Full logs of our runs are in output_logs
The current version of the framework uses Catapult HLS. The framework uses gcc, g++ and gdb.
Python dependencies:
python3.11 -m pip install pycparser openai anthropic pyyaml
An API key for the model in use is needed. It can be set as an environment variable (ANTHROPIC_API_KEY, OPENAI_API_KEY, DEEPSEEK_API_KEY)
The C/C++ code should be split into 3 files:
- includes file: should contain the include libraries and global variables that should not be refactored. We explicitly prompr the LLM to assume that the content of this file will be provided and should not be produced by the LLM.
- functions file: should contain all and only the functions that the LLM should refactor.
- test file: should contain a main function with one or more tests, the results should be printed to stdout.
One yaml file contains information about top function, file paths, type of interface (streaming or nor) and whether the code is hierarchical or not.
To run one of our benchmarks:
python3.11 src/tool.py inputs/${benchmark}/config_${benchmark}.yaml --model <model name> --opt_target <area/latency>
To get a list of the possible models and other options:
python3.11 src/tool.py -h