An implementation of the ACA on GPU platform

In this paper, an implementation of the ACA on GPU platform is presented, involving two parts: the matrix compression using the ACA and the batched matrix-vector products utilizing H-matrix form. Some numerical examples are provided to demonstrate the overall performance of the proposed implementation of the ACA algorithm on GPU platform through comparison with the 4-threaded CPU algorithm. In these testing examples, the speedup ratio of performing matrix compression using the ACA can achieve about from 50 to 100 for the single-precision case and from 25 to 50 for the double-precision case, and that the speedup ratio of executing batched batched matrix-vector products utilizing H-matrix form can achieve about from 10 to 30 for the single-precision case and from 6 to 17 for the double-precision case.