Contribution of the Resampling Stage to the Execution Time of Particle Filter

Particle filter is a serial Monte Carlo estimation algorithm. It represents the posterior probability density function with particles and their weights. As time progresses, the normalized weight of one of particles becomes nearly one, while the normalized weights of the remaining ones get close to zero. A common way to solve this problem, known as the degeneracy problem, is resampling. In resampling, the particles with larger weights are replicated, and the particles with smaller weights are eliminated. To tackle the numerical instability problem that is encountered by some of the resampling methods, the Metropolis resampling method is proposed by Murray and his co-workers. Unfortunately, Metropolis is liable to non-coalesced global memory access patterns on the GPU. In this work, we point to the Metropolis-C1 and Metropolis-C2 resampling methods which are proposed earlier. Then we examine the contribution of the stages of the particle filter to the total execution time by increasing the number of particles on a tracking application on the GPU. We use the Sampling Importance Resampling (SIR) method, which is a common particle filter. In the experiments, Metropolis resampling consumes the biggest portion of the execution time of the SIR particle filter. The share of Metropolis increases as the number of particles grows. It can be argued that this is because of non-coalesced global memory access patterns. To reach this conclusion it is sufficient to i) Compare the results of Metropolis, which has non-coalesced access patterns, with Metropolis-C1 and Metropolis-C2, which have confined non-coalesced access patterns, ii) See that the previous stages of the SIR particle filter are not subject to non-coalesced access patterns.