Hi, I just wanted to register my interest in NVHPC support. I have had success using the DFT–D4 library in the past with CPU versions of VASP 6.4.x and 6.5.x, and would love to use it in a GPU build of the latest VASP 6.5.1. I've been in conversations with staff at NERSC, who have previously built VASP with DFT–D4 for the Perlmutter machine. Their last successful build of VASP (OpenACC) was VASP 6.4.2 built against DFT–D4 3.4.0 under NVHPC 22.5 compilers. Since then, the NVHPC 22.5 SDK has been removed and the newer NVHPC SDKs (23.x, 24.5) appear unable to build even DFT–D4 alone (versions 3.4.0, 3.6.0, and 3.7.0), let alone linking to it from VASP.

I also use a second HPC cluster with the NVHPC SDK (22.3, 22.9, 23.1, 23.11). I've also had no luck building DFT–D4 there, even the old 3.4.0 version that built on Perlmutter.

Staff at NERSC and I have run into some of the issues mentioned above. But even after resolving the issue with -Mbackslash, there are still cryptic internal compiler errors that result for both of us. For example, attempting to build from the current nvhpc (commit fed77fa) with NVHPC 24.5 gives this error:

My attempt to build `nvhpc` branch with NVHPC 24.5

elvis@perlmutter ~/Downloads/dftd4> cmake --build _build
[215/306] Linking Fortran executable _deps/mctc-lib-build/app/mctc-convert
/usr/bin/ld: warning: /tmp/pgcudafatYYIIhwSQbUfGk.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[240/306] Linking Fortran executable _deps/mctc-lib-build/test/mctc-lib-tester
/usr/bin/ld: warning: /tmp/pgcudafatAMOIhoIrrgMux.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[272/306] Linking Fortran executable _deps/mstore-build/app/info/mstore-info
/usr/bin/ld: warning: /tmp/pgcudafatYrZIhwSTEUfPH.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[273/306] Linking Fortran executable _deps/mstore-build/app/fortranize/mstore-fortranize
/usr/bin/ld: warning: /tmp/pgcudafatYZZIhwSW8UfY7.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[274/306] Linking Fortran executable _deps/multicharge-build/app/multicharge
/usr/bin/ld: warning: /tmp/pgcudafatYUZIhwSWiUfYF.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[281/306] Linking Fortran executable _deps/multicharge-build/test/unit/multicharge-tester
/usr/bin/ld: warning: /tmp/pgcudafatyj3IhiOer04n1.o: missing .note.GNU-stack section implies executable stack
/usr/bin/ld: NOTE: This behaviour is deprecated and will be removed in a future version of the linker
[283/306] Building Fortran object CMakeFiles/dftd4-lib.dir/src/dftd4/disp.f90.o
FAILED: CMakeFiles/dftd4-lib.dir/src/dftd4/disp.f90.o include/dftd4_disp.mod
/opt/cray/pe/craype/2.7.32/bin/ftn -I../src/dftd4 -I../include -Iinclude -I../src/dftd4 -I_deps/mctc-lib-src/include -I_deps/mctc-lib-build/include -I_deps/multicharge-build/include -O2 -gopt -module include -fPIC -mp -c CMakeFiles/dftd4-lib.dir/src/dftd4/disp.f90-pp.f90 -o CMakeFiles/dftd4-lib.dir/src/dftd4/disp.f90.o
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 77)
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 102)
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 114)
  0 inform,   0 warnings,   3 severes, 0 fatal for get_dispersion
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 157)
  0 inform,   0 warnings,   1 severes, 0 fatal for get_properties
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 199)
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 213)
NVFORTRAN-S-0155-Ambiguous interfaces for generic procedure get_lattice_points (../src/dftd4/disp.f90: 221)
  0 inform,   0 warnings,   3 severes, 0 fatal for get_pairwise_dispersion
[285/306] Building Fortran object CMakeFiles/dftd4-lib.dir/src/dftd4/param.f90.o                                                                                                                                                                                                        ninja: build stopped: subcommand failed.

NERSC staff appear to have resolved that error by using the patch in #140:

which I was able to patch in source using the patch in:

#140

This required patching 3 files:
src/dftd4/disp.f90

src/dftd4/cutoff.f90

test/unit/test_model.f90 # <--in addition to those in the link above

but then ran into another internal error:

Finally, I received an Internal Compiler Error:

[291/296] Building Fortran object app/CMakeFiles/dftd4-exe.dir/driver.f90.o
FAILED: app/CMakeFiles/dftd4-exe.dir/driver.f90.o app/dftd4_driver.mod 
/opt/cray/pe/craype/2.7.30/bin/ftn -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/app -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/include -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/_build/include -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/_build/_deps/mctc-lib-src/include -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/_build/_deps/mctc-lib-build/include -I/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/_build/_deps/multicharge-build/include -Mbackslash -O2 -gopt -module app -mp -c app/CMakeFiles/dftd4-exe.dir/driver.f90-pp.f90 -o app/CMakeFiles/dftd4-exe.dir/driver.f90.o
NVFORTRAN-F-0000-Internal compiler error. Deferred-length character symbol must have descriptor    9125  (/global/cfs/cdirs/nstaff/elvis/VASP/DFTD4/dftd4/app/driver.f90: 327)
NVFORTRAN/x86-64 Linux 23.9-0: compilation aborted

Having access to DFT–D4 in the newest OpenACC ports of VASP would be helpful for those of us relying on dispersion corrections. As of last month, at NERSC alone there are around 850 registered VASP users, and VASP accounts for ~15% of total NERSC compute load, so supporting NVHPC would make a huge difference.

I'm not super experienced in debugging Fortran compiler errors, but I'm happy to do some testing on my end if that would be helpful.

DFTD4 support for nvhpc compiler #288

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions