Bug 1106

The radeon driver's RENDER acceleration code allocates offscreen memory using
xf86AllocateOffscreenLinear(). However, given the fact that the "size" argument
to that call is supposed to be in units of "pixels" (and thus depending on the
current screen's depth), it does this the wrong way:

Currently, this looks like this:

  tex_bytepp = PICT_FORMAT_BPP(format) >> 3;
  [...]
  dst_pitch = (width * tex_bytepp + 31) & ~31;
  size = dst_pitch * height;

  AllocateLinear(size);  

This is wrong because the result is the size of the texture in bytes, not screen
pixels. The "size" has to be calculated like this (where as bpp =
pScrn->bitsPerPixel >> 3):

  size = ((dst_pitch + bpp - 1) / bpp) * height;