A Comparative Study of Preconditioners for GPU-Accelerated Conjugate Gradient Solver

We compare two types of preconditioners for GPU-Accelerated conjugate gradient solver. For the standard IC preconditioner, we exploit level scheduling to increase multi-thread parallelism of sparse triangular solve on GPU. Meanwhile, we propose a novel reordering technique to maximize the coalescing of global memory accesses. For the approximate inverse preconditioner SSOR-AI, we extend it to second order approximation. Experiments indicate that our IC PCG runs 25% faster than using vendor implementation in CUSPARSE library and SSOR-AI PCG can be twice as fast as IC PCG.