[FFmpeg-devel] [PATCH] [3/3] SIPR 16k decoder
Ronald S. Bultje
Fri Jan 15 01:30:16 CET 2010
On Thu, Jan 14, 2010 at 7:25 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> On Thu, Jan 14, 2010 at 07:19:47PM -0500, Ronald S. Bultje wrote:
>> On Thu, Jan 14, 2010 at 7:05 PM, Vitor Sessak <vitor1001 at gmail.com> wrote:
>> > The rule I use is: always use floats, unless when you need to call
>> > ff_acelp_lspd2lpc(), which takes doubles as arguments (and also to
>> > avoiding converting buffers to double or back).
>> So let's change ff_acelp_lspd2lpc()? Why does it use doubles?
> dont forget testing PSNR & with yor ears if PSNR differs
> i remember this double vs float discussion and the last time the awnser IIRC
> was that floats are not accurate enough. I dont want to see the floats are
> faster on my system so change it discussion repeat until testing is forgotten
> and we put a possibly hard to debug bug in
My point is more or less that *ALL* voice decoders should use doubles
everywhere. The 16k code uses floats, takes a cos(f), casts it to a
double and suddenly we have a double. That's not right. It should've
used doubles from the beginning if the goal was to do a double.
And why on earth are LPCs in float if LSPs are in double?
More information about the ffmpeg-devel