Add PGI compiler toolchain. #1342

bartoldeman · 2015-08-07T02:56:17Z

This adds a PGI compiler toolchain, see also pull request for easybuilders/easybuild-easyblocks#658

hpcugentbot · 2015-08-07T02:56:18Z

Automatic reply from Jenkins: Can I test this?

boegel · 2015-08-07T06:22:43Z

easybuild/toolchains/compiler/pgi.py

@@ -0,0 +1,71 @@
+##
+# Copyright 2015 Bart Oldeman


transfer copyright to Ghent University, or see https://github.com/hpcugent/easybuild-framework/blob/master/CONTRIBUTING.md

nvm, triple-licensing also works

boegel · 2015-08-07T06:26:11Z

@georgets, @geimer, @rjeschmi: since you have looked into PGI support as well, please review? is a draft of what you were working on available somewhere?

boegel · 2015-08-07T07:03:56Z

Jenkins: ok to test

hpcugentbot · 2015-08-07T07:42:19Z

Refer to this link for build results (access rights to CI server needed):
https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/1945/
EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/1945/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

rjeschmi · 2015-08-08T02:40:41Z

I didn't get started on anything, but I ahve a need for this so will try to test.

hpcugentbot · 2015-08-08T03:14:19Z

Refer to this link for build results (access rights to CI server needed):
https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/1963/
EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/1963/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

rjeschmi · 2015-09-10T19:52:03Z

I have tested this and didn't have any problems.

@bartoldeman might want to rebase to develop

boegel · 2015-09-13T08:26:16Z

easybuild/toolchains/compiler/pgi.py

+    COMPILER_UNIQUE_OPTION_MAP = {
+        'i8': 'i8',
+        'r8': 'r8',
+        'optarch': '', # PGI by default generates code for the arch it is running on!


is there any way to disable this default behaviour, and specify the architecture to generate code for?

The target architecture can be specified using the -tp=<arch> option. PGI 14.1 supports the following architectures, newer versions may accept more (can't check atm):

px Generic x86 Processor k8 AMD64 Processor k8-64e AMD64 Processor rev E or later, 64-bit mode barcelona AMD Barcelona processor shanghai AMD Shanghai processor istanbul AMD Istanbul processor bulldozer AMD Bulldozer processor piledriver AMD Piledriver processor p7 Intel P7 Architecture (Pentium 4, Xeon, Centrino) core2 Intel Core-2 Architecture penryn Intel Penryn Architecture nehalem Intel Nehalem processor sandybridge Intel SandyBridge processor haswell Intel Haswell processor x64 Unified AMD/Intel mode; -tp=k8,p7

@geimer: thanks for the feedback!

It would be interesting to somehow enforce picking one of these in case the optarch toolchain option is set to False.
The default is True, which (apparently) corresponds to not specifying any specific flag with PGI, but every now and then False is specified explicitly, so we should not just ignore that.

Maybe we need to support a custom toolchain option for PGI-based toolchains, something like target-platform?

Or, we should reuse optarch for this? If optarch is True, we don't add any specific flags, if it's False we complain to specify the name of a target platform instead?

target-platform would be interesting for other compilers as well, however, I thought the recommended way of handling this is the --optarch=<arg> command-line option. Why not simply use a generic target like px or x64 if optarch is False? AFAIK, this is the default behavior of other compilers. For PGI you have to explicitly request it.

OK, that makes sense too, indeed. x64 seems like a decent default indeed, and people can always override via --optarch.

@bartoldeman: are you up for tackling this yourself, i.e. using -tp=x64 if optarch is set to False? See https://github.com/hpcugent/easybuild-framework/blob/master/easybuild/toolchains/compiler/craype.py#L100 for inspiration.

