Best Paper & Poster Award

Share it now
DASIP 2011
Best Paper Award

 
The Multi-Dataflow Composer Tool:
A Runtime Reconfigurable HDL Platform Composer
Francesca Palumbo, Nicola Carta and Luigi Raffo
(University of Cagliari)

Abstract
Dataflow model of computation is particularly suitable to close the gap between hardware architects and software developers. Such a gap affects time-to-market of nowadays systems, which need to withstand the high level of complexity required by modern market needs. In this context, this paper has a two-fold target. First of all, leveraging on the combination of the dataflow model of computation with a coarse-grained reconfigurable approach to hardware design, we aim at presenting a tool, the Multi-Dataflow Composer tool, able to close such a gap: automatically deriving HDL runtime reconfigurable platforms from a given set of applications. Besides, we want to demonstrate that our approach can also lead to a considerable on-chip area saving. Performance assessment will be carried on reporting the results for different applications in the image processing reference domain. Very promising synthesis trials, adopting a 90 nm CMOS technology, have been achieved.


 

DASIP 2011
Best Poster Award

Parallelization of an Ultrasound Reconstruction Algorithm for Non-Destructive Testing on Multicore CPU and GPU
Antoine Pédron (CEA),
Lionel Lacassagne
(CEA),
Franck Bimbard (
Université Paris Sud 11),
Stéphane Le Berre
(CEA)

Abstract
The CIVA software platform developed by CEA-LIST offers various simulation and data processing modules dedicated to non-destructive testing (NDT). In particular, ultrasonic imaging and reconstruction tools are proposed, in the purpose of localizing echoes and identifying and sizing the detected defects. Because of the complexity of data processed, computation time is now a limitation for the optimal use of available information. In this article, we present performance results on parallelization of one computationally heavy algorithm on general purpose processors (GPP) and graphic processing units (GPU). GPU implementation makes an intensive use of atomic intrinsics. Compared to initial GPP implementation, optimized GPP implementation runs up to x116 faster and GPU implementation up to x631. This shows that, even with irregular workloads, combining software optimization and hardware improvements, GPU give high performance.
Share it now