Acceleration of Digital Radar Landmass Simulation on Multi-core CPU and GPGPU Computer
(Room S320C)
29 Nov 17
3:00 PM
-
3:30 PM
Digital Radar Landmass Simulation (DRLMS) for the purpose of training radar operators is a challenging computationally-intensive task. To improve fidelity, the databases are continuously increasing in resolution and density. Moreover, the radar simulation involves increasingly sophisticated physics-based models. One of them is the application of the radar antenna radiation pattern illuminating a particular region composed of landmass, sea clutter, precipitation and targets. To achieve real-time, a number of approximations are used in current simulations, which includes modeling a narrow antenna radiation pattern and employing a coarse sampling of the illuminated region. However, this harms the simulation fidelity.
To avoid these approximations while respecting the real-time constraints, this paper proposes a multi-level parallelization approach of landmass simulation applicable to multi-core Central Processing Unit (CPU) and General-Purpose Graphics Processing Unit (GPGPU). At the first level, the processing is divided into two parallel pipelines: (1) the power accumulation, which involves database processing and (2) the antenna pattern convolution. At the second level, the convolution is divided into several threads running in parallel. Two implementations are compared: one on multi-core CPU and one on GPGPU.
As benefits for training, we improve considerably the simulation performances as we are capable to apply a more detailed radar antenna pattern and support more complex databases, both contributing to more realistic radar images. The improved DRLMS shows respectively, a speedup of 12x on multi-core CPU running 16 threads, and a speedup of around 250x on a contemporary high-end graphics card over a one-thread execution on CPU.