ABINIT + abinit-fallbacks on Lucia (six toolchains)

ABINIT + abinit-fallbacks on Lucia (six toolchains)

Status, 2026-10-08. ABINIT builds on all six toolchains of Lucia, against external libraries built from source with abinit-fallbacks. That includes BigDFT, AOCC, and every build check. With the changes of section 5, the four Cray toolchains need no configure argument besides the library paths. On Cray, ABINIT now also uses FFTW3 from the cray-fftw module, still without any extra argument (section 5.5). The test suite was not part of this work.

Versions:

• ABINIT 10.9.3 (development):
  • branch fix_hdf5, commit 11cc7b0a57 (sections 4.1–4.3);
  • then branch cray_autodetect on top of it, commits 0d793ca5da, 41f1b841a2 and 892bc37de3 (section 5; not published yet).
• abinit-fallbacks, fork tsaihsiaoyi/abinit-fallbacks:
  • branch my_develop, commit 3ddec17;
  • then branch bigdft_xc_fix on top of it, commits 46c9322, f36663b and 6266085 (section 3.3; not published yet).
• Cluster: Lucia (Cenaero), debug nodes with 2 × AMD EPYC 7763 (Zen 3, AVX2, no AVX-512), RHEL 8, Slurm 25.11.

1. Goal and rules

For each of the six compiler toolchains available on Lucia:

1. build all the ABINIT external libraries from source with abinit-fallbacks;
2. build ABINIT against exactly those libraries, with the smallest possible set of configure arguments.

The rules were the same for both steps:

• Only the compiler, MPI and math library come from cluster modules. HDF5, netCDF, LibXC, ELPA, Wannier90, BigDFT, … all come from the fallbacks, never from cray-hdf5, cray-netcdf or EasyBuild modules. The one later addition is cray-fftw, which ABINIT uses on the Cray toolchains. None of the fallbacks uses FFTW.
• Everything runs in a clean environment: env -i, then source /etc/profile (only for the module command), then the module lines of the toolchain. My ~/.bashrc activates a conda environment, and two hidden dependencies came from it before I switched to env -i:
  • netCDF-C picked up conda’s xml2-config;
  • BigDFT needed conda’s python.
• Every configure argument beyond the library paths must be justified by a failing attempt without it. Each attempt’s config.log is kept.
• Heavy work runs as Slurm jobs on the debug partition.

2. The six toolchains

Name	Modules	Compilers	MPI	Math library
cray_cray	Cray/24.07 PrgEnv-cray/8.4.0	CCE 18.0.0	Cray MPICH 8.1.30	Cray LibSci 24.07, cray-fftw 3.3.10.8
cray_gnu	Cray/24.07 PrgEnv-gnu/8.4.0	GCC 13.3	Cray MPICH 8.1.30	Cray LibSci 24.07, cray-fftw 3.3.10.8
cray_intel	Cray/24.07 PrgEnv-intel/8.4.0	icx 2022.2 + ifort 2021.7	Cray MPICH 8.1.30	Cray LibSci 24.07, cray-fftw 3.3.10.8
cray_aocc	Cray/24.07 PrgEnv-aocc/8.4.0	AOCC 4.1 (clang 16 + classic flang)	Cray MPICH 8.1.30	Cray LibSci 24.07, cray-fftw 3.3.10.8
eb_intel	EasyBuild/2025a intel/2025a	icx/ifx 2025.1	Intel MPI 2021.15	MKL 2025.1
eb_foss	EasyBuild/2025a foss/2025a	GCC 14.2	OpenMPI 5.0.7	FlexiBLAS 3.4.5 / OpenBLAS 0.3.29, ScaLAPACK 2.2.2

Each build directory has an env.sh with exactly these lines, for example:

module --force purge          # also unloads the sticky Cray/ and EasyBuild/ modules
module load Cray/24.07
module load PrgEnv-gnu/8.4.0
module load cray-fftw/3.3.10.8   # ABINIT build directories of the Cray toolchains only

cray-fftw/3.3.10.8 is the only version installed. Because craype-x86-milan is loaded, the module selects the Milan build of FFTW.

