Bug#1126042: hypre: "make checkpar" fails tolerance on arm64 and other arches
Drew Parsons
dparsons at debian.org
Tue Jan 20 20:36:10 GMT 2026
Source: hypre
Version: 3.0.0-2
Severity: serious
Justification: FTBFS
X-Debbugs-Cc: debian-arm at lists.debian.org
User: debian-arm at lists.debian.org
Usertags: arm64
Control: forwarded -1 https://github.com/hypre-space/hypre/issues/1457
In 2.31.0 (upstream PR #959) I added a test configuration option
HYPRE_CHECKPAR_TOL to allow controlling the checkpar test tolerance,
which was failing on many of the architectures that debian builds for
(arm64 i386 ppc64el riscv64 s390x hppa loong64 powerpc ppc64), see
Issue #955.
Setting HYPRE_CHECKPAR_TOL facilitated builds of hypre 2.*. But with
hyper 3.0.0 the discrepancy has widened, and the required tolerance
for some tests would need to be greater than 1e-1 (in hypre 2 these
arches passed by relaxing the tolerance from 1e-6 to 1e-3).
A tolerance greater than 1e-1 seems a little excessive, so I raise the
issue upstream $1457
Debian build logs showing the test error logs can be found at
https://buildd.debian.org/status/package.php?p=hypre (past logs can be
accessed via the "old" links). For instance the failing build for for
arm64 is
https://buildd.debian.org/status/fetch.php?pkg=hypre&arch=arm64&ver=3.0.0-2&stamp=1768915710&raw=0
An excerpt from the log is
Testing IJ ...
FAILED : TEST_ij/solvers.err (63510)
(651) - Final L2 norm of residual: 2.011802e-02
(651) + Final L2 norm of residual: 1.874054e-02 (err 1.38e-03)
(655) - Final L2 norm of error: 9.197532e+00
(655) + Final L2 norm of error: 8.649254e+00 (err 5.48e-01)
(656) - Final L2 norm of e0: 1.798116e+00
(656) + Final L2 norm of e0: 1.696116e+00 (err 1.02e-01)
(657) - Final L2 norm of e1: 5.728875e+00
(657) + Final L2 norm of e1: 5.417061e+00 (err 3.12e-01)
(658) - Final L2 norm of e2: 6.967164e+00
(658) + Final L2 norm of e2: 6.525966e+00 (err 4.41e-01)
...
make check is passing, it's only make checkpar that fails this way. The tests pass on amd64 (x86_64)
More information about the debian-science-maintainers
mailing list