From: Mohamed Tarek <mtarek16_at_gmail.com>
Date: Sun, 4 Jul 2010 18:35:38 +0300

Hello.

I have finally been able to completely convert the imdct_half function to
fixed point along with the trig tables needed by wma pro. Turns out the
problem with our version of that function is that the floating point trig
tables for wma pro are scaled down (presumably to save energy), while our
fixed point tables aren't, so the output is incorrect.

There may be a workaround for this issue to be able to use our imdct_half,
but for now I'll use the one I converted for wma pro so as not to lose more
time to it that I already did.

Currently the decoder has the following parts in fixed point : inverse
quantization and rescaling (which required just a small lookup table),
imdct, windowing.
And with those, our decoder has an error of 0.329% in magnitude compared to
ffmpeg's floating point one.

With this part over, I'm hoping I would have the whole decoder in fixed
point by the end of next week.

I'll clean up the code since it's a mess right now and upload a patch later.

--
MT

Received on 2010-07-04