A GPU optimization workflow for real-time execution of ultra-high frame rate computer vision applications.

This work proposes a GPU optimization methodology for real-time execution of ultra high frame rate applications with small frame sizes. While the use of GPUs for offline processing is well-established, real-time execution remains challenging due to the lack of real-time execution guarantees, especially for embedded GPUs. Our methodology introduces guidelines and a workflow by focusing on: (a) controlling latency by means of minimization of CPU-GPU interactions; (b) computation pruning; and (c) inter/intra-kernel optimizations. Furthermore, our approach takes advantage of multi-frame processing to attain significantly higher throughput at the cost of increased latency when the application permits such trade-offs. To evaluate our optimization methodology, we applied it to the monitoring and controlling of laser powder bed fusion machines, a widely used metal additive manufacturing technique. Results show that in the considered application, the required performance could be obtained on a Jetson Xavier AGX platform, and by sacrificing latency, significantly higher throughput was achieved. [ABSTRACT FROM AUTHOR]