Analysis And Characterization Of Popular OpenCL Kernels

Abstract

In all computing domains heterogeneous parallel computing platforms composed of CPUs, GPUs, FPGAs and DSPs are widening their user base. GPUs are emerging as a general purpose high performance computing device. GPGPU research is growing very fast as a cost e ective approach for accelerating data and compute intensive application which has made many new workloads available. It has been driven by the introduction of C-based programming environments such as NVIDIA's CUDA , OpenCL. OpenCL (Open Computing Language) is an open standard and emerging parallel programming model to write parallel applications for such heterogeneous platforms. However there is no systematic approach to analyse and characterize the workload to assist future work on microarchitecture design, application re-structuring and compiler optimizations. In this paper I will present the software architecture that produces interesting statistics to understand the dynamic behaviour of the GPGPU. In addition to that a set of metrics for evaluation of OpenCL kernels is proposed. ArchOCL is a software model that includes dynamic library that support all OpenCL application to produce statistics to analyse OpenCL application. Development of ArchOCL model includes implementa-tion of OpenCL APIs, Compiler enhancement to support OpenCL kernel compilation and Statistics collection.