An efficient work-distribution strategy for gridding radio-telescope data on GPUs

Abstract

This paper presents a novel work-distribution strategy for GPUs, that efficiently convolves radio-telescope data onto a grid, one of the most time-consuming processing steps to create a sky image. Unlike existing work-distribution strategies, this strategy keeps the number of device-memory accesses low, without incurring the overhead from sorting or searching within telescope data. Performance measurements show that the strategy is an order of magnitude faster than existing accelerator-based gridders. We compare CUDA and OpenCL performance for multiple platforms. Also, we report very good multi-GPU scaling properties on a system with eight GPUs, and show that our prototype implementation is highly energy efficient. Finally, we describe how a unique property of GPUs, fast texture interpolation, can be used as a potential way to improve image quality.

Faster GPU-based convolutional gridding via thread coarsening

“…As a workaround, we added the statement asm("") as a compiler-level memory barrier. Romein (2012) found that the majority of memory traffic was due to cache misses in reading the convolution GCF values. To avoid this problem, we have used a separable approximation to the GCF (Merry, 2016).…”

“…However, the irregular data access patterns make this a non-trivial task. One of the first really practical algorithms for GPU-accelerated gridding is due to Romein (2012). Despite being state of the art, it typically spends only about 25% of a GPU's compute power on the actual convolution operations.…”

The Square Kilometre Array Science Data Processor. Preliminary compute platform design

“…We can investigate computational efficiency of the most costly algorithmic components, an estimate based on our current undestanding of the required processing, on current day bestof-breed hardware. This shows very poor efficiency of at most 20% of R peak [1], [10].…”

A Scalable Hybrid Fpga/Gpu Fx Correlator

“…While this may be sufficient for some applications, in the general case other elements in the data path pose unsolved problems of scale owing to dependence on at least N 2 . The computation challenges associated with gridding irregularly spaced visibilities in preparation for FFT imaging Romein, (2012), and subtraction of sky models from correlator output in the visibility domain Mitchell, et al (2008), for example, will also need to be addressed.…”