dsyev crashes on a simple matrix

[ansobolev]

Hi,

Thanks for solving the previous issue. I can confirm that it has indeed been fixed. 

With MKL 2026.0 we encountered another issue: when trying to calculate the eigenvalues of a simple matrix with dsyev, we get floating divide by zero. The MWE is below:

program dsyev_example
  implicit none

  real*8 :: matrix(3,3), ev(3), work(8)
  integer :: info

  matrix(:, 1) = (/523.720904015851,   0.0,              0.0              /).
  matrix(:, 2) = (/-87.7269993344388,  509.765510517387, 0.0              /)
  matrix(:, 3) = (/-87.7269993344388, -114.410201478550, 509.765510517387 /)

  call dsyev( 'V', 'U', 3, matrix, 3, ev, work, 8, info )

  print *, "Eigenvalues:"
  print *, ev

end program dsyev_example

When compiling this with `-fpe0`, we get 

forrtl: error (73): floating divide by zero
Image              PC                Routine            Line        Source             
libc.so.6          0000126CD7845330  Unknown               Unknown  Unknown
libmkl_core.so.3   0000126CD8EFA69F  mkl_lapack_dlaq5      Unknown  Unknown
libmkl_core.so.3   0000126CD8FAB391  mkl_lapack_dsteqr     Unknown  Unknown
libmkl_core.so.3   0000126CD8FB21C7  mkl_lapack_dsyev      Unknown  Unknown
libmkl_intel_lp64  0000126CDFE6A1D6  dsyev_                Unknown  Unknown
a.out              00000000004053A5  Unknown               Unknown  Unknown
a.out              000000000040528D  Unknown               Unknown  Unknown
libc.so.6          0000126CD782A1CA  Unknown               Unknown  Unknown
libc.so.6          0000126CD782A28B  __libc_start_main     Unknown  Unknown
a.out              00000000004051A5  Unknown               Unknown  Unknown

MKL 2025.3 and 2026.0 without `-mpe0` give the same correct results:

 Eigenvalues:
   319.854583405053        599.221603728736        624.175704956055    

Thanks in advance.

[Fengrui]

Hi,

Thanks for reporting this issue. I can reproduce it on my side. The oneMKL 2025.3 release doesn't have this issue.

We are currently working on it. I will keep you posted.

Thanks,

Fengrui

[Fengrui]

It turns out to be a known issue. It has been fixed and fix will be included in the next oneMKL release.

[ansobolev]

Do you mean 2026.0.1 or 2026.1.0 by the next release?

[Fengrui]

According to the current plan, that would be the next update release, 2026.1.0.

[ansobolev]

Hi @Fengrui, I checked it and this issue seems to still persist in MKL 2026.1.

[Fengrui]

Apologies for the mistake. It is actually included in the 2026.2 release which is scheduled to launch early October.