topic dsyevr for large matrices and minimum work sizes in Intel® oneAPI Math Kernel Library

dsyevr for large matrices and minimum work sizes

duesler__brad — Sat, 23 Dec 2017 00:32:20 GMT

I'm getting an error when calling dsyevr_ for large matrices (n>=6144) and when using minimally required work sizes (liwork = 10n and lwork = 26n). The error message is: "Intel MKL ERROR: Parameter 12 was incorrect on entry to DSBRDB." This error message seems to occur for any matrix of size greater than or equal to 6144. The error does not occur when using optimal liwork and lwork parameters - so it's not really a big deal. However, I've seen dozens of people post about this error when calling dsyevr indirectly via Anaconda Python scipy library (original call scipy.linalg.eigvalsh), which currently uses minimum work sizes.

I reproduced this issue on Mac OS X and Ubuntu - here is my build of "test.c", which calls fortran function dsyevr on identity matrix of rank 6144:

cc -m64 -I/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/include test.c -Wl,--start-group /opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_intel_lp64.a /opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_sequential.a /opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_core.a -Wl,--end-group -lpthread -lm -ldl -o test

the call is dsyevr_("N", "A", "U", ...) with minimum required lwork & liwork... all other parameters follow Intel's generic example for dsyevr_

I was also able to run this code without error on other LAPACK implementations (OpenBLAS).

Any thoughts on what's going on here? Mostly just curious since optimal work sizes solves the problem.

Thank you!

Could you share this test.c

Gennady_F_Intel — Sun, 24 Dec 2017 05:11:23 GMT

Could you share this test.c with us to quick checking the problem with the latest versions? we had fixed the similar issue in mkl 2018

https://software.intel.com/en-us/articles/intel-math-kernel-library-intel-mkl-2018-bug-fixes-list ; - see MKLD-3350

Here it is - I've commented

duesler__brad — Sun, 24 Dec 2017 12:27:37 GMT

Here it is - I've commented out the optimal work parameters (which work), and replaced with the minimums which returns all zeros and error message "Parameter 12 was incorrect on entry to DSBRDB."

Thank you!

Yes, I have to confirm that

Gennady_F_Intel — Mon, 25 Dec 2017 04:01:37 GMT

Yes, I have to confirm that the problem exists with MKL 2018 update 1 version too. We will fix the issue into one of the next updates and keep you informed with the status of this fix. Thanks for bringing this topic.

Hello. Could you please check

Gennady_F_Intel — Sat, 09 Jun 2018 04:01:13 GMT

Hello. Could you please check the latest mkl 2018 u3 ( released at the end of May'18) and let us know how this update will work on your side.

I can confirm that the

Horta__Danilo — Tue, 03 Jul 2018 10:08:32 GMT

I can confirm that the problem persists.

the fix of the problem

Gennady_F_Intel — Fri, 05 Oct 2018 03:57:55 GMT

the fix of the problem available in MKL 2019.0.

here the output of the example you gave us:

test_dsyevr.exe

DSYEVR Example Program Results

lwork: 159744

liwork: 61440

The DSYEBVR algorithm PASSED to compute Eigenvalues.

The total number of eigenvalues found:6144

Major version: 2019

Minor version: 0

Update version: 0

Product status: Product

Build: 20180829

Processor optimization: Intel(R) Advanced Vector Extensions 2 (Intel(R) AVX2) enabled processors

===============================================================

Great, thank you!

duesler__brad — Fri, 05 Oct 2018 11:13:38 GMT

Great, thank you!