3. Step 1: abinit-fallbacks

3.1 Packages

All packages are installed as static libraries, one directory per toolchain ($FB below):

Package	Version
HDF5 (parallel)	1.14.6
netCDF-C (parallel)	4.9.3
netCDF-Fortran	4.6.2
LibXC	7.0.0
ELPA	2025.01.001
Wannier90	2.0.1.1
XMLF90	1.6.3
libPSML	2.1.0
BigDFT	abinit-1.7.1.33
AtomPAW	4.2.0.3 (program only)

3.2 Changes to the abinit-fallbacks build system

Upstream abinit-fallbacks did not work out of the box on these toolchains. The changes are in commits a46c150…3ddec17 of my fork:

• Compilers:
  • In a Cray PE, use the craype wrappers cc/CC/ftn.
  • Otherwise prefer the MPI wrappers.
• Linear algebra:
  • New libsci flavor: the craype wrappers link Cray LibSci implicitly, so no library has to be added.
  • BLACS is detected inside libscalapack (ScaLAPACK ≥ 2.0 has no separate libblacs).
• Cray Fortran:
  • -ef to get lower-case module file names (netcdf.mod, not NETCDF.mod).
  • A one-line patch to ELPA (a non-standard continuation line).
• ELPA:
  • Build only the library, not its hundreds of test programs.
  • Disabled without MPI.
  • SIMD kernels SSE/AVX/AVX2.
• netCDF-C: --disable-libxml2 (bundled tinyxml2) and --disable-dap.
• BigDFT:
  • PYTHON=/usr/bin/python3 (there is no python in a clean RHEL 8 environment).
  • Link the full static netCDF/HDF5 dependency list (nc-config --static --libs).
  • -h ipa1 for CCE 18 (internal compiler error, “Broken module found”).
  • -O1 for AOCC 4.1 (flang crashes on src/forces.f90).
• GCC ≥ 10: -fallow-argument-mismatch for BigDFT and Wannier90.
• XMLF90: module search paths needed by flang-based compilers.
• LibXC: optimisation tweaks for the Intel and Cray compilers.
• abinit-fallbacks-config: fixed the BigDFT path, and dropped a non-existent libatompaw.

3.3 BigDFT: three more fixes, needed to link it into ABINIT

These are in branch bigdft_xc_fix. Each one surfaced only once ABINIT was configured with --with-bigdft:

1. Private routines used from outside (patch bigdft-abinit-1.7.1.33-0001.patch, commit 46c9322). BigDFT’s src/modules/xc.f90 calls abi_libxc_functionals_getrefs and abi_libxc_functionals_constants_load, but the bundled libABINIT declares them private. The compilers then treat the calls as calls to external procedures that don’t exist, so every program linking libbigdft-1.a fails:

   undefined reference to `abi_libxc_functionals_constants_load_'
   

   The patch makes the two routines public.

2. Missing module files (patch -0002, commit f36663b). BigDFT installs its own .mod files but not those of the bundled libABINIT (abi_libxc_functionals.mod, …). Classic flang (AOCC) needs every module used indirectly, so the patch adds an install rule for them.

3. Fake ScaLAPACK (commit 6266085). BigDFT’s configure only tries -lscalapack. Without it, BigDFT compiles src/modules/blacs_fake.f90, stubs such as blacs_gridinit or pdgemm that just stop 'FAKE …'. In a program linking libbigdft-1.a, these stubs replace the real routines whenever the real library comes later on the link line. That is the case with the Cray wrappers, which add LibSci at the very end. Every ScaLAPACK run of ABINIT then stopped in FAKE BLABS_GRIDINIT. The fallbacks now pass --with-scalapack=-lsci_<PrgEnv>_mpi (LibSci) or -lmkl_scalapack_lp64 (MKL) to BigDFT.

3.4 Build

build.sh in each directory does, from any shell:

env -i HOME=$HOME USER=$USER LOGNAME=$LOGNAME TERM=$TERM PATH=/usr/bin:/bin \
    bash --noprofile --norc
source /etc/profile
source env.sh
../configure --prefix=$PWD      # compilers and math library are detected automatically
make
make install
./check.sh                      # writes build-summary.txt

check.sh verifies that:

• every package is installed;
• bin/abinit-fallbacks-config returns existing directories;
• no installed program depends on a shared HDF5, netCDF or LibXC library from the cluster.

