New user self-registration is disabled due to spam. Please email eigen-core-team @ lists.tuxfamily.org if you need an account.
Before reporting a bug, please make sure that your Eigen version is up-to-date!
Bug 1666 - Scale by power of two when computing stableNorm
Summary: Scale by power of two when computing stableNorm
Status: CONFIRMED
Alias: None
Product: Eigen
Classification: Unclassified
Component: Core - general (show other bugs)
Version: 3.4 (development)
Hardware: All All
: Normal Accuracy Problem
Assignee: Nobody
URL:
Whiteboard:
Keywords: accuracy, performance
Depends on:
Blocks:
 
Reported: 2019-01-23 14:23 UTC by Christoph Hertzberg
Modified: 2019-01-23 20:52 UTC (History)
3 users (show)



Attachments

Description Christoph Hertzberg 2019-01-23 14:23:24 UTC
Currently stableNorm scales by the actual inverse of the maximum element. This requires at least one division and is less accurate than scaling by the next power of two.
This also applies to the scaling done by most Eigenvalue or singular value decompositions.


Somewhat related SO question/answer:
https://stackoverflow.com/questions/54297525/fast-way-to-get-a-close-power-of-2-number-floating-point
Comment 1 Gael Guennebaud 2019-01-23 20:52:26 UTC
Good idea!

Note You need to log in before you can comment on or make changes to this bug.