Author jeff.allen
Recipients christian.heimes, eric.smith, jeff.allen, koobs, lemburg, mark.dickinson, ned.deily, python-dev, serhiy.storchaka, stutzbach
Date 2016-08-29.17:54:16
SpamBayes Score -1.0
Marked as misclassified Yes
Message-id <1472493257.0.0.351330948211.issue26040@psf.upfronthosting.co.za>
In-reply-to
Content
Ah, cunning: I can make sense of it in hex.

>>> hex(to_ulps(expected))
'0x3ff0000000000000'
>>> hex(to_ulps(got))
'0x3feffffffffffffc'
>>> hex( to_ulps(got) - to_ulps(expected) )
'-0x4'

... and what you've done with ulp then follows.

In my version a format like "{:d} ulps" was a bad idea when the error was a gross one, but your to_ulps is only piece-wise linear -- large differences are compressed.

I'm pleased my work has mostly survived: here's hoping the house build-bots agree. erfc() is perhaps the last worry, but math & cmath  pass on my machine.
History
Date User Action Args
2016-08-29 17:54:17jeff.allensetrecipients: + jeff.allen, lemburg, mark.dickinson, eric.smith, christian.heimes, ned.deily, stutzbach, python-dev, serhiy.storchaka, koobs
2016-08-29 17:54:17jeff.allensetmessageid: <1472493257.0.0.351330948211.issue26040@psf.upfronthosting.co.za>
2016-08-29 17:54:16jeff.allenlinkissue26040 messages
2016-08-29 17:54:16jeff.allencreate