Methodology: the included graphs measure the accuracy of the forward and backward transforms: an array is generated with random uniform values between -0.5 and 0.5, and the results of its transform are compared with either pyfftw (in long double precision) if available, or scipy if available, or numpy fft. The L2 curve measures the average square norm difference, and the L_∞ the maximum difference.

Note: for the R2C inverse transform, the result of the forward transform is used instead of re-using the random array (in order to have a proper half-Hermitian array), contrary to what is done for other transforms. This explains why the IFFT R2C maximum (L_∞) errors are larger.

Note 2: some "errors" for DCT may be due to unsupported sizes in VkFFT, which vary depending on the card and language used (amount of shared/local memory). So they just indicate a current limit for the transform sizes rather than a real error.

[Click on the highlighted cells for details and accuracy graphs vs the transform size]