[MRG] Fix pre-build checks to handle compilers from build_ext options #16193

jeremiedbb · 2020-01-24T12:38:02Z

There are (at least ?) 2 ways to specify the compiler to build the project:

CC=<compiler> python setup.py build_ext -i
python setupy.py build_ext --compiler=<compiler> -i

In the second option, <compiler> must be one of the predefined compilers in distutils or in numpy.distutils. For example, to build with icc, it's --compiler=intelem.

Currently the checks before building sklearn (compile a test program and compile a test program using openmp) ignore the compiler specified through the second option.

I also updated get_openmp_flag to better handle icc on linux (although it accepts -fopenmp, it's recommended to use -qopenmp).

I temporarily modified a ci job, Linux_Runs_pylatest_conda_mkl, to build sklearn with icc (installed through intel oneAPI) to check it works as expected. The install of icc is actually pretty fast and the job is not longer than the other ones. We could keep a job where we build with icc, not necessarily this one. wdyt ?

@oleksandr-pavlyk I'd like your feedback on this. Especially, do you use the same command to build sklearn with icc ?

…learn + support icc linux openmp flag

jeremiedbb · 2020-01-24T13:56:01Z

The doc test of SpectralCoclustering fails randomly. I don't know if it's specific to building with icc, but it occurs quite often.

oleksandr-pavlyk · 2020-01-24T14:22:17Z

We use CC=icc python setup.py config_cc --compiler=intelem install.

The CC=icc is necessary because config_cc somehow does not affect build_clib's notion of compiler.

Using ICC implies that numpy.distutils's compiler options are going to be used, and it likely uses -O3.

jeremiedbb · 2020-01-24T14:26:36Z

this command
python setup.py build_ext --compiler=intelem -i build_clib --compiler=intelem
allows to specify the compiler for both build_ext and build_clib without requiring to set CC.

NicolasHug · 2020-01-24T14:27:50Z

The doc test of SpectralCoclustering fails randomly. I don't know if it's specific to building with icc, but it occurs quite often.

I don't think it's related to this PR since it also occurred in #16175

jeremiedbb · 2020-01-24T14:31:25Z

Thanks for the info. I wonder if it occurs more often than with gcc because it's like in 2 out of 3 runs.

jeremiedbb · 2020-01-24T14:32:17Z

Using ICC implies that numpy.distutils's compiler options are going to be used, and it likely uses -O3.

actually it's also -O3 with gcc

oleksandr-pavlyk · 2020-01-24T14:37:36Z

sklearn/_build_utils/openmp_helpers.py

@@ -40,6 +40,8 @@ def get_openmp_flag(compiler):
        # export LDFLAGS="$LDFLAGS -Wl,-rpath,/usr/local/opt/libomp/lib
        #                          -L/usr/local/opt/libomp/lib -lomp"
        return []
+    elif sys.platform == "linux" and "icc" in compiler:


I think this applies to sys.platform == "darwin" as well.

For darwin we set -openmp. Is it outdated or preferable to use -qopenmp ?

also is icl something ? I remember putting icc or icl because I had the feeling that it could be a name of the icc executable but I actually never see that anywhere.

Yes, it is icl on Windows.

ogrisel

Some comments below.

Why would be the motivation to extend our CI to build with another compiler? To "shake" our test suite further to reveal more potential numerical stability issues?

The ICC installation overhead seems minimal when using the APT repo as done in this PR. However I am not sure what will be the long term maintenance overhead.

build_tools/azure/install.sh

build_tools/azure/test_docs.sh

jeremiedbb · 2020-01-27T15:54:30Z

Following the discussion in the meeting, I added a travis cron job where sklearn is built with icc instead of a regular azure job. I think it's ready to merge now.

thomasjpfan · 2020-01-27T17:30:54Z

.travis.yml

+      env:
+        - CHECK_WARNING="true"
+        - BUILD_WITH_ICC="true"
+      if: type = cron OR commit_message =~ /\[scipy-dev\]/


Do we want to use the same scipy-dev tag for this?

