[Ffmpeg-devel] [PATCH] wma optimizations

[Michael Niedermayer]

Hi

On Sat, Aug 19, 2006 at 02:51:03AM +0200, Benjamin Larsson wrote:
> With the new dsp vector functions in the wma decoder it shaved of about
> 120000 dezicycles from the windowing code loop.
> 
> Non simd before optimization:
> 138006 dezicycles in winmul, 4090 runs, 6 skipss dup=0 drop=0
> 81426 dezicycles in winspecial, 4091 runs, 5 skips
> 137845 dezicycles in winmul, 8172 runs, 20 skips dup=0 drop=0
> 81290 dezicycles in winspecial, 8177 runs, 15 skips
> 137565 dezicycles in winmul, 16350 runs, 34 skipsdup=0 drop=0
> 81051 dezicycles in winspecial, 16358 runs, 26 skips
> 
> Simd:
> 50668 dezicycles in winmul, 4092 runs, 4 skips/s dup=0 drop=0
> 42723 dezicycles in winspecial, 4091 runs, 5 skips
> 50486 dezicycles in winmul, 8186 runs, 6 skips
> 42602 dezicycles in winspecial, 8186 runs, 6 skips
> 50317 dezicycles in winmul, 16369 runs, 15 skips dup=0 drop=0
> 42464 dezicycles in winspecial, 16370 runs, 14 skips
> 
> A md5sum test on an amd64 box gave the same checksum for both versions.
> But I'm not sure that would be true for all kind of simd code, but the
> diffrent rounding should only lead to 1 bit differences.
> 
> Overall speedup is around 3%. (From 14.2 -> 13.8).

great, and patch looks ok

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

In the past you could go to a library and read, borrow or copy any book
Today you'd get arrested for mere telling someone where the library is