A Fast Solution of Hessenberg Reduction Algorithm Using Multi-Matrix Processing on GPU

In this letter, we propose a multi-matrix programming model in GPU computation to deal with Hessenberg reduction problem. Hessenberg reduction problem is the most important step in eigenvalue problem, which is computationally expensive. Conventional method using GPU is inefficient in GPU resource usage when it deals with small matrix. The proposed method computes as many matrices as possible which maximally utilizes the GPU resources. Therefore, the proposed method achieves higher memory transfer rate between CPU and GPU, and massive parallel computation that is well fit for GPU computation. Experimental results show that the proposed outperforms the optimized commercial packages.