Abstract
Current graphics hardware offers only very limited support for convolution operations, which is primarily intended for image processing. The input and output sample grids have to coincide, making it impossible to use these features for more general filtering tasks such as image or texture resampling. Furthermore, most hardware employs linear interpolation for texture reconstruction purposes, incurring noticeable artifacts. Higher-order interpolation via general convolution is able to remove most of these artifacts. Real-time applications currently do not consider higher-order filtering due to lack of hardware support. We present algorithms for extremely fast convolution on graphics hardware. This framework can be used for general convolution tasks, but is especially Suited to substituting the native bi-linear or tri-linear interpolation currently used for texture magnification, while still achieving frame rates of up to 100 frames per second for full screen filtering with bi-cubic interpolation.