Author mark.dickinson
Recipients mark.dickinson, rhettinger, serhiy.storchaka, skrah, steven.daprano, tim.peters
Date 2018-03-19.08:25:08
SpamBayes Score -1.0
Marked as misclassified Yes
Message-id <1521447909.06.0.467229070634.issue33089@psf.upfronthosting.co.za>
In-reply-to
Content
+1 for a single-rounded dot product. If we're allowed to assume IEEE 754, it's straightforward to code up something that's not too inefficient and gives correctly rounded results for "normal" cases, using a combination of Veltkamp splitting, Dekker multiplication, and fsum. The difficulties come in if you want to maintain correct rounding in cases where any of the partial products overflows or (especially awkwardly) underflows.

Also, if we can figure out how to do a correctly-rounded dot product, that gives us math.fma as a special case... (a*b + c = dot([a, c], [b, 1.0])). (Of course, that's a bit backwards, since usually you'd see fma as a primitive and *use* it in computing a dot product.)
History
Date User Action Args
2018-03-19 08:25:09mark.dickinsonsetrecipients: + mark.dickinson, tim.peters, rhettinger, steven.daprano, skrah, serhiy.storchaka
2018-03-19 08:25:09mark.dickinsonsetmessageid: <1521447909.06.0.467229070634.issue33089@psf.upfronthosting.co.za>
2018-03-19 08:25:09mark.dickinsonlinkissue33089 messages
2018-03-19 08:25:08mark.dickinsoncreate