Kalyan, At least one of my problems has been fixed in a newer version of 17.0: This version of the compiler (version 17.0 Build 20161005 [I think he pasted the wrong thing here]) crashes. However, I also checked the latest compiler build and it looks like the problem is fixed there: $ ./build.sh $ ifort -V Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0 Build 20161005 Copyright (C) 1985-2016 Intel Corporation. All rights reserved. So, please expect the fix available in the upcoming update package. I will escalate this to the enigneering team anyway to double check that. I will investigate the problem with wrong results next week and update you as well. Any idea of when we will get this? Thanks, Steve Steven C. Pieper, Senior Physicist Emeritus: [email protected] Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100 http://www.phy.anl.gov/theory/staff/scp.html On 10/06/16 22:01, Allen, Benjamin S. wrote:

License has been updated:

[bsallen@jlselogin2 ~]$ ifort -V Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0.0.042 Beta Build 20160315 Copyright (C) 1985-2016 Intel Corporation. All rights reserved.

ifort: NOTE: The Beta evaluation period for this product ends on 6-nov-2016 UTC.

Ben

...

On Oct 6, 2016, at 2:30 PM, Steven C. Pieper <[email protected]> wrote:

Hi Jeff,

I put !DIR$ NOVECTOR in front of every do loop (including all the nested do's). It still gives the wrong result. The optimization report acknowledges some of my !DIRS but only a few. Many of the other loops were unrolled, which I guess is not vectorization. But I don't know why -no-vector on the compile line has more effect.

Steve Steven C. Pieper, Senior Physicist Emeritus: [email protected]

Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100

http://www.phy.anl.gov/theory/staff/scp.html On 10/06/16 13:13, Hammond, Jeff R wrote:

...

Thanks, Steve.

I don’t know how many loops are in the relevant source file, but if you remove the -no-vec flag and bisect the code with “!DIR$ NOVECTOR” (https://software.intel.com/en-us/node/680017) on loops, you can localize the problem.

Martyn noted that one regression from beta to 17.0 release has already been fixed and will be in the update. It is possible that this will fix your problem. I have no ETA for 17.0 update 1 but my guess is that it will be available soon, based upon two important bug fixes I know about.

The license issue was addressed by Martyn already. Kumar and Ben have details. I assume they will update the license file, but in the event that your beta compiler stops working tomorrow, you can use the file-specific -no-vec workaround, or a loop-specific NOVECTOR if you can find where it is necessary.

Best,

Jeff

From: "Steven C. Pieper" <[email protected]> Organization: Physics, Argonne National Laboratory Reply-To: "[email protected]" <[email protected]> Date: Thursday, October 6, 2016 at 11:07 AM To: Jeff R Hammond <[email protected]>, "Kumaran, Kalyan" <[email protected]> Cc: "[email protected]" <[email protected]> Subject: Re: [JLSE-Intel-NDA] [EarlyScience] Production Intel compiler segfaults

Using -no-vec does fix the problem.

I have just submitted this and the compiler segfault bugs to the Intel site:

reference number is : 6000163313.

I included a 7 MB tar file which has a readme.

The jlse beta version still says:

[jlselogin1 ~] ifort -V Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0.0.042 Beta Build 20160315 Copyright (C) 1985-2016 Intel Corporation. All rights reserved.

ifort: NOTE: The Beta evaluation period for this product ends on 7-oct-2016 UTC.

Steve

Steven C. Pieper, Senior Physicist Emeritus: [email protected]

Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100

http://www.phy.anl.gov/theory/staff/scp.html On 10/05/16 18:35, Hammond, Jeff R wrote:

...

Martyn suggests you try with -no-vec. If that also fixes the issue, then you can narrow it down to the specific loop that breaks with novector directives.

Martyn said the beta license situation was fixed. I don’t yet know if any action is required on ALCF’s part.

Jeff

On 10/5/16, 1:04 PM, "Steven C. Pieper" <[email protected]> wrote:

...

Jeff and/or Kalyan:

I have isolated the wrong numerical result to one (large) subroutine. It is not the same one that causes the compiler to segfault; that subroutine is not involved in this wrong numerical result. I can compile everything with the production v17 using -O2 and openmp (I haven't tried more optimization yet). Then if I compile this one subroutine with -O0 and no openmp, it works. But if I compile it with -O2 and no openmp, it gives the wrong numbers. So this is not an OpenMP bug; it is a -O2 optimization error. This is a bit of a mess to make a demo for -- the subroutine is large and has a big table of input indices, etc., and it needs lots of other subroutines to have run before it is called.

As I said, all of this works fine with -O3, AVX512 with the v17-beta compiler.

I don't know what to do next. I can make an ugly package to give to you for Intel, but making a simple demo is probably too difficult. Is there any further update to the production v17 that could be installed for me to try?

