high performance computing on graphics processing units: hgpu.org

Recently, CPU-GPU heterogeneous parallelism has brought transformational performance milestones to static timing analysis (STA) algorithms. As the computing ecosystem continues to proliferate, performance portability has emerged as a new challenge when deploying the result to diverse heterogeneous computing platforms. Specifically, the optimal code written on a CPU-GPU architecture may not be optimal for other CPUGPU architectures, due to various performance, interoperability, and availability constraints. As a result, we introduce in this paper a learning-based framework to enhance the performance portability of a GPU-accelerated STA program. We parameterize important performance parameters and leverage a neural network model to adapt performance optimization to any given computing platforms. We have demonstrated the effectiveness of our framework in real STA applications.