[FFmpeg-devel] [PATCH] WMA Voice decoder
Ronald S. Bultje
Wed Jan 20 22:55:31 CET 2010
2010/1/20 M?ns Rullg?rd <mans at mansr.com>:
> 1. Use of double-precision FP
> ? Is this really necessary in all places? ?Even if the answer is that
> ? some existing function takes double arguments, it should be checked
> ? whether a float version of same would be good enough here.
All double is for LSF/LSP calculation, since that's what the acelp
function takes. I think we should just hit the shit and convert it all
to float, either before or after I commit, no matter whether it
changes PSNR by a percent or not.
In the end, these are low bitrate voice codecs. Shit goes in, shit
comes out. We're just adjusting the shit to contain slightly less
shitty after our decoding adds more noise to it, but it's still shitt.
> 2. FP conversions
> ? You seem to be doing a lot of int to float conversions. ?These take
> ? time, and it might be better to store some tables as floats
> ? instead, at least some of the small ones. ?Type punning is often
> ? useful too.
I think I only do it for the dq, because it decreases binary size
considerably. Any other place I missed?
> 3. memcpy, memset etc
> ? I noticed a number of calls to memcpy and memset. ?Are all of these
> ? really needed? ?Perhaps some could be done by pointer swapping
> ? instead.
I just removed one (or, well, merged two into one) in my
to-be-published patch. The others are unfortunately necessary, yes.
More information about the ffmpeg-devel