[FFmpeg-devel] [PATCH] address speed regression in h264 deblocking filter after PAFF

[Andreas Öman]

Hi,

Michael Niedermayer wrote:
> 
> hmm, have you benchmarked the functions individually with START/STOP_TIMER?
> which becomes slower?
> and after you know the funtion try a binary search to find it out more
> precissely
> of course theres a good chance this wont lead to anything ...

placed the timer around the deblock callback

av_always_inlined
   31429 dezicycles in deblock, 524256 runs, 32 skips

without av_always_inlined
   30299 dezicycles in deblock, 524108 runs, 180 skips

So I lose ~60Mcycles there.

Seem gcc also forces inline of all the filter_mb_edge* functions.
I tried to add 'av_noinline' to these, but it didn't change the
figures very much.

The sheer facts that the inline/noinline modification results varies
very much on different processor architectures makes me kind of
unmotivated to continue working on these type of "optimizations"

The time is probably better spent looking into better multithreading
support.

> btw, your video doesnt use different deblock functions for each frame right?

No, i've tried with various different content too :/