Léopold Cambier
Léopold Cambier
An even simpler (/more obvious) repro ``` from jax.numpy.fft import fft2 import numpy as np N = 7603 x = np.ones((N, N)) y = np.zeros((N, N)) y[0,0] = N*N f1...
Just to add some context: Imho, this bug has two distinct problems Roughly speaking, for a matrix `A` with a condition number of `k`, the inverse can only be computed...
Hi there! I'm working on cuFFT and following up from https://github.com/JuliaGPU/CUDA.jl/issues/1656 (internal NVIDIA bug# 3847913). Sorry to hijack your issues - I thought it would be easier and faster to...
> Are complex to real transform algorithm changes during the recent version? Yes, there are occasional changes. I doesn't surprise me that you didn't see this behavior in older versions....
This applies to 1D, 2D and 3D yes. If the imaginary part of `x[0]` and `x[N/2+1]` is not zero, computing a complex-to-real DFT doesn't really make sense since the output...
Because cuFFT FP64 transforms use (slightly) different algorithms, so the issue is not visible.