GPURepair: Automated Repair of GPU Kernels
This addresses the issue of debugging and optimizing GPU kernels for developers, though it is incremental as it builds on prior work like GPUVerify.
The paper tackles the problem of repairing errors in GPU kernels, such as data races and barrier divergence, by introducing GPURepair, a tool that can fix more kernels and uniquely remove redundant barriers and handle inter-block data races, as demonstrated through experiments on about 750 kernels.
This paper presents a tool for repairing errors in GPU kernels written in CUDA or OpenCL due to data races and barrier divergence. Our novel extension to prior work can also remove barriers that are deemed unnecessary for correctness. We implement these ideas in our tool called GPURepair, which uses GPUVerify as the verification oracle for GPU kernels. We also extend GPUVerify to support CUDA Cooperative Groups, allowing GPURepair to perform inter-block synchronization for CUDA kernels. To the best of our knowledge, GPURepair is the only tool that can propose a fix for intra-block data races and barrier divergence errors for both CUDA and OpenCL kernels and the only tool that fixes inter-block data races for CUDA kernels. We perform extensive experiments on about 750 kernels and provide a comparison with prior work. We demonstrate the superiority of GPURepair through its capability to fix more kernels and its unique ability to remove redundant barriers and handle inter-block data races.