[Ffmpeg-devel] VP3/Theora Perfection

Michael Niedermayer michaelni
Thu May 19 23:43:48 CEST 2005


Hi

On Thursday 19 May 2005 22:26, Mike Melanson wrote:
> Mike Melanson wrote:
> >     I think I understand. I will make some tests along these lines...
>
> 	Hmm, I ran some new tests and the improvement using my new code is even
> more dramatic:
>
> [original, no inline]
> 1218 dezicycles in unpack_token, 32760 runs, 8 skips
> 1196 dezicycles in unpack_token, 65515 runs, 21 skips
> [original, using -finline-limit=2000]
> 1341 dezicycles in unpack_token, 32760 runs, 8 skips
> 1314 dezicycles in unpack_token, 65523 runs, 13 skips
> [new]
> 733 dezicycles in unpack_token, 32745 runs, 23 skips
> 734 dezicycles in unpack_token, 65486 runs, 50 skips

hmm, same here, but not if the benchmarking is only done around 
unpack_dct_coeffs() in that later case its the other way around which is 
somewhat strange 


>
> Not sure why the -finline-limit version performed even worse. Though it
> is worth noting that inspection of the ASM did not show the function
> inlined.

hmm, indeed neither here, but unpack_vlcs becomes inlined


>
> 	My understanding of inlining is when the compiler puts the entire code
> for a function in place of a function call. Have I got that right?

yes

[...]
-- 
Michael





More information about the ffmpeg-devel mailing list