topic hello, thanks for report. in Intel® oneAPI Math Kernel Library

error while using zgetri

Tai_Q_ — Tue, 03 Oct 2017 01:50:50 GMT

Dear all,

I am running a program that has been running many times in a cluster.

Maybe because the cluster has been through software upgrade, there are errors while running the executable file a.out.

There is no problem for compiling and linking. Just error will show up while run the program halfway..

forrtl: error (65): floating invalid
Image              PC                Routine            Line        Source
libifcoremt.so.5   00002B6454D7A6D4 for__signal_handl     Unknown Unknown
libpthread-2.17.s 00002B6452C20370 Unknown               Unknown Unknown
libmkl_avx512_mic 00002B646F5370BE mkl_blas_avx512_m     Unknown Unknown
libmkl_avx512_mic 00002B646F544B61 mkl_blas_avx512_m     Unknown Unknown
libmkl_avx512_mic 00002B646F541935 mkl_blas_avx512_m     Unknown Unknown
libmkl_intel_thre 00002B644F2FF714 mkl_blas_ztrsm_ho     Unknown Unknown
libmkl_intel_thre 00002B644F319606 mkl_blas_ztrsm        Unknown Unknown
libmkl_core.so     00002B64515C1F74 mkl_lapack_ztrtri     Unknown Unknown
libmkl_core.so     00002B64514B032C mkl_lapack_zgetri     Unknown Unknown
libmkl_intel_lp64 00002B644E98683D ZGETRI                Unknown Unknown

Now we are using intel/17.0.4, impi/17.0.3.

      call ZGETRF( N_LEN_2, N_LEN_2, BQ , N_LEN_2, IPIV , INFO )
 
      call ZGETRI( N_LEN_2, BQ, N_LEN_2, IPIV, WORK, N_LEN_2, INFO )

The first subroutine

ZGETRF

is fine. But when it comes to the second function

ZGETRI. There is always a floating invalid error.

I just do not understand. Because the input of ZGETRI are just the output of ZGTRF.

*********updates********

I found the following on Intel® Math Kernel Library (Intel® MKL) 2017 Release Notes

Fixed irregular division by zero and invalid floating point exceptions 
in {C/Z}TRSM for Intel® Xeon Phi™ processor x200 (aka KNL) and Intel® Xeon® 
Processor supporting Intel® Advanced Vector Extensions 512 (Intel® AVX-512) code path

I found this maybe useful because my error message just mentioned

TRSM, Invalid floating AVX-512

********updates2********

It seems the error has something to do with the MKL library.

1. The code has been running for a long time.

2. I run the code in a low version MKL library, it works well.

I think the current MKL library which is 17.0.4 must has something not correct.

hello, thanks for report.

Gennady_F_Intel — Tue, 03 Oct 2017 04:23:28 GMT

hello, thanks for report.

1. what is the current version you are use? could you look at mkl_version.h file? and

2 >>> I run the code in a low version MKL library, it works well.

what is the previous version which works well?

3. Could you give us the reproducer?

--Gennady

Hi Gennady,

Tai_Q_ — Mon, 09 Oct 2017 05:35:04 GMT

Hi Gennady,

Those days I have been debugging my code.

As I said, this code has been used for many years.

Previous intel mkl version is 16.0.3, now is 17.0.4 (17.0.1 is also fine for this code).

Even running the same case as before will produce errors.

I think the 17.0.4 must have some changes in dealing with complex numbers.

Thanks!

Tai

Hi Tai,

Gennady_F_Intel — Mon, 09 Oct 2017 06:03:42 GMT

Hi Tai,

could you try :

1. to check the input paramaters of zgetri by setting verbose mode MKL_VERBOSE=1

2. to check if the problem is caused by multithreading

call mkl_set_num_threads(1)

call zgetri(……)

// and then try to restore the original # of threads

call mkl_set_num_threads(#orig num of threads)

3. Could you captured the input data of zgetri and create the stanalone reproducer based on these data?

--Gennady