3.5 Results

All six builds pass (RESULT: OK). They were run as Slurm jobs with 16 cores each, from commit 6266085:

Toolchain	Compilers detected	Math library detected	Time
cray_cray	cc / CC / ftn (CCE 18.0)	libsci	25 min
cray_gnu	cc / CC / ftn (GCC 13.3)	libsci	8 min
cray_intel	cc / CC / ftn (ifort 2021.7)	libsci	14 min
cray_aocc	cc / CC / ftn (AOCC 4.1)	libsci	16 min
eb_intel	mpiicx / mpiicpx / mpiifx (2025.1)	mkl (-qmkl=cluster)	13 min
eb_foss	mpicc / mpicxx / mpif90 (GCC 14.2)	OpenBLAS (pkg-config) + -lscalapack	12 min

The tuned options target the build node’s CPU: ELPA’s SIMD kernels, -march=native, and craype-x86-milan on Cray. On a different CPU, rebuild; for ELPA, --disable-elpa-simd gives generic kernels.

4. Step 2: ABINIT fix_hdf5 as it is

4.1 Baseline configure command

The same for every toolchain. It contains only the libraries to use, and no --prefix (ABINIT is built in place):

../configure \
  --with-hdf5=$FB/hdf5/default \
  --with-netcdf=$FB/netcdf4/default \
  --with-netcdf-fortran=$FB/netcdf4_fortran/default \
  --with-libxc=$FB/libxc/default \
  --with-elpa=$FB/elpa/default \
  --with-wannier90=$FB/wannier90/default \
  --with-xmlf90=$FB/xmlf90/default \
  --with-libpsml=$FB/libpsml/default

ABINIT’s configure takes only 20–60 s here, so each configure attempt was a short Slurm job, and its logs were archived before the next attempt.

4.2 Minimal extra arguments needed by fix_hdf5

Toolchain	Extra arguments	Configure attempts
eb_intel	none	1
eb_foss	--with-linalg-flavor=easybuild+elpa	2
cray_gnu	CC=cc CXX=CC FC=ftn LINALG_LIBS="-lsci_gnu_mpi -lsci_gnu"	3 (+1 experiment)
cray_cray	CC=cc CXX=CC FC=ftn LINALG_LIBS="-lsci_cray_mpi -lsci_cray"	3
cray_intel	CC=cc CXX=CC FC=ftn LINALG_LIBS="-lsci_intel_mpi -lsci_intel"	3
cray_aocc	as cray_*, plus FCFLAGS="-g -Mextend -Qunused-arguments -I$FB/elpa/default/include/elpa-2025.01.001/modules", plus 3 source fixes	5

Why each argument was needed:

CC=cc CXX=CC FC=ftn (Cray). Without them, ABINIT searches PATH for mpiicx mpiicc mpicc, mpiicpx mpiicpc mpic++ mpicxx and mpiifx mpiifort mpifort mpif90 mpif95 (config/m4/sd_arch_mpi.m4). In every PrgEnv it therefore picks cray-mpich’s own mpicc/mpic++/mpifort, not the craype wrappers. Those wrappers do not link LibSci and do not add the craype-x86-milan target options, with these consequences:

• cray_gnu: links the system /usr/lib64/libopenblas.so (RHEL’s OpenBLAS 0.3.15), and LAPACK is still not found;
• cray_cray: only the netlib flavor is tried, -lblas does not exist, and there is no linear algebra at all;
• cray_intel: MKL is selected (-qmkl=cluster) instead of LibSci: for Intel compilers ABINIT always tries MKL first, and PrgEnv-intel ships one;
• cray_aocc: configure stops at the Fortran checks (see below).

LINALG_LIBS="-lsci_<x>_mpi -lsci_<x>" (Cray). ABINIT had no LibSci flavor: an unknown name fails with no library settings for linear algebra flavor. Even with the wrappers, its auto-detection still picks system OpenBLAS (gnu), finds nothing (cray, aocc) or uses MKL (intel). Setting any LINALG_* variable switches ABINIT to “verify only” mode, and --with-elpa keeps working in that mode.

