Re: [mpeg2-dev] [PATCH] SSE2 idct optimizations

[Diego B. <di...@bi...>]

On Wed, Apr 02, 2008 at 01:52:24PM +0200, Christophe Massiot wrote:
> On Tue, Apr 01, 2008, Diego Biurrun wrote:
>  
> > > Does anyone know why the output of the SSE2 IDCT is different from the
> > > one of the MMXEXT IDCT ? Since the algorithm is supposed to be the same,
> > > they should have exactly identical results. This not being the case, it
> > > may indicate that some extra approximations are used, and the SSE2 code
> > > may not pass the ieee 1180 conformance test.
> > 
> > How much of a problem is this given that an IDCT is always a lossy
> > operation?
> 
> IDCT is lossy but MPEG-2 specifies that it mustn't be too lossy. So we
> must check that the output is within the allowed boundaries. If the
> results are identical to the ones of the MMXEXT version, then there is
> no need to do the check since the MMXEXT has already been checked to be
> compliant.
> 
> However, the results are different, so the checks must be done again.
> This is really not my area of expertise, I've asked to people who know
> about this but haven't had any answer yet (this is a good opportunity
> to remind them :). But if you speak x86 asm fluently (which I don't),
> maybe you can spot what's the reason why the results are different while
> implementing the same algorithm.

Alexander Strange told me he found the problem and provided me with an
updated patch.  Apparently

  static const int32_t rounder4_128[] ATTR_ALIGN(16) = rounder (0);                                     

should have been

  static const int32_t rounder4_128[] ATTR_ALIGN(16) = rounder_sse2 (0);        

Please try the attached version and let me know if this works as
expected.  Again, I could only do minimal compilation tests, I do not
have a machine with SSE.

Diego