On 9/10/25 09:31, Sébastien Villemot wrote:
Control: found -1 0.3.29+ds-3
On Fri, 5 Sep 2025 12:32:41 +0200 Matthias Klose <[email protected]>
wrote:
as a data point, 0.3.29 built fine in Ubuntu with IEEE long double.
I just rebuild 0.3.29+ds-3 on platti.debian.org, using the most recent
gcc-15, and it also fails (same 6 failed tests). A build log is
attached.
My understanding is that 0.3.30+ds-2 also built fine with IEEE long
double in Ubuntu, while it failed in Debian. So it looks like Ubuntu
and Debian behave differently with respect with this transition. Any
explanation for the cause of this discrepancy?
ppc64el on Ubuntu builds using -O3. But I doubt that would cause this
discrepancy.
Another thing to be noted is that some numerical error in the testsuite
are huge with 0.3.30+ds-2 in Debian (about 1e54 for
dgemmt:c_api_rowmajor_upper_M_100_K_50_a_trans_b_notrans,
dgemmt:c_api_colmajor_lower_M_50_K_100_a_notrans_b_trans,
dgemmt:lower_M_50_K_100_a_notrans_b_trans), while they are smaller, but
still large, with 0.3.29+ds-3.
