need a way to disable REAL16 support with configure

[jeffhammond]

I am trying to build e604107 on an AArch64 system with NVHPC Fortran 21.2-0.

It appears that REAL16 support in NVFortran isn't what Open-MPI is looking for, but I do not care about REAL16 support and would prefer to just not have it in the build.

I see this is not a new problem (#63) but cannot figure out how to disable REAL16 now.

$ make V=1 VERBOSE=1
/bin/sh ../../../../../libtool  --tag=FC   --mode=compile nvfortran   -c -o psizeof_f.lo  psizeof_f.f90
libtool: compile:  nvfortran -c psizeof_f.f90  -o .libs/psizeof_f.o
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 641)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_scalar
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 651)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r1
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 661)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r2
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 671)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r3
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 681)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r4
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 691)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r5
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 701)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r6
NVFORTRAN-S-0087-Non-constant expression where constant expression required (psizeof_f.f90: 711)
  0 inform,   0 warnings,   1 severes, 0 fatal for pmpi_sizeof_real16_r7
make: *** [Makefile:2369: psizeof_f.lo] Error 1

$ grep -i real16 psizeof_f.f90
      SUBROUTINE PMPI_Sizeof_real16_scalar(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16)::x
      END SUBROUTINE PMPI_Sizeof_real16_scalar
      SUBROUTINE PMPI_Sizeof_real16_r1(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r1
      SUBROUTINE PMPI_Sizeof_real16_r2(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r2
      SUBROUTINE PMPI_Sizeof_real16_r3(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r3
      SUBROUTINE PMPI_Sizeof_real16_r4(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,1,1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r4
      SUBROUTINE PMPI_Sizeof_real16_r5(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,1,1,1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r5
      SUBROUTINE PMPI_Sizeof_real16_r6(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,1,1,1,1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r6
      SUBROUTINE PMPI_Sizeof_real16_r7(x, size, ierror)
        USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
      REAL(REAL16), DIMENSION(1,1,1,1,1,1,*)::x
      END SUBROUTINE PMPI_Sizeof_real16_r7

Reproduce

You need to use GCC for C/C++ and NVIDIA Fortran to reproduce this.

$ ../configure CC=gcc CXX=g++ FC=nvfortran  --disable-man-pages  && make -j100

commit 1da3f93c7a0b8acff53637a3510cb745e71e1b16 (HEAD -> master, origin/master, origin/HEAD)
Merge: 23833935f3 a2cd6a95e2
Author: Jeff Squyres <jsquyres@cisco.com>
Date:   Tue Mar 16 10:26:37 2021 -0400

    Merge pull request #8626 from mkurnosov/scatter-bmtree-fix-tmpbuf
    
    coll/base: reduce memory consumption in Scatter

[jeffhammond]

I will add that this Fortran compiler does not support REAL*16 (https://forums.developer.nvidia.com/t/real-16/134046) and configure knows this, so I am confused why it's trying to compile such code anyways.

configure:52398: WARNING: *** Fortran REAL*16 does not have expected size!
configure:52400: WARNING: *** Expected 16, got 4
configure:52402: WARNING: *** Disabling MPI support for Fortran REAL*16
configure:53055: checking for C type matching bit representation of REAL*16
configure:53683: result: skipped (no REAL*16)
configure:53693: WARNING: MPI_REAL16 and MPI_COMPLEX32 support have been disabled

[bosilca]

Definitively not new, but your reference dates from 2014 and it related to a different topic (the MPI_Op not being correctly implemented and not the type support itself).

That put aside, the F90 code is autogenerated (gen-mpi-sizeof.pl) and the REAL16 part should not be inside. I tried on few machines without being able to replicate, but I do not have access to an AArch64 with the NVidia compilers. Could this file be a leftover from a prior build ? If not can you remove ompi/mpi/fortran/mpif-h/sizeof_f.f90 in build directory, and do a make V=1. We would need to know how gen-mpi-sizeof.pl is invoked.

[jeffhammond]

This is a brand new build on a system I started using today so it's clean.

NVHPC compilers are free. You can download and install using the shell commands provided on https://developer.nvidia.com/nvidia-hpc-sdk-downloads. You do not need GPUs attached to use them.

[jeffhammond]

Is REAL16 even standard Fortran? The 16-byte REAL is REAL128 not REAL16.

[jeffhammond]

The problem is that real2 is enabled, if that is the thing that translates to REAL16.

make[3]: Entering directory '~/MPI/ompi-github/build-gcc-nvf-arm/ompi/include'
../../../ompi/mpi/fortran/base/gen-mpi-sizeof.pl \
    --header=mpif-sizeof.h --ierror=mandatory \
    --maxrank=7 \
    --generate=1 \
    --real2=1 \
    --real16=0 \
    --complex4=0 \
    --complex32=0
../../../ompi/mpi/fortran/base/gen-mpi-mangling.pl \
    --caps 0 \
    --plain 0 \
    --single 1 \
    --double 0
ln -s ../../../opal/include/opal/opal_portable_platform.h mpi_portable_platform.h

[jeffhammond]

As #6205 appears to be the cause here, maybe @kawashima-fj can take a look. I just need a way to disable real2 support completely in Fortran. If --enable-alt-short-float does that, it is not clear from the description configure --help provides.

[bosilca]

There seems to be a real issue with the generator, as in the command line you provided the real16 is clearly turned off, so the code should not have been generated (independently of the real2 support).

[jeffhammond]

REAL16 here means 16-bit real not 16-byte real. The problem is that Open-MPI is enabling 2-byte floats when the NVF compiler doesn't support them, because they are not standard Fortran.

[bosilca]

You're right, short float enables the float16. What is configure saying about the support of short float ?

[ggouaillardet]

@jeffhammond can you give this a try

configure --enable-mpi-ext=affinity,cuda,pcollreq ...

(the default is --enable-mpi-ext=all feel free not to include some of these extensions is they cause any issue on your platform)

[jeffhammond]

Having extensions that require non-standard Fortran language features enabled by default seems like a bad decision. How many Fortran compilers support 2-byte floats?

[ggouaillardet]

First things first: did this help?

At first glance, it seems the configury logic of this extension fails to disqualify it (or at least the Fortran extension),
so that sounds like a bug that will be fixed.
(fwiw, installing nvidia compilers require root access - i currently do not have, so I am trying to setup a Linux VM on M1 ...)

[jeffhammond]

I install NVHPV compilers without root every day.

1. wget the tar
2. tar -xaf the tar
3. ./install
4. tell it to install in $HOME/someplace

That's all.

I'll test configure later. Had to be AFK before I could try it.

[ggouaillardet]

thanks for the info! nvidia website says to sudo install, so I simply assumed root was needed ...

[ggouaillardet]

it seems there is something wrong with our logic:

we only test Fortran REAL*2 and if it works, assumes that
USE, INTRINSIC :: iso_fortran_env, ONLY: REAL16
will work too.

with nvfortran, the former is supported, but the latter is not, and compilation fails ...