dfeast_sygv -4 error, BUT B IS POSITIVE DEFINITE!!!

[yera__rolando]

Hello all,

I am using the extended eigensolver routines specifically dfeast_sygv function, and I get the following error

==>INFO code =: -4
Intel MKL Extended Eigensolvers Error: Matrix B is not positive definite.

But the matrix B is positive definite!!!

B =

+2.222e-01      +0.000e+00      +1.111e-01      +0.000e+00      +1.111e-01      +0.000e+00      +5.556e-02      +0.000e+00
+0.000e+00      +2.222e-01      +0.000e+00      +1.111e-01      +0.000e+00      +1.111e-01      +0.000e+00      +5.556e-02
+1.111e-01      +0.000e+00      +2.222e-01      +0.000e+00      +5.556e-02      +0.000e+00      +1.111e-01      +0.000e+00
+0.000e+00      +1.111e-01      +0.000e+00      +2.222e-01      +0.000e+00      +5.556e-02      +0.000e+00      +1.111e-01
+1.111e-01      +0.000e+00      +5.556e-02      +0.000e+00      +2.222e-01      +0.000e+00      +1.111e-01      +0.000e+00
+0.000e+00      +1.111e-01      +0.000e+00      +5.556e-02      +0.000e+00      +2.222e-01      +0.000e+00      +1.111e-01
+5.556e-02      +0.000e+00      +1.111e-01      +0.000e+00      +1.111e-01      +0.000e+00      +2.222e-01      +0.000e+00
+0.000e+00      +5.556e-02      +0.000e+00      +1.111e-01      +0.000e+00      +1.111e-01      +0.000e+00      +2.222e-01

using dfeast_syev function give the following eigenvalues.

+5.556e-02
+5.556e-02
+1.667e-01
+1.667e-01
+1.667e-01
+1.667e-01
+5.000e-01
+5.000e-01

all positive.

Here go the matrix A for the general eigenvalues problem

A =

+4.569e+02      +0.000e+00      -3.046e+02      +0.000e+00      +7.616e+01      +0.000e+00      -2.285e+02      +0.000e+00
+0.000e+00      +4.569e+02      +0.000e+00      +7.616e+01      +0.000e+00      -3.046e+02      +0.000e+00      -2.285e+02
-3.046e+02      +0.000e+00      +4.569e+02      +0.000e+00      -2.285e+02      +0.000e+00      +7.616e+01      +0.000e+00
+0.000e+00      +7.616e+01      +0.000e+00      +4.569e+02      +0.000e+00      -2.285e+02      +0.000e+00      -3.046e+02
+7.616e+01      +0.000e+00      -2.285e+02      +0.000e+00      +4.569e+02      +0.000e+00      -3.046e+02      +0.000e+00
+0.000e+00      -3.046e+02      +0.000e+00      -2.285e+02      +0.000e+00      +4.569e+02      +0.000e+00      +7.616e+01
-2.285e+02      +0.000e+00      +7.616e+01      +0.000e+00      -3.046e+02      +0.000e+00      +4.569e+02      +0.000e+00
+0.000e+00      -2.285e+02      +0.000e+00      -3.046e+02      +0.000e+00      +7.616e+01      +0.000e+00      +4.569e+02

can anybody help me here???

I'm getting something wrong???

Im using linux (debian testing).

thanks in advance.

ps. Sorry about the english.

[Gennady_F_Intel]

Which version of mkl do you try? 

[yera__rolando]

Hi,

Im using MKL 2019 update 1 version, installed with parallel_studio_xe_2019_update1_cluster_edition package under student licence. Im also using the ICC version that comes in the package.

[Gennady_F_Intel]

Ok, could you get us the C or F reproducer to check this behavior on our side with the newest versions 2019.3?

[yera__rolando]

Hi,

Here go the src file.

[Gennady_F_Intel]

Thanks for noticing this case. This looks like the issue which will be escalated.  

[yera__rolando]

Hi,

I switch to the csr version "dfeast_scsrgv"  for the general eigenvalues problem and now I getting SIGSEGV error, the "dfeast_scsrev" works fine with B for the standard eigenvalues problem.

Here go the src file.

[yera__rolando]

There was a bug in the code of the previous src file (conformation of the rowIndex data), here go the fixed version, still the behavior of the "dfeast_scsrgv" remain the same.

[Gennady_F_Intel]

thanks for the update. 

[Gennady_F_Intel]

the matrix checker has been fixed in the latest 2019 u4 version of mkl.

[yera__rolando]

Thanks you very much!!!!