A Portable and Heterogeneous LU Factorization on IRIS

Here, the IRIS programming model is evaluated as a method to improve performance portability for heterogeneous systems that use LU matrix factorization. LU (lower-upper) factorization is considered one of the most important numerical linear algebra operations used in multiple high-performance computing and scientific applications. IRIS enables the separation of the algorithm’s definition from the tuning by using tasks + dependencies. This considerably reduces the effort required to achieve performance portability on heterogeneous systems. One IRIS code can use different settings depending on the underlying hardware features. Different configurations are evaluated on two different heterogeneous systems to achieve important speedups for the reference code with minimal changes to the source code.