LINALG_LIBS alone is not enough. I checked this on cray_gnu: without the wrappers, cray-mpich’s mpifort cannot find LibSci at all:

ld: cannot find -lsci_gnu_mpi: No such file or directory

--with-linalg-flavor=easybuild+elpa (eb_foss). The automatic detection finds EasyBuild’s OpenBLAS (through pkg-config) but no ScaLAPACK. Its MPI flavor netlib hard-codes -lblacs -lblacsCinit -lblacsF77init, which ScaLAPACK 2.2.2 does not have. The easybuild flavor uses -lopenblas -lscalapack. +elpa is needed because an explicit flavor no longer adds the ELPA check by itself (sd_math_linalg.m4).

FCFLAGS=... (AOCC). There are two problems:

• AOCC’s ftn --version prints AMD clang version 16.0.3 (CLANG: AOCC_4.1.0...). That matches ABINIT’s LLVM-flang test, which runs before its AOCC test, so ABINIT applies the new-flang hints -ffixed-line-length=132 -Qunused-arguments. Classic flang rejects the first one (clang-16: error: unknown argument: '-ffixed-line-length=132'), and configure stops with “cannot compile a simple Fortran program”. -Mextend is the classic-flang equivalent.
• Classic flang needs module files that are only used indirectly (elpa_api.mod), but ABINIT passes the ELPA module directory only to the directories that use linear algebra directly. Hence the extra -I.

The intended override, FCFLAGS_HINTS, does not work. configure.ac applies the user’s value (ABI_ENV_RECALL, line 136) before computing the vendor hints (ABI_FC_HINTS, line 403), which overwrite it. So the full FCFLAGS had to be given.

4.3 AOCC 4.1: three source-level problems

Even with the right flags, make stopped in ABINIT’s own sources:

File	Problem
shared/libpaw/src/m_paw_atom_solve.F90:7507	type(logical) (Fortran 2008 TYPE(intrinsic-type)) is not supported by classic flang: F90-S-0155-Derived type has not been declared - logical
src/62_ctqmc/m_CtqmcInterface.F90:148	Bug in flang’s integrated preprocessor: the nested macro MALLOC → ABI_MALLOC comes out as C(this%Hybrid_chains,(this%num_chains))
src/48_diago/m_chebfi2.F90:1386,1663,2029	Same preprocessor bug inside ABI_MALLOC_IFNOT

The preprocessor bug depends on the data: the same macros expand correctly in hundreds of other places, and in small test files. The fixes are equivalent rewrites (commit 0d793ca5da):

--- a/shared/libpaw/src/m_paw_atom_solve.F90
+++ b/shared/libpaw/src/m_paw_atom_solve.F90
@@ -7504,7 +7504,7 @@
 SUBROUTINE setcoretail(Grid,coreden,PAW,needvtau)
- type(logical),intent(in) :: needvtau
+ logical,intent(in) :: needvtau
  TYPE(GridInfo), INTENT(IN) :: Grid
--- a/src/62_ctqmc/m_CtqmcInterface.F90
+++ b/src/62_ctqmc/m_CtqmcInterface.F90
@@ -145,7 +145,7 @@
   this%num_chains = num_chains
-  MALLOC(this%Hybrid_chains, (this%num_chains))
+  ABI_MALLOC(this%Hybrid_chains, (this%num_chains))
   this%Hybrid => this%Hybrid_chains(1)
--- a/src/48_diago/m_chebfi2.F90
+++ b/src/48_diago/m_chebfi2.F90
@@ -1383,7 +1383,7 @@
-    ABI_MALLOC_IFNOT(nrowsLinalg,(num_proc))
+    if (.not. allocated(nrowsLinalg)) then; ABI_MALLOC(nrowsLinalg,(num_proc)); endif
     nrowsLinalg_ptr => nrowsLinalg
(same change at lines 1660 and 2026)

5. Branch cray_autodetect: no extra arguments on Cray, with BigDFT

5.1 Changes to ABINIT’s build system

Commit 41f1b841a2 makes configure do what the arguments of section 4.2 did:

• Cray PE compilers (config/m4/arch-mpi.m4). When CRAYPE_VERSION and PE_ENV are set and cc/CC/ftn exist, configure uses the craype wrappers unless CC/CXX/FC are given:

  checking for a Cray Programming Environment... yes (PrgEnv: GNU)
  configure: using the Cray compiler wrappers: CC=cc CXX=CC FC=ftn
  

