Failed check for miniapp_eigensolver with partial backtransformation

[albestro]

During last benchmark campaign some failed checks have been spotted for miniapp_eigensolver.

[2] 12.3568s dL (30097, 30097) (0, 1) (1024, 1024) 128 (4, 2) 64 GPU
ERROR: Max Diff / Max A: 6.02435e-14
--
[2] 18.3557s dL (40960, 40960) (0, 1) (512, 512) 128 (4, 2) 64 GPU
ERROR: Max Diff / Max A: 6.59472e-14
--
[2] 18.8038s dL (40960, 40960) (0, 1) (1024, 1024) 128 (4, 2) 64 GPU
ERROR: Max Diff / Max A: 6.15064e-14
--
[2] 2.74818s dL (10240, 10240) (0, 102) (512, 512) 128 (4, 2) 64 GPU
Warning: Max Diff / Max A: 3.59712e-14
--
[2] 3.24331s dL (10240, 10240) (0, 102) (1024, 1024) 128 (4, 2) 64 GPU
Warning: Max Diff / Max A: 4.86278e-14
--
[2] 6.89857s dL (20480, 20480) (0, 205) (512, 512) 128 (4, 2) 64 GPU
Warning: Max Diff / Max A: 6.93889e-14
--
[2] 7.65433s dL (20480, 20480) (0, 205) (1024, 1024) 128 (4, 2) 64 GPU
Warning: Max Diff / Max A: 6.90281e-14
--
[2] 11.7632s dL (30097, 30097) (0, 301) (512, 512) 128 (4, 2) 64 GPU
Warning: Max Diff / Max A: 1.02141e-13

Apparently it happens:

• for miniapp_eigensolver but not for miniapp_gen_eigensolver
• just for a number of eigenvalues less than tile size

We haven't investigated yet, but it might likely be something in the miniapp check rather than an actual numerical error. Moreover, it can't be excluded the same problem does not affect also miniapp_gen_eigensolver.

[RMeli]

just for a number of eigenvalues less than tile size

This is a case I haven't tested much, I'm afraid. I can have a look into this issue.

[msimberg]

@rasolca on slack:

https://github.com/eth-cscs/DLA-Future/blob/master/miniapp/miniapp_eigensolver.cpp#L344
n should be the size of the original matrix. The number of eigenvectors does not have any effect on the norm of the difference

here too:
https://github.com/eth-cscs/DLA-Future/blob/master/miniapp/miniapp_gen_eigensolver.cpp#L381