[XviD-devel] [RFC] Decoder speedup

[Edouard Gomez]

Michael Militzer (michael at xvid.org) wrote:
> I'm very sure that the ratio between zero and non zero coeffs will indeed
> be huge (probably 3% non-zero coeffs), but of course including the coeffs
> of not coded blocks (which however are of course not dequantized in the
> XviD SIMD implementation).
> 
> Regarding C implementation for dequant - well, I don't know the ffmpeg
> sources too well but I can't imagine how coeff decoding and dequant should
> be merged using SIMD instructions. So I'd guess a pure c implementation is
> used...

Yeah i realized that after i asked the question, unless they
have some sort of stack of 8 elements with multipliers,coeffs, bias and
positions inside the block, they can't do much SIMD acceleration.

Btw i tried the on the fly dequant for inter h263 (as intra
blocks are full of non zero coeffs, i'm not going to try this
optim on them):
BENCHMARKs: VC: 237,420s VO:   0,178s A: 0,000s Sys: 11,038s =  248,636s <-- normal
BENCHMARKs: VC: 227,911s VO:   0,181s A: 0,000s Sys: 10,955s =  239,048s <-- on the fly dequant + get_coeff optim

As we know the get_coeff optim accounts for ~2% speedup, the
on the fly dequant is around 2% too. Not quite impressive :-(

-- 
Edouard Gomez