Department of Computer Engineering
S E M I N A R
Effective Kernel Mapping for OpenCL Applications in Heterogeneous Platforms
Ömer Erdil Albayrak
Computer Engineering Department
Manycore systems are getting more and more popular due to their processing capabilities. Such systems may consist of multiple devices, and mapping of applications to those systems needs to be enhanced to maximize the utilization of underlying architecture. Especially in GPUs, kernel mapping becomes critical for multi-kernel applications as kernels may have different characteristics such as resource and data requirements. While some kernels can obtain good speed-up on GPUs, others may rather stay on CPU due to the fact that data transfer overhead dominates the total kernel execution time. Furthermore, the dependencies between kernels should also be considered to generate better mappings. Having said that CPU and GPU together may yield better performance when compared to CPU-only and GPU-only executions. We propose a novel profiling-based kernel mapping algorithm to assign each kernel to the proper device to improve the overall performance of an application. Initial experiments show that our approach can effectively map kernels to CPU and GPU, and outperforms CPU-only and GPU-only approaches.
Keywords: Heterogeneous, OpenCL, kernel mapping, GPGPU
DATE: 16 April, 2012, Monday @ 16:00