• libsci linear-algebra flavor (sd_math_linalg*.m4), tried first on a Cray PE.
  • With the wrappers, no library has to be named: BLAS, LAPACK, BLACS and ScaLAPACK are all “none required”.
  • With other compilers (e.g. FC=mpifort), the libsci_<PrgEnv> and libsci_<PrgEnv>_mpi libraries are linked explicitly, but only when they match the compiler in use.
• AOCC (lang-fortran.m4, configure.ac). AOCC flang is still treated as LLVM, because it needs the LLVM settings. But:
  • -ffixed-line-length=132 becomes -Mextend in its Fortran hints;
  • the linear-algebra Fortran flags (the ELPA modules) are added to FCFLAGS for every directory.
• BigDFT link order (sd_bigdft.m4). libyaml.a is now passed as -Wl,<path>/libyaml.a. The Cray ftn driver moves plain file arguments in front of all -l options. That put libyaml.a before libbigdft-1.a, so with CCE every BigDFT link failed with undefined reference to yaml_parser_initialize.

Commit 892bc37de3 fixes a real bug that only CCE detected. m_mklocl_realspace.F90 called BigDFT’s PSolver with the integer 0 where the interface expects a type(xc_info) (ftn-1279). The other compilers accepted it, so PSolver read an integer as an xc_info object. It now passes an object initialised with xc_init(xc, 0, XC_ABINIT, 1).

5.2 Configure arguments now

All six toolchains use the library paths of section 4.1, plus --with-bigdft=$FB/bigdft/default. Beyond that:

Toolchain	Extra arguments
cray_cray, cray_gnu, cray_intel, cray_aocc	none (FFTW3 included, see 5.5)
eb_intel	none
eb_foss	--with-linalg-flavor=easybuild+elpa (as before)

5.3 What “OK” means

Every build ends with a check.sh that verifies:

1. No .ac9 configuration file was read: not loading options (no config file available) in config.log.
2. The compilers are the expected ones: cc/CC/ftn on Cray, the Intel or OpenMPI wrappers on EasyBuild.
3. MPI works for C, C++ and Fortran, and MPI-IO is enabled.
4. The math library is the cluster’s:
   • linear algebra, ScaLAPACK and ELPA are all yes;
   • ldd abinit shows libsci_<x>(_mpi) on Cray, MKL on eb_intel, and EasyBuild OpenBLAS/FlexiBLAS + ScaLAPACK on eb_foss;
   • never /usr/lib64/libopenblas*.
5. Every -I/-L of HDF5, netCDF, LibXC, ELPA, Wannier90, XMLF90, libPSML and BigDFT points into $FB, and none of them (nor libyaml) appears as a shared library in ldd abinit.
6. abinit -b lists HAVE_MPI HAVE_MPI_IO HAVE_HDF5_MPI HAVE_NETCDF_MPI HAVE_NETCDF_FORTRAN_MPI HAVE_LIBXC HAVE_LINALG_SCALAPACK HAVE_LINALG_ELPA HAVE_WANNIER90 HAVE_LIBPSML HAVE_XMLF90 HAVE_BIGDFT.
7. make exits with 0.
8. On Cray only, FFTW comes from the cray-fftw module:
   • the FFT flavor is fftw3;
   • the FFTW3 flags point to $FFTW_ROOT;
   • abinit -b lists HAVE_FFTW3;
   • ldd abinit resolves libfftw3* to a cray-fftw file.

5.4 Results

All six builds pass every check (RESULT: OK). They used make -j 64 on one debug node:

Toolchain	Linear algebra (configure / ldd)	FFT	make
cray_cray	elpa+libsci / libsci_cray(_mpi).so.6	FFTW3 (cray-fftw 3.3.10.8, + FFTW3-MPI)	32.5 min
cray_gnu	elpa+libsci / libsci_gnu(_mpi).so.6	FFTW3 (cray-fftw 3.3.10.8, + FFTW3-MPI)	8.3 min
cray_intel	elpa+libsci / libsci_intel(_mpi).so.6 (no MKL)	FFTW3 (cray-fftw 3.3.10.8, + FFTW3-MPI)	20.2 min
cray_aocc	elpa+libsci / libsci_aocc(_mpi).so.6	FFTW3 (cray-fftw 3.3.10.8, + FFTW3-MPI)	26.2 min
eb_intel	elpa+mkl / MKL 2025.1 (libmkl_scalapack_lp64, libmkl_blacs_intelmpi_lp64, …)	DFTI	9.5 min
eb_foss	easybuild+elpa / OpenBLAS 0.3.29, ScaLAPACK 2.2.2, FlexiBLAS 3.4.5	FFTW3 (EasyBuild FFTW 3.3.10, found with pkg-config)	9.1 min

All times are clean builds. The four Cray builds are clean builds of the final source, with cray-fftw. For the two EasyBuild toolchains, the PSolver fix (one file) came after their clean build, so they were brought up to date with an incremental make, after which check.sh passes again.

