Use ssml from https://doi.org/10.1145/3061665, and fix typos #559
Conversation
Codecov Report
@@ Coverage Diff @@
## master #559 +/- ##
=======================================
Coverage 82.37% 82.37%
=======================================
Files 1894 1894
Lines 190677 190677
=======================================
Hits 157063 157063
Misses 33614 33614 Continue to review full report at Codecov.
|
| ! ssml = 1/S, where S was defined in https://doi.org/10.1145/355769.355771 | ||
| (minexponent(0._dp) - digits(0._dp)) * 0.5_dp)) | ||
| ! sbig = 1/S, where S was defined in https://doi.org/10.1145/355769.355771 | ||
| real(dp), parameter :: dsbig = real(radix(0._dp), dp)**( - ceiling( & |
There was a problem hiding this comment.
I'm still unsure that this formula is correct. It gives dsbig=5.0052077379577523E-147 while the Anderson's code uses the Blue's formula which gives 1.1113793747425387E-162. Further, some test cases in my Go port overflow with this dsbig value but do not with Anderson's value. It seems that the part - digits(0._dp) + 1 should be + digits(0._dp) - 1 to reproduce the value from Anderson's code.
There was a problem hiding this comment.
Right! I will fix it. Thanks again!
There was a problem hiding this comment.
Hi @vladimir-ch. Could you please review my changes once more. I think now everything is fine. Thanks!
…-scaling-constants Use ssml from https://doi.org/10.1145/3061665, and fix typos
…formance Libraries! This is the same bug fixed on Reference-LAPACK#559 for la_constants.f90.
4 participants
Description
Fixes #554. This PR:
la_constants.f90.* The ssml from https://doi.org/10.1145/3061665 is sufficient to scale both normalized and denormalized floating-point numbers.