• Have docker installed (used when running DynaRapid)
• Have python 3.10 or newer
• Have Vitis installed and sourced (We use version 2024.2)
• Set api_key environment variable: export OPENROUTER_API_KEY=<Add API Key> export OPENAI_API_KEY=<OPENAI KEY> This is if you would like to use any GPT model

git clone git@github.com:Preston-Walker/HLS-FAB.git
cd HLF-FAB
python3 -m venv venv
source venv/bin/activate
pip install -e .
<!-- bash build_rag_dir.sh -->

cd dataset
# Add just one kernel for the test run
cp -r examples/abs_int8 kernels
make dataset
cd ..

To set which LLM models to use, edit the dataset/llm_models.txt file.\

wget https://gist.githubusercontent.com/clavin-xlnx/7ff0b7238aa5f5de4ff2dac7392d5132/raw/a130d502374cfa2779e59f918a2248e61fec5363/Dockerfile.dynarapid
docker build -t my-llvm6-dynarapid-image .
docker run -v <Path to hlseval reposotory>/hlseval:/mount --name dynarapid_container -t my-llvm6-dynarapid-image

The /generate_vision_kernel skill uses Claude Code to generate generate.py files for kernels from the Xilinx Vitis Vision Library. Each generate.py describes a kernel's function signature, I/O buffers, a pure-numpy behavioral model, and example data for AI-assisted HLS code generation.

• Clone the Vitis Libraries repository:

git clone https://github.com/xilinx/vitis_libraries.git

• Have Claude Code installed

The skill needs to know where Vitis Libraries and Vitis HLS are on your system. You can configure this in one of two ways:

Option A — Environment variables (take priority):

export VITIS_LIBRARIES=/path/to/Vitis_Libraries
export VITIS_HLS=/path/to/Vitis_HLS/2024.2

Option B — Config file:

cp vision_skill_paths.cfg.example vision_skill_paths.cfg
# Edit vision_skill_paths.cfg with your paths

From within Claude Code, run the skill with a kernel name, all, or a path to a list file:

/generate_vision_kernel boxfilter
/generate_vision_kernel all
/generate_vision_kernel @kernels_todo.txt

Generated files are saved to dataset/kernels/<top_function_name>/generate.py.

Sample of run command:
make run_vitis_completions LOG=DEBUG N=5 K=1
Options:

• LOG option: Allows you to set the logging level of the ouput log file. Options: DEBUG, INFO, WARNING, ERROR, CRITICAL
• N option: Allows you to determine the number of generation attempts per kernel. Setting this to 1 would make an LLM call for each kernel without incorporating feedback. Setting N to a value of 2 or more will incorporate a feedback loop where the results from each attempt are passed back to the LLM to improve upon the result each time.
• K option: Allows you to determine the number of independent runs to do sequentially. Setting this to 2 would run N number of feedback attempts 2 times.

Sample of run command:
make run_dynarapid_completions LOG=DEBUG N=4 RAPIDWRIGHT=TRUE Options:

• LOG option: Allows you to set the logging level of the ouput log file. Options: DEBUG, INFO, WARNING, ERROR, CRITICAL
• N option: Allows you to determine the number of generation attempts per kernel. Setting this to 1 would make an LLM call for each kernel without incorporating feedback. Setting N to a value of 2 or more will incorporate a feedback loop where the results from each attempt are passed back to the LLM to improve upon the result each time.
• K option: Allows you to determine the number of independent runs to do sequentially. Setting this to 2 would run N number of feedback attempts 2 times.
• RAPIDWRIGHT option: If omitted, or set to FALSE, Dynamatic will run without RapidWright, resulting in a .dot file, but no placed and routed circut. Adding the DYNAMATIC option and setting it to anything other than FALSE will run both Dynamatic and RapidWright. The value TRUE is used here for clarity.