mkl_dcoomm changed behaviour from 10.1 to 10.2 (Update 2)

[AndrewC]

I have been tracking down a change in a regression case when attempting to upgrade from 10.1 to 10.2.

mkl_dcoomm is 'failing' under 10.2, whereas under 10.1 it worked as expected. As far as I can tell under 10.2 the routine is unhappy with the input and returning early. xerbla is not being called.
The sparse input is a symmetric matrix with fortran indexing i.e. matdescra='SUNF'

One thing that is not clear from the documentation is whether the symmetric sparse structure input to mkl_dcoomm is required to include the diagonal elements even if they are 0.0. I realize this is required for sparse input to the direct solvers, but there is nothing in the documentation to indicate this is a requirement for matrix/matrix or matrix/vector multiplcation routines.

I have submitted a premier support issue....

[Gennady_F_Intel]

Quoting - vasci_intel

I have been tracking down a change in a regression case when attempting to upgrade from 10.1 to 10.2.

mkl_dcoomm is 'failing' under 10.2, whereas under 10.1 it worked as expected. As far as I can tell under 10.2 the routine is unhappy with the input and returning early. xerbla is not being called.
The sparse input is a symmetric matrix with fortran indexing i.e. matdescra='SUNF'

One thing that is not clear from the documentation is whether the symmetric sparse structure input to mkl_dcoomm is required to include the diagonal elements even if they are 0.0. I realize this is required for sparse input to the direct solvers, but there is nothing in the documentation to indicate this is a requirement for matrix/matrix or matrix/vector multiplcation routines.

I have submitted a premier support issue....

at least on win32 the output results are the same for MKL 10.2 update 2 (Package ID: w_mkl_p_10.2.2.025)
and MKL 10.1 (Package ID: w_mkl_p_10.1.2.026).
--Gennady

[AndrewC]

Quoting - Gennady Fedorov (Intel)

at least on win32 the output results are the same for MKL 10.2 update 2 (Package ID: w_mkl_p_10.2.2.025)
and MKL 10.1 (Package ID: w_mkl_p_10.1.2.026).
--Gennady

I am sure Gennady will weigh in on this, but this turned out to be a real bug (that only occurs with larger size problems) and can be repoduced on Linux64/32 as well as Windows. Gennady has provided great support and my understanding is that this will be fixed in an Update 3. I does also affect mkl_dcscmm and friends.

[Gennady_F_Intel · Accepted solution]

Quoting - vasci_intel

I am sure Gennady will weigh in on this, but this turned out to be a real bug (that only occurs with larger size problems) and can be repoduced on Linux64/32 as well as Windows. Gennady has provided great support and my understanding is that this will be fixed in an Update 3. I does also affect mkl_dcscmm and friends.

Yes, this is the bug caused by the internal threading. The bug has been already fixed and, we plan, will be available next update 3. I will announce at the Forum when Update 3 is available.

[AndrewC]

Quoting - Gennady Fedorov (Intel)

Yes, this is the bug caused by the internal threading. The bug has been already fixed and, we plan, will be available next update 3. I will announce at the Forum when Update 3 is available.

The 11.1 Professional Compiler Update 3 has been released. Is the fix included?

[Shane_S_Intel]

Gennady likely meant MKL 10.2 Update 3 whichis planned toship with the "next" Compiler Pro update, 11.1 Update 4.

[AndrewC]

Quoting - Shane Story (Intel)

Gennady likely meant MKL 10.2 Update 3 whichis planned toship with the "next" Compiler Pro update, 11.1 Update 4.

Ok, but this has been a painful bug for us - for the third time, I have had to revert back to 10.1 from 10.2. We were very lucky our QA regression procedures caught this bug - unlike MKL regression testing which obviously needs better coverage!