GPU Accelerated Optimization for Large Data Sets from Pill Identification

Gundolf Haase and Andreas Kucher


parallel GPGPU computing, optimization, Quasi–Newton, polynomial curves, circular splines


This paper describes the fast minimization of a functional for a best curve approximation of data resulting from a pill identification problem. The curve approximation in RN uses polynomial curves as well as circular splines. Functional evaluations have been faster on the CPU with circular splines but the situation changes on the GPU in favor of polynomial curves. We used a line search with a fixed step size and a fixed reduction factor of it in the Quasi–Newton iteration. This naive step size control reduces code branches in the implemented optimization algorithm and yields a performance gain between 65 and 194 on the GPU in comparison to the CPU implementation. The MPI–parallelization of the code shows additional good speedup results.

Important Links:

Go Back