high performance computing on graphics processing units: hgpu.org

Effective detection of data races in parallel computing environments is essential for ensuring the correctness and performance of multi-threaded applications. This paper addresses the issue with OpenCL data racing analysis. Currently, for the data racing research, there is a well-established DataRaceBench benchmark, designed for OpenMP. In our research, we rewrite the OpenMP DataRaceBench benchmark for the OpenCL benchmark to provide a specialized benchmark suite for evaluating data race detection tools in OpenCL environments. In our analysis, we introduce a novel approach to detect data races in OpenCL programs by leveraging an LLVMbased analysis pass. To facilitate detailed analysis, annotations are inserted into the OpenCL kernels, which fetch data from the kernel side and return data to the host side. Then the detector on the host side utilized those events to analyze and track the interactions between different threads. Through vector clocks, we can partially order them. This methodology helps identify potential data races by analyzing patterns within these annotated sections. Experimental results demonstrate the efficacy of our approach, which can accurately detect data races across multiple threads in OpenCL by obtaining data via LLVM passes and analyzing them on the host. This work not only enhances the toolset for developers working with OpenCL but also contributes significantly to the field of parallel computing by providing a rigorous benchmarking tool for data race detection.