Toward a Multi-Target Linear Algebra Library for GPUs and Multicore CPUs
Abstract
Highly structured sparse matrices arise frequently from numerical discretizations of partial differential equations. Legolas++ is a C++ generic library designed for describing and manipulating such multi-level blocked matrices with the corresponding blocked vectors and algorithms. Legolas++ allows a very detailed description of the linear systems to be solved that can be used to generate efficient parallel implementations. We are working on enlarging the scope of Legolas++ from multi- core target to GPUs and cluster of GPUs.