Análisis del impacto de sentencias  condicionales y repetitivas en la arquitectura NVIDIA G80

Snaider Carrillo; Jakob Siegel; Xiaoming Li

Authors

Snaider Carrillo University of Ulster
Jakob Siegel University of Delaware
Xiaoming Li University of Delaware

Abstract

AbstractIn this paper, we propose two novel techniques to transform control state- ments so they can be executed effciently on the NVIDIA G80 architecture.
Our techniques called loop splitting and branch splitting smartly increase
code redundancy, which might be deemed as “de-optimization” for CPU;
but for a GPU framework these techniques improve the occupancy of a
program on the GPU device and therefore improve its performance. We
demonstrate our optimizations on an artifcial benchmark and the results
show that these techniques are very effcient and, depending on the problem
layout, can lead to an increase in occupancy and a drastic improvement
in performance compared to non-split version of the same algorithm.

Author Biographies

Snaider Carrillo, University of Ulster

Beng in Electronics Engineering from Universidad del Norte, Barranquilla
(Colombia). MSc in Electronics Engineering from Pontificia
Universidad Javeriana, Bogotá (Colombia). PhD Student in Computer
Science at University of Ulster (Londonderry - UK).

Jakob Siegel, University of Delaware

BEng in Software Engineering from HS Essligen: University of
Applied Science (Essligen - Germany). MSc in Electrical and Computer
Engineering from University of Delaware (Newark - USA). PhD Student
in Electrical and Computer Engineering at University of Delaware (Newark
- USA).

Xiaoming Li, University of Delaware

BSc and MSc in Computer Science from Nanjing University
(Shanghai - China), and PhD in Computer Science at the University of
Illinois at Urbana-Champaign (Urbana-USA). Assistant Professor in the
Department of Electrical and Computer Engineering at University of
Delaware (Newark - USA).