BLASTed

A collection of sparse matrix containers and manipulators, along with certain linear algebra operations. Note that this is not a complete sparse BLAS library; instead, the focus is on providing thread-parallel low-level kernels required to implement preconditioners for large sparse linear systems of algebraic equations, in an object-oriented framework. In case of a distributed-memory parallel solver, the operations implemented here are meant to be used in the subdomains' local preconditioners.

The main contribution of this library is a practical implementation of asynchronous iterations as preconditioners. These include scalar and point-block versions of 'asynchronous symmetric Gauss-Seidel' iterations developed as part of this work, an independent implementation of an asynchronous incomplete LU factorization preconditioner, as well as an asynchronous point-block ILU factorization preconditioner with asynchronous block-triangular solves. The advantages of point-block asynchronous ILU preconditioners, especially for compressible fluid flow problems, are documented in this preprint submitted to SIAM.

Building

The following programs and libraries are required:

• CMake version 3.0 or later
• Eigen version 3.3.4 or later (an even more recent version is needed for the GCC 7 series) [EIGEN3_ROOT]
• Boost [BOOST_ROOT]
• PETSc 3.8 or above and MPI are required to build the PETSc interface. [PETSC_DIR, PETSC_ARCH] Note that for the release build of BLASTed, PETSc must be configured with --with-memalign=64 on x86_64 platforms.

Quick-build: Assuming the programs and libraries mentioned above are in standard locations, in your PATH or LD_LIBRARY_PATH environment variables, or you have otherwise made them available by setting the environment variables mentioned in square brackets above, cd to the top-level directory of the BLASTed sources and try:

mkdir build && cd build
../build_example.sh
make -j4
make test

If this does not work, you may need to run cmake manually and change some settings.

Long version: Assuming that you are in the top-level BLASTed directory, to configure a release version with AVX vectorization and the PETSc interface, type

mkdir build && cd build
cmake -DAVX=1 -DWITH_PETSC=1 -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx -DCMAKE_BUILD_TYPE=Release ..

where mpicc and mpicxx are the C and C++ MPI compiler wrappers you want to use. This will build the library for use with PETSc with available block sizes 4 and 5 (the default block sizes). To build without the PETSc interface, -DWITH_PETSC should be removed. -DBUILD_BLOCK_SIZE=10 can be specified to additionally build the block solver operations for a block size of 10, for instance. See the beginning of the top-level CMakeLists.txt file for all the options. To build,

make -j4

and to run the tests,

make test

A C++ compiler with C++ 14 support is required. The build is known to work with GCC 7.3, 8.1, 8.2, 9.1, Intel 2017, 2018, and Clang 6.0, 7.0 in a GNU/Linux environment. To build in other enviroments, tweaking the CMakeLists.txt file will be required. This build process generates a shared object build/lib/libblasted_petsc.so that should be linked to your application.

To build the Doxygen documentation,

cd path/to/BLASTed/doc
doxygen blasted_doxygen.cfg

This will build HTML documentation in a subdirectory called html in the current directory.

Finally, from the build directory, one can issue

make tags

to generate a tags file for easier navigation of the source code in Vim. If you are using Spacemacs, this is not needed; just use the shortcut SPC p G.

Usage

For the most part, one would want to use the library as a plugin for PETSc. This is explained in the usage documentation.

---