[XviD-devel] zero-biased dequant/idct

Michael Militzer xvid-devel@xvid.org
Mon, 29 Jul 2002 12:44:53 +0200


Hi,

----- Original Message -----
From: "daniel smith" <danielsmith@astroboymail.com>
To: <xvid-devel@xvid.org>
Sent: Sunday, July 28, 2002 7:31 PM
Subject: [XviD-devel] zero-biased dequant/idct


> i just remembered a paper i skimmed over that i
> wanted to try out to speed decoding (i will
> benchmark this afternoon unless someone's
> already tried it):

AFAIK, noone tried it yet...

> - trap the number of bits required to store coeffs
> during mb decoding
> - if coeff bitcount for block is below a threshold,
> call versions of dequant_inter / idct which include
> a zero-byte skip check for every 8 bytes of
> block[64]
>
> yes we lose time with branch mispredictions, but
> is the zeroed content of inter blocks high enough
> to justify?

Hm, that's of course hard to estimate, but I think one can be at least sure
that there won't be any huge advantages of this approach. However there
might be some mild speed improvement, so if you'd like to try this out, I
would appreciate it. btw: did the paper where you read about this idea
present any numbers about speed improvements? Maybe this approach is more
successful for c-code only? We already learned that an early-stop check
within sad[8,16] was not useful...

> this could perhaps aid encoding speed as well,
> using the sum of quantized coeffs as a trigger
> during dequant/idct stage.

I'm pretty sure that an encoding speed-up won't be noticeable: IDCT +
dequant take up only about 4% of total encoding time...

bye
Michael