Practical considerations for GPU-accelerated CT

The introduction of programmability into commodity graphics hardware (GPUs) has enabled their use much beyond their native domain of computer graphics, in many areas of high performance computing. We have shown in previous work that many types of CT algorithms, both iterative and non-iterative, can also greatly benefit from the high degree of SIMD (same instruction multiple data) parallelism these platforms provide. In this paper, we extend this work by describing how one can deal with a number of challenges that frequently arise in practical application settings using the Feldkamp algorithm: large data, angle-dependent projection geometry, and the need for higher accuracy without compromising speed. For this, we combine our fast hardware-native 8-bit interpolation scheme with a higher precision dual-pass mechanism. This latest version of our RapidCT system runs on the most current GPU hardware, nearly eight times faster than the previous version