VASP porting and parallel optimization on GPU like accelerator
Hu Shibiao, Peng Wu, Wudi Cao, Jiu Liu, Xudong Tang · 2023
Vienna ab initio Simulation Package(VASP) is software for performing electronic structure calculations and quantum mechanical molecular dynamics simulations, and is widely used in materials simulation and computational material science research. Currently VASP is accelerated on NVIDIA GPUs via the OpenACC programming model, which cannot be directly compiled and used by domestic GPU like accelerator platform. We use the HIP programming model to port VASP to GPU like accelerator. Refer to the OpenACC port of VASP, we use the HIP API for device and data management, write kernel to implement cyclic accelerated computation, and use mathematical libraries such as hipBLAS and hipFFT to support the computation. After the porting was completed we compare the computed results of the HIP port version of VASP with the CPU version and the OpenACC version. We passed the validation of the correctness of the HIP ported version. We also analyze the principle of All-to-All parallel communication and propose optimization strategies such as aggregated communication and merged copy for the case of multiple bands communication, which are implemented on the HIP port version. We choose the B.hR105 testcase to test the All-to-All before and after optimization. The results demonstrate a 25.57% performance improvement after optimization.