Performance Issues in HPF Implementations of Sensor-Based Applications

	       D. O'Hallaron, J. Subhlok, and J. Webb.

Abstract: Sensor-based computations are an important and often
overlooked application domain for HPF.  These applications typically
perform regular operations on dense arrays, and often have latency and
throughput requirements that can only be achieved with parallel
machines.  We have written a number of sensor-based applications
using a dialect of subset HPF that was developed at Carnegie Mellon.
The applications include FFT, synthetic aperture radar, narrowband
tracking radar, multibaseline stereo, and magnetic resonance imaging.
We have found that good performance is possible for these applications
on commercial machines such as the Intel Paragon.  In the paper we
identify three core operations that are key to achieving good
performance for sensor-based computations: parallel loops, index
permutations, and reductions and we discuss the implications for HPF
compilers. We also introduce some simple tests that HPF programmers
and implementors can use to measure the efficiency of the loops,
reductions, and permutations produced by an HPF compiler.

@article	(fxsensorbased97,
author	=	"D. O'Hallaron and J. Webb and J. Subhlok",
title	=	"Performance Issues in {H}igh {P}erformance {F}ortran Implementations of Sensor-based Applications",
journal	=	"Scientific Programming" ,
month	=	"Spring" ,
year	=	"1997" ,
volume	=	"6" ,
number	=	"1" ,
pages	=	"59--72" ,
)