Abstract
Practical aperture synthesis imaging algorithms work by iterating between estimating the sky brightness distribution and a comparison of a prediction based on this estimate with the measured data ("visibilities"). Accuracy in the latter step is crucial but is made difficult by irregular and non-planar sampling of data by the telescope. In this work we present a GPU implementation of 3d de-gridding which accurately deals with these two difficulties and is designed for distributed operation. We address the load balancing issues caused by large variation in visibilities that need to be computed. Using CUDA and NVidia GPUs we measure performance up to 1.2 billion visibilities per second.
Introduction
In radio interferometry, telescopes measure "visibilities"—samples of the Fourier transform of the sky's brightness. Creating an image requires iterative algorithms like CLEAN. A core component is degridding: interpolating a regular grid (the sky model) back to the irregular, non-planar coordinates of the telescope's data collection. This is computationally intensive due to irregular sampling, the 3D nature of the "w-term" (baseline dimension), and significant load balancing challenges where data density varies across the sky.
Methodology
The implementation utilizes NVIDIA's CUDA platform and Gridding Convolution Functions (GCF) for 3D interpolation. Four kernel strategies were evaluated:
- D1 (Optimized y-reduction): Uses a single warp (32 threads) for a batch of visibilities, employing shared memory and a two-phase deconvolution (y-axis accumulation followed by parallel reduction).
- D2 (Basic): One threadblock processes one visibility.
- D3 (Dynamic Parallelism): Employs NVIDIA’s Dynamic Parallelism to launch kernels based on subgrid needs, targeting load-balancing.
- D4 (z-reduction): Uses 64 threads per block with reduction in the z-direction (w component) first, requiring inter-warp synchronization.
Performance Evaluation
The kernels were stress-tested using simulated visibility distributions in two scenarios: constant visibilities per subgrid to measure peak throughput, and varying visibilities to simulate real-world unbalanced observations.
Results and Discussion
The experiments revealed a distinct "crossover" point in kernel efficiency. As shown in the performance plots, for lower visibility counts (below 262,000 per subgrid), the D4 (z-reduction) kernel is superior. However, for higher densities, the D1 (y-reduction) kernel achieves better performance, eventually reaching a peak throughput of 1.2 billion visibilities per second.
D1's efficiency stems from confining operations to a single CUDA warp, avoiding expensive synchronization barriers. While Dynamic Parallelism (D3) improved the basic kernel's handling of load imbalance, it did not surpass the manually optimized memory management of D1 and D4.
Conclusion
This work demonstrates a high-performance 3D degridding implementation capable of handling the massive data rates expected in future arrays like the Square Kilometre Array (SKA). Future work includes integrating these kernels into full imaging pipelines like WSCLEAN and exploring cross-platform performance on non-NVIDIA hardware.
Summary updated