);
+    TRACE_EXIT("Permedia2SetupForCPUToScreenColorExpandFill");
 }
 
 static void
@@ -1034,6 +994,7 @@ Permedia2SubsequentCPUToScreenColorExpandFill(
     GLINTPtr pGlint = GLINTPTR(pScrn);
     int dwords = ((w + 31) >> 5) * h;
 
+    TRACE_ENTER("Permedia2SubsequentCPUToScreenColorExpandFill");
     GLINT_WAIT(8);
     Permedia2LoadCoord(pScrn, x, y, w, h);
 
@@ -1047,6 +1008,7 @@ Permedia2SubsequentCPUToScreenColorExpandFill(
 					pGlint->FrameBufferReadMode, Render);
     GLINT_WRITE_REG((dwords - 1)<<16 | 0x0D, OutputFIFO);
     GLINT_WAIT(dwords);
+    TRACE_EXIT("Permedia2SubsequentCPUToScreenColorExpandFill");
 }
 
 static void
@@ -1059,6 +1021,7 @@ Permedia2SetupForScanlineCPUToScreenColorExpandFill(
     GLINTPtr pGlint = GLINTPTR(pScrn);
     int dobackground = 0;
 
+    TRACE_ENTER("Permedia2SetupForScanlineCPUToScreenColorExpandFill");
     if (bg != -1) dobackground |= ForceBackgroundColor;
     pGlint->BackGroundColor = bg;
     pGlint->ForeGroundColor = fg;
@@ -1075,11 +1038,12 @@ Permedia2SetupForScanlineCPUToScreenColorExpandFill(
 	pGlint->FrameBufferReadMode = FastFillEnable;
   	GLINT_WRITE_REG(UNIT_DISABLE, ColorDDAMode);
         GLINT_WRITE_REG(fg, FBBlockColor);
-	GLINT_WRITE_REG(0,RasterizerMode);
+	GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     } else {
         GLINT_WRITE_REG(UNIT_ENABLE, ColorDDAMode);
         GLINT_WRITE_REG(fg, ConstantColor);
-	GLINT_WRITE_REG(BitMaskPackingEachScanline|dobackground,RasterizerMode);
+	GLINT_WRITE_REG(BitMaskPackingEachScanline|dobackground|
+				pGlint->RasterizerSwap,RasterizerMode);
 	if (dobackground) {
 	    GLINT_WRITE_REG(bg, Texel0);
 	    pGlint->FrameBufferReadMode = TextureEnable;
@@ -1088,6 +1052,7 @@ Permedia2SetupForScanlineCPUToScreenColorExpandFill(
 	}
     }
     LOADROP(rop);
+    TRACE_EXIT("Permedia2SetupForScanlineCPUToScreenColorExpandFill");
 }
 
 static void
@@ -1098,6 +1063,8 @@ Permedia2SubsequentScanlineCPUToScreenColorExpandFill(
 ){
     GLINTPtr pGlint = GLINTPTR(pScrn);
 
+    TRACE_ENTER("Permedia2SubsequentScanlineCPUToScreenColorExpandFill");
+	
     pGlint->dwords = ((w + 31) >> 5); /* dwords per scanline */
  
     pGlint->cpucount = h;
@@ -1108,6 +1075,7 @@ Permedia2SubsequentScanlineCPUToScreenColorExpandFill(
     GLINT_WRITE_REG(((pGlint->dwords*h)-1)<<16 | 0x0D, OutputFIFO);
     GLINT_WAIT(pGlint->dwords);
     pGlint->cpucount--;
+    TRACE_EXIT("Permedia2SubsequentScanlineCPUToScreenColorExpandFill");
 }
 
 static void
@@ -1137,6 +1105,8 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
     int dwords, height, mode;
     Bool SecondPass = FALSE;
 
+    TRACE_ENTER("Permedia2WriteBitmap");
+
     w += skipleft;
     x -= skipleft;
     dwords = (w + 31) >> 5;
@@ -1159,15 +1129,16 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
 	mode = FastFillEnable;
 	GLINT_WRITE_REG(UNIT_DISABLE, ColorDDAMode);
-    	GLINT_WRITE_REG(0, RasterizerMode);
+    	GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     } else {
 	mode = 0;
 	GLINT_WRITE_REG(UNIT_ENABLE, ColorDDAMode);
-	GLINT_WRITE_REG(BitMaskPackingEachScanline,RasterizerMode);
+	GLINT_WRITE_REG(BitMaskPackingEachScanline|
+		pGlint->RasterizerSwap,RasterizerMode);
     }
 
     if(bg == -1) {
-	/* >>>>> set fg <<<<<<<< */
+	/* === set fg === */
 	REPLICATE(fg);
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
 	    GLINT_WRITE_REG(fg, FBBlockColor);
@@ -1175,8 +1146,8 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
 	    GLINT_WRITE_REG(fg, ConstantColor);
 	}
     } else if(rop == GXcopy) {
-	/* >>>>> set bg <<<<<<< */
- 	/* >>>>> draw rect (x,y,w,h) */
+	/* === set bg === */
+ 	/* === draw rect (x,y,w,h) === */
 	REPLICATE(bg);
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
 	    GLINT_WRITE_REG(bg, FBBlockColor);
@@ -1184,7 +1155,7 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
 	    GLINT_WRITE_REG(bg, ConstantColor);
 	}
 	GLINT_WRITE_REG(PrimitiveRectangle | XPositive | YPositive |mode,Render);
-	/* >>>>>> set fg <<<<<< */
+	/* === set fg === */
 	REPLICATE(fg);
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
 	    GLINT_WRITE_REG(fg, FBBlockColor);
@@ -1193,7 +1164,7 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
 	}
     } else {
 	SecondPass = TRUE;
-	/* >>>>> set fg <<<<<<< */
+	/* === set fg === */
 	REPLICATE(fg);
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
 	    GLINT_WRITE_REG(fg, FBBlockColor);
@@ -1202,8 +1173,8 @@ Permedia2WriteBitmap(ScrnInfoPtr pScrn,
 	}
     }
 
-   /* >>>>>>>>> initiate transfer (x,y,w,h).  Skipleft pixels on the
-	left edge will be clipped <<<<<< */
+   /* === initiate transfer (x,y,w,h).  Skipleft pixels on the
+	left edge will be clipped === */
 
 SECOND_PASS:
     GLINT_WRITE_REG(PrimitiveRectangle | XPositive | YPositive | mode | SyncOnBitMask, Render);
@@ -1221,24 +1192,25 @@ SECOND_PASS:
 
     if(SecondPass) {
 	SecondPass = FALSE;
-	/* >>>>>> invert bitmask and set bg <<<<<<<< */
+	/* === invert bitmask and set bg === */
 	REPLICATE(bg);
 	GLINT_WAIT(3);
     if ((pScrn->bitsPerPixel != 24) && (rop == GXcopy)) {
-    	    GLINT_WRITE_REG(InvertBitMask, RasterizerMode);
+   	    GLINT_WRITE_REG(InvertBitMask|pGlint->RasterizerSwap,RasterizerMode);
 	    GLINT_WRITE_REG(bg, FBBlockColor);
 	} else {
-    	    GLINT_WRITE_REG(InvertBitMask|BitMaskPackingEachScanline, 
-						RasterizerMode);
+    	    GLINT_WRITE_REG(InvertBitMask|BitMaskPackingEachScanline|
+    	    			pGlint->RasterizerSwap, RasterizerMode);
 	    GLINT_WRITE_REG(bg, ConstantColor);
 	}
 	goto SECOND_PASS;
     }
 
     GLINT_WAIT(1);
-    GLINT_WRITE_REG(0, RasterizerMode);
+    GLINT_WRITE_REG(pGlint->RasterizerSwap, RasterizerMode);
     Permedia2DisableClipping(pScrn);
     SET_SYNC_FLAG(infoRec);
+    TRACE_EXIT("Permedia2WriteBitmap");
 }
 
 
@@ -1263,7 +1235,7 @@ Permedia2WritePixmap8bpp(
 
     GLINT_WAIT(3);
     DO_PLANEMASK(planemask);
-    GLINT_WRITE_REG(0, RasterizerMode);
+    GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     if (rop == GXcopy) {      
 	GLINT_WRITE_REG(pGlint->pprod | FBRM_Packed, FBReadMode);
     } else {
@@ -1271,7 +1243,9 @@ Permedia2WritePixmap8bpp(
     }
 
     dwords = (w + 3) >> 2;
-    if((!(x&3)) && (!(w&3))) FastTexLoad = TRUE;	
+#if X_BYTE_ORDER == X_LITTLE_ENDIAN
+    if((!(x&3)) && (!(w&3))) FastTexLoad = TRUE;
+#endif	
     if((rop != GXcopy) || (planemask != ~0))
 	FastTexLoad = FALSE;
 
@@ -1417,9 +1391,10 @@ Permedia2WritePixmap16bpp(
     unsigned short* srcpword;
     Bool FastTexLoad;
 
+    TRACE_ENTER("Permedia2WritePixmap16bpp");
     GLINT_WAIT(3);
     DO_PLANEMASK(planemask);
-    GLINT_WRITE_REG(0, RasterizerMode);
+    GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     if (rop == GXcopy) {      
 	GLINT_WRITE_REG(pGlint->pprod | FBRM_Packed, FBReadMode);
     } else {
@@ -1428,7 +1403,9 @@ Permedia2WritePixmap16bpp(
 
 	FastTexLoad = FALSE;
 	dwords = (w + 1) >> 1;
+#if X_BYTE_ORDER == X_LITTLE_ENDIAN
 	if((!(x&1)) && (!(w&1))) FastTexLoad = TRUE;
+#endif
 	if((rop != GXcopy) || (planemask != ~0))
 		FastTexLoad = FALSE;
 
@@ -1552,6 +1529,7 @@ Permedia2WritePixmap16bpp(
 
     Permedia2DisableClipping(pScrn);
     SET_SYNC_FLAG(infoRec);
+    TRACE_EXIT("Permedia2WritePixmap16bpp");
 }
 
 static void
@@ -1573,7 +1551,7 @@ Permedia2WritePixmap24bpp(
     CARD32* srcp;
 
     GLINT_WAIT(3);
-    GLINT_WRITE_REG(0, RasterizerMode);
+    GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     GLINT_WRITE_REG(UNIT_DISABLE, ColorDDAMode);
     if (rop == GXcopy) {      
 	GLINT_WRITE_REG(pGlint->pprod, FBReadMode);
@@ -1675,7 +1653,7 @@ Permedia2WritePixmap32bpp(
 
     GLINT_WAIT(3);
     DO_PLANEMASK(planemask);
-    GLINT_WRITE_REG(0, RasterizerMode);
+    GLINT_WRITE_REG(pGlint->RasterizerSwap,RasterizerMode);
     if (rop == GXcopy) {      
 	GLINT_WRITE_REG(pGlint->pprod, FBReadMode);
     } else {
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c b/xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c
index 27068af1b..0604fa045 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c
@@ -28,7 +28,7 @@
  * 
  * GLINT 500TX / MX accelerated options.
  */
-/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c,v 1.15 1999/07/04 06:39:02 dawes Exp $ */
+/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/glint/tx_accel.c,v 1.19 2000/01/21 01:12:18 dawes Exp $ */
 
 #include "xf86.h"
 #include "xf86_OSproc.h"
@@ -36,11 +36,8 @@
 
 #include "xf86PciInfo.h"
 #include "xf86Pci.h"
-#define PSZ 8
-#include "cfb.h"
-#undef PSZ
-#include "cfb16.h"
-#include "cfb32.h"
+
+#include "fb.h"
 
 #include "miline.h"
 
@@ -92,13 +89,7 @@ static void TXSetupForSolidLine(ScrnInfoPtr pScrn, int color, int rop,
 				unsigned int planemask);
 static void TXSubsequentHorVertLine(ScrnInfoPtr pScrn, int x1, int y1,
 				int len, int dir);
-static void TXSubsequentSolidBresenhamLine8bpp(ScrnInfoPtr pScrn,
-        			int x, int y, int dmaj, int dmin, int e, 
-				int len, int octant);
-static void TXSubsequentSolidBresenhamLine16bpp(ScrnInfoPtr pScrn,
-        			int x, int y, int dmaj, int dmin, int e, 
-				int len, int octant);
-static void TXSubsequentSolidBresenhamLine32bpp(ScrnInfoPtr pScrn,
+static void TXSubsequentSolidBresenhamLine(ScrnInfoPtr pScrn,
         			int x, int y, int dmaj, int dmin, int e, 
 				int len, int octant);
 static void TXPolylinesThinSolidWrapper(DrawablePtr pDraw, GCPtr pGC,
@@ -108,7 +99,7 @@ static void TXPolySegmentThinSolidWrapper(DrawablePtr pDraw, GCPtr pGC,
 
 #define MAX_FIFO_ENTRIES		15
 
-static void
+void
 TXInitializeEngine(ScrnInfoPtr pScrn)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
@@ -218,17 +209,8 @@ TXAccelInit(ScreenPtr pScreen)
     infoPtr->PolylinesThinSolidFlags = 0;
     infoPtr->SetupForSolidLine = TXSetupForSolidLine;
     infoPtr->SubsequentSolidHorVertLine = TXSubsequentHorVertLine;
-    switch(pScrn->bitsPerPixel) {
-	case 8:		infoPtr->SubsequentSolidBresenhamLine = 
-				TXSubsequentSolidBresenhamLine8bpp;
-			break;
-	case 16:	infoPtr->SubsequentSolidBresenhamLine = 
-				TXSubsequentSolidBresenhamLine16bpp;
-			break;
-	case 32:	infoPtr->SubsequentSolidBresenhamLine = 
-				TXSubsequentSolidBresenhamLine32bpp;
-			break;
-    }
+    infoPtr->SubsequentSolidBresenhamLine = 
+					TXSubsequentSolidBresenhamLine;
     infoPtr->PolySegmentThinSolid = TXPolySegmentThinSolidWrapper;
     infoPtr->PolylinesThinSolid = TXPolylinesThinSolidWrapper;
 
@@ -710,7 +692,7 @@ TXWriteBitmap(ScrnInfoPtr pScrn,
     TXLoadCoord(pScrn, x, y, x+w, h, 0, 1);
 
     if(bg == -1) {
-	/* >>>>> set fg <<<<<<<< */
+	/* === set fg === */
 	REPLICATE(fg);
 	GLINT_WAIT(3);
 	if (rop == GXcopy) {
@@ -878,6 +860,7 @@ TXPolylinesThinSolidWrapper(
     XAAInfoRecPtr infoRec = GET_XAAINFORECPTR_FROM_GC(pGC);
     GLINTPtr pGlint = GLINTPTR(infoRec->pScrn);
     pGlint->CurrentGC = pGC;
+    pGlint->CurrentDrawable = pDraw;
     if(infoRec->NeedToSync) (*infoRec->Sync)(infoRec->pScrn);
     XAAPolyLines(pDraw, pGC, mode, npt, pPts);
 }
@@ -892,6 +875,7 @@ TXPolySegmentThinSolidWrapper(
     XAAInfoRecPtr infoRec = GET_XAAINFORECPTR_FROM_GC(pGC);
     GLINTPtr pGlint = GLINTPTR(infoRec->pScrn);
     pGlint->CurrentGC = pGC;
+    pGlint->CurrentDrawable = pDraw;
     if(infoRec->NeedToSync) (*infoRec->Sync)(infoRec->pScrn);
     XAAPolySegment(pDraw, pGC, nseg, pSeg);
 }
@@ -929,11 +913,10 @@ TXSubsequentHorVertLine(ScrnInfoPtr pScrn,int x,int y,int len,int dir)
 }
 
 static void 
-TXSubsequentSolidBresenhamLine8bpp( ScrnInfoPtr pScrn,
+TXSubsequentSolidBresenhamLine( ScrnInfoPtr pScrn,
         int x, int y, int dmaj, int dmin, int e, int len, int octant)
 {
     GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
     int dxdom, dy;
 
     if(dmaj == dmin) {
@@ -955,87 +938,9 @@ TXSubsequentSolidBresenhamLine8bpp( ScrnInfoPtr pScrn,
 	return;
     }
 
