[FFmpeg-devel] [PATCH] SPARC VIS simple_idct try#6

Balatoni Denes dbalatoni
Wed Aug 29 01:13:17 CEST 2007


Wednesday 29 August 2007 00:13-kor Michael Niedermayer ezt ?rta:
> > > Indeed, I didn't take that into account. So if I fix that 25% and the
> > > clamping part, will you accept the patch?
> >
> > Better yet: that would be 4 instructions. How about I gain 4 clocks in
> > some other way instead - how, let it be my secret. Okay?
> hmm no but you have to do that secret optimization too now at minimum for
> it to be considered for svn

4 instructions dealing with f46 (fzero and ldd in the macros) can be 
eliminated, if f60 or f62 is also put to use.

> ill investigate the register shortage vs. avoidable load/stores vs. latency
> after (the unlikely) case that you do correct the undisputed
> suboptimalities

Argh, I checked what it would involve to fix the 25% overlap. Half of the 
registers would have to be carefully renamed all over the place.

While the suboptimalities are not disputed, whether the benefit from fixing 
them outweights the cost (in time, and code beuty) is heavily disputed.

> [...]


More information about the ffmpeg-devel mailing list