Fast Tomographic Reconstruction on Parallel Hardware

Tomographic reconstruction is the mathematical procedure of approximating a function f, based on the integrals of f along a set of line sections. The need for fast tomographic reconstruction arises for example in the challenging problem of real time control of some plasma parameters in a fusion reactor. In this paper, we present a fast algorithm for tomographic reconstruction. A good property of our approach is that it fits well to hardware with two levels of parallelism (e.g. a GPU cluster). We also propose an objective evaluation method for measuring the quality of reconstruction on real datasets where f is unknown. We will demonstrate that our algorithm is able to perform more than 50 000 reconstructions per second at reasonably good quality, running on a relatively cheap hardware.