This PR adds CUDA and HIP support to FSAI according to a static pattern generation algorithm. The resulting method can also be used as a preconditioner for BoomerAMG. A detailed list of changes is given below:
* Add par_fsai_device.c
* Add hypre_FSAIApply
* Add function to dump local linear systems in dense format
* Implement static FSAI pattern computation via powers of A
* Improve filtering of candidate pattern
* Improve local linear systems extraction
* Add option for a 125pt matrix (27pt squared)
* Add options to control sizes of the memory pools with umpire
* Add hypre_GpuProfiling calls
* Improve candidate pattern truncation times
* Add max_nnz_row member and its private and public functions to FSAI
* Use max_nnz_row in FSAISetupDevice
* Add num_levels member and its private and public functions to FSAI
* Add threshold member and its public/private functions to FSAI
* Expose FSAI algorithm type to BoomerAMG
* Expose options to control FSAI setup
* Add cuSOLVER variables and calls
* Add batched dense linear solver calls to FSAI
* Improve execution time for generating random numbers
* Show FSAI parameters when amg_print_level >= 1
* Improve output of FSAIPrintStats
* Implement warp calls
* Add hypre_mask type and hypre_ballot_sync wrapper function
* Add hypre_popc and hypre_ffs wrapper functions
* Implement warp_allreduce_max calls
* Change: hypreDevice -> hypre_*Device
* Add rocSOLVER calls
* Apply astyle
* Remove redundant line
* Updated gcc compiler flags for strict-checking build option to throw floating point conversion warnings
* Several minor edits to clean up floating point conversion warnings and minor bugs.
* Updated saved files to reflect changes.
Thir PR adds a factorized sparse approximate inverse (FSAI) implementation on hypre, which can be used as a standalone solver, preconditioner to Krylov methods, or complex smoother to BoomerAMG. Particularly, we consider the adaptive algorithm version, where the sparsity pattern of the lower triangular factor G is built dynamically, i.e., during an iterative procedure that tries to find the best nonzero positions for a given row of G. This implementation was performed on top of the IJ interface. It uses the diagonal portion of A for constructing G, i.e., it's a block-Jacobi method in the MPI sense. List of additional changes:
* Add caliper instrumentation to FSAI.
* Add ZeroGuess option to FSAI.
* Performance optimizations.
* Add OpenMP support to FSAI.
* Make internal BLAS/LAPACK functions thread-safe.
* Update CMake build.
* Add new test cases: beam_tet_dof459_np1, beam_hex_dof459_np2, and beam_tet_dof2475_np4.
* Add documentation for FSAI.
Co-authored-by: Heather Switzer <switzer4@lassen36.coral.llnl.gov>
Co-authored-by: heatherms27 <hmswitzer@email.wm.edu>
Co-authored-by: Sarah Osborn <30503782+osborn9@users.noreply.github.com>
