A Framework for Resource-Constrained Rate-Optimal Software Pipelining

The rapid advances in high-performancecomputer architectureand compilationtechniques provide both challenges and opportunitiesto exploitthe rich solution space of software pipelined loop schedules. In this paper, we develop a framework to construct a software pipelinedloop schedulewhich runs on the given architecture(with a fixed number of processorresources) at the maximum possible iteration rate (a la rate-optimal)while minimizingthe number of buffers-a close approximationto minimizing the number of registers. The main contributionsof this paper are: First,we demonstratethat such problem can be describedby a simple mathematical formulationwith preciseoptimizationobjectivesunder a periodic linear schedulingframework. The mathematicalformulation provides a clear picturewhich permitsone to visualizethe overall solution space (for rate-optimalschedules) under different sets of constraints. Secondly, we show that a precise mathematicalformulationand its solution does make a significantper...

• Type: Conference
• Year: 1994
• Where: CONPAR
• Authors: Ramaswamy Govindarajan, Erik R. Altman, Guang R. Gao
• Keywords: CONPAR 1994, Distributed And Parallel Computing, Exploitthe Rich Solution, Software Pipelined Loop, Solution Space