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfbBresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth >> 2, 
+    fbBres(pGlint->CurrentDrawable, pGlint->CurrentGC, 0,
                 (octant & XDECREASING) ? -1 : 1, 
                 (octant & YDECREASING) ? -1 : 1, 
                 (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
+                x, y, dmin + e, dmin, -dmaj, len);
 }
-
-static void 
-TXSubsequentSolidBresenhamLine16bpp( ScrnInfoPtr pScrn,
-        int x, int y, int dmaj, int dmin, int e, int len, int octant)
-{
-    GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
-    int dxdom, dy;
-
-    if(dmaj == dmin) {
-	GLINT_WAIT(7);
-	if(octant & YDECREASING) {
-	    dy = -1;
-	} else {
-	    dy = 1;
-	}
-
-	if(octant & XDECREASING) {
-	    dxdom = -1;
-	} else {
-	    dxdom = 1;
-	}
-
-        TXLoadCoord(pScrn, x, y, 0, len, dxdom, dy);
-	GLINT_WRITE_REG(PrimitiveLine, Render);
-	return;
-    }
-
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfb16BresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth >> 1, 
-                (octant & XDECREASING) ? -1 : 1, 
-                (octant & YDECREASING) ? -1 : 1, 
-                (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
-}
-
-static void 
-TXSubsequentSolidBresenhamLine32bpp( ScrnInfoPtr pScrn,
-        int x, int y, int dmaj, int dmin, int e, int len, int octant)
-{
-    GLINTPtr pGlint = GLINTPTR(pScrn);
-    cfbPrivGCPtr devPriv;
-    int dxdom, dy;
-
-    if(dmaj == dmin) {
-	GLINT_WAIT(7);
-	if(octant & YDECREASING) {
-	    dy = -1;
-	} else {
-	    dy = 1;
-	}
-
-	if(octant & XDECREASING) {
-	    dxdom = -1;
-	} else {
-	    dxdom = 1;
-	}
-
-        TXLoadCoord(pScrn, x, y, 0, len, dxdom, dy);
-	GLINT_WRITE_REG(PrimitiveLine, Render);
-	return;
-    }
-
-    devPriv = cfbGetGCPrivate(pGlint->CurrentGC);
-
-    cfb32BresS(devPriv->rop, devPriv->and, devPriv->xor, 
-                (unsigned long*)pGlint->FbBase, pScrn->displayWidth, 
-                (octant & XDECREASING) ? -1 : 1, 
-                (octant & YDECREASING) ? -1 : 1, 
-                (octant & YMAJOR) ? Y_AXIS : X_AXIS,
-                x, y, dmin + e, dmin, dmin - dmaj, len);
-}
-
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/Imakefile b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/Imakefile
index 4fe4392cb..e0bb04760 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/Imakefile
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/Imakefile
@@ -16,18 +16,11 @@ DRIINCLUDES = -I$(SERVERSRC)/GL/dri -I$(LIBSRC)/GL/dri
 DRIDEFINES = $(GLX_DEFINES)
 #endif
 
-#
-# Uncomment these only if you have the proprietary sources
-#
-PRIVSRCS = tdfx_priv.c
-PRIVOBJS = tdfx_priv.o
-PRIVDEFINES = -DPROP_3DFX
-
 SRCS = tdfx_driver.c tdfx_io.c tdfx_hwcurs.c tdfx_accel.c \
-	tdfx_dga.c $(DRISRCS) $(PRIVSRCS)
+	tdfx_dga.c tdfx_priv.c tdfx_sli.c $(DRISRCS) 
 
 OBJS = tdfx_driver.o tdfx_io.o tdfx_hwcurs.o tdfx_accel.o \
-	tdfx_dga.o $(DRIOBJS) $(PRIVOBJS)
+	tdfx_dga.o tdfx_priv.o tdfx_sli.o $(DRIOBJS)
 
 #if defined(XF86DriverSDK)
 INCLUDES = -I. -I../../include
@@ -43,7 +36,7 @@ INCLUDES = -I. -I$(XF86COMSRC) -I$(XF86OSSRC) \
 	   $(DRIINCLUDES)
 #endif
 
-DEFINES = $(DRIDEFINES) $(PRIVDEFINES)
+DEFINES = -DPROP_3DFX $(DRIDEFINES) 
 
 #if MakeHasPosixVariableSubstitutions
 SubdirLibraryRule($(OBJS))
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.h b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.h
index 10bf7e4aa..66e9c0d65 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.h
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.h
@@ -35,6 +35,12 @@
 struct _TDFXRec;
 typedef struct _TDFXRec *TDFXPtr;
 
+#ifdef XF86DRI
+#define PIXMAP_CACHE_LINES 128
+#else
+#define PIXMAP_CACHE_LINES 512
+#endif
+
 #ifdef PROP_3DFX
 #include "tdfx_priv.h"
 #else
@@ -57,6 +63,7 @@ extern void FillPrivateDRI(TDFXPtr pTDFX, TDFXDRIPtr pTDFXDRI);
 #define TDFX_DEBUG_CMDS
 #define TRACECURS 1
 #define REGDEBUG 1
+#define TRACEREG
 #endif
 
 #ifdef TRACE
@@ -92,6 +99,14 @@ typedef char (*TDFXReadIndexedByteFunc)(TDFXPtr pTDFX, int addr,
 					char index);
 typedef void (*TDFXWriteWordFunc)(TDFXPtr pTDFX, int addr, int value);
 typedef int (*TDFXReadWordFunc)(TDFXPtr pTDFX, int addr);
+typedef void (*TDFXWriteChipIndexedByteFunc)(TDFXPtr pTDFX, int chip, 
+					     int addr, char index, char value);
+typedef char (*TDFXReadChipIndexedByteFunc)(TDFXPtr pTDFX, int chip, 
+					    int addr, char index);
+typedef void (*TDFXWriteChipWordFunc)(TDFXPtr pTDFX, int chip, 
+				      int addr, int value);
+typedef int (*TDFXReadChipWordFunc)(TDFXPtr pTDFX, int chip,
+				    int addr);
 typedef void (*TDFXSyncFunc)(ScrnInfoPtr pScrn);
 typedef void (*TDFXBufferFunc)(TDFXPtr pTDFX, int which);
 
@@ -119,21 +134,26 @@ typedef struct TextureData_t {
   struct TextureData_t *next;
 } TextureData;
 
+#define MAXCHIPS 4
+
 typedef struct _TDFXRec {
-  unsigned char *MMIOBase;
+  unsigned char *MMIOBase[MAXCHIPS];
   unsigned char *FbBase;
-  unsigned int PIOBase;
+  unsigned char *myFbBase;
+  unsigned int PIOBase[MAXCHIPS];
   long FbMapSize;
+  int pixelFormat;
   int stride;
   int cpp;
   int maxClip;
   int MaxClock;
-  int Chipset;
-  int LinearAddr;
-  int MMIOAddr;
-  EntityInfoPtr pEnt;
+  int ChipType;
   pciVideoPtr PciInfo;
-  PCITAG PciTag;
+  int LinearAddr[MAXCHIPS];
+  int MMIOAddr[MAXCHIPS];
+  EntityInfoPtr pEnt;
+  int numChips;
+  PCITAG PciTag[MAXCHIPS];
   int HasSGRAM;
   int PciCnt;
   int PrevDrawState;
@@ -162,6 +182,8 @@ typedef struct _TDFXRec {
   TDFXReadIndexedByteFunc readControl;
   TDFXWriteWordFunc writeLong;
   TDFXReadWordFunc readLong;
+  TDFXWriteChipWordFunc writeChipLong;
+  TDFXReadChipWordFunc readChipLong;
   TDFXSyncFunc sync;
   int syncDone;
   int scanlineWidth;
@@ -197,8 +219,8 @@ typedef struct {
 
 #define TDFX2XCUTOFF 135000
 
-#ifndef PCI_CHIP_VOODOOEXP
-#define PCI_CHIP_VOODOOEXP 9
+#ifndef PCI_CHIP_VOODOO5
+#define PCI_CHIP_VOODOO5 9
 #endif
 
 extern Bool TDFXAccelInit(ScreenPtr pScreen);
@@ -210,7 +232,7 @@ extern Bool TDFXDRIFinishScreenInit(ScreenPtr pScreen);
 extern Bool TDFXDGAInit(ScreenPtr pScreen);
 extern void TDFXCursorGrabMemory(ScreenPtr pScreen);
 extern void TDFXSetLFBConfig(TDFXPtr pTDFX);
-extern void TDFXSendNOPPrivate(TDFXPtr pTDFX);
+extern void TDFXSendNOPPrivate(ScrnInfoPtr pScrn);
 
 extern Bool TDFXSwitchMode(int scrnIndex, DisplayModePtr mode, int flags);
 extern void TDFXAdjustFrame(int scrnIndex, int x, int y, int flags);
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.man b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.man
new file mode 100644
index 000000000..a647ac301
--- /dev/null
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.man
@@ -0,0 +1,30 @@
+.\" $XFree86: xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx.cpp,v 1.1 2000/03/03 01:05:43 dawes Exp $ 
+.TH TDFX 4 "Version 4.0"  "XFree86"
+.SH NAME
+tdfx \- 3Dfx video driver
+.SH SYNOPSIS
+.B "Section ""Device"""
+.br
+.BI "  Identifier """  devname """"
+.br
+.B  "  Driver ""tdfx"""
+.br
+\ \ ...
+.br
+.B EndSection
+.SH DESCRIPTION
+.B tdfx 
+is an XFree86 driver for 3Dfx video cards.
+THIS MAN PAGE NEEDS TO BE FILLED IN.
+.SH SUPPORTED HARDWARE
+The
+.B tdfx
+driver supports...
+.SH CONFIGURATION DETAILS
+Please refer to XF86Config(5x) for general configuration
+details.  This section only covers configuration details specific to this
+driver.
+.SH "SEE ALSO"
+XFree86(1), XF86Config(5x), xf86config(1), Xserver(1), X(1)
+.SH AUTHORS
+Authors include: ...
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_accel.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_accel.c
index a02fd5308..e6db51242 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_accel.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_accel.c
@@ -79,13 +79,17 @@ TDFXNeedSync(ScrnInfoPtr pScrn) {
   pTDFX->syncDone=FALSE;
 }
 
-static void
+void
 TDFXFirstSync(ScrnInfoPtr pScrn) {
   TDFXPtr pTDFX = TDFXPTR(pScrn);
 
   if (!pTDFX->syncDone) {
-    pTDFX->sync(pScrn);
+    if (pTDFX->directRenderingEnabled) {
+      DRILock(screenInfo.screens[pScrn->scrnIndex], 0);
+      TDFXSwapContextPrivate(screenInfo.screens[pScrn->scrnIndex]);
+    }
     pTDFX->syncDone=TRUE;
+    pTDFX->sync(pScrn);
   }
 }
 
@@ -96,12 +100,15 @@ TDFXCheckSync(ScrnInfoPtr pScrn) {
   if (pTDFX->syncDone) {
     pTDFX->sync(pScrn);
     pTDFX->syncDone=FALSE;
+    if (pTDFX->directRenderingEnabled) {
+      TDFXLostContext(screenInfo.screens[pScrn->scrnIndex]);
+      DRIUnlock(screenInfo.screens[pScrn->scrnIndex]);
+    }
   }
 }
 
 void
 TDFXSelectBuffer(TDFXPtr pTDFX, int which) {
-#ifdef XF86DRI
   int fmt;
 
   TDFXMakeRoom(pTDFX, 4);
@@ -116,14 +123,20 @@ TDFXSelectBuffer(TDFXPtr pTDFX, int which) {
     TDFXWriteLong(pTDFX, SST_2D_SRCFORMAT, fmt);
     break;
   case TDFX_BACK:
-    fmt=((pTDFX->stride+127)/128)|(3<<16); /* Tiled 16bpp */
+    if (pTDFX->cpp==2)
+      fmt=((pTDFX->stride+127)/128)|(3<<16); /* Tiled 16bpp */
+    else
+      fmt=((pTDFX->stride+127)/128)|(5<<16); /* Tiled 32bpp */
     TDFXWriteLong(pTDFX, SST_2D_DSTBASEADDR, pTDFX->backOffset|BIT(31));
     TDFXWriteLong(pTDFX, SST_2D_DSTFORMAT, fmt);
     TDFXWriteLong(pTDFX, SST_2D_SRCBASEADDR, pTDFX->backOffset|BIT(31));
     TDFXWriteLong(pTDFX, SST_2D_SRCFORMAT, fmt);
     break;
   case TDFX_DEPTH:
-    fmt=((pTDFX->stride+127)/128)|(3<<16); /* Tiled 16bpp */
+    if (pTDFX->cpp==2)
+      fmt=((pTDFX->stride+127)/128)|(3<<16); /* Tiled 16bpp */
+    else
+      fmt=((pTDFX->stride+127)/128)|(5<<16); /* Tiled 32bpp */
     TDFXWriteLong(pTDFX, SST_2D_DSTBASEADDR, pTDFX->depthOffset|BIT(31));
     TDFXWriteLong(pTDFX, SST_2D_DSTFORMAT, fmt);
     TDFXWriteLong(pTDFX, SST_2D_SRCBASEADDR, pTDFX->depthOffset|BIT(31));
@@ -131,14 +144,27 @@ TDFXSelectBuffer(TDFXPtr pTDFX, int which) {
     break;
   default:
   }
-#endif  
 }
 
 void
 TDFXSetLFBConfig(TDFXPtr pTDFX) {
-  TDFXWriteLongMMIO(pTDFX, LFBMEMORYCONFIG, (pTDFX->backOffset>>12) |
-		    SST_RAW_LFB_ADDR_STRIDE_4K | 
-		    ((pTDFX->stride+127)/128)<<SST_RAW_LFB_TILE_STRIDE_SHIFT);
+  if (pTDFX->ChipType<=PCI_CHIP_VOODOO3) {
+    TDFXWriteLongMMIO(pTDFX, LFBMEMORYCONFIG, (pTDFX->backOffset>>12) |
+		      SST_RAW_LFB_ADDR_STRIDE_4K | 
+		      ((pTDFX->stride+127)/128)<<SST_RAW_LFB_TILE_STRIDE_SHIFT);
+  } else {
+    int chip;
+    int stride, bits;
+    if (pTDFX->cpp==2) stride=pTDFX->stride;
+    else stride=4*pTDFX->stride/pTDFX->cpp;
+    bits=pTDFX->backOffset>>12;
+    for (chip=0; chip<pTDFX->numChips; chip++) {
+      TDFXWriteChipLongMMIO(pTDFX, chip, LFBMEMORYCONFIG, (bits&0x1FFF) |
+			    SST_RAW_LFB_ADDR_STRIDE_4K | 
+			    ((bits&0x6000)<<10) |
+			    ((stride+127)/128)<<SST_RAW_LFB_TILE_STRIDE_SHIFT);
+    }
+  }
 }
 
