iree Use mlir's map_nested_foreach_thread_to_gpu

Use mlir's map_nested_foreach_thread_to_gpu_blocks

Open grypp opened this issue 2 years ago • 1 comments

MLIR's transform dialect now has capability to map scf.foreach_thread to GPU blocks. https://reviews.llvm.org/D134190

This PR makes iree specific transform dialect operation (iree.foreach_thread_to_workgroup) to use mlir's new feature.

It also renames workgroup to gridDim in order to have same naming convention with the mlir's gpu dialect.

Sep 23 '22 15:09 grypp

I have some tricking changes that I am trying to land in #10527, they will conflict with this, sorry .. Can you please rebase on top?

Sep 23 '22 16:09 nicolasvasilache

It also renames workgroup to gridDim in order to have same naming convention with the mlir's gpu dialect.

This is not a good change, IREE uses Vulkan terminology pervasively for CPU, GPU and mobile. MLIR's GPU dialect uses CUDA terminology.

Please revert this change.

Sep 26 '22 08:09 nicolasvasilache