Fast, GPU-based Computation of Large Point-Spread Function Sets for the Human Eye using the Extended Nijboer-Zernike Approach

Abstract

The point-spread function (PSF) is the diffraction pattern of an infinitesimal light source and plays an important role in the study and simulation of human vision. It forms the backbone of a multitude of vision-rendering algorithms, as it can be used to obtain the necessary kernels for convolution. Its computation is often performed via ray-tracing or the fast Fourier transform (FFT), but recently we also demonstrated that the Extended Nijboer-Zernike (ENZ) approach can be a more efficient alternative, which reduces the computation time of large PSF sets to just a few minutes. In this paper, we present a significantly faster, GPU-based computation scheme of the ENZ approach to further improve the computation process for such large PSF sets. Our algorithm works by reformulating the core $V_{n}^{m}$ function to reusable subterms that are efficient to accumulate in parallel. We demonstrate that our proposed method leads to substantial performance improvements and facilitates the interactive exploration of visual aberrations when paired with our existing vision simulation algorithm.