 Bool
@@ -263,8 +289,11 @@ static void TDFXMakeRoomNoProp(TDFXPtr pTDFX, int size) {
   }
 }
 
-static void TDFXSendNOPNoProp(TDFXPtr pTDFX)
+static void TDFXSendNOPNoProp(ScrnInfoPtr pScrn)
 {
+  TDFXPtr pTDFX;
+
+  pTDFX=TDFXPTR(pScrn);
   TDFXMakeRoomNoProp(pTDFX, 1);
   TDFXWriteLongMMIO(pTDFX, SST_2D_COMMAND, SST_2D_NOP);
 }  
@@ -278,7 +307,7 @@ void TDFXSync(ScrnInfoPtr pScrn)
   TDFXTRACEACCEL("TDFXSync\n");
   pTDFX=TDFXPTR(pScrn);
 
-  TDFXSendNOPNoProp(pTDFX);
+  TDFXSendNOPNoProp(pScrn);
   i=0;
   do {
     stat=TDFXReadLongMMIO(pTDFX, 0);
@@ -408,7 +437,7 @@ TDFXSubsequentScreenToScreenCopy(ScrnInfoPtr pScrn, int srcX, int srcY,
   }
   if ((srcY>=dstY-32 && srcY<=dstY)||
       (srcY>=pTDFX->prevBlitDest.y1-32 && srcY<=pTDFX->prevBlitDest.y1)) {
-    TDFXSendNOP(pTDFX);
+    TDFXSendNOP(pScrn);
   }
   pTDFX->sync(pScrn);
 
@@ -606,8 +635,8 @@ TDFXNonTEGlyphRenderer(ScrnInfoPtr pScrn, int x, int y, int n,
 	  SSTCP_SRCXY|SSTCP_COLORFORE|SSTCP_COMMAND);
   TDFXWriteLong(pTDFX, SST_2D_CLIP1MIN, ((pbox->y1&0x1FFF)<<16) |
 		(pbox->x1&0x1FFF));
-  TDFXWriteLong(pTDFX, SST_2D_CLIP1MAX, (((pbox->y2+1)&0x1FFF)<<16) |
-		((pbox->x2+1)&0x1FFF));
+  TDFXWriteLong(pTDFX, SST_2D_CLIP1MAX, ((pbox->y2&0x1FFF)<<16) |
+		(pbox->x2&0x1FFF));
   TDFXWriteLong(pTDFX, SST_2D_SRCFORMAT, SST_2D_PIXFMT_1BPP |
 		SST_2D_SOURCE_PACKING_DWORD);
   TDFXWriteLong(pTDFX, SST_2D_SRCXY, 0);
@@ -734,7 +763,7 @@ TDFXSubsequentScreenToScreenColorExpandFill(ScrnInfoPtr pScrn, int x, int y,
   pTDFX->DrawState|=DRAW_STATE_CLIP1CHANGED;
 
   if (srcy>=pTDFX->prevBlitDest.y1-8 && srcy<=pTDFX->prevBlitDest.y1) {
-    TDFXSendNOP(pTDFX);
+    TDFXSendNOP(pScrn);
   }
 
   if (pTDFX->cpp==1) fmt=(1<<16)|pTDFX->stride; 
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_dri.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_dri.c
index f8f0a6cb1..11ea6f3f0 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_dri.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_dri.c
@@ -47,10 +47,97 @@ TDFXInitVisualConfigs(ScreenPtr pScreen)
 
   switch (pScrn->bitsPerPixel) {
   case 8:
+  case 16:
+    numConfigs = 16;
+
+    if (!(pConfigs = (__GLXvisualConfig*)xnfcalloc(sizeof(__GLXvisualConfig),
+						   numConfigs))) {
+      return FALSE;
+    }
+    if (!(pTDFXConfigs = (TDFXConfigPrivPtr)xnfcalloc(sizeof(TDFXConfigPrivRec),
+						     numConfigs))) {
+      xfree(pConfigs);
+      return FALSE;
+    }
+    if (!(pTDFXConfigPtrs = (TDFXConfigPrivPtr*)xnfcalloc(sizeof(TDFXConfigPrivPtr),
+							 numConfigs))) {
+      xfree(pConfigs);
+      xfree(pTDFXConfigs);
+      return FALSE;
+    }
+    for (i=0; i<numConfigs; i++) 
+      pTDFXConfigPtrs[i] = &pTDFXConfigs[i];
+
+    i=0;
+    depth=1;
+    for (db = 0; db <=1; db++) {
+      for (depth = 0; depth<=1; depth++) {
+	for (accum = 0; accum <= 1; accum++) {
+	  for (stencil = 0; stencil <= 1; stencil++) {
+	    pConfigs[i].vid = -1;
+	    pConfigs[i].class = -1;
+	    pConfigs[i].rgba = TRUE;
+	    pConfigs[i].redSize = 5;
+	    pConfigs[i].greenSize = 6;
+	    pConfigs[i].blueSize = 5;
+	    pConfigs[i].redMask = 0x0000F800;
+	    pConfigs[i].greenMask = 0x000007E0;
+	    pConfigs[i].blueMask = 0x0000001F;
+	    pConfigs[i].alphaMask = 0;
+	    if (accum) {
+	      pConfigs[i].accumRedSize = 16;
+	      pConfigs[i].accumGreenSize = 16;
+	      pConfigs[i].accumBlueSize = 16;
+	      pConfigs[i].accumAlphaSize = 0;
+	    } else {
+	      pConfigs[i].accumRedSize = 0;
+	      pConfigs[i].accumGreenSize = 0;
+	      pConfigs[i].accumBlueSize = 0;
+	      pConfigs[i].accumAlphaSize = 0;
+	    }
+	    if (db)
+	      pConfigs[i].doubleBuffer = TRUE;
+	    else
+	      pConfigs[i].doubleBuffer = FALSE;
+	    pConfigs[i].stereo = FALSE;
+	    pConfigs[i].bufferSize = 16;
+	    if (depth) {
+	      if (pTDFX->cpp>2)
+		pConfigs[i].depthSize = 24;
+	      else 
+		pConfigs[i].depthSize = 16;
+	    } else {
+	      pConfigs[i].depthSize = 0;
+	    }
+	    if (stencil)
+	      pConfigs[i].stencilSize = 8;
+	    else
+	      pConfigs[i].stencilSize = 0;
+	    pConfigs[i].auxBuffers = 0;
+	    pConfigs[i].level = 0;
+	    if (stencil || accum)
+	      pConfigs[i].visualRating = GLX_SLOW_VISUAL_EXT;
+	    else
+	      pConfigs[i].visualRating = GLX_NONE_EXT;
+	    pConfigs[i].transparentPixel = 0;
+	    pConfigs[i].transparentRed = 0;
+	    pConfigs[i].transparentGreen = 0;
+	    pConfigs[i].transparentBlue = 0;
+	    pConfigs[i].transparentAlpha = 0;
+	    pConfigs[i].transparentIndex = 0;
+	    i++;
+	  }
+	}
+      }
+    }
+    if (i!=numConfigs) {
+      ErrorF("Incorrect initialization of visuals\n");
+      return FALSE;
+    } else ErrorF("Created %d visuals\n", numConfigs);
+    break; /* 16bpp */
+
   case 24:
   case 32:
-    break;
-  case 16:
     numConfigs = 8;
 
     pConfigs = (__GLXvisualConfig*) xnfcalloc(sizeof(__GLXvisualConfig), numConfigs);
@@ -74,66 +161,74 @@ TDFXInitVisualConfigs(ScreenPtr pScreen)
       pTDFXConfigPtrs[i] = &pTDFXConfigs[i];
 
     i=0;
-    depth=1;
-    for (accum = 0; accum <= 1; accum++) {
-      for (stencil = 0; stencil <= 1; stencil++) {
-        for (db=0; db<=1; db++) {
-          pConfigs[i].vid = -1;
-          pConfigs[i].class = -1;
-          pConfigs[i].rgba = TRUE;
-          pConfigs[i].redSize = 5;
-          pConfigs[i].greenSize = 6;
-          pConfigs[i].blueSize = 5;
-          pConfigs[i].redMask = 0x0000F800;
-          pConfigs[i].greenMask = 0x000007E0;
-          pConfigs[i].blueMask = 0x0000001F;
-          pConfigs[i].alphaMask = 0;
-          if (accum) {
-            pConfigs[i].accumRedSize = 16;
-            pConfigs[i].accumGreenSize = 16;
-            pConfigs[i].accumBlueSize = 16;
-            pConfigs[i].accumAlphaSize = 0;
-          } else {
-            pConfigs[i].accumRedSize = 0;
-            pConfigs[i].accumGreenSize = 0;
-            pConfigs[i].accumBlueSize = 0;
-            pConfigs[i].accumAlphaSize = 0;
-          }
-          if (db)
-            pConfigs[i].doubleBuffer = TRUE;
-          else
-            pConfigs[i].doubleBuffer = FALSE;
-          pConfigs[i].stereo = FALSE;
-          pConfigs[i].bufferSize = 16;
-          if (depth)
-            pConfigs[i].depthSize = 16;
-          else 
-            pConfigs[i].depthSize = 0;
-          if (stencil)
-            pConfigs[i].stencilSize = 8;
-          else
-            pConfigs[i].stencilSize = 0;
-          pConfigs[i].auxBuffers = 0;
-          pConfigs[i].level = 0;
-          if (stencil)
-            pConfigs[i].visualRating = GLX_SLOW_VISUAL_EXT;
-          else
-            pConfigs[i].visualRating = GLX_NONE_EXT;
-          pConfigs[i].transparentPixel = 0;
-          pConfigs[i].transparentRed = 0;
-          pConfigs[i].transparentGreen = 0;
-          pConfigs[i].transparentBlue = 0;
-          pConfigs[i].transparentAlpha = 0;
-          pConfigs[i].transparentIndex = 0;
-          i++;
-        }
+    for (db = 0; db <=1; db++) {
+      for (depth = 0; depth<=1; depth++) {
+         /*stencil = depth;*/  /* Z and stencil share the same memory */
+	for (accum = 0; accum <= 1; accum++) {
+           /*for (stencil = 0; stencil <=1; stencil++) {*/
+           stencil = depth;
+	    pConfigs[i].vid = -1;
+	    pConfigs[i].class = -1;
+	    pConfigs[i].rgba = TRUE;
+	    pConfigs[i].redSize = 8;
+	    pConfigs[i].greenSize = 8;
+	    pConfigs[i].blueSize = 8;
+	    pConfigs[i].alphaSize = (pScrn->bitsPerPixel==32) ? 8 : 0;
+	    pConfigs[i].redMask   = 0x00ff0000;
+	    pConfigs[i].greenMask = 0x0000ff00;
+	    pConfigs[i].blueMask  = 0x000000ff;
+	    pConfigs[i].alphaMask = (pScrn->bitsPerPixel==32) ? 0xff000000 : 0;
+	    if (accum) {
+	      pConfigs[i].accumRedSize = 16;
+	      pConfigs[i].accumGreenSize = 16;
+	      pConfigs[i].accumBlueSize = 16;
+	      pConfigs[i].accumAlphaSize = (pScrn->bitsPerPixel==32) ? 16 : 0;
+	    } else {
+	      pConfigs[i].accumRedSize = 0;
+	      pConfigs[i].accumGreenSize = 0;
+	      pConfigs[i].accumBlueSize = 0;
+	      pConfigs[i].accumAlphaSize = 0;
+	    }
+	    if (db)
+	      pConfigs[i].doubleBuffer = TRUE;
+	    else
+	      pConfigs[i].doubleBuffer = FALSE;
+	    pConfigs[i].stereo = FALSE;
+	    pConfigs[i].bufferSize = 16;
+	    if (depth) {
+	      if (pTDFX->cpp > 2)
+		pConfigs[i].depthSize = 24;
+	      else 
+		pConfigs[i].depthSize = 16;
+	    } else {
+	      pConfigs[i].depthSize = 0;
+	    }
+	    if (stencil)
+	      pConfigs[i].stencilSize = 8;
+	    else
+	      pConfigs[i].stencilSize = 0;
+	    pConfigs[i].auxBuffers = 0;
+	    pConfigs[i].level = 0;
+	    if (accum)
+	      pConfigs[i].visualRating = GLX_SLOW_VISUAL_EXT;
+	    else
+	      pConfigs[i].visualRating = GLX_NONE_EXT;
+	    pConfigs[i].transparentPixel = 0;
+	    pConfigs[i].transparentRed = 0;
+	    pConfigs[i].transparentGreen = 0;
+	    pConfigs[i].transparentBlue = 0;
+	    pConfigs[i].transparentAlpha = 0;
+	    pConfigs[i].transparentIndex = 0;
+	    i++;
+         /*}*/
+	}
       }
     }
     if (i!=numConfigs) {
       xf86DrvMsg(pScreen->myNum, X_ERROR,
                  "Incorrect initialization of visuals\n");
       return FALSE;
-    }
+    } else ErrorF("Created %d visuals\n", numConfigs);
     break;
   }
   pTDFX->numVisualConfigs = numConfigs;
@@ -143,6 +238,26 @@ TDFXInitVisualConfigs(ScreenPtr pScreen)
   return TRUE;
 }
 
+static void
+TDFXDoWakeupHandler(int screenNum, pointer wakeupData, unsigned long result,
+		    pointer pReadmask)
+{
+  ScreenPtr pScreen = screenInfo.screens[screenNum];
+  ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];
+
+  TDFXNeedSync(pScrn);
+}
+
+static void 
+TDFXDoBlockHandler(int screenNum, pointer blockData, pointer pTimeout,
+		  pointer pReadmask)
+{
+  ScreenPtr pScreen = screenInfo.screens[screenNum];
+  ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];
+
+  TDFXCheckSync(pScrn);
+}
+
 Bool TDFXDRIScreenInit(ScreenPtr pScreen)
 {
   ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];
@@ -150,7 +265,15 @@ Bool TDFXDRIScreenInit(ScreenPtr pScreen)
   DRIInfoPtr pDRIInfo;
   TDFXDRIPtr pTDFXDRI;
 
-  if (pScrn->bitsPerPixel != 16) return FALSE;
+  switch (pScrn->bitsPerPixel) {
+  case 8:
+    return FALSE;
+  case 16:
+    break;
+  case 24:
+  case 32:
+    if (pTDFX->ChipType<=PCI_CHIP_VOODOO3) return FALSE;
+  }
 
 #if XFree86LOADER
     /* Check that the GLX, DRI, and DRM modules have been loaded by testing
@@ -187,11 +310,16 @@ Bool TDFXDRIScreenInit(ScreenPtr pScreen)
   pDRIInfo->ddxDriverMajorVersion = TDFX_MAJOR_VERSION;
   pDRIInfo->ddxDriverMinorVersion = TDFX_MINOR_VERSION;
   pDRIInfo->ddxDriverPatchVersion = TDFX_PATCHLEVEL;
-  pDRIInfo->frameBufferPhysicalAddress = pTDFX->LinearAddr;
+  pDRIInfo->frameBufferPhysicalAddress = pTDFX->LinearAddr[0];
   pDRIInfo->frameBufferSize = pTDFX->FbMapSize;
   pDRIInfo->frameBufferStride = pTDFX->stride;
   pDRIInfo->ddxDrawableTableEntry = TDFX_MAX_DRAWABLES;
 
