bug when used openmp

[nvarini]

Dear all, I found a bug with mkl openmp on DSYEV. My code read a matrix of 250x250 and perform a dsyev on it. I print out the eigenvectors and eigenvalues and there are difference in the eigenvector when I use 1, 2, 4 and 8 threads.
I compile with this command:
ifort -openmp -o prova prova.f90 -Wl,--start-group $MKL_LIB/libmkl_intel_lp64.a $MKL_LIB/libmkl_intel_thread.a $MKL_LIB/libmkl_core.a -Wl,--end-group

It's possible to fix this bug?

Best regard

Nicola

[Gennady_F_Intel]

Nikola,

what magnitude of these differences?

--Gennady

[nvarini]

Hi Gennady, this is the first two eigenvector with 1 thread

5.206780874519512E-003 -2.512377498876253E-003 -0.560713995878056
-6.678309445599694E-002 1.730343130346750E-003 8.894833897540037E-002

while with 2 thread the first two eigenvector are

-3.956413109619314E-002 -8.250475621178963E-003 -0.514086076699900
-0.128043245900761 1.957126952260208E-002 7.716034330953539E-002

whit 4 thread

-3.923533964555784E-002 -8.579275778466389E-003 -0.514086076699883
-0.128043245900755 1.957126952259620E-002 7.716034330954206E-002

with 8 thread

-3.923533964555784E-002 -8.579275778466389E-003 -0.514086076699883
-0.128043245900755 1.957126952259620E-002 7.716034330954206E-002




--Nicola

[Gennady_F_Intel]

Hi Nicola,

Thanks for the update. If I am not mistaken, we had the similar problem in one of MKL 10.2 Update 2.

What MKL version you are using?

--Gennady

[nvarini]

In fact I use mkl 10.2.2, there are some patch for this problem?

Nicola

[Gennady_F_Intel]

yes, the latest Update 5 will be available soon, but you cancheck this problem with Update 3 or 4.

Please see here the list of fixed errors. See Update 3:

DPD200141616

Eigenvector bug in multi-threaded code of ?STEQR