[XviD-devel] new GMC, GME refinement

[skal]

	Hi All,

On Sat, 2003-04-12 at 10:18, Christoph Lampert wrote:
> 
> Hi,
> 
> I just added Skal's GMC code to my local source tree (I don't have CVS
> access from here). It's nice an clean, thank you, Skal. 
> Btw. @skal: there was a tiny bug in the code you sent (which took me two
> days to find): If GMC points to an address outside the VOP (left or top),
> you set the offset to 0. That should be -1 or -stride, because everything
> outside must be the edge pixel, including the value at "Offset+1" 
> This is not visible when decoding with the same code, but ffmpeg showed
> smearing errors. 

	oops, sorry about that bug :(

	Actually, I had this pitfall in mind when I wrote the function
	but the code is buggy: in fact the residual multipliers
	'ri' and 'rj' are initialized by default to '16', whereas
	it should be '16<<16', so that the leftmost sample (offset+0)
	is weighted full (16) and the rightmost one (offset+1)
	 unweighted. I don't know how you corrected it exactly but the
	fix should be 'ri=16<<16; rj=16<16' at lines 128,132,194 and 203
	of the original code.

	Now, thinking of it there's another bug in the 1warp-point
	clipping interpolation. I forgot the case where the
	block stands across the edge boundary, but not fully
	inside the (replicated edges). The code for setting up
	the clipping should be:

  if ((uint32_t)vo<=(uint32_t)H) Offset  = (vo>>4)*stride;
  else if (vo>H)                 Offset  = ( H>>4)*stride;
  else if (vo<=-16)    		 Offset  =-16*stride;
  else                           Offset  = (vo>>4)*stride;
  if ((uint32_t)uo<=(uint32_t)W) Offset += (uo>>4);
  else if (uo>W)                 Offset += ( W>>4);
  else if (uo<=-16)        	 Offset -= 16;
  else                           Offset += (uo>>4);

	for the 16x16 luma pred, and:

  int32_t Offset;
  if ((uint32_t)vo<=(uint32_t)H) Offset  = (vo>>4)*stride;
  else if (vo>H)                 Offset  = ( H>>4)*stride;
  else if (vo<=-8)               Offset  =-8*stride; 
  else                           Offset  = (vo>>4)*stride;
  if ((uint32_t)uo<=(uint32_t)W) Offset += (uo>>4);
  else if (uo>W)		 Offset += ( W>>4);
  else if (uo<=-8)               Offset -= 8; 
  else                           Offset += (uo>>4);

	for the 8x8 chroma one.


	Note that for the 2-3 warp-point interpolation
	there's few FPS to be gained by branching to a
	non-clipping inner loop whenever the whole
	block lies inside the picture without crossing
	the edges. Then for the clipped version, making
	special cases for left/right/top/bottom clipping
	grasps some more ticks here and there, but nothing
	magical. I'll have a look when you commit your
	GMC code...

	For the 2-warp-pts constants setup problem, it seems
	ok to me. I'll cross-check the calculation in the
	ISO.


	bye!
			Skal