Futhark.Optimise.MergeGPUBodies

This module implements an optimization pass that merges GPUBody kernels to eliminate memory transactions and reduce the number of kernel launches. This is useful because the Futhark.Optimise.ReduceDeviceSyncs pass introduces GPUBody kernels that only execute single statements.

To merge as many GPUBody kernels as possible, this pass reorders statements with the goal of bringing as many GPUBody statements next to each other in a sequence. Such sequence can then trivially be merged.

Documentation

mergeGPUBodies :: Pass GPU GPU

An optimization pass that reorders and merges GPUBody statements to eliminate memory transactions and reduce the number of kernel launches.