gcc_autotuning

Autotuning for GCC command line flags

This directory contains scripts to autotune the GCC environment using black box optimization strategies.

• Usage
  • Running autotuning experiments
  • Overview of results
• Reproducing the experiments from our paper

Usage

There are two scripts, tune.py and info.py. The tuning script is used to run the autotuning experiments and record log the results to CSV files. The info script reads and aggregates results over many multiple CSV files.

Running autotuning experiments

Invoke the tuning script using:

python -m gcc_autotuning.tune <arguments>

Where <arguments> is a list of configuration arguments. Key configuration options are:

Argument	Description
--gcc_bin	Binary to use for GCC. Use docker:<image> for docker.
--gcc_benchmark	List of benchmarks to autotune. Defaults to the 12 CHStone benchmarks.
--search	Types of autotuning to perform.
--gcc_search_budget	Maximum number of compilations per benchmark.
--gcc_search_reptitions	Number of times to repeat each search.
--objective	Objective function to optimize for. One of {obj_size,asm_size}.
--output_dir	Base output directory. Default: ~/logs/compiler_gym/gcc_autotuning.

For example, to run a genetic algorithm search with a budget of 100 compilations on two CHStone benchmarks:

python -m gcc_autotuning.tune \
    --gcc_bin=docker:gcc:11.2.0 \
    --seed=204 \
    --search=genetic \
    --gcc_search_budget=100 \
    --gcc_search_repetitions=3 \
    --gcc_benchmark=benchmark://chstone-v0/aes,benchmark://chstone-v0/sha

See --helpfull for a full list of arguments.

Overview of results

Summarize autotuning results using:

python -m gcc_autotuning.info

This aggregates results in the default ~/logs/compiler_gym/gcc_autotuning directory. Specify one or more different directories as command line arguments, e.g.:

python -m gcc_autotuning.info /path/to/logs/dir/a ~/logs_dir_b

Reproducing the experiments from our paper

To reproduce the experiments in Section VII.D of our paper, run:

python -m gcc_autotuning.tune \
    --gcc_bin=docker:gcc:11.2.0 \
    --seed=204 \
    --search=random,hillclimb,genetic \
    --gcc_search_budget=1000 \
    --gcc_search_repetitions=3 \
    --gcc_benchmark=benchmark://chstone-v0/adpcm,benchmark://chstone-v0/aes,benchmark://chstone-v0/blowfish,benchmark://chstone-v0/dfadd,benchmark://chstone-v0/dfdiv,benchmark://chstone-v0/dfmul,benchmark://chstone-v0/dfsin,benchmark://chstone-v0/gsm,benchmark://chstone-v0/jpeg,benchmark://chstone-v0/mips,benchmark://chstone-v0/motion,benchmark://chstone-v0/sha

Then use the info.py script to print the results and compare against Table V in the paper:

python -m gcc_autouning.info