The effect of a spoke surface error on a phase mask in a computational imaging system was analyzed by combining the similarity of the point spread function (PSF) and the peak signal-to-noise ratio (PSNR) of de-convoluted images. The spoke surface error was applied on a phase mask with different peak-to-valley (P-V) values with various numbers of spoke rings in simulation. The minimum requirement of PSF similarity will be determined by a given PSNR threshold, which relates the defocus aberration. Finally, it can be concluded that a low-spatial frequency surface error is critical for a cubic phase mask in a computational imaging system with lower P-V error.