[XviD-devel] New Qpel code

[Edouard Gomez]

Christoph Lampert (chl at math.uni-bonn.de) wrote:
> Feature freeze was 2 month ago. Do we really need new code just for 
> speedup, possibly introducing new errors before releasing 1.0? 

The only i dislike with this code is the fact we have now:
 1/ code for ia32 only
 2/ another piece of code for the non ia32

Moreover, i  had a look at  the new code,  and it is impossible  to code
equivalent functions  to plug them  in a XVID_QP_FUNCS because  with the
old functions (transfer8x8_... interpolate16x16_...) we need a temporary
storage area, that the new type of prototype does not provide. 

Another point i dislike is the fact that we could have removed the nasty
switch  easily,  just  by  not  using  a  structure  for  qpel  function
pointers. I mean something like that:

typedef struct XVID_QP_PASS xvid_qp_funcs_t[16];

Then  somewhere we  would  have  filled the  16  pointers, each  pointer
replacing the equivalent  calls we can find in a  switch case in current
code.

Then the big switch becomes simply:
qp_funcs[(dx&3) | ((dy&3)<<2)](dst, src, tmp, 16, stride, rounding);

I  use that  a  lot in  mbtransquant[1],  where i  removed  most of  the
branching  using function  arrays.   As xvid  already  uses pointers  to
functions, that does not slow things down.

PS: while checking the code, i just noticed it's a trick i have ready in
a patchset but is neither commited to  my tree nor the CVS. Btw i use it
extensively to remove branches when i know the branches just decide what
function has to be called.

--
Edouard Gomez