With *** it's straightforward, but Arrow syntax doesn't usually desugar to it. Therefore the more aggressive optimization attempt is >>>, although it will often have data dependencies that will prevent parallelism.
It should be measured with a few realistic benchmarks whether this is actually an optimization, or even a pessimization.
Another place to check for optimization gains would be the particle filter in rhine-bayes.
In https://github.com/turion/rhine/pull/299/files/f8f72050f4c69ad74864370738e9ef90a86ce3eb#diff-6421dff49f75037ffc55f3a3ceeed8653e196c490d41d5601a55d30a5d69525b I proposed, that when dunai MSFs are replaced by automata, we could by default use GHC parallelism (
parandpseq) whenever automata are combined, with***,&&&,<*>, or even with>>>/..With
***it's straightforward, but Arrow syntax doesn't usually desugar to it. Therefore the more aggressive optimization attempt is>>>, although it will often have data dependencies that will prevent parallelism.It should be measured with a few realistic benchmarks whether this is actually an optimization, or even a pessimization.
Another place to check for optimization gains would be the particle filter in
rhine-bayes.