Questions about jnp ffts

[cwoolfo1]

Does Jax ffts parallelize across cores automatically? Do they use pencil or slab decompositions?

[dfm]

This question might be better asked on the openxla/xla repo because JAX just lowers the FFT to whatever XLA provides. I believe (but don't quote me!) that on CPU XLA will use the multithreaded FFT from DUCC, and on NVIDIA GPUs it'll use cuFFT. I'm not sure what decompositions those implementations use!