Contents

The SPIKE algorithm [1, 2] is an efficient generic divide-and-conquer algorithm for solving banded systems. With partitioning the matrices into several blocks, sufficient concurrency can be exploited on GPUs. We implemented SPIKE::GPU, a solver which exploits the truncated SPIKE as pre-conditioner and then the pre-conditioned result is refined with BiCGStab. Our current results show that SPIKE::GPU can perform more than two times as fast as the banded linear system solver in Intel’s Math Kernel Library (MKL) and up to three times if the kernel is manually tuned