lcav / mkl_fft Goto Github PK

For now, only contiguous up to 2D array are supported. Addind support for arbitrary strides would bring:

• support for arbitrary dimension and arbitrary axis
• support for arbitrary views, including non-contiguous

This problem comes from limitation of LIBMKL which DFTI_DISTANCE parameter is just a scalar. For arbitrary strides, it needs to be a 2D vector. Thus a for loop is required.

Is there any way to perform 3D or nD FFTs with mkl_fft, even if in a hacky way? I need 3D FFTs for my application, and mkl_fft just so much faster than pyfftw or the mkl-optimized anaconda distribution.

Great job, by the way!