GPU Computing


Geometry processing requires efficient linear (and sometimes non-linear) solvers. We studied massively parallel implementations of sparse iterative solvers, and implemented them on the GPU [High Performance Computing Conference, 2007](second best student paper award), [International J. of Parallel, Emergent and Distributed Systems, 2008]. Our "Concurrent Number Cruncher" algorithm is a sparse Jacobi pre-conditioned conjugate gradient with several optimizations (block compressed row storage, register blocking).




* Software library: OpenNL/CNC (referenced in NVIDIA TESLA Software tools: 3rd party, Jacobi pre-conditioned Conjugate Gradient Soliver)