Created by W.Langdon from gp-bibliography.bib Revision:1.8243
The first part of this thesis proposes techniques to characterize how a processor exploits instruction-level parallelism. Based on a formal model, we explore ways to infer a processor’s port mapping from throughput measurements, i.e., how it splits instructions into micro-operations and how these are executed on the processor’s functional units. Our techniques enable accurate port mapping inference for processors that prior methods could not reason about.
In the second part, we introduce AnICA, a method to analyze inconsistencies between performance models. AnICA takes inspiration from differential testing and abstract interpretation to systematically characterize differences in the outputs of basic block throughput predictors. It can summarize thousands of inconsistencies in a few dozen descriptions that provide high-level insights into the differing behaviors of such predictors. These results have lead to improvements in the scheduling models of the widely used LLVM compiler infrastructure.",
Supervisor: Sebastian Hack",
Genetic Programming entries for Fabian Ritter