Fast, feature-rich weakly-compressible SPH on GPU: coding strategies and compiler choices

Abstract: GPUSPH was the first implementation of the weakly-compressible Smoothed Particle Hydrodynamics method to run entirely on GPU using CUDA. Version 5, released in June 2018, features a radical restructuring of the code, offering a more structured implementation of several features and specialized optimization of most heavy-duty computational kernels. While these improvements have led to a measurable performance boost (ranging from 15\% to 30\% depending on the test case and hardware configuration), it has also uncovered some of the limitations of the official CUDA compiler (\texttt{nvcc}) offered by NVIDIA, especially in regard to developer friendliness. This has led to an effort to support alternative compilers, particularly Clang, with surprising performance gains.