A Directive-Based Heterogeneous SPH Solver for Portable Performance on CPUs and GPUs

Smoothed Particle Hydrodynamics (SPH) is powerful but computationally intensive. While GPU acceleration via low-level APIs like CUDA offers immense speedups, it often leads to high maintenance costs, reduced portability, and limitations in numerical precision (FP32). We present a high-performance SPH framework that achieves performance portability across multi-core CPUs and GPUs using directive-based programming (OpenACC and OpenMP). Our approach implements key algorithmic optimizations for neighbor searching and memory access, enabling directives to effectively parallelize the code with minimal refactoring. We demonstrate two orders of magnitude of speedups for double-precision (FP64) multiphase simulations on GPUs compared to optimized multi-core CPU runs. Furthermore, we showcase the framework's portability by successfully migrating the solver from OpenACC to OpenMP. This work provides a practical pathway for developing efficient, maintainable, and portable SPH codes for modern heterogeneous HPC architectures.