Smoothed-Particle Hydrodynamics with ARM Scalable Vector Extension

Sammanfattning

This bachelor degree project investigates how intense Computation Fluid Dynamics methods, specifically the computationally demanding naive neighbour search in Smoothed Particle Hydrodynamics, can benefit from vectorization using Arm’s Scalable Vector Extension. Two C++ versions of a 2D naive neighbour search were implemented: one scalar and one vectorized utilizing SVE intrinsics. The vectorized version aims to accelerate the search by concurrently comparing distances between a particle and multiple other particles. Performance, in terms of clock cycles, was evaluated using the Gem5 simulator for SVE vector lengths from 128 to 2048 bits across various particle counts (16 to 65,536). The vectorized implementation demonstrated significant performance improvements for vector lengths of 256 bits and above, achieving up to an 11.14x speedup compared to the scalar version (at 2048-bit vectors with 65,536 particles). However, at the 128-bit vector length, the vectorization introduced overhead, resulting in a higher clock cycle count than the scalar baseline. The study confirms that for this naive neighbour search, increasing SVE vector length (beyond 128-bits) substantially reduces clock cycles. Finally, the results indicate that the speedups achieved from different vector lengths increased with problem size.