318 / 2020-01-06 11:16:00

GPU-accelerated parallel optimization for sparse regularization

block successive convex approximation; LASSO

We prove the concept that the block successive convex approximation algorithm can be configured in a flexible manner to adjust for implementations on modern parallel hardware architecture. A shuffle order update scheme and a all-close termination criterion are considered for efficient performance and convergence comparisons. Four different implementations are studied and compared. Simulation results on hardware show the condition of using shuffle order and selection of block numbers and implementations.