High-performance libraries

Various high-performance libraries for dense linear algebra, fast Fourier transforms, etc. are available via the module system. Many libraries are provided both as single-threaded and multithreaded versions, multithreaded modules are designated with omp in the module version. For pure MPI applications and applications calling libraries from multiple threads it is recommended to use a single- threaded library.

Availability of libraries may depend on the loaded compiler suite and MPI environment, use module avail for finding out available libraries. See the documentation of the library for instructions on how to build against it. Note that most modules set LIBRARY_PATH and LD_LIBRARY_PATH environment variables so that the -llibrary linker flag is often enough. Most modules set also <library>_INSTALL_ROOT environment variables that can be utilized in custom build scripts. As an example, the fftw library can be used as follows:

module load fftw
<compiler_command> -o myprog myprog.o -lfftw3

and the directory containing include, lib, etc. is found under the FFTW_INSTALL_ROOT environment variable.

Libraries on Roihu-CPU

Selected libraries available on Roihu-CPU:

Dense linear algebra: openblas, amdblis
Dense distributed linear algebra: netlib-scalapack
Fast Fourier transforms: fftw

Libraries on Roihu-GPU

Selected libraries available on Roihu-GPU:

For Grace CPU:
Dense linear algebra: nvhpc (includes NVPL), openblas, netlib-lapack
Dense distributed linear algebra: nvhpc (includes NVPL), netlib-scalapack
Fast fourier transforms: nvhpc (includes NVPL), fftw
For Hopper GPU:
CUDA (module cuda/nvhpc) includes libraries such as cublas, cufft, cusolver, ...