LLVM’s OpenMP Loop Transformations to Enhance Locality in OpenMP Loop Scheduling
DescriptionThis talk will present OpenMP loop transformations, e.g., tiling, in a way that complements existing and novel dynamic schedules for OpenMP loops and OpenMP tasking. It can be used for auto-tuning for obtaining the optimal composition of multiple loop transformations and parameters without the cost of node-level dynamic scheduling, in particular to maximize spatial and temporal data locality. The loop transformations were first introduced in OpenMP 5.1 and will be further developed in OpenMP 6.0. We also briefly discuss iteration space reflection and mapping to different levels of the hardware hierarchy of parallelism. The work is motivated by research activities in the US DoE project PROTEAS-TUNE.
TimeWednesday, June 2815:00 - 15:30 CEST
Computer Science, Machine Learning, and Applied Mathematics