Skip to Main content Skip to Navigation
Theses

System-level approaches for fixed-point refinement of signal processing algorithms

Karthick Parashar 1
1 CAIRN - Energy Efficient Computing ArchItectures with Embedded Reconfigurable Resources
IRISA-D3 - ARCHITECTURE, Inria Rennes – Bretagne Atlantique
Abstract : The fixed-point refinement problem is a combinatorial optimization problem whose search space grows exponentially. It is known to be NP-hard in complexity. Scalability issues involved in performing fixed-point refinement are the central theme of this thesis. A divide-and-conquer technique, where a given system is decimated to smaller sub-systems organized in a hierarchy is at the heart of this approach. This paves way for fast accuracy evaluation and the proposed hierarchical word-length optimization problem. Due to the reduction in number of variables, the convergence of hierarchical optimization problem to a solution is much faster than in the classical case. The single noise source (SNS) model has been proposed to study the quantization error statistics. Instead of just focusing on the average noise-power and mean of the errors due to quantization, it also provides analytical formulae for deriving statistical parameters of the random process generating quantization errors due to fixed-point simulation. In the presence of un-smooth operations such as QAM-slicing, Min() or Max() etc., it is inevitable to use fixed-point simulation. A technique for analytical evaluation of quantization error statistics in the presence of un-smooth quantizers applicable for feed-forward networks is also proposed. In order to address systems with feedback involving un-smooth operations, a hybrid technique that makes use of the SNS model to accelerate fixed-point simulation is proposed. A convex-optimization framework is proposed as an improved heuristic to solve the word-length optimization problem. This not only improves the quality of the solution but also solves the problem much faster than classical iterative approaches. Application of the proposed techniques has resulted in improved reduction in system costs even and a reduction of several orders of magnitude in the over all time required for fixed-point refinement.
Complete list of metadatas

Cited literature [117 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00783806
Contributor : Olivier Sentieys <>
Submitted on : Friday, February 1, 2013 - 4:46:52 PM
Last modification on : Friday, March 6, 2020 - 1:30:20 AM
Document(s) archivé(s) le : Monday, June 17, 2013 - 6:17:30 PM

Identifiers

  • HAL Id : tel-00783806, version 1

Citation

Karthick Parashar. System-level approaches for fixed-point refinement of signal processing algorithms. Signal and Image processing. Université Rennes 1, 2012. English. ⟨tel-00783806⟩

Share

Metrics

Record views

549

Files downloads

1580