[CELT-dev] Modifications of kiss_fft/compatibility with original kiss_fft code?
jmvalin at jmvalin.ca
Sun Feb 20 16:14:13 PST 2011
It's more than just a reorganization of the code/data. The original
fixed-point kiss-fft supported either 16-bit precision or 32-bit
precision. Instead, the modified version we use in CELT uses 32-bit data
with 16-bit twiddles.
On 11-02-20 06:30 PM, Timothy B. Terriberry wrote:
>> code, I would like to check whether the changes, that were made to the
>> original kiss_fft code, change the actual behaviour of the kiss_fft
>> function or whether they were only made for speed improvement or
>> cosmetic reasons. In other words, could I still replace the functions
>> "kiss_fft_alloc" and "kiss_fft" with the original code (apart from the
>> fact that there are now functions for the inverse transformation in celt)?
> They were made mostly for data organization reasons (i.e., to allow
> static data and sharing of twiddles between FFTs of different sizes).
> The actual transforms should be the same. I'd need to double-check for
> fixed-point, but in my decoder re-implementation, I was able to start
> from the current kisfft release and achieve compatibility with the float
> code without changing the actual calculations involved.
> celt-dev mailing list
> celt-dev at xiph.org
More information about the celt-dev