2022 07

The SLATE project made a new release, 2022.07.00, focused on improving least squares and QR performance. By moving more operations to the GPU, tall-skinny QR performance improved by 5.5x for a 10M x 1024 problem on Summit. A Cholesky QR routine that targets tall-skinny problems was also added, which likewise showed 5x improvement over previous methods and provides another option for solving least squares problems. Other recent changes include performance improvements in BLAS, LU, Cholesky, and mixed precision solvers. BLAS++ and LAPACK++ portability layers were updated, adding support for Cholesky, LU, and QR factorizations on the GPU, using NVIDIA cuSolver or AMD rocSolver. The repository is at https://bitbucket.org/icl/slate.