Measurement and Analysis of GPU-Accelerated OpenCL Computations on Intel GPUs

Graphics Processing Units (GPUs) have become a key technology for accelerating node performance in supercomputers and exascale systems. Porting CPU functions to GPU kernels is not a trivial task since the execution model for GPUs differs from CPUs. Developers need to use performance tools for GPU-accelerated systems to assess an application’s GPU computations. This thesis describes extensions to Rice University’s HPCToolkit profiler that support measurement and analysis of Intel’s DPC++ programming model for GPU-accelerated systems atop the industry-standard OpenCL framework for heterogeneous parallelism on Intel GPUs. The Intel DPC++ extension to HPCToolkit supports four techniques for performance analysis. First, GPU utilization depicts kernel’s utilization of GPU resources. Second, CPU-GPU blame shifting identifies code regions on CPUs where GPUs are idle. Third, inefficiency checks are made for Intel GPU kernels. Fourth, fine-grained performance metrics are collected for OpenCL GPU kernels. These capabilities are illustrated with two case studies on Intel GPUs