Productive Programming Environment

Moving Heterogeneous GPU Computing into the Mainstream with Directive-Based, High-Level Programming Models

 

Motivation:

GPU-based heterogeneous systems have emerged as promising alternatives for high-performance computing.

Recently, several directive-based GPU programming models have been proposed to provide better productivity than existing ones, such as CUDA and OpenCL.

Current Issues:

  • Directive-based models provide different levels of abstraction.
  • Programming efforts required to conform to their models and optimize the performance also vary.
  • Existing models do not scale well

Goal:

  • Understanding the differences in these new models will give valuable insights on the current issues and future research directions for the productive GPU computing

 

Information That GPU Directives Can Provide:

Table showing the types of information that directive-based GPU models can provide.

Papers

Seyong Lee and Rudolf Eigenmann, OpenMPC: Extended OpenMP for Efficient Programming and Tuning for GPUs, Int J. Computational Science and Engineering, Vol. 8, No. 1, 2013

Seyong Lee and Jeffrey S. Vetter, Evaluation of Directive-Based GPU Programming Models for Productive Exascale Computing, SC12: ACM/IEEE International Conference for High Performance Computing, Networking, Storage, and Analysis, November 2012

Seyong Lee and Jeffrey S. Vetter, Moving Heterogeneous GPU Computing into the Mainstream with Directive-Based, High-Level Programming Models (Position Paper), DOE Exascale Research Conference, April 2012

Related Project

OpenARC: Open Accelerator Research Compiler

Vancouver: Designing a Next-Generation Software Infrastructure for Productive Heterogeneous Exascale Computing

OpenACC: Directives for Accelerators

OpenMPC: Extended OpenMP Programming and Tuning for GPUs

Software Download

OpenARC Compiler

OpenMPC Compiler ( Release 0.3.2RC1, Release Note)