Done below in _set_compiler_flags, thanks @bartoldeman.

ocaisa · 2015-10-23T14:05:38Z

I spoke to someone from NVIDIA yesterday and he mentioned that C++ support in PGI depends on the underlying GCC. I think it might be wise to build up the PGI toolchain in the same way as the intel toolchain (bintuils -> GCC + bintuils -> PGI) to side step problems from the underlying OS (like RedHAt shipping an ancient GCC and a binutils that doesn't support modern architectures)...I guess this doesn't matter here but it does in the associated easyconfigs.

boegel · 2015-10-23T14:54:02Z

@ocaisa: already done that way, see easybuilders/easybuild-easyconfigs#1833

ocaisa · 2015-10-23T14:57:02Z

eb --delete-lazy-comment

boegel · 2015-10-23T14:57:26Z

@ocaisa: please make that a separate PR ;)

boegel · 2016-01-09T22:22:40Z

Other stuff is coming in that is built on top of this (cfr. #1537), so we should really try and get this merged.

@bartoldeman: we haven't heard from you in a while on this, are you still willing to follow up on this, or should we take it from here?

bartoldeman · 2016-04-05T00:19:38Z

I rebased this PR and updated with the change from @jhein32 so pgf77 is used for F77.

hpcugentbot · 2016-04-05T01:03:18Z

EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/2927/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

damianam · 2016-04-07T07:48:15Z

easybuild/toolchains/pgi.py

+@author: Bart Oldeman (McGill University, Calcul Quebec, Compute Canada)
+"""
+
+from easybuild.toolchains.compiler.pgi import Pgi


Shouldn't this import and use GCCcore as well? PGI compilers need an underlying GCC and I guess also binutils. Without GCCcore here any software installed there won't be available for PGI-based toolchains when using minimal toolchains.

yes, this means just catching up with what is happening in easybuild... I put this in the same way as for iccifort.py.

Thanks @bartoldeman!

hpcugentbot · 2016-04-07T19:26:56Z

EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/2929/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

hpcugentbot · 2016-04-12T14:36:03Z

EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/2932/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

bartoldeman · 2016-04-13T16:20:30Z

Here is the story with numactl (I mentioned that at the conf call):
any program that you compile with pgcc (with or without -mp) is linked to libnuma:

$ pgcc hello.c
$ ldd a.out 
        linux-vdso.so.1 =>  (0x00007ffc6676b000)
        libpgmp.so => /sb/software/CentOS-6/eb/software/Core/PGI/16.3-GCC-4.9.3-2.25/linux86-64/16.3/lib/libpgmp.so (0x00002ae2cc3ca000)
        libnuma.so.1 => /usr/lib64/libnuma.so.1 (0x0000003b96e00000)
        libpthread.so.0 => /lib64/libpthread.so.0 (0x0000003b95a00000)
        libpgc.so => /sb/software/CentOS-6/eb/software/Core/PGI/16.3-GCC-4.9.3-2.25/linux86-64/16.3/lib/libpgc.so (0x00002ae2cc66c000)
        libm.so.6 => /lib64/libm.so.6 (0x0000003b95e00000)
        libc.so.6 => /lib64/libc.so.6 (0x0000003b95600000)
        /lib64/ld-linux-x86-64.so.2 (0x0000003b95200000)

If the PGI installer detects /usr/lib64/libnuma.so.1 it will use that, if not, it will symlink linux86-64/16.3/lib/libnuma.so and ....so.1 to linux86-64/16.3/lib/libpgnuma.so. EB has its own numactl however. Perhaps we need numactl built with GCCcore?

valtandor · 2016-04-21T02:11:53Z

How is this toolchain coming along? Does it need some additional testing?

boegel · 2016-04-21T06:40:45Z

@valtandor things are slowly aligning to get this merged, see also easybuilders/easybuild#211

A main issue on my end is that I need to set up the license that was provided to me by PGI, for testing purposes. That's less straightforward than I was hoping it would be, but I'm getting there...

boegel · 2016-05-02T17:57:24Z

@bartoldeman do you mind updating this with latest develop, to trigger Travis testing your PR as well (we're working on moving from Jenkins to Travis, and I've enabled Travis tests as required to be able to merge PRs)?

Let me know if you need help with that, I can easily send you a PR for it and that's easier for you.

…to iccifort.py.

hpcugentbot · 2016-05-09T18:17:14Z

EasyBuild framework unit test suite PASSed (see https://jenkins1.ugent.be/job/easybuild-framework-pr-builder/3009/console for more details).

