Automatic partitioning of parallel programs on heterogeneous multicore processors

Abstract

With the advent of hybrid CPUs integrated with GPUs, achieving peak performance and energy efficiency necessitates the effective utilization of all available resources. This project explores workload partitioning strategies for heterogeneous multicore processors, leveraging tools like Intel PCM for CPU analysis and OpenCL for GPU-based computations. A key focus was on implementing custom convolution functions and analyzing metrics such as frequency scaling and power consumption. Significant challenges were encountered, including device compatibility issues and optimization of custom functions. Results indicate that OpenCL enables successful GPU and CPU operations, but limitations with simultaneous device detection and slower custom implementations compared to PyTorch underline the need for further refinement. Future directions include addressing these challenges, optimizing workload distribution, and expanding scalability to larger datasets.