This repository is archived, the current effort to productionize finch is finch-tensor-lite

An out-of-tree MLIR dialect

This is an example of an out-of-tree MLIR dialect along with a finch opt-like tool to operate on that dialect.

Install

0. Do a recursive clone

git clone --recursive https://github.com/finch-tensor/Finch-mlir.git

Or if you have already cloned this repository, run

git submodule init --update --recursive

This will take some time as it perform a shallow-clone of LLVM.

1. Install prerequisites to build the dialect and the Python bindings:

1a. Use your system package manager or conda to install a C/C++ compiler and CCache.

1b. Install the build-time Python requirements:

# Generate the build requirements dynamically from `pyproject.toml` and install them.
./scripts/install_build_reqs.sh

2. Build the wheel or install the package:

python -m build --no-isolation

OR

pip install --no-build-isolation .

In theory, the --no-[build-]isolation flag is unnecessary, but it caches the LLVM build which can greatly speed-up the buid process on repeated runs.

Run Tests

One can run the tests once the wheel is built and/or the package is installed with

cmake --build build --target check-finch

Assuming the CMake configure step doesn't need to run.