You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
openblas was called by numpy and the issue looks similar to #2728 where the segfault occurred in dgemm_oncopy_HASWELL.
Extract from the stack backtrace:
#5 0x0000ffffab2fe3b0 in zgemm_oncopy_EMAG8180 () from /usr/lib/aarch64-linux-gnu/libopenblas.so.0
#6 0x0000ffffaaac8430 in zgetrf_single () from /usr/lib/aarch64-linux-gnu/libopenblas.so.0
#7 0x0000ffffaa064f3c in zgesv_ () from /usr/lib/aarch64-linux-gnu/liblapack.so.3
#8 0x0000ffffaa00f330 in ?? () from /usr/lib/python3/dist-packages/numpy/linalg/_umath_linalg.cpython-38-aarch64-linux-gnu.so
You are probably right that this is a rerun of #2728 - I appear to have overlooked that common_arm64.h has the same disparity of BUFFERSIZE setting between models that caused the problems in DYNAMIC_ARCH builds on x86_64. (Need to check now if simple phone hardware supports the settings used for servers)
Hi,
a sagemath test segfaulted in openblas 0.3.10, libopenblas.so.0 / zgemm_oncopy_EMAG8180 when building on arm64, see https://buildd.debian.org/status/fetch.php?pkg=sagemath&arch=arm64&ver=9.2%7Ebeta12-1&stamp=1599922222&raw=0
openblas was called by numpy and the issue looks similar to #2728 where the segfault occurred in dgemm_oncopy_HASWELL.
Extract from the stack backtrace:
In Debian this is reported as https://bugs.debian.org/970558
Best,
Tobias
The text was updated successfully, but these errors were encountered: