Publication: Cheaper adjoints by reversing address computations
Introduction
Applications
Tools
Research Groups
Workshops
Publications
   List Publications
   Advanced Search
   Info
   Add Publications
My Account
About
Impress

Cheaper adjoints by reversing address computations

- Article in a journal -
 

Author(s)
Laurent Hascoët , Jean Utke , Uwe Naumann

Published in
Scientific Programming

Year
2008

Publisher
IOS Press

Abstract
The reverse mode of automatic differentiation is widely used in science and engineering. A severe bottleneck for the performance of the reverse mode, however, is the necessity to recover certain intermediate values of the program in reverse order. Among these values are computed addresses, which traditionally are recovered through forward recomputation and storage in memory. We propose an alternative approach for recovery that uses inverse computation based on dependency information. Address storage constitutes a significant portion of the overall storage requirements. An example illustrates substantial gains that the proposed approach yields, and we show use cases in practical applications.

AD Theory and Techniques
Code Optimization, Data Flow Analysis, Recomputation, Reverse Mode

BibTeX
@ARTICLE{
         Hascoet2008Cab,
       title = "Cheaper adjoints by reversing address computations",
       author = "Laurent Hasco{\"e}t and Jean Utke and Uwe Naumann",
       publisher = "IOS Press",
       year = "2008",
       journal = "Scientific Programming",
       volume = "16",
       number = "1",
       pages = "81--92",
       doi = "10.3233/SPR-2008-0238",
       abstract = "The reverse mode of automatic differentiation is widely used in science and
         engineering. A severe bottleneck for the performance of the reverse mode, however, is the necessity
         to recover certain intermediate values of the program in reverse order. Among these values are
         computed addresses, which traditionally are recovered through forward recomputation and storage in
         memory. We propose an alternative approach for recovery that uses inverse computation based on
         dependency information. Address storage constitutes a significant portion of the overall storage
         requirements. An example illustrates substantial gains that the proposed approach yields, and we
         show use cases in practical applications.",
       ad_theotech = "Code Optimization, Data Flow Analysis, Recomputation, Reverse Mode"
}


back
  

Contact:
autodiff.org
Username:
Password:
(lost password)