This repository is meant to help MLIR pass writers to generate MLIR files that often come from ML/HPC sources, to allow testing and developing their passes (as well as creating unit tests) with the current state of the known front-ends.
Warning: This is very much work-in-progress, don't expect half of it to work.
Each front-end has a prepare script, for example:
cd torch-script
./prepare.sh
It also has models ready to use (after prepared):
source venv/bin/activate
python models/linear.py
Results are in <front-end-name>/mlir.
If you want to build an existing front-end, you need to:
- Add a new directory for the front-end
- Add a
prepare.shscript that installs it and prepares the virtualenv - Add a
generate.shscript that generates all the MLIR files - Add models as Python files in
models, run the export, and save them as MLIR files inmlir