This pull request is now ready for review/testing.

Please try and find someone who can tackle this; contact @boegel if you're not sure what to do.

boegel · 2016-05-09T18:20:49Z

All green, so this is good to go...

I'd like to get easybuilders/easybuild-easyconfigs#2899 merge-ready first though.

It pretty much should be, but we've hit a snag with the module naming scheme @bartoldeman is using, causing the OpenMPI sanity check to fail. I'd like to figure that out first...

@ocaisa, @damianam: have you guys ever used PGI with ToolchainMNS, or another hierarchical module naming scheme?

boegel · 2016-05-10T19:28:07Z

#1754 fixes the bug that was holding this PR back, and is merged now, so this is good to go as well

Thanks a lot for your work on this @bartoldeman, and for your patience in following this up; and of course also thanks to everyone else who providing feedback and testing!

boegel reviewed Aug 7, 2015
View reviewed changes

boegel reviewed Sep 13, 2015
View reviewed changes

boegel added this to the v2.4.0 milestone Sep 14, 2015

boegel modified the milestones: v2.4.0, v2.5.0 Oct 28, 2015

boegel modified the milestones: v2.5.0, v2.6.0 Dec 14, 2015

jgphpc mentioned this pull request Jan 8, 2016

CrayPGI toolchain support #1536

Closed

boegel mentioned this pull request Jan 9, 2016

trying to fix 1342_pgisupport #1537

Closed

bartoldeman force-pushed the pgi branch from 4690479 to 33b87c5 Compare April 5, 2016 00:18

damianam reviewed Apr 7, 2016
View reviewed changes

boegel mentioned this pull request Apr 15, 2016

PGI support (roadmap) easybuilders/easybuild#211

Closed

10 tasks

geimer mentioned this pull request Apr 18, 2016

Score-P easyblock update (REVIEW) easybuilders/easybuild-easyblocks#889

Merged

7 tasks

jgphpc mentioned this pull request Apr 20, 2016

PrgEnv-pgi support #1729

Merged

bartoldeman and others added 9 commits May 9, 2016 13:15

Add PGI compiler toolchain.

b32fdc7

Change copyright, add some comments.

ebd3a75

Fix capitalization for PGI

ef396f6

PGI: add support for --optarch=GENERIC

abc917e

Also define $FC in build environment

0d60482

PGI: Set -tp=x64 if optarch is set to False.

ef2887c

PGI: change COMPILER_F77 to pgf77 as suggested by @jhein32.

a3f9af3

PGI: Use GCCcore as subtoolchain, as suggested by @damianam, similar …

8d0dc88

…to iccifort.py.

PGI: add more references as requested by @boegel.

8f96aa1

bartoldeman force-pushed the pgi branch from a4f2495 to 8f96aa1 Compare May 9, 2016 17:15

boegel merged commit e2eefd1 into easybuilders:develop May 10, 2016

Add PGI compiler toolchain. #1342

Add PGI compiler toolchain. #1342

Conversation

bartoldeman commented Aug 7, 2015 • edited by boegel Loading

hpcugentbot commented Aug 7, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

boegel commented Aug 7, 2015

boegel commented Aug 7, 2015

hpcugentbot commented Aug 7, 2015

rjeschmi commented Aug 8, 2015

hpcugentbot commented Aug 8, 2015

rjeschmi commented Sep 10, 2015

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ocaisa commented Oct 23, 2015

boegel commented Oct 23, 2015

ocaisa commented Oct 23, 2015

boegel commented Oct 23, 2015

boegel commented Jan 9, 2016

bartoldeman commented Apr 5, 2016

hpcugentbot commented Apr 5, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

hpcugentbot commented Apr 7, 2016

hpcugentbot commented Apr 12, 2016

bartoldeman commented Apr 13, 2016

valtandor commented Apr 21, 2016

boegel commented Apr 21, 2016

boegel commented May 2, 2016

hpcugentbot commented May 9, 2016

boegel commented May 9, 2016

boegel commented May 10, 2016

bartoldeman commented Aug 7, 2015 •

edited by boegel

Loading