STRUMPACK Versions Save

• Fix for sparse solve with MAGMA with 1 MPI rank
• suppress several warnings
• small bugfix in BLR partitioning in sparse solver
• SYCL CMake compilation fix
• In StructuredMatrix interface, construct_from_elements now takes point geometry for faster HSS compression
• In StructuredMatrix interface, BLR compress_and_factor is now supported

• Memory leak fix from MPI_Datatypes
• Set RPATH in strumpack library
• Add counting of subnormal number in the sparse factors
• Changes in the BLR compression tolerances in the sparse solver, using a scaled absolute tolerance
• Fixes in the sparse solver using MAGMA, resetting error codes which show up in larger systems, but are not actually errors.
• Set SOVERSION
• Update CMake for HIP, using enable_language(HIP), requires CMake 3.21

Workaround for SLATE <= 20220700

Small bugfix

ROCm compilation fix

• Bugfix in matrix equilibration code
• Several bugfixes, especially for SLATE and large problems on GPU
• Sparse triangular solve on the GPU when using MAGMA
• Other MAGMA fixes for the sparse direct solver (MAGMA still optional)
• New HSS random sketching operators based on sparse Johnson Lindestrauss
• Fix for HODLRMatrix construction from elements, or blocks
• Compilation fixes for NVHPC compiler
• SYCL updates
• Add lapmr routines, which are not available in Mac LAPACK implementations
• Support newer ( >= 1.0) ZFP versions
• Fixes for clang 15
• Add NDBFS GPU matrix ordering code

• Many bugfixes and general improvements.
• Important fixes in the GPU code, and in the usage of SLATE (GPU capable ScaLAPACK replacement).
• The default ordering now uses METIS_NodeND, instead of the (undocumented) METIS_NodeNDP routine. This can impact performance, or for some problems lead to stack overflow, but for others it drastically reduces memory usage. The old behavior can be restored with --sp_enable_METIS_NodeNDP.

• Fix for setting CUDA/HIP device when there are multiple, but MPI was not initialized
• Memory leak fix in distributed memory GPU code
• Fixed small memory leaks from MPI datatypes
• Change in BLR algorithm selection options
• Changed default blocksize for 2D block cyclic distribution when using SLATE to 512
• Add 64bit support in the matching (MC64)
• Fix installation of Fortran modules

• Change default sparsity reducing ordering to use METIS_NodeNDP (from METIS_NodeND)
• Significant performance improvements in the GPU code for the direct solver, from the NERSC December 2021 Hackathon
• Performance fix in symbolic phase (affecting only some MPI implementations)
• Bump minimum CMake version to 3.17
• CMake fix for Perlmutter
• Compilation fix for GCC 8
• Add support for single precision HODLR/Butterfly (now requiring ButterflyPACK >= 2.1.0)