+  pDRIInfo->wrap.ValidateTree = 0;
+  pDRIInfo->wrap.PostValidateTree = 0;
+  pDRIInfo->wrap.BlockHandler = TDFXDoBlockHandler;
+  pDRIInfo->wrap.WakeupHandler = TDFXDoWakeupHandler;
+
   if (SAREA_MAX_DRAWABLES < TDFX_MAX_DRAWABLES)
     pDRIInfo->maxDrawableTableEntry = SAREA_MAX_DRAWABLES;
   else
@@ -261,7 +389,7 @@ Bool TDFXDRIScreenInit(ScreenPtr pScreen)
   }
 
   pTDFXDRI->regsSize=TDFXIOMAPSIZE;
-  if (drmAddMap(pTDFX->drmSubFD, (drmHandle)pTDFX->MMIOAddr, 
+  if (drmAddMap(pTDFX->drmSubFD, (drmHandle)pTDFX->MMIOAddr[0], 
 		pTDFXDRI->regsSize, DRM_REGISTERS, 0, &pTDFXDRI->regs)<0) {
     TDFXDRICloseScreen(pScreen);
     return FALSE;
@@ -303,9 +431,7 @@ TDFXCreateContext(ScreenPtr pScreen, VisualPtr visual,
 		  drmContext hwContext, void *pVisualConfigPriv,
 		  DRIContextType contextStore)
 {
-  /*ScrnInfoPtr pScrn = xf86Screens[pScreen->myNum];*/
-  /*TDFXPtr pTDFX = TDFXPTR(pScrn);*/
-  /*TDFXConfigPrivPtr pTDFXConfig = (TDFXConfigPrivPtr)pVisualConfigPriv;*/
+  TDFXConfigPrivPtr pTDFXConfig = (TDFXConfigPrivPtr)pVisualConfigPriv;
   TDFXDRIContextPtr ctx;
 
   ctx=(TDFXDRIContextPtr)contextStore;
@@ -351,14 +477,21 @@ TDFXDRISwapContext(ScreenPtr pScreen, DRISyncType syncType,
 		   DRIContextType oldContextType, void *oldContext,
 		   DRIContextType newContextType, void *newContext)
 {
-  if ((syncType==DRI_3D_SYNC) && (oldContextType==DRI_2D_CONTEXT) &&
-      (newContextType==DRI_2D_CONTEXT)) { /* Entering from Wakeup */
-    TDFXSwapContextPrivate(pScreen);
-  }
+#if 0
+  ScrnInfoPtr pScrn;
+
+  pScrn = xf86Screens[pScreen->myNum];
   if ((syncType==DRI_2D_SYNC) && (oldContextType==DRI_NO_CONTEXT) &&
       (newContextType==DRI_2D_CONTEXT)) { /* Exiting from Block Handler */
+    TDFXCheckSync(pScrn);
     TDFXLostContext(pScreen);
   }
+  if ((syncType==DRI_3D_SYNC) && (oldContextType==DRI_2D_CONTEXT) &&
+      (newContextType==DRI_2D_CONTEXT)) { /* Entering from Wakeup */
+    TDFXSwapContextPrivate(pScreen);
+    TDFXNeedSync(pScrn);
+  }
+#endif
 }
 
 static void
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_driver.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_driver.c
index 9f1483629..8b0c2dc38 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_driver.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_driver.c
@@ -153,7 +153,7 @@ DriverRec TDFX = {
   TDFX_VERSION,
   TDFX_DRIVER_NAME,
 #if 0
-  "Accelerated driver for 3dfx Voodoo Banshee and Voodoo3 cards",
+  "Accelerated driver for 3dfx cards",
 #endif
   TDFXIdentify,
   TDFXProbe,
@@ -165,14 +165,14 @@ DriverRec TDFX = {
 static SymTabRec TDFXChipsets[] = {
   { PCI_CHIP_BANSHEE, "3dfx Banshee"},
   { PCI_CHIP_VOODOO3, "3dfx Voodoo3"},
-  { PCI_CHIP_VOODOOEXP, "3dfx VoodooExp"},
+  { PCI_CHIP_VOODOO5, "3dfx Voodoo5"},
   { -1, NULL }
 };
 
 static PciChipsets TDFXPciChipsets[] = {
   { PCI_CHIP_BANSHEE, PCI_CHIP_BANSHEE, RES_SHARED_VGA },
   { PCI_CHIP_VOODOO3, PCI_CHIP_VOODOO3, RES_SHARED_VGA },
-  { PCI_CHIP_VOODOOEXP, PCI_CHIP_VOODOOEXP, RES_SHARED_VGA },
+  { PCI_CHIP_VOODOO5, PCI_CHIP_VOODOO5, RES_SHARED_VGA },
   { -1, -1, RES_UNDEFINED }
 };
 
@@ -182,13 +182,15 @@ static PciChipsets TDFXPciChipsets[] = {
 typedef enum {
   OPTION_NOACCEL,
   OPTION_SW_CURSOR,
-  OPTION_USE_PIO
+  OPTION_USE_PIO,
+  OPTION_NO_SLI
 } TDFXOpts;
 
 static OptionInfoRec TDFXOptions[] = {
   { OPTION_NOACCEL, "NoAccel", OPTV_BOOLEAN, {0}, FALSE },
   { OPTION_SW_CURSOR, "SWcursor", OPTV_BOOLEAN, {0}, FALSE },
   { OPTION_USE_PIO, "UsePIO", OPTV_BOOLEAN, {0}, FALSE},
+  { OPTION_NO_SLI, "NoSLI", OPTV_BOOLEAN, {0}, FALSE},
   { -1, NULL, OPTV_NONE, {0}, FALSE}
 };
 
@@ -415,9 +417,7 @@ TDFXProbe(DriverPtr drv, int flags) {
 				  TDFXChipsets, TDFXPciChipsets,
 				  devSections, numDevSections,
 				  drv, &usedChips);
-  if (devSections)
-    xfree(devSections);
-  devSections=NULL;
+
   if (numUsed<=0) return FALSE;
 
   if (flags & PROBE_DETECT)
@@ -445,6 +445,7 @@ TDFXProbe(DriverPtr drv, int flags) {
     xf86ConfigActivePciEntity(pScrn, usedChips[i], TDFXPciChipsets, 0, 0, 0, 0, 0);
   }
   xfree(usedChips);
+  if (devSections) xfree(devSections);
 
   return foundScreen;
 }
@@ -458,7 +459,7 @@ TDFXCountRam(ScrnInfoPtr pScrn) {
   pTDFX = TDFXPTR(pScrn);
   TDFXTRACE("TDFXCountRam start\n");
   memSize=0;
-  if (pTDFX->PIOBase) {
+  if (pTDFX->PIOBase[0]) {
     CARD32 
       partSize,                 /* size of SGRAM chips in Mbits */
       nChips,                   /* # chips of SDRAM/SGRAM */
@@ -488,7 +489,7 @@ TDFXCountRam(ScrnInfoPtr pScrn) {
     /* determine memory size from strapping pins (dramInit0 and dramInit1) */
     dramInit0_strap = pTDFX->readLong(pTDFX, DRAMINIT0);
 
-    if (pTDFX->PciInfo->chipType<=5) { /* Banshee/V3 */
+    if (pTDFX->ChipType<=PCI_CHIP_VOODOO3) { /* Banshee/V3 */
       if (memType == MEM_TYPE_SDRAM) {
 	memSize = 16;
       } else {
@@ -510,8 +511,6 @@ TDFXCountRam(ScrnInfoPtr pScrn) {
       partSize=1<<((dramInit0_strap&0x38000000)>>28);
       banks=((dramInit0_strap&BIT(30))==0) ? 2 : 4;
       memSize=nChips*partSize*banks;
-      ErrorF("reg=%x nChips=%d partSize=%d banks=%d memSize=%d\n", 
-	     dramInit0_strap, nChips, partSize, banks, memSize);
     }
     TDFXTRACEREG("dramInit0 = %x dramInit1 = %x\n", dramInit0_strap, dramInit1_strap);
     TDFXTRACEREG("MemConfig %d chips %d size %d total\n", nChips, partSize, memSize);
@@ -543,6 +542,83 @@ TDFXProbeDDC(ScrnInfoPtr pScrn, int index)
     }
 }
 
+static int TDFXCfgToSize(int cfg)
+{
+  if (cfg<4) return 0x8000000<<cfg;
+  return 0x4000000>>(cfg-4);
+}
+
+static int TDFXSizeToCfg(int size)
+{
+  switch (size) {
+  case 0x40000000: return 3;
+  case 0x20000000: return 2;
+  case 0x10000000: return 1;
+  case 0x08000000: return 0;
+  case 0x04000000: return 4;
+  case 0x02000000: return 5;
+  case 0x01000000: return 6;
+  case 0x00800000: return 7;
+  case 0x00400000: return 8;
+  default:
+    return -1;
+  }
+}
+
+static void
+TDFXFindChips(ScrnInfoPtr pScrn, pciVideoPtr match)
+{
+  TDFXPtr pTDFX;
+  pciVideoPtr *ppPci;
+
+  pTDFX=TDFXPTR(pScrn);
+  pTDFX->numChips=0;
+  pTDFX->ChipType=match->chipType;
+  for (ppPci = xf86GetPciVideoInfo(); *ppPci != NULL; ppPci++) {
+    if ((*ppPci)->bus == match->bus && 
+	(*ppPci)->device == match->device) {
+      pTDFX->PciTag[pTDFX->numChips] = pciTag((*ppPci)->bus, 
+					      (*ppPci)->device,
+					      (*ppPci)->func);
+      pTDFX->PIOBase[pTDFX->numChips] = (*ppPci)->ioBase[2]&0xFFFFFFFC;
+      pTDFX->numChips++;
+    }
+  }
+}
+
+static void
+TDFXInitChips(ScrnInfoPtr pScrn)
+{
+  TDFXPtr pTDFX;
+  int i, cfgbits, initbits;
+  int mem0base, mem1base, mem0size, mem0bits, mem1size, mem1bits;
+
+  pTDFX=TDFXPTR(pScrn);
+  cfgbits=pciReadLong(pTDFX->PciTag[0], CFG_PCI_DECODE);
+  mem0base=pciReadLong(pTDFX->PciTag[0], CFG_MEM0BASE);
+  mem1base=pciReadLong(pTDFX->PciTag[0], CFG_MEM1BASE);
+  mem0size=32*1024*1024; /* Registers are always 32MB */
+  mem1size=pScrn->videoRam*1024*2; /* Linear mapping is 2x memory */
+  mem0bits=TDFXSizeToCfg(mem0size);
+  mem1bits=TDFXSizeToCfg(mem1size)<<4;
+  cfgbits=(cfgbits&~(0xFF))|mem0bits|mem1bits;
+  for (i=0; i<pTDFX->numChips; i++) {
+    initbits=pciReadLong(pTDFX->PciTag[i], CFG_INIT_ENABLE);
+    initbits|=BIT(10);
+    pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE, initbits);
+    pTDFX->MMIOAddr[i]=mem0base+i*mem0size;
+    pciWriteLong(pTDFX->PciTag[i], CFG_MEM0BASE, 0xFFFFFFFF);
+    pciWriteLong(pTDFX->PciTag[i], CFG_MEM0BASE, pTDFX->MMIOAddr[i]);
+    pTDFX->MMIOAddr[i]&=0xFFFFFF00;
+    pTDFX->LinearAddr[i]=mem1base+i*mem1size;
+    pciWriteLong(pTDFX->PciTag[i], CFG_MEM1BASE, 0xFFFFFFFF);
+    pciWriteLong(pTDFX->PciTag[i], CFG_MEM1BASE, pTDFX->LinearAddr[i]);
+    pTDFX->LinearAddr[i]&=0xFFFFFF00;
+    pciWriteLong(pTDFX->PciTag[i], CFG_PCI_DECODE, cfgbits);
+    initbits&=~BIT(10);
+    pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE, initbits);
+  }
+}
 
 /*
  * TDFXPreInit --
@@ -552,7 +628,8 @@ TDFXProbeDDC(ScrnInfoPtr pScrn, int index)
  *
  */
 static Bool
-TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
+TDFXPreInit(ScrnInfoPtr pScrn, int flags) 
+{
   TDFXPtr pTDFX;
   ClockRangePtr clockRanges;
   int i;
@@ -560,6 +637,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
   char *mod=0, *reqSym=0;
   int flags24;
   rgb defaultWeight = {0, 0, 0};
+  pciVideoPtr match;
 
   TDFXTRACE("TDFXPreInit start\n");
   if (pScrn->numEntities != 1) return FALSE;
@@ -594,9 +672,8 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
     }
   }
 
-  pTDFX->PciInfo = xf86GetPciInfoForEntity(pTDFX->pEnt->index);
-  pTDFX->PciTag = pciTag(pTDFX->PciInfo->bus, pTDFX->PciInfo->device,
-			 pTDFX->PciInfo->func);
+  match=pTDFX->PciInfo=xf86GetPciInfoForEntity(pTDFX->pEnt->index);
+  TDFXFindChips(pScrn, match);
 
   if (xf86RegisterResources(pTDFX->pEnt->index, 0, ResNone))
       return FALSE;
@@ -610,6 +687,14 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
     return FALSE;
   }
   
+  /* The vgahw module should be loaded here when needed */
+  if (!xf86LoadSubModule(pScrn, "vgahw")) return FALSE;
+
+  xf86LoaderReqSymLists(vgahwSymbols, NULL);
+
+  /* Allocate a vgaHWRec */
+  if (!vgaHWGetHWRec(pScrn)) return FALSE;
+
   /* Set pScrn->monitor */
   pScrn->monitor = pScrn->confScreen->monitor;
 
@@ -650,14 +735,6 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
     }
   }
 
-  /* The vgahw module should be loaded here when needed */
-  if (!xf86LoadSubModule(pScrn, "vgahw")) return FALSE;
-
-  xf86LoaderReqSymLists(vgahwSymbols, NULL);
-
-  /* Allocate a vgaHWRec */
-  if (!vgaHWGetHWRec(pScrn)) return FALSE;
-
   /* We use a programamble clock */
   pScrn->progClock = TRUE;
 
@@ -681,7 +758,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
 	       pTDFX->pEnt->device->chipID);
   } else {
     from = X_PROBED;
-    pScrn->chipset = (char *)xf86TokenToString(TDFXChipsets, pTDFX->PciInfo->chipType);
+    pScrn->chipset = (char *)xf86TokenToString(TDFXChipsets, match->chipType);
   }
   if (pTDFX->pEnt->device->chipRev >= 0) {
     xf86DrvMsg(pScrn->scrnIndex, X_CONFIG, "ChipRev override: %d\n",
@@ -691,11 +768,11 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
   xf86DrvMsg(pScrn->scrnIndex, from, "Chipset: \"%s\"\n", pScrn->chipset);
 
   if (pTDFX->pEnt->device->MemBase != 0) {
-    pTDFX->LinearAddr = pTDFX->pEnt->device->MemBase;
+    pTDFX->LinearAddr[0] = pTDFX->pEnt->device->MemBase;
     from = X_CONFIG;
   } else {
-    if (pTDFX->PciInfo->memBase[1] != 0) {
-      pTDFX->LinearAddr = pTDFX->PciInfo->memBase[1];
+    if (match->memBase[1] != 0) {
+      pTDFX->LinearAddr[0] = match->memBase[1];
       from = X_PROBED;
     } else {
       xf86DrvMsg(pScrn->scrnIndex, X_ERROR, 
@@ -708,11 +785,11 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
 	     (unsigned long)pTDFX->LinearAddr);
 
   if (pTDFX->pEnt->device->IOBase != 0) {
-    pTDFX->MMIOAddr = pTDFX->pEnt->device->IOBase;
+    pTDFX->MMIOAddr[0] = pTDFX->pEnt->device->IOBase;
     from = X_CONFIG;
   } else {
-    if (pTDFX->PciInfo->memBase[0]) {
-      pTDFX->MMIOAddr = pTDFX->PciInfo->memBase[0];
+    if (match->memBase[0]) {
+      pTDFX->MMIOAddr[0] = match->memBase[0];
       from = X_PROBED;
     } else {
       xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
@@ -724,8 +801,8 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
   xf86DrvMsg(pScrn->scrnIndex, from, "MMIO registers at addr 0x%lX\n",
 	     (unsigned long)pTDFX->MMIOAddr);
 
