>
> Hi all!
>
> 11/3 is holiday in Japan, and I have plenty of time...
>
> check this new patch containing new MDCT/subband filtering and FFT.
> This patch can be applied to both of the original LAME 3.37beta(3.50stable)
> and LAME3.37 with my yesterday Huffman coding patch.
>
> On my architechture, the benchmark result is
> lame 3.37 (normal) 32.54
> + Humman patch 30.37
> + MDCT + FFT 27.90
> ---
> Takehiro TOMINAGA // may the source be with you!
I applied the Huffman patch, it will be in 3.51.
The MDCT + FFT patch is more difficult. They are
noticably faster for me too, and I would like to get them
into lame, but:
The before and after patch results are not identical. Maybe
just rouding errors, but it also looks like there are
some errors syncing the psycho-acoustics and the MDCT encoding?
At least the frame analyzer shows them out of sync by about
1 granule.
Since there are so many changes, it is hard to track down
what is causing the differences. Takehiro got rid of a lot
of buffering, but at the expense of making lame.c more
complicated: Right now you call l3psycho_anal with the
exact granule that should be analyzed. l3psycho_anal then does
its own buffering and padding to center this granule in
the analysis window. This makes it much easier to
make sure the psycho acoustics and the MDCTs are properly
synced - there were many hard to find bugs in the ISO code
related to this.
I tried to put in just the MDCT patches, but also couldn't
reproduce the original output. Takehiro: is the delay of
your MDCT different than the original? The current MDCT
routine has a delay of 240 (filterbanks) + 288 (MDCT's).
So when you feed it 576 samples, the output is from a
576 sample window that is 528 samples earlier in the
input stream.
Mark
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
