Personal tools
You are here: Home Events ICSA COLLOQUIUM TALK-Rescheduled


— filed under:

Talk and presentation by Dominik Grewe, University of Edinburgh. Title: Automatically Mapping Parallel Programs to GPU-based Heterogeneous Systems. RESCHEDULED TALK FROM 15th Nov

  • Colloquium Series
When Nov 20, 2012
from 02:00 PM to 03:00 PM
Where IF. 4.31/4.33
Add event to calendar vCal


GPU-based heterogeneous systems have become ubiquitous in all areas of computing, from high-performance to mobile computing. Programming these systems, however, is challenging and requires a deep understanding of the system's architecture. Enabling mainstream use of these systems requires us to overcome two challenges: picking the most suitable device for a task and optimizing the code for that target device.

In this talk I will present two approaches that each tackle one of these challenges. The first is a machine-learning based approach to solve the task mapping problem. Using static code features our model determines on how to partition the task's work across the processors in a heterogeneous system. In the second piece of work we developed a method for generating OpenCL code from data-parallel OpenMP programs and optimizing this code for GPUs. Because not all programs are equally suited for GPU computing we also deploy a machine-learning model to decide whether or not to use the GPU. On the NAS Parallel Benchmark suite our approach outperforms a hand-coded OpenCL implementation by a factor of 1.63 and the closest related work by a factor of 10.


Document Actions