-  if (pTDFX->PciInfo->ioBase[2]) {
-    pTDFX->PIOBase = pTDFX->PciInfo->ioBase[2]&0xFFFFFFFC;
+  if (match->ioBase[2]) {
+    pTDFX->PIOBase[0] = match->ioBase[2]&0xFFFFFFFC;
   } else {
     xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 	       "No valid PIO address in PCI config space\n");
@@ -733,7 +810,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
     return FALSE;
   }
   xf86DrvMsg(pScrn->scrnIndex, X_PROBED, "PIO registers at addr 0x%lX\n",
-	     (unsigned long)pTDFX->PIOBase);
+	     (unsigned long)pTDFX->PIOBase[0]);
 
   /* We have to use PIO to probe, because we haven't mappend yet */
   TDFXSetPIOAccess(pTDFX);
@@ -746,6 +823,8 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
     from = X_CONFIG;
   }
 
+  TDFXInitChips(pScrn);
+
   /* Multiple by two because tiled access requires more address space */
   pTDFX->FbMapSize = pScrn->videoRam*1024*2;
   xf86DrvMsg(pScrn->scrnIndex, from, "VideoRAM: %d kByte Mapping %d kByte\n",
@@ -780,12 +859,12 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
       pTDFX->MaxClock = pTDFX->pEnt->device->dacSpeeds[0];
     from = X_CONFIG;
   } else {
-    switch (pTDFX->PciInfo->chipType) {
+    switch (pTDFX->ChipType) {
     case PCI_CHIP_BANSHEE:
       pTDFX->MaxClock = 270000;
       break;
     case PCI_CHIP_VOODOO3:
-      switch(pTDFX->PciInfo->subsysCard) {
+      switch(match->subsysCard) {
       case PCI_SUBDEVICE_ID_VOODOO3_2000:
 	pTDFX->MaxClock = 300000;
 	break;
@@ -797,7 +876,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
 	break;
       }
       break;
-    case PCI_CHIP_VOODOOEXP:
+    case PCI_CHIP_VOODOO5:
       pTDFX->MaxClock = 350000;
       break;
     }
@@ -892,21 +971,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
       xf86SetDDCproperties(pScrn,xf86PrintEDID(pMon));
   }
 
-  
-  /*  We wont be using the VGA access after the probe */
-  if (!xf86ReturnOptValBool(TDFXOptions, OPTION_USE_PIO, FALSE)) {
-    resRange vgaio[] = { {ResShrIoBlock,0x3B0,0x3BB},
-			 {ResShrIoBlock,0x3C0,0x3DF},
-			 _END };
-    resRange vgamem[] = {{ResShrMemBlock,0xA0000,0xAFFFF},
-			 {ResShrMemBlock,0xB8000,0xBFFFF},
-			 {ResShrMemBlock,0xB0000,0xB7FFF},
-			 _END };
-
-    pTDFX->usePIO=FALSE;
-    xf86SetOperatingState(vgaio, pTDFX->pEnt->index, ResUnusedOpr);
-    xf86SetOperatingState(vgamem, pTDFX->pEnt->index, ResDisableOpr);
-  } else {
+  if (xf86ReturnOptValBool(TDFXOptions, OPTION_USE_PIO, FALSE)) {
     pTDFX->usePIO=TRUE;
   }
 
@@ -916,7 +981,7 @@ TDFXPreInit(ScrnInfoPtr pScrn, int flags) {
 static Bool
 TDFXMapMem(ScrnInfoPtr pScrn)
 {
-  int mmioFlags;
+  int mmioFlags, i;
   TDFXPtr pTDFX;
 
   TDFXTRACE("TDFXMapMem start\n");
@@ -924,15 +989,18 @@ TDFXMapMem(ScrnInfoPtr pScrn)
 
   mmioFlags = VIDMEM_MMIO | VIDMEM_READSIDEEFFECT;
 
-  pTDFX->MMIOBase = xf86MapPciMem(pScrn->scrnIndex, mmioFlags, 
-				      pTDFX->PciTag, 
-				      pTDFX->MMIOAddr,
-				      TDFXIOMAPSIZE);
-  if (!pTDFX->MMIOBase) return FALSE;
+  for (i=0; i<pTDFX->numChips; i++) {
+    pTDFX->MMIOBase[i] = xf86MapPciMem(pScrn->scrnIndex, mmioFlags, 
+				       pTDFX->PciTag[i], 
+				       pTDFX->MMIOAddr[i],
+				       TDFXIOMAPSIZE);
+
+    if (!pTDFX->MMIOBase[i]) return FALSE;
+  }
 
   pTDFX->FbBase = xf86MapPciMem(pScrn->scrnIndex, VIDMEM_FRAMEBUFFER,
-				pTDFX->PciTag,
-				pTDFX->LinearAddr,
+				pTDFX->PciTag[0],
+				pTDFX->LinearAddr[0],
 				pTDFX->FbMapSize);
   if (!pTDFX->FbBase) return FALSE;
 
@@ -943,12 +1011,16 @@ static Bool
 TDFXUnmapMem(ScrnInfoPtr pScrn)
 {
   TDFXPtr pTDFX;
+  int i;
 
   TDFXTRACE("TDFXUnmapMem start\n");
   pTDFX = TDFXPTR(pScrn);
 
-  xf86UnMapVidMem(pScrn->scrnIndex, (pointer)pTDFX->MMIOBase, TDFXIOMAPSIZE);
-  pTDFX->MMIOBase=0;
+  for (i=0; i<pTDFX->numChips; i++) {
+    xf86UnMapVidMem(pScrn->scrnIndex, (pointer)pTDFX->MMIOBase[i], 
+		    TDFXIOMAPSIZE);
+    pTDFX->MMIOBase[i]=0;
+  }
 
   xf86UnMapVidMem(pScrn->scrnIndex, (pointer)pTDFX->FbBase, pTDFX->FbMapSize);
   pTDFX->FbBase = 0;
@@ -1220,7 +1292,7 @@ SetupGfxPLL(int freq) {
   TDFXTRACE("SetupGfxPLL start\n");
   pTDFX=TDFXPTR();
   tdfxReg=(vgaTDFXPtr)vgaNewVideoState;
-  if (pTDFX->PciInfo->chipType==PCI_CHIP_BANSHEE)
+  if (pTDFX->chipType==PCI_CHIP_BANSHEE)
     tdfxReg->gfxpll=CalcPLL(freq, &f_out, 1);
   else
     tdfxReg->gfxpll=CalcPLL(freq, &f_out, 0);
@@ -1393,8 +1465,9 @@ TDFXModeInit(ScrnInfoPtr pScrn, DisplayModePtr mode)
 #endif
   DoRestore(pScrn, &hwp->ModeReg, &pTDFX->ModeReg, FALSE);
 #ifdef XF86DRI
-  if (pTDFX->directRenderingEnabled)
+  if (pTDFX->directRenderingEnabled) {
     DRIUnlock(screenInfo.screens[pScrn->scrnIndex]);
+  }
 #endif
 
   return TRUE;
@@ -1476,20 +1549,17 @@ TDFXLoadPalette24(ScrnInfoPtr pScrn, int numColors, int *indices, LOCO *colors,
 #define TILE_HEIGHT 32
 
 static int
-calcBufferStride(int xres, Bool tiled)
+calcBufferStride(int xres, Bool tiled, int cpp)
 {
   int strideInTiles;
 
   if (tiled == TRUE) {
     /* Calculate tile width stuff */
-    strideInTiles = (xres << 1) >> 7;
-    if ((xres << 1) & (TILE_WIDTH - 1))
-      strideInTiles++;
+    strideInTiles = (xres+TILE_WIDTH-1)/TILE_WIDTH;
     
-    return (strideInTiles * TILE_WIDTH);
-
+    return strideInTiles*cpp*TILE_WIDTH;
   } else {
-    return (xres << 1);
+    return xres*cpp;
   }
 } /* calcBufferStride */
 
@@ -1510,26 +1580,26 @@ calcBufferHeightInTiles(int yres)
 } /* calcBufferHeightInTiles */
 
 static int
-calcBufferSizeInTiles(int xres, int yres) {
+calcBufferSizeInTiles(int xres, int yres, int cpp) {
   int bufSizeInTiles;           /* Size of buffer in tiles */
 
   bufSizeInTiles =
-    calcBufferHeightInTiles(yres) * (calcBufferStride(xres, TRUE) >> 7);
+    calcBufferHeightInTiles(yres) * (calcBufferStride(xres, TRUE, cpp) >> 7);
 
   return bufSizeInTiles;
 
 } /* calcBufferSizeInTiles */
 
 static int
-calcBufferSize(int xres, int yres, Bool tiled)
+calcBufferSize(int xres, int yres, Bool tiled, int cpp)
 {
   int stride, height, bufSize;
 
   if (tiled) {
-    stride = calcBufferStride(xres, tiled);
+    stride = calcBufferStride(xres, tiled, cpp);
     height = TILE_HEIGHT * calcBufferHeightInTiles(yres);
   } else {
-    stride = xres << 1;
+    stride = xres*cpp;
     height = yres;
   }
 
@@ -1555,7 +1625,15 @@ static void allocateMemory(ScrnInfoPtr pScrn) {
   /* Remove one scanline for page alignment */
   memRemaining-=4095;
   /* Remove the back and Z buffers */
-  screenSizeInTiles=calcBufferSize(pScrn->virtualX, pScrn->virtualY, TRUE);
+  if (pTDFX->cpp!=3) {
+    screenSizeInTiles=calcBufferSize(pScrn->virtualX, pScrn->virtualY,
+				     TRUE, pTDFX->cpp);
+  }
+  else {
+    /* cpp==3 needs to bump up to 4 */
+    screenSizeInTiles=calcBufferSize(pScrn->virtualX, pScrn->virtualY,
+				     TRUE, 4);
+  }
   memRemaining-=screenSizeInTiles*2;
 
   /* Give all the rest to textures, rounded down to a page */
@@ -1613,18 +1691,22 @@ TDFXScreenInit(int scrnIndex, ScreenPtr pScreen, int argc, char **argv) {
   if (!TDFXMapMem(pScrn)) return FALSE;
   pScrn->memPhysBase = (int)pTDFX->LinearAddr;
 
-  if (!pTDFX->usePIO) {
-    TDFXSetMMIOAccess(pTDFX);
-    hwp->IOBase = ((hwp->readMiscOut(hwp) & 0x01) ? 
-      VGA_IOBASE_COLOR : VGA_IOBASE_MONO) + (unsigned long)pTDFX->MMIOBase -
-      0x300;
-  } else {
-    vgaHWGetIOBase(hwp);
-  }
+  if (!pTDFX->usePIO) TDFXSetMMIOAccess(pTDFX);
+  vgaHWGetIOBase(hwp);
   if (!vgaHWMapMem(pScrn)) return FALSE;
 
   allocateMemory(pScrn);
 
+  if (pTDFX->numChips>1) {
+    if (xf86ReturnOptValBool(TDFXOptions, OPTION_NO_SLI, FALSE)) {
+      TDFXSetupSLI(pScrn, FALSE, 0);
+    } else {
+      TDFXSetupSLI(pScrn, TRUE, 0);
+    }
+  }
+
+  TDFXSetLFBConfig(pTDFX);
+
 #ifdef PROP_3DFX
   if (!TDFXInitPrivate(pScreen)) {
     xf86DrvMsg(pScrn->scrnIndex, X_ERROR, "Failed to initialize private\n");
@@ -1772,7 +1854,6 @@ TDFXScreenInit(int scrnIndex, ScreenPtr pScreen, int argc, char **argv) {
     }
     if (pTDFX->directRenderingEnabled) {
 	xf86DrvMsg(pScrn->scrnIndex, X_INFO, "direct rendering enabled\n");
-	TDFXSetLFBConfig(pTDFX);
     } else {
 	xf86DrvMsg(pScrn->scrnIndex, X_INFO, "direct rendering disabled\n");
     }
@@ -1839,6 +1920,7 @@ TDFXEnterVT(int scrnIndex, int flags) {
   if (pTDFX->directRenderingEnabled) {
     pScreen = screenInfo.screens[scrnIndex];
     DRIUnlock(pScreen);
+    TDFXLostContext(pScreen);
   }
 #endif
   if (!TDFXModeInit(pScrn, pScrn->currentMode)) return FALSE;
@@ -1865,6 +1947,7 @@ TDFXLeaveVT(int scrnIndex, int flags) {
   pTDFX = TDFXPTR(pScrn);
   if (pTDFX->directRenderingEnabled) {
     DRILock(pScreen, 0);
+    TDFXSwapContextPrivate(pScreen);
   }
 #endif
 }
@@ -2011,3 +2094,4 @@ TDFXDisplayPowerManagementSet(ScrnInfoPtr pScrn, int PowerManagementMode,
   pTDFX->writeLong(pTDFX, DACMODE, dacmode);
 }
 #endif
+
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_io.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_io.c
index 62b883a16..73a8b4604 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_io.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_io.c
@@ -39,27 +39,35 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #include "tdfx.h"
 
