This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
math tests: reduced precision on Opteron 6272
- From: Aurelien Jarno <aurelien at aurel32 dot net>
- To: libc-alpha at sourceware dot org
- Date: Wed, 29 Jun 2016 23:12:45 +0200
- Subject: math tests: reduced precision on Opteron 6272
- Authentication-results: sourceware.org; auth=none
Hi all,
I have encountered some tests failures when building a 32-bit libc on an
AMD Opteron 6272. It concerns the math/test-double, math/test-double-finite
and math/test-idouble for the cacos_downward, cacos_towardzero,
cacosh_downward, cacosh_towardzero functions. The results are 2 ULP
instead of 1 ULP for the real part:
| Failure: Test: Real part of: cacos_downward (1.0 + 0x1.fp-129 i)
| Result:
| is: 5.3356352769822323e-20 0x1.f7efbeb8d4f10p-65
| should be: 5.3356352769822335e-20 0x1.f7efbeb8d4f12p-65
| difference: 1.2037062152420224e-35 0x1.0000000000000p-116
| ulp : 2.0000
| max.ulp : 1.0000
I have not been able to reproduce the issue on other CPUs, nor in 64-bit.
This is likely because on x86-64 we require 2 ULP instead of 1 ULP on
i386 for these functions.
Would it make sense to regenerate the libm-test-ulps on this CPU?
Thanks,
Aurelien
--
Aurelien Jarno GPG: 4096R/1DDD8C9B
aurelien@aurel32.net http://www.aurel32.net