BUGS.txt

Known bugs & issues

These show up when using testing run_tests.py --tol 100.


Nov 2017
When compiling MAGMA with CUDA 9.0RC, the compiler issues numerous warnings:
    dereferencing type-punned pointer will break strict-aliasing rules.
This appears to be an issue with CUDA, not with MAGMA.


Nov 2017
When using make.inc.mkl-gcc-ilp64 and OMP_NUM_THREADS > 1, the LAPACK matrix
norm function lange returns wrong results, causing some MAGMA testers to fail.
Bug reported to Intel.


Nov 2016
zheevd and zhetrd (all versions, uplo=lower) -- segfault
Using OpenBLAS 0.2.19; appears to be in zhemv_l function.
With previous versions of OpenBLAS, also crashed in ssyevd.
Crashes do NOT occur when using Intel MKL.
Suspect OpenBLAS bug.


Jan 2016
[cz]heevr                               -- accuracy failures, xerbla errors
Fails when computing only eigenvalues (jobz=N).


Jan 2016
[sdcz]geqlf, geqrf, lange, larfb, gesdd, gesvd  --  accuracy failures
Several routines fail accuracy tests for very tall or very wide matrices
(e.g., -N 2000000,10 -N 10,2000000). These may simply be not normalized correctly.  


Jan 2016
[sdcz]geqr2x_gpu --version 2 and 4      --  accuracy failures
Fail orthogonality checks, |I - Q^H*Q|.


Aug 2015
[sdcz]trsm_batched                      --  accuracy failures
These result from badly conditioned triangular matrices.
The tester needs to be updated to take the conditioning into account in the check.


Aug 2015
[s c ]geqrf_batched                     --  transient NAN failures
Occasionally gets NaN results, possibly due to race condition.


Jan 2015
[sdcz]sy/hemm_mgpu                      --  accuracy failures
In a multi-GPU run, some GPUs report failed results for some offsets.
It does not appear to affect the eigenvalue routines where it is used.
Suspect that the tester needs to be updated to match the actual routine output.


Aug 2015
[sd  ]gegqr_gpu --version 3             --  MGS accuracy failure, N=95
[sd  ]gegqr_gpu --version 4             --  Cholesky QR fails (not SPD), N=95
These occur for the same matrix, so are probably related to a deficiency, e.g.,
poor conditioning, in that matrix.