-#define minb(p) MMIO_IN8(pTDFX->MMIOBase, (p))
-#define moutb(p,v) MMIO_OUT8(pTDFX->MMIOBase, (p),(v))
-#define minl(p) MMIO_IN32(pTDFX->MMIOBase, (p))
-#define moutl(p,v) MMIO_OUT32(pTDFX->MMIOBase, (p),(v))
+#define minb(b, p) MMIO_IN8((b), (p))
+#define moutb(b, p,v) MMIO_OUT8((b), (p),(v))
+#define minl(b, p) MMIO_IN32((b), (p))
+#define moutl(b, p,v) MMIO_OUT32((b), (p),(v))
 
 static void TDFXWriteControlPIO(TDFXPtr pTDFX, int addr, char index, char val) {
-  outb(pTDFX->PIOBase+addr, index);
-  outb(pTDFX->PIOBase+addr+1, val);
+  outb(pTDFX->PIOBase[0]+addr, index);
+  outb(pTDFX->PIOBase[0]+addr+1, val);
 }
 
 static char TDFXReadControlPIO(TDFXPtr pTDFX, int addr, char index) {
-  outb(pTDFX->PIOBase+addr, index);
-  return inb(pTDFX->PIOBase+addr+1);
+  outb(pTDFX->PIOBase[0]+addr, index);
+  return inb(pTDFX->PIOBase[0]+addr+1);
+}
+
+void TDFXWriteChipLongPIO(TDFXPtr pTDFX, int chip, int addr, int val) {
+  outl(pTDFX->PIOBase[chip]+addr, val);
+}
+
+int TDFXReadChipLongPIO(TDFXPtr pTDFX, int chip, int addr) {
+  return inl(pTDFX->PIOBase[chip]+addr);
 }
 
 static void TDFXWriteLongPIO(TDFXPtr pTDFX, int addr, int val) {
-  outl(pTDFX->PIOBase+addr, val);
+  TDFXWriteChipLongPIO(pTDFX, 0, addr, val);
 }
 
 static int TDFXReadLongPIO(TDFXPtr pTDFX, int addr) {
-  return inl(pTDFX->PIOBase+addr);
+  return TDFXReadChipLongPIO(pTDFX, 0, addr);
 }
 
 void TDFXSetPIOAccess(TDFXPtr pTDFX) {
@@ -69,24 +77,34 @@ void TDFXSetPIOAccess(TDFXPtr pTDFX) {
   pTDFX->readControl=TDFXReadControlPIO;
   pTDFX->writeLong=TDFXWriteLongPIO;
   pTDFX->readLong=TDFXReadLongPIO;
+  pTDFX->readChipLong=TDFXReadChipLongPIO;
+  pTDFX->writeChipLong=TDFXWriteChipLongPIO;
 }
 
 static void TDFXWriteControlMMIO(TDFXPtr pTDFX, int addr, char index, char val) {
-  moutb(addr, index);
-  moutb(addr+1, val);
+  moutb(pTDFX->MMIOBase[0], addr, index);
+  moutb(pTDFX->MMIOBase[0], addr+1, val);
 }
 
 static char TDFXReadControlMMIO(TDFXPtr pTDFX, int addr, char index) {
-  moutb(addr, index);
-  return minb(addr+1);
+  moutb(pTDFX->MMIOBase[0], addr, index);
+  return minb(pTDFX->MMIOBase[0], addr+1);
+}
+
+void TDFXWriteChipLongMMIO(TDFXPtr pTDFX, int chip, int addr, int val) {
+  moutl(pTDFX->MMIOBase[chip], addr, val);
+}
+
+static int TDFXReadChipLongMMIO(TDFXPtr pTDFX, int chip, int addr) {
+  return minl(pTDFX->MMIOBase[chip], addr);
 }
 
 void TDFXWriteLongMMIO(TDFXPtr pTDFX, int addr, int val) {
-  moutl(addr, val);
+  TDFXWriteChipLongMMIO(pTDFX, 0, addr, val);
 }
 
 int TDFXReadLongMMIO(TDFXPtr pTDFX, int addr) {
-  return minl(addr);
+  return TDFXReadChipLongMMIO(pTDFX, 0, addr);
 }
 
 void TDFXSetMMIOAccess(TDFXPtr pTDFX) {
@@ -96,5 +114,7 @@ void TDFXSetMMIOAccess(TDFXPtr pTDFX) {
   pTDFX->readControl=TDFXReadControlMMIO;
   pTDFX->writeLong=TDFXWriteLongMMIO;
   pTDFX->readLong=TDFXReadLongMMIO;
+  pTDFX->writeChipLong=TDFXWriteChipLongMMIO;
+  pTDFX->readChipLong=TDFXReadChipLongMMIO;
 }
 
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_priv.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_priv.c
index 7082547df..0d09b79dd 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_priv.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_priv.c
@@ -23,28 +23,41 @@
    C    - D-1  : Z buffer
 */
 
-static void TDFXSendNOPPrivate2D(TDFXPtr pTDFX)
+void TDFXSendNOPPrivate3D(ScrnInfoPtr pScrn)
 {
+  TDFXPtr pTDFX;
+
+  if (!pTDFX->syncDone) {
+    TDFXFirstSync(pScrn);
+    return;
+  }
+  pTDFX=TDFXPTR(pScrn);
   TDFXAllocateSlots(pTDFX, 2);
-  SET_3DPK4_HEADER(1, 0x48);
+  SET_3DPK4_HEADER(1, 0x48<<3);
   WRITE_FIFO(pTDFX, 0, 0);
 }
 
-static void TDFXSendNOPPrivate3D(TDFXPtr pTDFX)
+void TDFXSendNOPPrivate2D(ScrnInfoPtr pScrn)
 {
+  TDFXPtr pTDFX;
+
+  pTDFX=TDFXPTR(pScrn);
   TDFXAllocateSlots(pTDFX, 2);
   SET_PKT2_HEADER(SSTCP_COMMAND);
   WRITE_FIFO(pTDFX, SST_2D_COMMAND, SST_2D_NOP|SST_2D_GO);
 }  
 
-void TDFXSendNOPPrivate(TDFXPtr pTDFX)
+void TDFXSendNOPPrivate(ScrnInfoPtr pScrn)
 {
-  TDFXSendNOPPrivate2D(pTDFX);
-  TDFXSendNOPPrivate3D(pTDFX);
+  TDFXSendNOPPrivate2D(pScrn);
+  TDFXSendNOPPrivate3D(pScrn);
 }
 
-static void InstallFifo(TDFXPtr pTDFX)
+void InstallFifo(ScrnInfoPtr pScrn)
 {
+  TDFXPtr pTDFX;
+
+  pTDFX=TDFXPTR(pScrn);
   /* Install the fifo */
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_BASEADDR0, pTDFX->fifoOffset>>12);
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_BUMP0, 0);
@@ -54,7 +67,7 @@ static void InstallFifo(TDFXPtr pTDFX)
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_AMAX0, pTDFX->fifoOffset-4);
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_DEPTH0, 0);
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_HOLECNT0, 0);
-  if (pTDFX->PciInfo->chipType == PCI_CHIP_BANSHEE) 
+  if (pTDFX->ChipType == PCI_CHIP_BANSHEE) 
     TDFXWriteLongMMIO(pTDFX, SST_FIFO_FIFOTHRESH, (0x9<<5) | 0x2);
   else 
     TDFXWriteLongMMIO(pTDFX, SST_FIFO_FIFOTHRESH, (0xf<<5) | 0x8);
@@ -65,14 +78,16 @@ static void InstallFifo(TDFXPtr pTDFX)
   pTDFX->fifoPtr = pTDFX->fifoBase;
   pTDFX->fifoSlots = (pTDFX->fifoSize>>2) - 1;
   pTDFX->fifoEnd = pTDFX->fifoBase+pTDFX->fifoSlots;
-  TDFXSendNOPPrivate(pTDFX);
+  TDFXSendNOPPrivate(pScrn);
 }
 
-static void TDFXResetFifo(TDFXPtr pTDFX)
+void TDFXResetFifo(ScrnInfoPtr pScrn)
 {
+  TDFXPtr pTDFX;
   int oldValue;
   CARD32 start_sec, end_sec, dummy;
 
+  pTDFX=TDFXPTR(pScrn);
   ErrorF("Resetting FIFO\n");
   /* Shut down the fifo */
   TDFXWriteLongMMIO(pTDFX, SST_FIFO_BASESIZE0, 0);
@@ -91,7 +106,7 @@ static void TDFXResetFifo(TDFXPtr pTDFX)
     xf86getsecs(&end_sec, &dummy);
   } while (end_sec-start_sec<2);
   TDFXWriteLongMMIO(pTDFX, MISCINIT1, oldValue);
-  InstallFifo(pTDFX);
+  InstallFifo(pScrn);
 }
 
 /*
@@ -111,7 +126,7 @@ static void TDFXSyncFifo(ScrnInfoPtr pScrn)
 
   TDFXTRACEACCEL("TDFXSyncFifo start\n");
   pTDFX=TDFXPTR(pScrn);
-  TDFXSendNOPPrivate(pTDFX);
+  TDFXSendNOPPrivate(pScrn);
   i=0;
   cnt=0;
   start_sec=0;
@@ -125,7 +140,7 @@ static void TDFXSyncFifo(ScrnInfoPtr pScrn)
       } else {
 	xf86getsecs(&end_sec, &dummy);
 	if (end_sec-start_sec>3) {
-	  TDFXResetFifo(pTDFX);
+	  TDFXResetFifo(pScrn);
 	  start_sec=0;
         }
       }
@@ -153,7 +168,7 @@ Bool TDFXInitPrivate(ScreenPtr pScreen)
   pTDFX->fifoMirrorPtr = pTDFX->fifoMirrorBase;
 #endif
   pTDFX->sync=TDFXSyncFifo;
-  InstallFifo(pTDFX);
+  InstallFifo(pScrn);
   return TRUE;
 }
 
@@ -213,7 +228,7 @@ void TDFXSwapContextPrivate(ScreenPtr pScreen)
       (sPriv->fifoRead<pTDFX->fifoOffset) ||
       (sPriv->fifoRead>(int)pTDFX->fifoOffset+pTDFX->fifoSize)) {
     ErrorF("Invalid offsets passed between client and X server\n");
-    ResetFifo(pTDFX);
+    ResetFifo(pScrn);
   } else {
     pTDFX->fifoPtr = (unsigned int *)(pTDFX->FbBase+sPriv->fifoPtr);
     pTDFX->fifoRead = (unsigned int *)(pTDFX->FbBase+sPriv->fifoRead);
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_sli.c b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_sli.c
new file mode 100644
index 000000000..fd442c49d
--- /dev/null
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfx_sli.c
@@ -0,0 +1,768 @@
+
+#include "xf86.h"
+#include "xf86_ansic.h"
+#include "xf86_OSproc.h"
+#include "xf86Pci.h"
+#include "tdfx.h"
+
+#define AACLKOUTDEL 0x2
+#define CFGSWAPALGORITHM 0x1
+
+/* #define RD_ABORT_ERROR */
+#define H3VDD
+
+static Bool TDFXDisableSLI(TDFXPtr pTDFX)
+{
+  int i;
+  int v;
+
+  for (i=0; i<pTDFX->numChips; i++) {
+    v=pciReadLong(pTDFX->PciTag[i], CFG_INIT_ENABLE);
+    pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE,
+		 v&~(CFG_SNOOP_MEMBASE0 | CFG_SNOOP_EN | CFG_SNOOP_MEMBASE0_EN |
+		     CFG_SNOOP_MEMBASE1_EN | CFG_SNOOP_SLAVE |
+		     CFG_SNOOP_FBIINIT_WR_EN | CFG_SWAP_ALGORITHM |
+		     CFG_SWAP_QUICK));
+    v=pciReadLong(pTDFX->PciTag[i], CFG_SLI_LFB_CTRL);
+    pciWriteLong(pTDFX->PciTag[i], CFG_SLI_LFB_CTRL, 
+		 v&~(CFG_SLI_LFB_CPU_WR_EN | CFG_SLI_LFB_DPTCH_WR_EN | 
+		     CFG_SLI_RD_EN));
+#ifdef H3VDD
+    pTDFX->writeChipLong(pTDFX, i, SST_3D_SLICTRL, 0);
+    pTDFX->writeChipLong(pTDFX, i, SST_3D_AACTRL, 0);
+#endif
+    v=pciReadLong(pTDFX->PciTag[i], CFG_AA_LFB_CTRL);
+    pciWriteLong(pTDFX->PciTag[i], CFG_AA_LFB_CTRL,
+		 v&~(CFG_AA_LFB_CPU_WR_EN | CFG_AA_LFB_DPTCH_WR_EN |
+		     CFG_AA_LFB_RD_EN));
+    v=pciReadLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC);
+    pciWriteLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC,
+		 (v&~CFG_VGA_VSYNC_OFFSET) |
+		 (0 << CFG_VGA_VSYNC_OFFSET_PIXELS_SHIFT) |
+		 (0 << CFG_VGA_VSYNC_OFFSET_CHARS_SHIFT) |
+		 (0 << CFG_VGA_VSYNC_OFFSET_HXTRA_SHIFT));
+    pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0, 0);
+    pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1, 0);
+    pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2, 0);
+
+    if (pTDFX->numChips>1) {
+      v=pTDFX->readChipLong(pTDFX, i, PCIINIT0);
+      pTDFX->writeChipLong(pTDFX, i, PCIINIT0,
+			   (v&~(SST_PCI_DISABLE_IO|SST_PCI_DISABLE_MEM|
+				SST_PCI_RETRY_INTERVAL)) |
+			   (0<<SST_PCI_RETRY_INTERVAL_SHIFT) |
+			   SST_PCI_FORCE_FB_HIGH);
+    } else {
+      v=pTDFX->readChipLong(pTDFX, i, PCIINIT0);
+      pTDFX->writeChipLong(pTDFX, i, PCIINIT0,
+			   (v&~(SST_PCI_DISABLE_IO|SST_PCI_DISABLE_MEM|
+				SST_PCI_RETRY_INTERVAL)) |
+			   (0<<SST_PCI_RETRY_INTERVAL_SHIFT));
+    }
+
+#if 0
+    if (i>0) {
+      pTDFX->writeChipLong(pTDFX, i, DACMODE, 
+			   SST_DAC_DPMS_ON_VSYNC | SST_DAC_DPMS_ON_HSYNC);
+      v=pTDFX->readChipLong(pTDFX, i, VIDPROCCFG);
+      pTDFX->writeChipLong(pTDFX, i, VIDPROCCFG, v&~SST_VIDEO_PROCESSOR_EN);
+    }
+#endif
+  }
+  return TRUE;
+}
+
+Bool TDFXSetupSLI(ScrnInfoPtr pScrn, Bool sliEnable, int aaSamples)
+{
+  TDFXPtr pTDFX;
+  int i, sliLines, sliLinesLog2, nChipsLog2, v;
+  int sli_renderMask, sli_compareMask, sli_scanMask;
+  int sliAnalog, dwFormat;
+
+  pTDFX=TDFXPTR(pScrn);
+  if (pScrn->depth == 24 || pScrn->depth==32) {
+    if ((aaSamples == 4) && (pTDFX->numChips>1)) {
+      pTDFX->pixelFormat=GR_PIXFMT_AA_4_ARGB_8888;
+    } else if (aaSamples >= 2) {
+      pTDFX->pixelFormat=GR_PIXFMT_AA_2_ARGB_8888;
+    } else {
+      pTDFX->pixelFormat=GR_PIXFMT_ARGB_8888;
+    }
+  } else if (pScrn->depth == 16) {
+    if ((aaSamples == 4) && (pTDFX->numChips>1)) {
+      pTDFX->pixelFormat=GR_PIXFMT_AA_4_RGB_565;
+    } else if (aaSamples >= 2) {
+      pTDFX->pixelFormat=GR_PIXFMT_AA_2_RGB_565;
+    } else {
+      pTDFX->pixelFormat=GR_PIXFMT_RGB_565;
+    }
+  } else if (pScrn->depth == 8) {
+    pTDFX->pixelFormat=GR_PIXFMT_I_8;
+  }
+  if (!sliEnable && !aaSamples) { /* Turn off */
+    return TDFXDisableSLI(pTDFX);
+  }
+
+  if (pScrn->virtualY>768) sliLinesLog2=5;
+  else sliLinesLog2=4;
+  sliLines=1<<sliLinesLog2;
+  if (pScrn->virtualY*pScrn->virtualX>1600*1024) sliAnalog=1;
+  else sliAnalog=0;
+  /* XXX We need to avoid SLI in double scan modes somehow */
+
+  switch (pTDFX->numChips) {
+  case 1:
+    nChipsLog2=0;
+    break;
+  case 2:
+    nChipsLog2=1;
+    break;
+  case 4:
+    nChipsLog2=2;
+    break;
+  default:
+    return FALSE;
+    /* XXX Huh? Unsupported configuration */
+  }
+
+  for (i=0; i<pTDFX->numChips; i++) {
+    /* Do we want to set these differently for a VIA board? */
+    v=pTDFX->readChipLong(pTDFX, i, PCIINIT0);
+    v=(v&~(SST_PCI_RETRY_INTERVAL|SST_PCI_FORCE_FB_HIGH)) |
+		     SST_PCI_READ_WS | SST_PCI_WRITE_WS |
+		     SST_PCI_DISABLE_IO | SST_PCI_DISABLE_MEM |
+		     (5<<SST_PCI_RETRY_INTERVAL_SHIFT);
+    pTDFX->writeChipLong(pTDFX, i, PCIINIT0,
+			 (v&~(SST_PCI_RETRY_INTERVAL|SST_PCI_FORCE_FB_HIGH)) |
+			 SST_PCI_READ_WS | SST_PCI_WRITE_WS |
+			 SST_PCI_DISABLE_IO | SST_PCI_DISABLE_MEM |
+			 (5<<SST_PCI_RETRY_INTERVAL_SHIFT));
+    v=pTDFX->readChipLong(pTDFX, i, TMUGBEINIT);
+    pTDFX->writeChipLong(pTDFX, i, TMUGBEINIT, 
+			 (v&~(SST_AA_CLK_DELAY | SST_AA_CLK_INVERT)) |
+			 (AACLKOUTDEL<<SST_AA_CLK_DELAY_SHIFT) | 
+			 SST_AA_CLK_INVERT);
+
+    if (pTDFX->numChips>1) {
+      v=pciReadLong(pTDFX->PciTag[i], CFG_INIT_ENABLE);
+      pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE, v |
+		   (CFGSWAPALGORITHM << CFG_SWAPBUFFER_ALGORITHM_SHIFT) |
+		   CFG_SWAP_ALGORITHM | ((!i)? CFG_SWAP_MASTER : 0));
+      if (!i) {
+	v=pciReadLong(pTDFX->PciTag[i], CFG_INIT_ENABLE);
+	pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE, 
+			 v | CFG_SNOOP_EN);
+	v=pciReadLong(pTDFX->PciTag[i], CFG_PCI_DECODE);
+      } else {
+	v=pciReadLong(pTDFX->PciTag[i], CFG_INIT_ENABLE);
+	v=(v & ~CFG_SNOOP_MEMBASE0) | CFG_SNOOP_EN | 
+	  CFG_SNOOP_MEMBASE0_EN | CFG_SNOOP_MEMBASE1_EN |
+	  CFG_SNOOP_SLAVE | CFG_SNOOP_FBIINIT_WR_EN |
+	  (((pTDFX->MMIOAddr[0]>>22)&0x3ff)<<CFG_SNOOP_MEMBASE0_SHIFT) |
+	  ((pTDFX->numChips>2)? CFG_SWAP_QUICK : 0);
+	pciWriteLong(pTDFX->PciTag[i], CFG_INIT_ENABLE, v);
+	v=pciReadLong(pTDFX->PciTag[i], CFG_PCI_DECODE);
+	v=(v & ~CFG_SNOOP_MEMBASE1) |
+	  ((pTDFX->LinearAddr[0]>>22)&0x3ff)<<CFG_SNOOP_MEMBASE1_SHIFT;
+	pciWriteLong(pTDFX->PciTag[i], CFG_PCI_DECODE, v);
+      }
+    }
+
+    if (sliEnable && aaSamples<4) {
+      /* SLI is on and we're using less than 4 AA samples */
+      sli_renderMask = (pTDFX->numChips-1) << sliLinesLog2;
+      sli_compareMask = i << sliLinesLog2;
+      sli_scanMask = sliLines - 1;
+      v = (sli_renderMask << CFG_SLI_LFB_RENDERMASK_SHIFT) |
+	(sli_compareMask << CFG_SLI_LFB_COMPAREMASK_SHIFT) |
+	(sli_scanMask << CFG_SLI_LFB_SCANMASK_SHIFT) |
+	(nChipsLog2 << CFG_SLI_LFB_NUMCHIPS_LOG2_SHIFT) |
+	CFG_SLI_LFB_CPU_WR_EN | CFG_SLI_LFB_DPTCH_WR_EN;
+#ifndef RD_ABORT_ERROR
+      v|=CFG_SLI_RD_EN;
+#endif
+      pciWriteLong(pTDFX->PciTag[i], CFG_SLI_LFB_CTRL, v);
+
+#ifdef H3VDD
+      pTDFX->writeChipLong(pTDFX, i, SST_3D_SLICTRL,
+			   (sli_renderMask << SLICTL_3D_RENDERMASK_SHIFT) |
+			   (sli_compareMask << SLICTL_3D_COMPAREMASK_SHIFT) |
+			   (sli_scanMask << SLICTL_3D_SCANMASK_SHIFT) |
+			   (nChipsLog2 << SLICTL_3D_NUMCHIPS_LOG2_SHIFT) |
+			   SLICTL_3D_EN);
+#endif
+    } else if (!sliEnable && aaSamples) {
+      /* SLI is off and AA is on */
+      sli_renderMask = 0;
+      sli_compareMask = 0;
+      sli_scanMask = 0;
+      pciWriteLong(pTDFX->PciTag[i], CFG_SLI_LFB_CTRL,
+		   (sli_renderMask << CFG_SLI_LFB_RENDERMASK_SHIFT) |
+		   (sli_compareMask << CFG_SLI_LFB_COMPAREMASK_SHIFT) |
+		   (sli_scanMask << CFG_SLI_LFB_SCANMASK_SHIFT) |
+		   (0x0 << CFG_SLI_LFB_NUMCHIPS_LOG2_SHIFT));
+#ifdef H3VDD
+      pTDFX->writeChipLong(pTDFX, i, SST_3D_SLICTRL,
+			   (sli_renderMask << SLICTL_3D_RENDERMASK_SHIFT) |
+			   (sli_compareMask << SLICTL_3D_COMPAREMASK_SHIFT) |
+			   (sli_scanMask << SLICTL_3D_SCANMASK_SHIFT) |
+			   (0 << SLICTL_3D_NUMCHIPS_LOG2_SHIFT));
+#endif
+    } else {
+      /* SLI is on && aaSamples=4 */
+      sli_renderMask = ((pTDFX->numChips>>1)-1) << sliLinesLog2;
+      sli_compareMask = (i>>1) << sliLinesLog2;
+      sli_scanMask = sliLines - 1;
+      v = (sli_renderMask << CFG_SLI_LFB_RENDERMASK_SHIFT) |
+	(sli_compareMask << CFG_SLI_LFB_COMPAREMASK_SHIFT) |
+	(sli_scanMask << CFG_SLI_LFB_SCANMASK_SHIFT) |
+	((nChipsLog2-1) << CFG_SLI_LFB_NUMCHIPS_LOG2_SHIFT) |
+	CFG_SLI_LFB_CPU_WR_EN | CFG_SLI_LFB_DPTCH_WR_EN;
+#ifndef RD_ABORT_ERROR
+      v|=CFG_SLI_RD_EN;
+#endif
+      pciWriteLong(pTDFX->PciTag[i], CFG_SLI_LFB_CTRL, v);
+#ifdef H3VDD
+      pTDFX->writeChipLong(pTDFX, i, SST_3D_SLICTRL,
+			   (sli_renderMask << SLICTL_3D_RENDERMASK_SHIFT) |
+			   (sli_compareMask << SLICTL_3D_COMPAREMASK_SHIFT) |
+			   (sli_scanMask << SLICTL_3D_SCANMASK_SHIFT) |
+			   ((nChipsLog2-1) << SLICTL_3D_NUMCHIPS_LOG2_SHIFT) |
+			   SLICTL_3D_EN);
+#endif
+    }
+
+    TDFXSetLFBConfig(pTDFX);
+    if (pTDFX->cpp==2) dwFormat = CFG_AA_LFB_RD_FORMAT_16BPP;
+    else dwFormat = CFG_AA_LFB_RD_FORMAT_32BPP;
+    if (pTDFX->numChips==2 && !sliEnable && aaSamples==2)
+      dwFormat|=CFG_AA_LFB_RD_DIVIDE_BY_4;
+    /* Thess are wrong, because we don't know where the secondary buffers
+       are located */
+    pTDFX->writeChipLong(pTDFX, i, CFG_AA_LFB_CTRL, 
+			 (pScrn->videoRam<<10 /* 2nd buf */ << CFG_AA_BASEADDR_SHIFT) |
+			 CFG_AA_LFB_CPU_WR_EN | CFG_AA_LFB_DPTCH_WR_EN |
+			 CFG_AA_LFB_RD_EN | dwFormat |
+			 ((aaSamples==4)?CFG_AA_LFB_RD_DIVIDE_BY_4:0));
+    pTDFX->writeChipLong(pTDFX, i, CFG_AA_ZBUFF_APERTURE,
+			 ((pTDFX->depthOffset>>12)<<CFG_AA_DEPTH_BUFFER_BEG_SHIFT) |
+			 ((pScrn->videoRam>>2)<<CFG_AA_DEPTH_BUFFER_END_SHIFT));
+
+    if (pTDFX->numChips>1 && i && (aaSamples || sliEnable)) {
+      int vsyncOffsetPixels, vsyncOffsetChars, vsyncOffsetHXtra;
+
+      if (aaSamples || (pTDFX->numChips==4 && sliEnable && aaSamples==4 &&
+			sliAnalog && i==3)) {
+	vsyncOffsetPixels=7;
+	vsyncOffsetChars=4;
+	vsyncOffsetHXtra=0;
+      } else {
+	vsyncOffsetPixels=7;
+	vsyncOffsetChars=5;
+	vsyncOffsetHXtra=0;
+      }
+      v=pciReadLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC);
+      pciWriteLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC, 
+		   (v&~CFG_VGA_VSYNC_OFFSET) |
+		   (vsyncOffsetPixels << CFG_VGA_VSYNC_OFFSET_PIXELS_SHIFT) |
+		   (vsyncOffsetChars << CFG_VGA_VSYNC_OFFSET_CHARS_SHIFT) |
+		   (vsyncOffsetHXtra <<
+		    CFG_VGA_VSYNC_OFFSET_HXTRA_SHIFT));
+    }
+    if (pTDFX->numChips==1 && aaSamples) {
+      /* 1 chip 2 AA */
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		   CFG_ENHANCED_VIDEO_EN | 
+		   CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		   CFG_VIDEO_OTHERMUX_SEL_PIPE<<CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT |
+		   CFG_DIVIDE_VIDEO_BY_2);
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		   0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT |
+		   0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT |
+		   0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT |
+		   0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT);
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		   0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT |
+		   0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT);
+    } else if (pTDFX->numChips==2 && !sliEnable && aaSamples==4 &&
+	       !sliAnalog) {
+      /* 2 chips 4 digital AA */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN | 
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE_PLUS_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     (CFG_ENHANCED_VIDEO_EN |
+		      CFG_ENHANCED_VIDEO_SLV |
+		      CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		      (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		       CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		      CFG_DIVIDE_VIDEO_BY_1));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==2 && !sliEnable && aaSamples==4 && sliAnalog) {
+      /* 2 chips 4 analog AA */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     (CFG_ENHANCED_VIDEO_EN |
+		      CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		      (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		       CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		      CFG_DIVIDE_VIDEO_BY_4));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_DAC_HSYNC_TRISTATE |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==2 && sliEnable && !aaSamples && !sliAnalog) {
+      /* 2 chips 2 digital SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     (CFG_ENHANCED_VIDEO_EN |
+		      (CFG_VIDEO_OTHERMUX_SEL_AAFIFO <<
+		       CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		      (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		       CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		      CFG_DIVIDE_VIDEO_BY_1));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     ((0x0<<sliLinesLog2) << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((i<<sliLinesLog2) <<
+		      CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips>=2 && sliEnable && !aaSamples && sliAnalog) {
+      /* 2 or 4 chips 2/4 analog SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (((pTDFX->numChips-1)<<sliLinesLog2) << 
+		      CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (((pTDFX->numChips-1)<<sliLinesLog2) << 
+		      CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((i<<sliLinesLog2) <<
+		      CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==2 && sliEnable && aaSamples==2 && !sliAnalog) {
+      /* 2 chips 2 AA 2 digital SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((0x1<<sliLinesLog2) <<
+		      CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((i<<sliLinesLog2) <<
+		      CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==2 && !sliEnable && aaSamples==2 && !sliAnalog) {
+      /* 2 chips 2 digital AA */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE_PLUS_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==2 && !sliEnable && aaSamples==2 && sliAnalog) {
+      /* 2 chips 2 analog AA */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_DAC_HSYNC_TRISTATE |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+      }
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		   (0x0 << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		   (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		   (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		   (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		   (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		   (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+    } else if (pTDFX->numChips>=2 && sliEnable && aaSamples==2 && sliAnalog) {
+      /* 2 or 4 chips 2 AA 2 or 4 analog SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (((pTDFX->numChips-1)<<sliLinesLog2) << 
+		      CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_2);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (((pTDFX->numChips-1)<<sliLinesLog2) << 
+		      CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==4 && sliEnable && !aaSamples && !sliAnalog) {
+      /* 4 chips 4 digital SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     (CFG_VIDEO_OTHERMUX_SEL_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_SLI_AAFIFO_COMPARE_INV |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((0x0<<sliLinesLog2) <<
+		      CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==4 && sliEnable && aaSamples==2 && !sliAnalog) {
+      /* 4 chips 2 AA 4 digital SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_SLI_AAFIFO_COMPARE_INV |
+		     CFG_DIVIDE_VIDEO_BY_2);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     (0x0 << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (((pTDFX->numChips-1)<<sliLinesLog2) <<
+		      CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     ((i<<sliLinesLog2) << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==4 && sliEnable && aaSamples==4 && !sliAnalog) {
+      /* 4 chips 4 AA 2 digital SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE_PLUS_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else if (i==1 || i==3) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_DAC_HSYNC_TRISTATE |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_1);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((((i+1)>>2)<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((0xff<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE_PLUS_AAFIFO <<
+		      CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0xff << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    } else if (pTDFX->numChips==4 && sliEnable && aaSamples==4 && sliAnalog) {
+      /* 4 chips 4 AA 2 analog SLI */
+      if (!i) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((0x0<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else if (i==1 || i==3) {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_DAC_HSYNC_TRISTATE |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((((i+1)>>2)<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((((i+1)>>2)<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      } else {
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		     CFG_ENHANCED_VIDEO_EN |
+		     CFG_ENHANCED_VIDEO_SLV |
+		     CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY |
+		     (CFG_VIDEO_OTHERMUX_SEL_PIPE <<
+		      CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT) |
+		     CFG_DIVIDE_VIDEO_BY_4);
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL1,
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_COMPAREMASK_FETCH_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_RENDERMASK_CRT_SHIFT) |
+		     ((0x1<<sliLinesLog2) << CFG_SLI_COMPAREMASK_CRT_SHIFT));
+	pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL2,
+		     (0x0 << CFG_SLI_RENDERMASK_AAFIFO_SHIFT) |
+		     (0x0 << CFG_SLI_COMPAREMASK_AAFIFO_SHIFT));
+      }
+    }
+    if (pTDFX->numChips==4 && sliEnable && aaSamples==4 && i==3) {
+      v=pciReadLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC);
+      pciWriteLong(pTDFX->PciTag[i], CFG_SLI_AA_MISC, 
+		   v | CFG_AA_LFB_RD_SLV_WAIT);
+    }
+    if (i) {
+      v=pciReadLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0);
+      pciWriteLong(pTDFX->PciTag[i], CFG_VIDEO_CTRL0,
+		   v|CFG_VIDPLL_SEL);
+      v=pTDFX->readChipLong(pTDFX, i, MISCINIT1);
+      pTDFX->writeChipLong(pTDFX, i, MISCINIT1, v|SST_POWERDOWN_DAC);
+    }
+  }
+  return TRUE;
+}
diff --git a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfxdefs.h b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfxdefs.h
index bb82da4cd..c6224d78f 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfxdefs.h
+++ b/xc/programs/Xserver/hw/xfree86/drivers/tdfx/tdfxdefs.h
@@ -176,7 +176,184 @@
 #define SST_3D_NOP 0
 
