nabu.processing.fft_cuda module

nabu.processing.fft_cuda.init_skcuda()[source]
class nabu.processing.fft_cuda.SKCUFFT(shape, dtype, r2c=True, axes=None, normalize='rescale', **backend_options)[source]

Bases: _BaseFFT

Base class for Fast Fourier Transform (FFT).

Parameters:

  • shape (list of int) – Shape of the input data

  • dtype (str or numpy.dtype) – Data type of the input data

  • r2c (bool, optional) – Whether to use real-to-complex transform for real-valued input. Default is True.

  • axes (list of int, optional) –

    Axes along which FFT is computed.

    • For 2D transform: axes=(1,0)

    • For batched 1D transform of 2D image: axes=(-1,)

  • normalize (str, optional) –

    Whether to normalize FFT and IFFT. Possible values are:

    • ”rescale”: in this case, Fourier data is divided by “N” before IFFT, so that IFFT(FFT(data)) = data. This corresponds to numpy norm=None i.e norm=”backward”.

    • ”ortho”: in this case, FFT and IFFT are adjoint of eachother, the transform is unitary. Both FFT and IFFT are scaled with 1/sqrt(N).

    • ”none”: no normalizatio is done : IFFT(FFT(data)) = data*N

  • backend_options (dict, optional) – Parameters to pass to CudaProcessing or OpenCLProcessing class.

implem = 'skcuda'
backend = 'cuda'
ProcessingCls

alias of CudaProcessing

fft(array, output=None)[source]
ifft(array, output=None)[source]
class nabu.processing.fft_cuda.VKCUFFT(shape, dtype, r2c=True, axes=None, normalize='rescale', **backend_options)[source]

Bases: _BaseVKFFT

Cuda FFT, using VKFFT backend

Base class for Fast Fourier Transform (FFT).

Parameters:

  • shape (list of int) – Shape of the input data

  • dtype (str or numpy.dtype) – Data type of the input data

  • r2c (bool, optional) – Whether to use real-to-complex transform for real-valued input. Default is True.

  • axes (list of int, optional) –

    Axes along which FFT is computed.

    • For 2D transform: axes=(1,0)

    • For batched 1D transform of 2D image: axes=(-1,)

  • normalize (str, optional) –

    Whether to normalize FFT and IFFT. Possible values are:

    • ”rescale”: in this case, Fourier data is divided by “N” before IFFT, so that IFFT(FFT(data)) = data. This corresponds to numpy norm=None i.e norm=”backward”.

    • ”ortho”: in this case, FFT and IFFT are adjoint of eachother, the transform is unitary. Both FFT and IFFT are scaled with 1/sqrt(N).

    • ”none”: no normalizatio is done : IFFT(FFT(data)) = data*N

  • backend_options (dict, optional) – Parameters to pass to CudaProcessing or OpenCLProcessing class.

implem = 'vkfft'
backend = 'cuda'
ProcessingCls

alias of CudaProcessing

vkffs_cls

alias of VkFFTApp

nabu.processing.fft_cuda.has_vkfft(safe=True)[source]

Determine whether pyvkfft is available. For Cuda GPUs, vkfft relies on nvrtc which supports a narrow range of Cuda devices. Unfortunately, it’s not possible to determine whether vkfft is available before creating a Cuda context. So we create a process (from scratch, i.e no fork), do the test within, and exit. This function cannot be tested from a notebook/console, a proper entry point has to be created (if __name__ == “__main__”).

nabu.processing.fft_cuda.has_skcuda(safe=True)[source]

Determine whether scikit-cuda/CUFFT is available. Currently, scikit-cuda will create a Cuda context for Cublas, which can mess up the current execution. Do it in a separate thread.

nabu.processing.fft_cuda.get_fft_class(backend='vkfft')[source]
nabu.processing.fft_cuda.get_available_fft_implems()[source]