topic Hi Fabio, in Intel® oneAPI Math Kernel Library

LAPACKE_dsyev on Xeon Phi - eigenvalues on xeon phi

Fábio_A_ — Tue, 28 Jun 2016 00:58:01 GMT

Dear MKL Forum,

I am testing a Xeon Phi x100 family with 5GB using automatic offload, MKL composer_xe_2013.1.117 and icc 13.0.1, a function as cblas_dgemm is working well with xeon phi. However the function LAPACKE_dsyev is not using xeon Phi. The documentation says the ?syev should work with a N bigger then 8000, my matrix has 10000. There is any different configuration to calculate eigenvalues in xeon phi ?

Best regards

Hi Fabio,

Ying_H_Intel — Tue, 28 Jun 2016 01:59:01 GMT

Hi Fabio,

Is it possible to try later version (MKL 11.3. update 3, composer 2016 update 3)? The mkl in 2013.1.117 is about MKL 11.0.1, there were much modify since the version.

I check the release notes:

https://software.intel.com/en-us/articles/intel-mkl-110-release-notes

https://software.intel.com/en-us/articles/intel-mkl-111-release-notes

It mentioned some known issues like

There is no LAPACK Automatic Offload (AO) on Windows* OS for ilp64 interface. To enable AO LAPACK offload use 32-bit integer and lp64 interface

Service functions like mkl_mic_enable, mkl_mic_set_offload_report etc. are missing .

So please try latest version if possible, or please tell some details about your system, command line and the result if set

mkl_mic_set_offload_report=2.

Best Regards,

Ying

Dear Ying,

Fabio_A_1 — Tue, 28 Jun 2016 17:28:24 GMT

Dear Ying,

Now i am using the icc 16.0.3 and MKL 11.3.3- (i believe is the lastest). The DGEMM operation runs fine in the Xeo Phi, however the dsyev executes on the host.

The output from offload report:

[MKL] [MIC --] [AO Function] DGEMM
[MKL] [MIC --] [AO DGEMM Workdivision] 0.29 0.71
[MKL] [MIC 00] [AO DGEMM CPU Time] 11.060836 seconds
[MKL] [MIC 00] [AO DGEMM MIC Time] 8.325781 seconds

In the dsyev operation, the N is 10000,. Please, there is any way to know why the operation is not on the Xeon phi ?

Best regards

Dear Fabio,

Ying_H_Intel — Tue, 05 Jul 2016 06:23:32 GMT

Dear Fabio,

Thank you for your test. We can reproduce the problem. Our developer will investigate it and get back later if any result.

Best Regards,

Ying

Fabio, the issue is

Gennady_F_Intel — Wed, 13 Jul 2016 04:07:42 GMT

Fabio, the issue is reproduced on our side and escalated. The preliminary plan to provide the fix of the problem the next ( nearest) MKL update. We will let you know when this will happen. thanks.

Fabio, the latest 11.3.3

Gennady_F_Intel — Wed, 13 Jul 2016 08:01:09 GMT

Fabio, the latest 11.3.3 contains the fix of the problem. Here is the mkl_lapacke.h from 11.3.3 :

LAPACK_DECL
lapack_int LAPACKE_mkl_stppack( int matrix_layout, char uplo, char trans, lapack_int n,
float* ap, lapack_int i, lapack_int j, lapack_int rows,
lapack_int cols, const float* a, lapack_int lda );

lapack_int LAPACKE_mkl_stpunpack( int matrix_layout, char uplo, char trans, lapack_int n,
const float* ap, lapack_int i, lapack_int j, lapack_int rows,
lapack_int cols, float* a, lapack_int lda );

and etc ......

Dear Gennady Fedorov,

Fábio_A_ — Thu, 14 Jul 2016 01:14:54 GMT

Dear Gennady Fedorov,

I updated MKL to 11.3.3.210 and the with the function DGEMM still working fine, however the function

LAPACKE_dsyev(LAPACK_ROW_MAJOR,'V', 'U', m, temp, m,w); the M size is 10000

The processing still is out of Xeon Phi, this my compiling options:

LINK= -lpthread -ldl -Wl,--start-group ${MKLROOT}/lib/intel64/libmkl_intel_ilp64.a ${MKLROOT}/lib/intel64/libmkl_core.a ${MKLROOT}/lib/intel64/libmkl_intel_thread.a -Wl,--end-group -lpthread -lm -ldl

INC = -I. -L. -D INTEL -D OMP -O3 -qopenmp -Wall -DMKL_ILP64 -qopenmp -I${MKLROOT}/include

This is the offload report:

[MKL] [MIC --] [AO Function] DGEMM
[MKL] [MIC --] [AO DGEMM Workdivision] 0.00 1.00
[MKL] [MIC 00] [AO DGEMM CPU Time] 13.124382 seconds
[MKL] [MIC 00] [AO DGEMM MIC Time] 10.678612 seconds
[MKL] [MIC 00] [AO DGEMM CPU->MIC Data] 1600000000 bytes
[MKL] [MIC 00] [AO DGEMM MIC->CPU Data] 3200000000 bytes

Please, there is any way to find out why the LAPACKE_dsyev is not running on xeon phi ?

Best regards

sorry - this is wrong thread.

Gennady_F_Intel — Thu, 14 Jul 2016 03:39:31 GMT

sorry - this is wrong thread. I mean another issue which were fixed in update 3.

Please, there any new on this

Fábio_A_ — Tue, 01 Nov 2016 21:03:55 GMT

Please, there any new on this issue ?

Hi Fabio,

Ying_H_Intel — Wed, 02 Nov 2016 01:57:45 GMT

Hi Fabio,

The issue should be targeted to be fixed in 2017 update 1, which should be available for download in the two days. You are welcomed to try it out

Thanks

Ying

Hi Fabio,

Ying_H_Intel — Wed, 16 Nov 2016 01:18:24 GMT

Hi Fabio,

Any chance to try the MKL 2017 update 1?

Thanks

Ying

Dear Ying,

Fabio_A_1 — Tue, 29 Nov 2016 23:02:53 GMT

Dear Ying,

Now i am using the MKL 2017 update 1, some of the LAPACKE_dsyev was on Phi:

[MKL] [MIC --] [AO Function] DSYRDB
[MKL] [MIC --] [AO DSYRDB Workdivision] -1.00 -1.00
[MKL] [MIC 00] [AO DSYRDB CPU Time] 41.617044 seconds
[MKL] [MIC 00] [AO DSYRDB MIC Time] 13.513522 seconds
[MKL] [MIC 00] [AO DSYRDB CPU->MIC Data] 2792328424 bytes
[MKL] [MIC 00] [AO DSYRDB MIC->CPU Data] 796024744 byte

Thank all for the solution.

Best regards.