High Performance Code Generation for Stencil Computation on Heterogeneous Multi-device Architectures

Pei Li 1 Elisabeth Brunet 1 Raymond Namyst 2, 3
3 RUNTIME - Efficient runtime systems for parallel architectures
Inria Bordeaux - Sud-Ouest, UB - Université de Bordeaux, CNRS - Centre National de la Recherche Scientifique : UMR5800
Abstract : Heterogeneous architectures have been widely used in the domain of high performance computing. On one hand, it allows a designer to use multiple types of computing units and each able to execute the tasks that it is best suited for to increase performance; on the other hand, it brings many challenges in programming for novice users, especially for heterogeneous systems with multi-devices. In this paper, we propose the code generator STEPOCL that generates OpenCL host program for heterogeneous multi-device architecture. In order to simplify the analyzing process, we ask user to provide the description of input and kernel parameters in an XML file, then our generator analyzes the description and generates automatically the host program. Due to the data partition and data exchange strategies, the generated host program can be executed on multi-devices without changing any kernel code. The experiment of iterative stencil loop code (ISL) shows that our tool is efficient. It guarantees the minimum data exchanges and achieves high performance on heterogeneous multi-device architecture.
Complete list of metadatas

Cited literature [8 references]  Display  Hide  Download

https://hal.inria.fr/hal-00925481
Contributor : Elisabeth Brunet <>
Submitted on : Wednesday, January 8, 2014 - 10:35:12 AM
Last modification on : Thursday, September 12, 2019 - 3:40:03 PM
Long-term archiving on : Tuesday, April 8, 2014 - 10:30:54 PM

File

HPCC2013.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-00925481, version 1

Citation

Pei Li, Elisabeth Brunet, Raymond Namyst. High Performance Code Generation for Stencil Computation on Heterogeneous Multi-device Architectures. HPCC-15th IEEE International Conference on High Performance Computing and Communications, Nov 2013, Zhangjiajie, China. ⟨hal-00925481⟩

Share

Metrics

Record views

306

Files downloads

619