An Optimized Framework for Matrix Factorization on the New Sunway Many-core Platform

Matrix factorization functions are used in many areas and often play an important role the overall performance of applications. In LAPACK library, matrix implemented with blocked algorithm, shifting most workload to high-performance Level-3 BLAS functions. But non-blocked part, panel factorization, becomes bottleneck, especially for small- medium-size matrices that common cases real On new Sunway many-core platform, bottleneck can be alleviated by keeping LDM factorization. Therefore, we propose a framework implementing on facilitating in-LDM The provides template class wrapper functions, which integrates inter-CPE communication Level-1 Level-2 flexible interfaces accommodate different partitioning schemes. With framework, writing code data residing space done much higher productivity. We three ( dgetrf , dgeqrf dpotrf ) based compared our work CPE_BLAS version, uses original implementation linked optimized library runs CPE mesh. Using favorable partitioning, part achieves speedup up 26.3, 19.1, 18.2 For whole function, is carefully tuned recursion added specific optimization some subroutines Overall, obtained average 9.76 10.12 4.16 version. Based current class, extended support more categories linear algebra