Register pressure aware scheduling for high level synthesis

Variations of list scheduling became the de-facto standard of scheduling straight line code in software compilers, a trend faithfully inherited by high-level synthesis solutions. Due to its nature, list scheduling is oblivious of the tightly coupled register pressure; a dangling fundamental problem...

Full description

Saved in:
Bibliographic Details
Published in16th Asia and South Pacific Design Automation Conference (ASP-DAC 2011) pp. 461 - 466
Main Authors Beidas, R, Wai Sum Mong, Jianwen Zhu
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.01.2011
Subjects
Online AccessGet full text
ISBN9781424475155
1424475155
ISSN2153-6961
DOI10.1109/ASPDAC.2011.5722234

Cover

More Information
Summary:Variations of list scheduling became the de-facto standard of scheduling straight line code in software compilers, a trend faithfully inherited by high-level synthesis solutions. Due to its nature, list scheduling is oblivious of the tightly coupled register pressure; a dangling fundamental problem that has been attacked by the compiler community for decades, and which results, in case of highlevel synthesis, in excessive instantiations of registers and accompanying steering logic. To alleviate this problem, we propose a synthesis framework called soft scheduling, which acts as a resource unconstrained pre-scheduling stage that restricts subsequent scheduling to minimize register pressure. This optimization objective is formulated as a live range minimization problem, a measure shown to be proportional to register pressure, and optimally solved in polynomial time using minimum cost network flow formulation. Unlike past solutions in the compiler community, which try to reduce register pressure by local serialization of subject instructions, the proposed solution operates on the entire basic block or hyperblock and systematically handles instruction chaining subject to the same objective. The application of the proposed solution to a set of real-life benchmarks results in a register pressure reduction ranging, on average, between 11% and 41% depending on the compilation and synthesis configurations with minor 2% to 4% increase in schedule latency.
ISBN:9781424475155
1424475155
ISSN:2153-6961
DOI:10.1109/ASPDAC.2011.5722234