Accelerated Gauss-Huard Algorithm on Hybrid GPU-CPU: Look-Ahead with the Delayed Algorithm Approach

In this paper, we tackle a significant bottleneck - the panel factorization step - in the Gauss-Huard algorithm through a novel parallel computing approach. We address the open question in the literature regarding the feasibility of applying look-ahead in this context. Our strategy combines the use of the delayed Gauss-Huard algorithm with random butterfly transformations instead of the traditional partial pivoting. The proposed technique not only allows for the application of look-ahead but also enhances the overall efficiency of the Gauss-Huard algorithm in a parallel computing environment, presenting possibilities for further optimizations.