Acceleration of dRMSD calculation and efficient usage of GPU caches

In this paper, we introduce the GPU acceleration of dRMSD algorithm, used to compare different structures of a molecule. Comparing to multithreaded CPU implementation, we have reached 13.4× speedup in clustering and 62.7× speedup in I:I dRMSD computation using mid-end GPU. The dRMSD computation exposes strong memory locality and thus is compute-bound. Along with conservative implementation using shared memory, we have decided to implement variants of the algorithm using GPU caches to maintain memory locality. Our implementation using cache reaches 96.5% and 91.6% of shared memory performance on Fermi and Maxwell, respectively. We have identified several performance pitfalls related to cache blocking in compute-bound codes and suggested optimization techniques to improve the performance.