[Paper Review] Fully Parallel Particle Learning for GPGPUs and Other Parallel Devices

Kenichiro McAlinn, Teruo Nakatsuma|arXiv (Cornell University)|Dec 7, 2012

Target Tracking and Data Fusion in Sensor Networks25 references3 citations

TL;DR

This paper proposes a fully parallel resampling algorithm for particle filters that executes all particle filtering steps—likelihood computation, CDF construction, resampling, and propagation—entirely on GPUs, eliminating CPU-GPU data transfers. The method achieves up to 248× speedup on the CDF step and 45.3× on propagation, demonstrating superior performance over sequential CPU implementations while maintaining exact resampling via a cut-point method.

ABSTRACT

We develop a novel parallel resampling algorithm for fully parallelized particle filters, which is designed with GPUs (graphics processing units) or similar parallel computing devices in mind. With our new algorithm, a full cycle of particle filtering (computing the value of the likelihood for each particle, constructing the cumulative distribution function (CDF) for resampling, resampling the particles with the CDF, and propagating new particles for the next cycle) can be executed in a massively and completely parallel manner. One of the advantages of our algorithm is that every single numerical computation or memory access related to the particle filtering is executed solely inside the GPU in parallel, and no data transfer between the GPU's device memory and the CPU's host memory occurs unless for further processing, so that it can circumvent the limited memory bandwidth between the GPU and the CPU. To demonstrate the advantage of our parallel algorithm, we conducted a Monte Carlo experiment in which we apply the parallel algorithm as well as conventional sequential algorithms for estimation of a simple state space model via particle learning, and compare them in terms of execution time. The results show that the parallel algorithm is far superior to the sequential algorithm.

Motivation & Objective

• To address the computational bottleneck of particle filtering in high-dimensional, nonlinear, non-Gaussian state space models.
• To eliminate CPU-GPU data transfers during particle filtering cycles, minimizing bandwidth limitations.
• To develop a fully parallel resampling algorithm that is both exact and portable across parallel architectures.
• To demonstrate significant performance gains over sequential CPU-based particle filtering using GPU acceleration.

Proposed method

• The algorithm performs all particle filtering steps—likelihood evaluation, CDF construction, resampling, and propagation—within GPU device memory, avoiding host-CPU memory transfers.
• It uses a cut-point method for resampling that enables exact sampling from the cumulative distribution function (CDF) in parallel.
• The method avoids device-specific GPU features, ensuring portability to multi-GPU or grid computing systems.
• Resampling is implemented via parallel prefix sums and exclusive scan operations to construct selection indices efficiently.
• The algorithm is designed for single-instruction-multiple-data (SIMD) execution on massively parallel architectures like GPGPUs.
• A Monte Carlo experiment compares execution times between the GPU-parallel and CPU-sequential implementations on a simple state space model.

Experimental results

Research questions

• RQ1 Can a fully parallel particle filtering algorithm be designed to eliminate CPU-GPU data transfers while maintaining exact resampling?
• RQ2 How does the performance of a fully parallel GPU-based particle filter compare to a conventional sequential CPU-based implementation?
• RQ3 What is the impact of different computational steps (CDF, propagation, resampling) on overall execution time in a GPU environment?
• RQ4 Can the proposed algorithm be implemented without relying on device-specific GPU functionalities?

Key findings

• The proposed GPU-based algorithm achieved a 248× speedup on the CDF construction step compared to the sequential CPU implementation.
• The propagation step saw a 45.3× speedup, demonstrating strong parallel efficiency on GPU architectures.
• The resampling step achieved an 11.9× speedup, with a 2.7× gain even when excluding sorting overhead, outperforming stratified resampling on CPU.
• The Store step—data transfer from GPU to CPU—accounted for 15–20% of total execution time, highlighting memory transfer as a bottleneck.
• The algorithm’s performance advantage is most pronounced for large particle counts and long time horizons, where initialization time is negligible.
• The method is fully portable and can be extended to multi-GPU or grid computing systems without architectural dependencies.