I put the same name to avoid having options that we will not remember :)
But I'm ok to put a different name. icc-build ?

icc or icc-build both are fine to me.

oleksandr-pavlyk · 2020-01-27T21:18:20Z

When I try to run the build command, I get:

(sk-dev2) [15:14:23 vmlin scikit-learn_src]$ python setup.py build_ext --compiler=intelem -i -j 4 build_clib --compiler=intelem
Partial import of sklearn during the build process.
/localdisk/miniconda3_latest/envs/sk-dev2/lib/python3.6/distutils/dist.py:261: UserWarning: Unknown distribution option: 'project_urls'
  warnings.warn(msg)
/localdisk/miniconda3_latest/envs/sk-dev2/lib/python3.6/distutils/dist.py:261: UserWarning: Unknown distribution option: 'python_requires'
  warnings.warn(msg)
/localdisk/miniconda3_latest/envs/sk-dev2/lib/python3.6/distutils/dist.py:261: UserWarning: Unknown distribution option: 'install_requires'
  warnings.warn(msg)
C compiler: gcc -pthread -Wno-unused-result -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wformat -Wformat-security -fstack-protector-all -D_FORTIFY_SOURCE=2 -fpic -fPIC -O3 -Wformat -Wforma
t-security -fstack-protector-all -D_FORTIFY_SOURCE=2 -fpic -fPIC -O3 -fPIC

compile options: '-c'
gcc: test_program.c
gcc -pthread objects/test_program.o -o test_program
C compiler: gcc -pthread -Wno-unused-result -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wformat -Wformat-security -fstack-protector-all -D_FORTIFY_SOURCE=2 -fpic -fPIC -O3 -Wformat -Wforma
t-security -fstack-protector-all -D_FORTIFY_SOURCE=2 -fpic -fPIC -O3 -fPIC

compile options: '-c'
extra options: '-fopenmp'
gcc: test_program.c
gcc -pthread objects/test_program.o -o test_program -fopenmp
Compiling sklearn/__check_build/_check_build.pyx because it changed.
....

So it is reporting that GCC is being used to test openmp flags. This is unexpected. The only way I found to work around this is to set CC=icc.

Edit: I realized I built master branch, let me try with your branch.

oleksandr-pavlyk · 2020-01-27T21:25:07Z

It behaves as advertised in the branch, however I found that

python setup.py config_cc --compiler=intelem build_ext --compiler=intelem -i -j 4 build_clib --compiler=intelem

which used to execute in master (albeit using gcc) is now throwing an error:

distutils.errors.DistutilsArgError: invalid command 'config_cc'

jeremiedbb · 2020-02-04T09:45:55Z

@oleksandr-pavlyk Thanks for testing it and for the feedback.

distutils.errors.DistutilsArgError: invalid command 'config_cc'

I fixed that. But I don't think you need config_cc anymore now by specifying the compiler to build_ext and build_clib.

ogrisel

LGTM. Could you please add a short section at the end of the advanced installation page in the doc to summarize the main commands to build with ICC (under Linux with the oneapi repo).

ogrisel · 2020-02-08T17:24:52Z

Thanks for the doc. LGTM again.

oleksandr-pavlyk · 2020-02-11T16:01:25Z

BTW, please be advised that oneAPI is currently governed by a beta license, which precludes distribution of produced binaries. The license will change once oneAPI is released.

agramfort

I think it's very valuable to test with ICC and document how to build sklearn with it.

