PyTorch implementation of the percent point function (PPF) for a Gaussian distribution, i.e., inverse CDF of a Gaussian distribution.
The goal is to enable using the PPF function in CUDA to go back and forth from a Gaussian to a Uniform distribution.
The function emulates the behavior of scipy.special.ndtri
but on either GPU or CPU.Tested on Google Colab with a T100, process got a 6x improvement on GPU (performance is worse on CPU):