spcl/daceml

Consider flattening maps for elementwise operations

tbennun opened this issue · 1 comments

Otherwise, parallelism opportunities are missed as blockDim.x is usually small on its own.

Will close once we port daceml-samples things