mkdir testing
cd testing
/regalloc-testing/scripts/test_file.sh /regalloc-testing/benchmarks/fannkuch-redux.c
All of the results should be placed in a results.txt file in the same directory.
Once you have the results.txt file, you can also use the evaluate_no_weights.py
script to automatically evaluate how well the model performed in terms of the
average absolute difference between the regalloc score and the actual performance
relative to the baseline and in terms of how often if correctly found the polarity.
python3 /reglloc-testing/scripts/evaluate_no_weights.py results.txt
This will print the calculated information out to STDOUT.
The docker image that is described in Dockerfile contains all the tooling
need to run all the experiments within this repository. To build it, you
also need the ml-compiler-opt
repository checked out. Build the development docker image in that repository,
following the instructions in experimental/docker/README.md. This will give
an image tagged mlgo-development. Then you can build the Docker image
present in this repository:
cd /path/to/regalloc-cost-model-evaluation
docker build -t regalloc-cost-model-evaluation .This will take a while and consume quite a bit of disk space as it does quite a few things to set up a development environment that allows for a lot of experimentation.
Starting the docker container is pretty simple:
docker run -it --name regalloc-testing --privileged regalloc-cost-model-evaluationNote the presence of the --privileged flag. Some of the functionality present
in the scripts within this repository will not work if you do not run the
container with this flag (particularly perf recording).