On Fri, Apr 29, 2011 at 10:24 AM, Ronald S. Bultje <[email protected]> wrote: > Hi, > > On Fri, Apr 29, 2011 at 9:05 AM, Diego Biurrun <[email protected]> wrote: >> On Fri, Apr 29, 2011 at 02:37:35PM +0000, Ronald S. Bultje wrote: >>> >>> On Fri, Apr 29, 2011 at 2:08 PM, Diego Biurrun <[email protected]> wrote: >>> > int/unsigned is the natural memory access type for CPUs, using sized types >>> > for temporary variables, counters and similar just increases code size and >>> > can possibly cause a slowdown. >>> > --- >>> > libavcodec/alpha/simple_idct_alpha.c | 6 +- >>> > libavcodec/jfdctfst.c | 18 ++-- >>> > libavcodec/jfdctint.c | 18 ++-- >>> > libavcodec/vorbis.c | 9 +- >>> > libavcodec/vorbis.h | 2 +- >>> > libavcodec/vorbisdec.c | 227 >>> > ++++++++++++++++------------------ >>> > 6 files changed, 134 insertions(+), 146 deletions(-) >>> >>> Don't touch the alpha code unless you have access to one. >> >> Mans reviewed that stuff and suggested the change, so I trust it. >> FATE will complain if it fails to compile (I don't see how) and >> then I can revert that part. It's how I did the last changes to >> Alpha code and for this platform I think it is a valid approach. >>
[...] >> >>> Regardless, speed tests are >>> missing to confirm no significant slowdown. >> >> On the contrary there is something like ~5% speedup: >> >> time ./ffmpeg -i ~/tmp/FaiF_0x0D_NDAs.ogg -f null -v 0 - > > And for the JFDCT? > Seems like the logical thing to do is split this patch three ways so each part can be reverted individually. _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
