DBCSR is a library designed to efficiently perform sparse matrix matrix multiplication, among other operations. It is MPI and OpenMP parallel and can exploit GPUs via CUDA.
You absolutely need:
- GNU make
- a Fortran compiler which supports at least Fortran 2003 (respectively 2008+TS when using the C-bindings)
- a LAPACK implementation (reference, OpenBLAS-bundled and MKL have been tested)
- a BLAS implementation (reference, OpenBLAS-bundled and MKL have been tested)
- a Python version installed (2.7 and 3.5+ have been tested)
Optionally, you can install libxsmm.
To build with CUDA support you further need:
- CUDA Toolkit
- a C++ compiler which supports at least C++11 standard
We test against GNU and Intel compilers.
Download either a release tarball or clone the latest version from Git using:
git clone --recursive https://github.com/cp2k/dbcsr.git
Run
make help
to list all possible targets.
Update the provided Makefile.inc to fit your needs (read the documentation inside the file for further explanations) and then run
make <target>
Support for cmake
is still considered experimental and may not cover all functionalities yet.
If you are using cmake
to build DBCSR, please make sure you can reproduce any errors using the plain Makefile
before reporting them.
Some examples on how to use the library (which is the only current documentation) are available under the examples directory (see readme).
You can compile with
make CINT=1
to generate the C interface. Make sure your Fortran compiler supports F2008 standard (including the TS) by updating the flag in the Makefile.inc.