[SciPy-Dev] Accuracy of single-precision FFT

[Sturla Molden]

Den 25.06.2010 18:38, skrev Anne Archibald:
> In the long term, if numpy is to become an array library plus a python
> wrapper, the FFT code must migrate to C. I don't anticipate this being
> too bad either, but it amounts to extending FFTPACK.
>    

FFTs in NumPy are already written in C.

FFTs in SciPy are written in Fortran 77.

I was suggesting to use NumPy's FFTs in SciPy, instead of have two 
almost identical FFT libraires (one C and one Fortran), as we do today.

The Fortran FFTPACK is used in SciPy because of single precision, 
complex numbers and DCT. But we can get around that easily by compiling 
fftpack_lite.c with other definitions of Treal than double.

We don't even need special Fortran code for DCT. It just takes an FFT 
and two or three extra lines of Python.

Why do we use Fortran FFTPACK at all?


Sturla