Accelerating Lattice QCD Multigrid on GPUs Using Fine-Grained Parallelization

摘要

The past decade has witnessed a dramatic acceleration of lattice quantumchromodynamics calculations in nuclear and particle physics. This has been dueto both significant progress in accelerating the iterative linear solvers usingmulti-grid algorithms, and due to the throughput improvements brought by GPUs.Deploying hierarchical algorithms optimally on GPUs is non-trivial owing to thelack of parallelism on the coarse grids, and as such, these advances have notproved multiplicative. Using the QUDA library, we demonstrate that by exposingall sources of parallelism that the underlying stencil problem possesses, andthrough appropriate mapping of this parallelism to the GPU architecture, we canachieve high efficiency even for the coarsest of grids. Results are presentedfor the Wilson-Clover discretization, where we demonstrate up to 10x speedupover present state-of-the-art GPU-accelerated methods on Titan. Finally, welook to the future, and consider the software implications of our findings.