DBCSR¶
To configure, build, and install DBCSR for CP2K, one may proceed as shown below.
git clone --recursive https://github.com/cp2k/dbcsr.git
cd dbcsr
wget --content-disposition https://github.com/hfp/xconfigure/raw/main/configure-get.sh
chmod +x configure-get.sh
./configure-get.sh dbcsr
There are four flavors, which differ by the accelerator backend ("cpu" or "gpu") and by the kind of library (static or "shared"). The build relies on the GNU Compiler Collection and on MPI compiler wrappers (mpicc, mpicxx, mpif90). LIBXS is used for host-side small matrix multiplications, and the GPU flavors rely on OpenCL and LIBXSTREAM. LIBXS, LIBXSMM (optional), and LIBXSTREAM are looked up next to DBCSR (e.g., ../libxs), via LIBXSROOT (LIBXSMMROOT, LIBXSTREAMROOT), and in $HOME (e.g., $HOME/libxs), and the found locations are printed. If built with GNU Make, the matching lib/<target> of LIBXS, LIBXSMM, or LIBXSTREAM is used (e.g., libxs/lib/x86_64). LIBXSMM is enabled if found. Other locations can be supplied using the CMAKE_PREFIX_PATH environment variable.
./configure-dbcsr-gpu-shared.sh
cmake --build build-<arch>-gnu-gpu-shared --target install -j $(nproc)
Here, <arch> names the target as determined by the compiler (e.g., sapphirerapids or neoverse-v2), which allows to share the folder structure among different systems (XCFG_ARCH overrides the name). The above installs into ../dbcsr/<arch>/gnu-gpu-shared, and similarly for configure-dbcsr-cpu.sh (gnu-cpu), configure-dbcsr-cpu-shared.sh (gnu-cpu-shared), and configure-dbcsr-gpu.sh (gnu-gpu). An optional first argument changes the "gnu" part of the name, and further arguments are passed to CMake, e.g., to select another GPU backend:
./configure-dbcsr-gpu.sh gnu -DUSE_ACCEL=cuda -DWITH_GPU=H100
BLAS and LAPACK are found by CMake (e.g., OpenBLAS or Intel MKL). With NVPL=1, NVPL's BLAS and LAPACK are used instead (e.g., on NVIDIA Grace, NVPL_ROOT if not in a system location), and CP2K must be configured with NVPL=1 as well to link the same BLAS library.