[FFmpeg-devel] [PATCH] MMX VP3 Loop Filter

[David Conrad]

On Oct 4, 2008, at 4:54 AM, Diego Biurrun wrote:

> On Sat, Oct 04, 2008 at 12:21:42AM -0400, David Conrad wrote:
>>
>> This was adapted from libtheora and gives an overall speedup of about
>> 5%.
>
> Do you have benchmark numbers comparing FFmpeg and libtheora?

Using Elephant's Dream 1080p (forgot link) as a test file,  libtheora  
was around 30% faster than FFmpeg last time I checked. This is mainly  
due to a better method of decoding DCT coefficients (FFmpeg spends  
about 25% of the decode time skipping blocks to find the block a  
coefficient belongs to) and more cache-friendly code after that.

I'm working on porting those optimizations to FFmpeg, then I'll do  
more in-depth benchmarks again.