Bug 27670 - too large error for j0/j1/y1 and their dbl-96 and float128 friends
Summary: too large error for j0/j1/y1 and their dbl-96 and float128 friends
Status: NEW
Alias: None
Product: glibc
Classification: Unclassified
Component: math (show other bugs)
Version: unspecified
: P2 normal
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2021-03-31 04:47 UTC by Paul Zimmermann
Modified: 2021-04-02 04:39 UTC (History)
0 users

See Also:
Host:
Target:
Build:
Last reconfirmed:

Attachments
Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description Paul Zimmermann 2021-03-31 04:47:11 UTC 
When adding a patch to reduce the errors in the binary32 format (reference will be added later here when this patch is in), some entries that yield large (but allowed) errors for binary32 yield too large errors for wider formats.

To avoid a "make check" failure, "xfail" entries have been added:

j1 0x1.2f28eap+7
j1 0x1.a1d20ap+6
y0 0x1.33eaacp+5
y0 0x1.a681cep-1
y1 0x1.065194p+7
y1 0x1.c1badep+0

See also #16492.
Comment 1 Paul Zimmermann 2021-03-31 04:51:13 UTC 
I forgot:

j0 0x2.602774p+0
Comment 2 Paul Zimmermann 2021-03-31 05:38:29 UTC 
large errors for y0 are moved to #16492
Comment 3 Paul Zimmermann 2021-03-31 05:47:19 UTC 
cf https://members.loria.fr/PZimmermann/papers/accuracy.pdf for inputs with large errors. Here are the inputs for the version from February 5, 2021, corresponding to GNU libc 2.33.

dbl-64:
j0 0x1.33d152e971b4p+1 (4.51e14 ulps)
j1 -0x1.ea75575af6f09p+1 (4.47e14 ulps)
y1 0x1.193bed4dff243p+1 (5.56e15 ulps)

ldbl-96:
j0 -0x2.67a2a5d2e367f784p+0l (9.79e17 ulps)
j1 0x3.d4eaaeb5ede115p+0l (3.38e18 ulps)
y1 0xb.bfc89c6a1903022p+0l (4.61e18 ulps)

float128:
j0 -0x8.a75ab6666f64eae68f8eb383dad8p+0 (4.10e32 ulps)
j1 0x3.d4eaaeb5ede114ff552b1726d4ep+0 (2.46e33 ulps)
y1 0x2.3277da9bfe485c85c35e5bcc806p+0 (3.47e33 ulps)
Comment 4 Paul Zimmermann 2021-04-02 04:39:31 UTC 
The commit that reduces the errors in the binary32 format is 9acda61
(April 1st, 2021)