 /* 3D Registers */
-#define SST_3D_OFFSET           0x200000
-#define SST_3D_COMMAND SST_3D_OFFSET+0x120
+#define SST_3D_OFFSET           	0x200000
+#define SST_3D_COMMAND 			SST_3D_OFFSET+0x120
+#define SST_3D_SLICTRL 			SST_3D_OFFSET+0x20C
+#define SST_3D_AACTRL			SST_3D_OFFSET+0x210
+#define SST_3D_CHIPMASK			SST_3D_OFFSET+0x214
+
+/* NAPALM REGISTERS */
+#define CFG_MEM0BASE			16
+#define CFG_MEM1BASE			20
+#define CFG_INIT_ENABLE        		64
+#define CFG_PCI_DECODE         		72
+#define CFG_VIDEO_CTRL0        		128
+#define CFG_VIDEO_CTRL1        		132
+#define CFG_VIDEO_CTRL2        		136
+#define CFG_SLI_LFB_CTRL       		140
+#define CFG_AA_ZBUFF_APERTURE  		144
+#define CFG_AA_LFB_CTRL        		148
+#define CFG_SLI_AA_MISC        		172
+
+/* Pixel Formats */
+#define GR_PIXFMT_I_8                   0x0001
+#define GR_PIXFMT_AI_88                 0x0002
+#define GR_PIXFMT_RGB_565               0x0003
+#define GR_PIXFMT_ARGB_1555             0x0004
+#define GR_PIXFMT_ARGB_8888             0x0005
+#define GR_PIXFMT_AA_2_RGB_565          0x0006
+#define GR_PIXFMT_AA_2_ARGB_1555        0x0007
+#define GR_PIXFMT_AA_2_ARGB_8888        0x0008
+#define GR_PIXFMT_AA_4_RGB_565          0x0009
+#define GR_PIXFMT_AA_4_ARGB_1555        0x000a
+#define GR_PIXFMT_AA_4_ARGB_8888        0x000b
+
+/* pciInit0 */
+#define SST_PCI_STALL_ENABLE            BIT(0)
+#define SST_PCI_LOWTHRESH_SHIFT         2
+#define SST_PCI_LOWTHRESH               (0xF << SST_PCI_LOWTHRESH_SHIFT)
+#define SST_PCI_HARDCODE_BASE           BIT(7)
+#define SST_PCI_READ_WS                 BIT(8)
+#define SST_PCI_WRITE_WS                BIT(9)
+#define SST_PCI_DISABLE_IO              BIT(11)
+#define SST_PCI_DISABLE_MEM             BIT(12)
+#define SST_PCI_RETRY_INTERVAL_SHIFT    13
+#define SST_PCI_RETRY_INTERVAL          (0x1F << SST_PCI_RETRY_INTERVAL_SHIFT)
+#define SST_PCI_INTERRUPT_ENABLE        BIT(18)
+#define SST_PCI_TIMEOUT_ENABLE          BIT(19)
+#define SST_PCI_FORCE_FB_HIGH           BIT(26)
+
+#define SST_AA_CLK_INVERT               BIT(20)
+#define SST_AA_CLK_DELAY_SHIFT          21
+#define SST_AA_CLK_DELAY                (0xF<<SST_AA_CLK_DELAY_SHIFT)
+
+#define CFG_SWAP_ALGORITHM_VSYNC	(0x00)
+#define CFG_SWAP_ALGORITHM_SYNCIN	(0x01)
+#define CFG_SWAPBUFFER_ALGORITHM_SHIFT	(25)
+
+/* CFG_INIT_ENABLE */
+#define CFG_UPDATE_MEMBASE_LSBS		BIT(10)
+#define CFG_SNOOP_EN			BIT(11)
+#define CFG_SNOOP_MEMBASE0_EN		BIT(12)
+#define CFG_SNOOP_MEMBASE1_EN		BIT(13)
+#define CFG_SNOOP_SLAVE			BIT(14)
+#define CFG_SNOOP_MEMBASE0_SHIFT	15
+#define CFG_SNOOP_MEMBASE0		(0x3FF<<CFG_SNOOP_MEMBASE0_SHIFT)
+#define CFG_SWAP_ALGORITHM		BIT(25)
+#define CFG_SWAP_MASTER			BIT(26)
+#define CFG_SWAP_QUICK			BIT(27)
+#define CFG_MULTI_FUNCTION_DEV		BIT(28)
+#define CFG_LFB_RD_CACHE_DISABLE	BIT(29)
+#define CFG_SNOOP_FBIINIT_WR_EN		BIT(30)
+#define CFG_SNOOP_MEMBASE0_DECODE_SHIFT	10
+#define CFG_SNOOP_MEMBASE0_DECODE	(0xF<<CFG_SNOOP_MEMBASE0_DECODE_SHIFT)
+#define CFG_SNOOP_MEMBASE1_DECODE_SHIFT	14
+#define CFG_SNOOP_MEMBASE1_DECODE	(0xF<<CFG_SNOOP_MEMBASE1_DECODE_SHIFT)
+#define CFG_SNOOP_MEMBASE1_SHIFT	18
+#define CFG_SNOOP_MEMBASE1		(0x3FF<<CFG_SNOOP_MEMBASE1_SHIFT)
+
+/* CFG_VIDEO_CTRL0 */
+#define CFG_ENHANCED_VIDEO_EN		BIT(0)
+#define CFG_ENHANCED_VIDEO_SLV		BIT(1)
+#define CFG_VIDEO_TV_OUTPUT_EN		BIT(2)
+#define CFG_VIDEO_LOCALMUX_SEL		BIT(3)
+#define CFG_VIDEO_LOCALMUX_DESKTOP_PLUS_OVERLAY	BIT(3)
+#define CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT 	4
+#define CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT	6
+#define CFG_VIDEO_OTHERMUX_SEL_TRUE	(0x3<<CFG_VIDEO_OTHERMUX_SEL_TRUE_SHIFT)
+#define CFG_VIDEO_OTHERMUX_SEL_FALSE	(0x3<<CFG_VIDEO_OTHERMUX_SEL_FALSE_SHIFT)
+#define CFG_VIDEO_OTHERMUX_SEL_PIPE	0
+#define CFG_VIDEO_OTHERMUX_SEL_PIPE_PLUS_AAFIFO	1
+#define CFG_VIDEO_OTHERMUX_SEL_AAFIFO 	2
+#define CFG_SLI_FETCH_COMPARE_INV	BIT(8)
+#define CFG_SLI_CRT_COMPARE_INV		BIT(9)
+#define CFG_SLI_AAFIFO_COMPARE_INV	BIT(10)
+#define CFG_VIDPLL_SEL                 	BIT(11)
+#define CFG_DIVIDE_VIDEO_SHIFT		12
+#define CFG_DIVIDE_VIDEO		(0x7<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_1		(0x0<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_2		(0x1<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_4		(0x2<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_8		(0x3<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_16		(0x4<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_DIVIDE_VIDEO_BY_32		(0x5<<CFG_DIVIDE_VIDEO_SHIFT)
+#define CFG_ALWAYS_DRIVE_AA_BUS		BIT(15)
+#define CFG_VSYNC_IN_DEL_SHIFT		16
+#define CFG_VSYNC_IN_DEL		(0xF<<CFG_VSYNC_IN_DEL_SHIFT)
+#define CFG_DAC_VSYNC_TRISTATE		BIT(24)
+#define CFG_DAC_HSYNC_TRISTATE		BIT(25)
+
+/* CFG_VIDEO_CTRL1 */
+#define CFG_SLI_RENDERMASK_FETCH_SHIFT	0
+#define CFG_SLI_RENDERMASK_FETCH	(0xFF<<CFG_SLI_RENDERMASK_FETCH_SHIFT)
+#define CFG_SLI_COMPAREMASK_FETCH_SHIFT 8
+#define CFG_SLI_COMPAREMASK_FETCH	(0xFF<<CFG_SLI_COMPAREMASK_FETCH_SHIFT)
+#define CFG_SLI_RENDERMASK_CRT_SHIFT	16
+#define CFG_SLI_RENDERMASK_CRT		(0xFF<<CFG_SLI_RENDERMASK_CRT_SHIFT)
+#define CFG_SLI_COMPAREMASK_CRT_SHIFT	24 
+#define CFG_SLI_COMPAREMASK_CRT		(0xFF<<CFG_SLI_COMPAREMASK_CRT_SHIFT)
+
+/* CFG_VIDEO_CTRL2 */
+#define CFG_SLI_RENDERMASK_AAFIFO_SHIFT 0
+#define CFG_SLI_RENDERMASK_AAFIFO	(0xFF<<CFG_SLI_RENDERMASK_AAFIFO_SHIFT)
+#define CFG_SLI_COMPAREMASK_AAFIFO_SHIFT 8
+#define CFG_SLI_COMPAREMASK_AAFIFO	(0xFF<<CFG_SLI_COMPAREMASK_AAFIFO_SHIFT)
+
+/* CFG_SLI_LFB_CTRL */
+#define CFG_SLI_LFB_RENDERMASK_SHIFT	0
+#define CFG_SLI_LFB_RENDERMASK		(0xFF<<CFG_SLI_LFB_RENDERMASK_SHIFT)
+#define CFG_SLI_LFB_COMPAREMASK_SHIFT	8
+#define CFG_SLI_LFB_COMPAREMASK		(0xFF<<CFG_SLI_LFB_COMPAREMASK_SHIFT)
+#define CFG_SLI_LFB_SCANMASK_SHIFT	16	
+#define CFG_SLI_LFB_SCANMASK		(0xFF<<CFG_SLI_LFB_SCANMASK_SHIFT)
+#define CFG_SLI_LFB_NUMCHIPS_LOG2_SHIFT 24
+#define CFG_SLI_LFB_NUMCHIPS_LOG2	(0x3<<CFG_SLI_LFB_NUMCHIPS_LOG2_SHIFT)
+#define CFG_SLI_LFB_CPU_WR_EN		BIT(26)
+#define CFG_SLI_LFB_DPTCH_WR_EN		BIT(27)
+#define CFG_SLI_RD_EN			BIT(28)
+
+/* CFG_AA_ZBUFF_APERTURE */
+#define CFG_AA_DEPTH_BUFFER_BEG_SHIFT	0
+#define CFG_AA_DEPTH_BUFFER_BEG		(0x7FFF<<CFG_AA_DEPTH_BUFFER_BEG_SHIFT)
+#define CFG_AA_DEPTH_BUFFER_END_SHIFT	16
+#define CFG_AA_DEPTH_BUFFER_END		(0xFFFF<<CFG_AA_DEPTH_BUFFER_END_SHIFT)
+
+/* CFG_AA_LFB_CTRL */
+#define CFG_AA_BASEADDR_SHIFT		0
+#define CFG_AA_BASEADDR			(0x3FFFFFF<<CFG_AA_BASEADDR_SHIFT)
+#define CFG_AA_LFB_CPU_WR_EN		BIT(26)
+#define CFG_AA_LFB_DPTCH_WR_EN		BIT(27)
+#define CFG_AA_LFB_RD_EN		BIT(28)
+#define CFG_AA_LFB_RD_FORMAT_SHIFT	29
+#define CFG_AA_LFB_RD_FORMAT		(0x3<<CFG_AA_LFB_RD_FORMAT_SHIFT)
+#define CFG_AA_LFB_RD_FORMAT_16BPP	(0x0<<CFG_AA_LFB_RD_FORMAT_SHIFT)
+#define CFG_AA_LFB_RD_FORMAT_15BPP	(0x1<<CFG_AA_LFB_RD_FORMAT_SHIFT)
+#define CFG_AA_LFB_RD_FORMAT_32BPP	(0x2<<CFG_AA_LFB_RD_FORMAT_SHIFT)
+#define CFG_AA_LFB_RD_DIVIDE_BY_4	BIT(31)
+
+/* CFG_SLI_AA_MISC */
+#define CFG_VGA_VSYNC_OFFSET_SHIFT	0
+#define CFG_VGA_VSYNC_OFFSET		(0x1ff<<CFG_VGA_VSYNC_OFFSET_SHIFT)
+#define CFG_VGA_VSYNC_OFFSET_PIXELS_SHIFT	0
+#define CFG_VGA_VSYNC_OFFSET_CHARS_SHIFT	3
+#define CFG_VGA_VSYNC_OFFSET_HXTRA_SHIFT	6 
+#define CFG_HOTPLUG_SHIFT		9
+#define CFG_HOTPLUG_TRISTATE		(0x0<<CFG_HOTPLUG_SHIFT)
+#define CFG_HOTPLUG_DRIVE0		(0x2<<CFG_HOTPLUG_SHIFT)
+#define CFG_HOTPLUG_DRIVE1		(0x3<<CFG_HOTPLUG_SHIFT)
+#define CFG_AA_LFB_RD_SLV_WAIT		BIT(12)
+
+/* SLICTL_3D_CTRL */
+#define SLICTL_3D_RENDERMASK_SHIFT	0
+#define SLICTL_3D_RENDERMASK		(0xFF<<SLICTL_3D_RENDERMASK_SHIFT)
+#define SLICTL_3D_COMPAREMASK_SHIFT	8
+#define SLICTL_3D_COMPAREMASK		(0xFF<<SLICTL_3D_COMPAREMASK_SHIFT)
+#define SLICTL_3D_SCANMASK_SHIFT	16	
+#define SLICTL_3D_SCANMASK		(0xFF<<SLICTL_3D_SCANMASK_SHIFT)
+#define SLICTL_3D_NUMCHIPS_LOG2_SHIFT 	24
+#define SLICTL_3D_NUMCHIPS_LOG2		(0x3<<SLICTL_3D_NUMCHIPS_LOG2_SHIFT)
+#define SLICTL_3D_EN			BIT(26)
+
+#define SST_POWERDOWN_DAC               BIT(8)
 
 #endif
diff --git a/xc/programs/Xserver/hw/xfree86/os-support/linux/drm/kernel/mga_state.c b/xc/programs/Xserver/hw/xfree86/os-support/linux/drm/kernel/mga_state.c
index 5a3da6743..c9c6ef3df 100644
--- a/xc/programs/Xserver/hw/xfree86/os-support/linux/drm/kernel/mga_state.c
+++ b/xc/programs/Xserver/hw/xfree86/os-support/linux/drm/kernel/mga_state.c
@@ -338,7 +338,7 @@ static void mgaG200EmitPipe(drm_mga_private_t * dev_priv)
 		   (u32) (dev_priv->WarpIndex[pipe].
 			  phys_addr | WIA_wmode_start | WIA_wagp_agp));
 
-	PRIMADVANCE(dev_priv);
+	PRIMADVANCE( dev_priv );
 }
 
 static void mgaEmitState(drm_mga_private_t * dev_priv)