A Fine-Grained Pipelined Implementation of LU Decomposition on SIMD Processors

Abstract : The LU decomposition is a widely used method to solve the dense linear algebra in many scientific computation applications. In recent years, the single instruction multiple data (SIMD) technology has been a popular method to accelerate the LU decomposition. However, the pipeline parallelism and memory bandwidth utilization are low when the LU decomposition mapped onto SIMD processors. This paper proposes a fine-grained pipelined implementation of LU decomposition on SIMD processors. The fine-grained algorithm well utilizes data dependences of the native algorithm to explore the fine-grained parallelism among all the computation resources. By transforming the non-coalesced memory access to coalesced version, the proposed algorithm can achieve the high pipeline parallelism and the high efficient memory access. Experimental results show that the proposed technology can achieve a speedup of 1.04x to 1.82x over the native algorithm and can achieve about 89% of the peak performance on the SIMD processor.