parker-research/verilog-eval

Verilog evaluation benchmark for large language model

VerilogEval: Evaluating Large Language Models for Verilog Code Generation

This is an evaluation harness for the VerilogEval problem solving dataset described in the paper "VerilogEval: Evaluating Large Language Models for Verilog Code Generation".

This evaluation dataset consists of 156 problems from the Verilog instructional website HDLBits. We provide two sets of problem descriptions: machine generated and manually converted to text-only format.

Installation

We closely follow guidance from HumanEval.

Make sure to use Python 3.7 or later:

$ conda create -n codex python=3.7
$ conda activate codex

Install ICARUS Verilog:

$ git clone https://github.com/steveicarus/iverilog.git && cd iverilog \
        && git checkout 01441687235135d1c12eeef920f75d97995da333 \
        && sh ./autoconf.sh && ./configure && make -j4\
        && make install

It is recommended to use the provided Dockerfile which already pre-installed ICARUS Verilog Simulator. Using the docker container you would still need to complete the following step.

Check out and install this repository:

$ git clone https://github.com/NVlabs/verilog-eval
$ pip install -e verilog-eval

Usage

This program would make system calls to iverilog and vvp to simulate untrusted model-generated code. Users are strongly encouraged not to do so outside of a robust security sandbox. The execution call in execution.py is deliberately commented out to ensure users read this disclaimer before running code in a potentially unsafe manner. See the comment in execution.py for more information and instructions.

After following the above instructions to enable execution, generate samples and save them in the following JSON Lines (jsonl) format, where each sample is formatted into a single line like so:

{"task_id": "Corresponding VerilogEval task ID", "completion": "Completion only without the prompt"}

We provide examples under data/example to illustrate the format and help with debugging.

To evaluate the samples, run

$ evaluate_functional_correctness samples.jsonl --problem_file data/VerilogEval_Human.jsonl

Reading samples...
3120it [00:00, 16077.44it/s]
Running test suites...
100%|...| 3120/3120 [00:32<00:00, 97.47it/s]
Killing all hanging simulation process.
Writing results to samples.jsonl_results.jsonl...
100%|...| 3120/3120 [00:00<00:00, 30608.13it/s]
{'pass@1': ..., 'pass@5': ..., 'pass@10': ...}

The user must specify --problem_file input argument. We provide two sets of problem evaluations data/VerilogEval_Machine.jsonl and data/VerilogEval_Human.jsonl. We also provide problem description files used to sample Verilog code completions in descriptions directory.

This script provides more fine-grained information in a new file ending in <input_path>_results.jsonl. Each row now contains whether the completion passed along with the execution result which is one of "passed", "timed out", or "failed".

As a quick sanity-check, the example samples should yield 0.5 pass@1. The results can be verified against the provided output in data/example/ExampleSolution.jsonl_reference.jsonl.

$ evaluate_functional_correctness data/example/ExampleSolution.jsonl --problem_file=data/example/ExampleEval.jsonl

Reading samples...
6it [00:00, 221.60it/s]
Running example suites...
100%|...| 6/6 [00:00<00:00, 142.09it/s]
Killing all hanging simulation process.
Writing results to data/example/ExampleSolution.jsonl_results.jsonl...
100%|...| 6/6 [00:00<00:00, 19941.22it/s]
{'pass@1': 0.5}

Because there is no unbiased way of estimating pass@k when there are fewer samples than k, the script does not evaluate pass@k for these cases. To evaluate with other k values, pass --k=<comma-separated-values-here>. For other options, see

$ evaluate_functional_correctness --help

However, we recommend that you use the default values for the rest.

Issues

Problem descriptions in descriptions/VerilogDescription_Machine.jsonl are machine generated and we can not guarantee the absence of ambiguity and errors. We do not plan to maintain description correctness.

Functional correctness are evaluated through comparing simulation outputs using ICARUS Verilog. The evaluation of Verilog syntax is limited by the simulator, which might not include all features of Verilog HDL IEEE-1364 standard.

Citation

Please cite using the following bibtex entry:

@inproceedings{liu2023verilogeval,
  title={{VerilogEval:} Evaluating Large Language Models for Verilog Code Generation},
  author={Liu, Mingjie and Pinckney, Nathaniel and Khailany, Brucek and Ren, Haoxing},
  booktitle={2023 IEEE/ACM International Conference on Computer-Aided Design (ICCAD)}, 
  year={2023}
}