Harmonic-summing Module of SKA on FPGA--Optimising the Irregular Memory Accesses

TL;DR

This paper explores FPGA-based optimization techniques for the harmonic-summing module in SKA's pulsar search pipeline, focusing on irregular memory access management to improve energy efficiency and performance.

Contribution

It proposes and evaluates two FPGA approaches for harmonic-summing, including data preloading strategies, demonstrating energy efficiency advantages over GPUs.

Findings

Preloading all necessary points with input reordering is fastest.

FPGAs consume significantly less energy than GPUs for the same kernels.

FPGA approaches can effectively optimize irregular memory accesses in pulsar search.

Abstract

The Square Kilometre Array (SKA), which will be the world's largest radio telescope, will enhance and boost a large number of science projects, including the search for pulsars. The frequency domain acceleration search is an efficient approach to search for binary pulsars. A significant part of it is the harmonic-summing module, which is the research subject of this paper. Most of the operations in the harmonic-summing module are relatively cheap operations for FPGAs. The main challenge is the large number of point accesses to off-chip memory which are not consecutive but irregular. Although harmonic-summing alone might not be targeted for FPGA acceleration, it is a part of the pulsar search pipeline that contains many other compute-intensive modules, which are efficiently executed on FPGA. Hence having the harmonic-summing also on the FPGA will avoid off-board communication, which could destroy other acceleration benefits. Two types of harmonic-summing approaches are investigated in this paper: 1) storing intermediate data in off-chip memory and 2) processing the input signals directly without storing. For the second type, two approaches of caching data are proposed and evaluated: 1) preloading points that are frequently touched 2) preloading all necessary points that are used to generate a chunk of output points. OpenCL is adopted to implement the proposed approaches. In an extensive experimental evaluation, the same OpenCL kernel codes are evaluated on FPGA boards and GPU cards. Regarding the proposed preloading methods, preloading all necessary points method while reordering the input signals is faster than all the other methods. While in raw performance a single FPGA board cannot compete with a GPU, in terms of energy dissipation, GPU costs up to 2.6x times more energy than that of FPGAs in executing the same NDRange kernels.