Utilities for building and optimizing a computational graph for algorithmic decomposition
This document describes a utility to construct and evaluate an optimized execution graph from the tapefile generated by the ADOL-C automatic differentiation software. It describes the format of the ADOL-C tapefile, the data structures used in building and storing the graph, and the optimizations performed in transforming the computation trace stored in the tape into an efficient graph representation. In particular, we eliminate assignments, increase granularity by hoisting'' chains of unary operations, and remove so-called dead roots -- intermediate values that have no influence on the dependent. Examples show that the optimized graphs contain up to 50% fewer nodes than a graph that would be an exact analogue of the ADOL-C tape. We also describe an attempt at generating compiled code for the graph evaluation as an alternative to interpretative approaches to evaluating the graph. 3 refs., 11 figs., 5 tabs.
- Research Organization:
- Argonne National Lab., IL (USA)
- Sponsoring Organization:
- USDOE; National Science Foundation (NSF); USDOE, Washington, DC (USA); National Science Foundation, Washington, DC (USA)
- DOE Contract Number:
- W-31109-ENG-38
- OSTI ID:
- 5617477
- Report Number(s):
- ANL/MCS-TM-148; ON: DE91013035
- Country of Publication:
- United States
- Language:
- English
Similar Records
Developing Mango Graph Studio and its Applications for Bioinformatics and Systems Biology (SBIR Phase I Grant Final Technical Report)
Multiprocessor execution of functional programs