[XviD-devel] Question about bvop decoding

[Edouard Gomez]

Edouard Gomez (ed.gomez at free.fr) wrote:
> we do spend an
> incredible amount of time in decoder_bf_interpolate_mbinter
> (something like 14% of cpu time) though it's just a function
> that calls more CPU intensive functions like qpel/halfpel
> interpolation functions and so on.

I added some debug code in decoder, and i got this:
xvid: interpolated 19507156 blocks out of 637555200 total blocks.

so interpolated blocks represent 3.05% of total blocks. If I assume
these blocks require something like 1 or 2 times more cpu cyles than
usual forward/backward blocks (they require the interpolation of
both backward and forward refs, and then call the usual inter
decoding function). This should max at 9% total CPU time, including
all interpolation, compensation code. But as i said,
decoder_bf_interpolate_mbinter represents 14% of total CPU
cycles + all calls to sub functions (hard to tell how much, as
these functions are also called by other inter block decoding
functions, and oprofile lacks call graph output)

I'd like to be completly wrong... buit who's going to tell me
where ?

-- 
Edouard Gomez