Fortran + OSX + mpich (3.4) issues

[hhslepicka]

• [x] I read the conda-forge documentation and could not find the solution for my problem there.

Issue: One of my recipes started failing when building with MPICH and OSX. The previous build worked out fine so something probably changed along the way and started resulting into errors. This is the most recent build: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=270799&view=logs&j=b4588902-138a-5967-ecc7-b3fc381bfda2&t=5a7a20e7-b634-5369-ebb8-6b51f51eb32a&l=1152 The previous one that worked is: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=230160&view=results

I tracked down to be related to the new mpich == 3.4. Downgrading to 3.3.2 seems to solve the problem and produce a good build.

Environment (conda list):

Available at the CI link above.

Details about conda and system ( conda info ):

Available at the CI link above

[leofang]

I begin wondering if it's because when updating to 3.4 a patch was applied too aggressively...see https://github.com/conda-forge/mpich-feedstock/pull/53#issuecomment-761956468.

I just pushed a PR (#57) to limit it to OS X Arm64 builds, for which the patch was intended, but I am not sure if there is a way to verify it before it's merged...Thoughts? @conda-forge/mpich

[leofang]

ref: https://github.com/conda-forge/genesis2-feedstock/pull/4

[hhslepicka]

@leofang in theory if the patch was just for osx arm64 it should be fine to merge your PR if tests are passing.

[leofang]

@hhslepicka @awvwgk #57 is in. Could you try rebuilding without pinning mpich and see if it's fixed?

[awvwgk]

Sure can do.

Edit: Still fails: https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=270964&view=logs&j=b4588902-138a-5967-ecc7-b3fc381bfda2&t=5a7a20e7-b634-5369-ebb8-6b51f51eb32a

[hhslepicka]

@leofang I will give it a try now and see how it goes.

[hhslepicka]

No deal with the unpin... https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=271074&view=logs&j=b4588902-138a-5967-ecc7-b3fc381bfda2&t=5a7a20e7-b634-5369-ebb8-6b51f51eb32a&l=1152

Should I also add mpich-mpicc to my recipe as a dependency when osx? (ref. https://github.com/conda-forge/mpich-feedstock/issues/55#issuecomment-763453553)

[leofang]

@awvwgk @hhslepicka I feel it's likely cmake's fault, though I can't say for certain. In your feedstocks' previous successful builds, cmake was on 3.18 but now it's 3.19. Could you try to add

- mpich-mpifort  # [osx]

to the build section alongside other compilers, and see if it works? The goal is to make cmake see the Fortran compiler wrapper (mpifort).

[hhslepicka]

it seems odd to be cmake's fault given that the build was successful with the old mpich (3.3.2). Even using CMake 3.19 which is the case for my latest successful build after the mpich pin.

I will try to add the mpich-mpifort as suggested.

[leofang]

hmmm that's a good point...🤔

btw I am reading cmake's doc, and wondering if setting a few env vars like MPI_HOME would help or not: https://cmake.org/cmake/help/latest/module/FindMPI.html

[awvwgk]

I found CMake 3.19 has several issues with IntelMPI recently, so this might be as well CMake related.

Edit: See https://github.com/dftbplus/dftbplus/issues/699

[hhslepicka]

I am trying now with the unpinned mpich and pinned cmake to 3.18. I will report back with what I can find.

[hhslepicka]

Ok... pinning cmake to 3.18 builds properly. Sorry for the false alarm here... I will pursue the issue with the CMake feedstock or cmake directly.

[leofang]

@hhslepicka but pinning at 3.18 still fails: https://github.com/conda-forge/genesis2-feedstock/pull/5/commits/284dcde332328a15164651ccf4157f2031bd499f

[hhslepicka]

My bad.. I looked into the wrong build 💯 x 🤦 the build still fails

[awvwgk]

Can confirm, its not solvable by downgrading CMake for me as well.

[leofang]

Thanks @hhslepicka @awvwgk for testing. I called for attention in CF's gitter channel, let's see if someone can probe the issue. I can also look into it later today, though my knowledge in cmake is fairly limited 😅

[isuruf]

Try updating mpich to 3.4.1. 3.4 introduced a opencl option that trips cmake and in 3.4.1 it was removed.

[leofang]

@isuruf @hhslepicka @awvwgk Looks like 3.4.1 doesn't solve the problem...The error is still the same:

-- Found MPI_C: $PREFIX/lib/libmpi.dylib (found version "3.1") 
-- Could NOT find MPI_Fortran (missing: MPI_Fortran_WORKS) 
-- Could NOT find MPI (missing: MPI_Fortran_FOUND) (found version "3.1")
    Reason given by package: MPI component 'CXX' was requested, but language CXX is not enabled.  

CMake Error at CMakeLists.txt:46 (message):
  MPI not found, specify the MPI Fortran compiler with MPI_Fortran_COMPILER
  variable

https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=271731&view=logs&j=b4588902-138a-5967-ecc7-b3fc381bfda2&t=5a7a20e7-b634-5369-ebb8-6b51f51eb32a&l=1118

[leofang]

btw the Open MPI Fortran builds worked fine:

-- Found MPI_C: $PREFIX/lib/libmpi.dylib (found version "3.1") 
-- Found MPI_Fortran: $PREFIX/lib/libmpi_usempif08.dylib (found version "3.1") 
-- Found MPI: TRUE (found version "3.1")  

https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=271731&view=logs&j=f781b17d-3cc8-5223-204f-29a64ca9ef56&t=4545a821-75b7-516c-9203-a809cfd0e138&l=1114

[isuruf]

Try adding --disable-opencl to configure script of mpich.

[hhslepicka]

@isuruf trying your suggestion here: https://github.com/conda-forge/mpich-feedstock/pull/59

[hhslepicka]

It looks like #59 did the trick for my package. @awvwgk do you mind to check yours as well and reply back so we can close this Issue?

[awvwgk]

59 did it. Build is passing now on OSX. Thanks a lot for your help.

[leofang]

Glad to know it's fixed. Thanks @isuruf @hhslepicka @awvwgk @dalcinl!

[hhslepicka]

Thank you all for the support! Closing this issue now.

[leofang]

@isuruf How did you find out the OpenCL issue? Is there an issue to link to, or should we report to upstream?