thx @jeremiedbb

…cikit-learn#16193)

compile test program with same compiler as will be used to compile sk…

5ec4417

…learn + support icc linux openmp flag

jeremiedbb added the Build / CI label Jan 24, 2020

jeremiedbb requested a review from ogrisel January 24, 2020 12:38

jeremiedbb added 2 commits January 24, 2020 13:59

trigger ci

c57451c

trigger ci

256be7f

oleksandr-pavlyk reviewed Jan 24, 2020

View reviewed changes

openmp flag for icc on linux and macos

7d0d40e

ogrisel reviewed Jan 27, 2020

View reviewed changes

build_tools/azure/install.sh Outdated Show resolved Hide resolved

build_tools/azure/test_docs.sh Outdated Show resolved Hide resolved

jeremiedbb added 9 commits January 27, 2020 15:19

icc on cron job

3068d00

tst [scipy-dev]

a8d4df9

[scipy-dev] tst

23eb9f2

[scipy-dev] cln

5180542

[scipy-dev] fix yml

fa29257

[scipy-dev] tst

b3b3a71

tst [scipy-dev]

386de12

[scipy-dev] tst

40b3415

[scipy-dev] fix install script

874ed98

jeremiedbb changed the title ~~[WIP] Fix pre-build checks to handle compilers from build_ext options~~ [MRG] Fix pre-build checks to handle compilers from build_ext options Jan 27, 2020

thomasjpfan reviewed Jan 27, 2020

View reviewed changes

support config_cc

11a2b55

oleksandr-pavlyk approved these changes Feb 4, 2020

View reviewed changes

ogrisel approved these changes Feb 6, 2020

View reviewed changes

instructions to build with ICC using oneAPI on Linux

661fa48

agramfort approved these changes Jul 10, 2020

View reviewed changes

glemaitre merged commit cac1672 into scikit-learn:master Jul 10, 2020

jeremiedbb mentioned this pull request Jul 10, 2020

[MRG] BLD fix CI #17887

Merged

jayzed82 pushed a commit to jayzed82/scikit-learn that referenced this pull request Oct 22, 2020

CI Fix pre-build checks to handle compilers from build_ext options (s…

61eb83d

…cikit-learn#16193)

Uh oh!

[MRG] Fix pre-build checks to handle compilers from build_ext options #16193

[MRG] Fix pre-build checks to handle compilers from build_ext options #16193

Uh oh!

Conversation

jeremiedbb commented Jan 24, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jeremiedbb commented Jan 24, 2020

Uh oh!

oleksandr-pavlyk commented Jan 24, 2020

Uh oh!

jeremiedbb commented Jan 24, 2020

Uh oh!

NicolasHug commented Jan 24, 2020

Uh oh!

jeremiedbb commented Jan 24, 2020

Uh oh!

jeremiedbb commented Jan 24, 2020

Uh oh!

oleksandr-pavlyk Jan 24, 2020

Choose a reason for hiding this comment

Uh oh!

jeremiedbb Jan 24, 2020

Choose a reason for hiding this comment

Uh oh!

jeremiedbb Jan 24, 2020

Choose a reason for hiding this comment

Uh oh!

oleksandr-pavlyk Jan 24, 2020

Choose a reason for hiding this comment

Uh oh!

ogrisel left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

jeremiedbb commented Jan 27, 2020

Uh oh!

thomasjpfan Jan 27, 2020

Choose a reason for hiding this comment

Uh oh!

jeremiedbb Feb 4, 2020

Choose a reason for hiding this comment

Uh oh!

ogrisel Feb 6, 2020

Choose a reason for hiding this comment

Uh oh!

oleksandr-pavlyk commented Jan 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

oleksandr-pavlyk commented Jan 27, 2020

Uh oh!

jeremiedbb commented Feb 4, 2020

Uh oh!

ogrisel left a comment

Choose a reason for hiding this comment

Uh oh!

ogrisel commented Feb 8, 2020

Uh oh!

oleksandr-pavlyk commented Feb 11, 2020

Uh oh!

agramfort left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jeremiedbb commented Jan 24, 2020 •

edited

Loading

oleksandr-pavlyk commented Jan 27, 2020 •

edited

Loading