|Summary:||Add vectorization of sqrt for doubles|
|Component:||Core - vectorization||Assignee:||Nobody <eigen.nobody>|
|Severity:||enhancement||CC:||chtz, gael.guennebaud, jacob.benoit.1|
|Hardware:||x86 - SSE|
|Bug Depends on:|
Description panda-34 2013-08-08 20:34:26 UTC
Comment 1 Christoph Hertzberg 2013-08-13 14:41:45 UTC
I fully agree. I guess the main problem with the fast implementation are wrong results for negative values (should be NaN), +inf (should be inf) and denormalized values (rsqrt treats them as 0.0f, resulting in inf which leads to a NaN after the refinement).
Comment 2 Gael Guennebaud 2013-08-19 15:43:39 UTC
Related discussion on the ML (old!): http://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2009/03/msg00114.html And I agree with Christoph, the main pb is not on the accuracy but on the handling of negative, inf and denorms.
Comment 3 Gael Guennebaud 2013-08-19 16:01:59 UTC
https://bitbucket.org/eigen/eigen/commits/97d9a7772106/ Changeset: 97d9a7772106 User: ggael Date: 2013-08-19 16:02:27 Summary: Fix bug 642: add vectorization of sqrt for doubles, and make sqrt really safe if EIGEN_FAST_MATH is disabled
Comment 4 Nobody 2019-12-04 12:32:56 UTC
-- GitLab Migration Automatic Message -- This bug has been migrated to gitlab.com's GitLab instance and has been closed from further activity. You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.com/libeigen/eigen/issues/642.