15 April 2020
DFT and GW towards the exascale: porting MaX codes on to GPUs
MaX developers are working to support the European HPC community. As of today, a number of MaX codes as well as core libraries have been released production-ready with GPU-support on heterogeneous architectures.
This result, obtained through a deep refactoring of the main codes, is key to better exploit the potential offered by the future pre-exascale and exascale EuroHPC machines. The porting on GPUs has been done by adopting different GPU-aware programming models and libraries: CUDA-Fortran for Quantum ESPRESSO, Yambo and FLEUR, CUDA and OpenCL for BigDFT, CUDA and ROCm for SIRIUS and the DBCSR, COSMA and SpFFT libs. Noticeably, we have addressed the possibility of delivering GPU acceleration as a library feature, like for the case of PSolver and the forthcoming libconv (BigDFT) and DBCSR--SpFFT (CP2K, CSCS).