OMPAR

OMPAR is a compiler-oriented tool designed to identify and generate parallelization opportunities for serial code. It consists of the following pipeline:

1. OMPify: Detects opportunities for parallelization in code.
2. MonoCoder: Generates the appropriate OpenMP pragmas when a for loop is identified as beneficial for parallelization. Note: The weights for OMPify are not included in the repository and will be provided upon request.

Figure 1: OMPar workflow using a simple pi code example, comparing it with other compilers. Source-to-source automatic compilers (such as AutoPar) generate the necessary pragma, while HPC compilers (such as ICPC) generate a binary parallel output. In contrast, OMPar relies on two LLMs: one for classifying parallelization needs (OMPify) and one for generating the full pragma (MonoCoder-OMP). Both were trained on a large corpus of codes. The evaluation checks if the code compiles, performs with increasing threads, and verifies outputs.

Building OMPar

To build OMPAR, ensure that CUDA 12.1 is supported on your system. Follow these steps:

Clone the repository:

git clone https://github.com/Scientific-Computing-Lab/OMPar
cd OMPar

Create and activate the Conda environment:

conda create -n ompar_env python=3.11 -f environment.yml
conda activate ompar_env

Build the parser:

cd parser
./build.sh

Usage

To use OMPar, you need to download the Ompify model weights from here.

Here’s an example of how to use OMPAR:

code="""for(int i = 0; i <= 1000; i++){ partial_Sum += i; }"""device='cuda'iftorch.cuda.is_available() else'cpu'ompar=OMPAR(model_path=main_args.model_weights, device=device, args=main_args)
pragma=ompar.auto_comp(code)

To run additional use cases, execute the following command:

python run_ompar.py --model_weights /path/to/OMPify

OMPar Evaluation

The following table shows the performance of OMPar on HeCBench test set of 770 loops. OMPar accurately predicted the pragma in 74% of the test loop.

Test setup	TP	FP	TN	FN	Precision	Recall	Accuracy
OMPar accuracy with ground-truth label	311	127	262	70	71%	81%	74%
AutoPar accuracy with ground-truth label	63	25	365	317	71%	17%	56%
ICPC accuracy with ground-truth label	95	11	285	379	90%	25%	62%
OMPar accuracy with compile and run check	407	31	262	70	92%	85%	86%
AutoPar accuracy with compile and run check	24	25	365	356	49%	6%	50%
ICPC accuracy with compile and run check	68	5	312	385	93%	15%	49%

The results for OMPar, AutoPar, and ICPC can be reproduced using the information provided in the evaluation folder. Specifically, the data subfolder contains the code to gather the HeCBench serial codes, while the icpc, autoPar, and ompar subfolders describe the methods for running these codes.