Steve

Steven C. Pieper, Senior Physicist Emeritus:

[email protected]

Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100

http://www.phy.anl.gov/theory/staff/scp.html

On 10/05/16 13:40, Hammond, Jeff R wrote:

...

On 10/5/16, 11:04 AM, "[email protected] on behalf of Kumaran, Kalyan"

<[email protected] on behalf of [email protected]> wrote:

...

Steve, are you saying the Beta compilers have been removed, or will be removed in 2 days, from the JLSE cluster? If this is the case I can request Ben to keep it though I have not see any message about removing it.

I asked Martyn about this just now. Will post the substance of his response as soon as I get it.

...

Please do report bugs on the GA compilers. Without these I fear it will never get fixed.

Please report every bug to IPS. It¹s possible that the same bug will be identified by a different codes, but it is not good to rely upon this.

Jeff

...

-Kumar

...

On Oct 5, 2016, at 12:51 PM, Steven C. Pieper <[email protected]> wrote:

Tim, I didn't know where to send it. I just want it to be worked on. The non-beta V17 compiler is on jlse -- that is what is failing for me.

Some more recent info: Only this one subroutine file causes the compiler to segfault.

However much more seriously, I am getting the wrong numerical results using the non-beta compiler. They are wrong in that they differ from both the v17 beta and xlf on cetus. v17 beta and cetus give the same results. I'm trying now to narrow down where the wrong results are coming from.

Can we keep the beta compiler going for more than 2 more days????

Steve

Steven C. Pieper, Senior Physicist Emeritus:

[email protected]

Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100

http://www.phy.anl.gov/theory/staff/scp.html

On 10/05/16 12:45, Williams, Timothy J. wrote:

...

Hi Steve,

Did you mean to send this to early-users-discuss or intel-nda?

I think we should now have the non-beta V17 compilers on the white boxes. Will try to find out.

Tim

------- Timothy J. Williams, Ph.D. --- Argonne Leadership Computing Facility --- [email protected] --- 630-252-1154 -------

...

On Oct 5, 2016, at 11:25 AM, Steven C. Pieper <[email protected]>

wrote:

I am trying to recompile everything using the production intel compiler rather than the beta version which I previously used.

With the beta version, it works:

[jlselogin2 mc] source

/soft/compilers/intel_beta/compilers_and_libraries/linux/bin/compiler va rs.sh intel64 [jlselogin2 mc] ifort -V Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0.0.042 Beta Build 20160315 Copyright (C) 1985-2016 Intel Corporation. All rights reserved.

ifort: NOTE: The Beta evaluation period for this product ends on 7-oct-2016 UTC.

[jlselogin2 mc] ifort -FR -g -O2 -xMIC-AVX512 -qopenmp -lpthread -mcmodel=medium -r8 -WB -traceback -fpe3 -I../common/ -I../mc/ -fpp -DSPIN_SUF= -DPARAMNUC=nclude/paramnuc.12c-11_states-pn -DM0_DIV=1 -DDOCASE=0 -DPARAMCASE=nclude/paramcase.12c-11_states-pn -c make_3pi.f [jlselogin2 mc] <<compile is OK>

The production verson segfaults:

[jlselogin2 mc] source

/soft/compilers/intel/compilers_and_libraries/linux/bin/compilervars. sh intel64[jlselogin2 mc] ifort -V Intel(R) Fortran Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0.0.098 Build 20160721 Copyright (C) 1985-2016 Intel Corporation. All rights reserved.

[jlselogin2 mc] ifort -FR -g -O2 -xMIC-AVX512 -qopenmp -lpthread -mcmodel=medium -r8 -WB -traceback -fpe3 -I../common/ -I../mc/ -fpp -DSPIN_SUF= -DPARAMNUC=nclude/paramnuc.12c-11_states-pn -DM0_DIV=1 -DDOCASE=0 -DPARAMCASE=nclude/paramcase.12c-11_states-pn -c make_3pi.f catastrophic error: **Internal compiler error: segmentation violation signal raised** Please report this error along with the circumstances in which it occurred in a Software Problem Report. Note: File and line given may not be explicit cause of this error. compilation aborted for make_3pi.f (code 1)

If I remove the -xMIC-AVX512 then the production version compiles without error.

Lots of other subroutines are compiling OK.

The files are all in ~b21541/gfmc/mc/

I see the beta version is due to expire in 2 days. Can this be extended?

Steve

-- Steven C. Pieper, Senior Physicist Emeritus:

[email protected]

Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100

http://www.phy.anl.gov/theory/staff/scp.html

_______________________________________________ EarlyScience mailing list

[email protected]https://lists.alcf.anl.gov/mailman/listinfo/earlyscience

---

Intel-nda mailing list

[email protected]https://lists.jlse.anl.gov/mailman/listinfo/intel-nda

---

Intel-nda mailing list

[email protected]https://lists.jlse.anl.gov/mailman/listinfo/intel-nda

---

Intel-nda mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/intel-nda