5.5 FFTW on Cray: cray-fftw

Before cray-fftw was added to the Cray env.sh, configure already tried FFTW3 first, with -lfftw3_mpi -lfftw3. The test failed (cannot find -lfftw3_mpi), so ABINIT fell back to its internal Goedecker FFT. Loading the module was enough to change that, with no change to ABINIT and no configure argument:

• cray-fftw adds its fftw3.pc to PKG_CONFIG_PATH, and ABINIT’s FFTW3 macro (sd_fftw3.m4) uses pkg-config when it is available:

  checking for fftw3 via pkg-config... yes
  checking whether the FFTW3 library works... yes
  checking whether the FFTW3 library supports threads... yes
  checking whether the FFTW3 MPI library works... yes
  checking for the actual FFT flavor to use... fftw3
  

• The craype wrappers add -I$FFTW_INC. They also link the module’s libraries, fftw3_mpi and fftw3_threads included. So the fftw3-mpi.f03 test passes and HAVE_FFTW3_MPI is defined, although the pkg-config flags give only -lfftw3 -lfftw3f.
• This works the same with CCE, GCC, Intel and AOCC. ABINIT includes the FFTW Fortran 2003 interface files (fftw3.f03, fftw3-mpi.f03), which use only iso_c_binding. They are therefore independent of the Fortran compiler, and no compiler-specific FFTW build is needed.

One detail matters at run time. The binary is linked against the Milan build (x86_milan) because the module is loaded. Without LD_LIBRARY_PATH, however, the loader takes libfftw3*.so.mpi31.3 from /opt/cray/pe/lib64, which is listed in ld.so.conf. On Lucia those files point to the x86_rome build of the same version. To run with the Milan build, use Cray’s usual setting:

export LD_LIBRARY_PATH=$CRAY_LD_LIBRARY_PATH:$LD_LIBRARY_PATH

check.sh prints both resolutions.

6. Pitfalls worth knowing

• Automake rebuild rules. ABINIT’s Makefiles have no “maintainer mode”. If a file of config/m4 is newer than configure, every make reruns aclocal/autoconf/automake in the shared source tree, and concurrent builds collide (autom4te: cannot rename …).
  • aclocal does not rewrite aclocal.m4 when only the content of an included m4 file changes. So after ./autogen.sh, touch aclocal.m4, config.h.in, configure and the Makefile.in files, in that order.
• MPI launch inside a Slurm job (Slurm 25.11, no default MPI plugin). Each library needs its own plugin, or the ranks start as separate singletons:
  • Cray MPICH: srun --mpi=cray_shasta;
  • Intel MPI: srun --mpi=pmi2 with I_MPI_PMI_LIBRARY=/usr/lib64/libpmi2.so;
  • OpenMPI 5: srun --mpi=pmix.

7. Reproducing one build

For example cray_gnu, from a build directory inside the ABINIT source tree on branch cray_autodetect:

env -i HOME=$HOME USER=$USER LOGNAME=$LOGNAME TERM=$TERM PATH=/usr/bin:/bin \
    bash --noprofile --norc
source /etc/profile
module --force purge
module load Cray/24.07 PrgEnv-gnu/8.4.0 cray-fftw/3.3.10.8
FB=~/program/abinit-fallbacks/_build_cray_gnu      # the fallbacks built in step 1

../configure \
  --with-hdf5=$FB/hdf5/default --with-netcdf=$FB/netcdf4/default \
  --with-netcdf-fortran=$FB/netcdf4_fortran/default --with-libxc=$FB/libxc/default \
  --with-elpa=$FB/elpa/default --with-wannier90=$FB/wannier90/default \
  --with-xmlf90=$FB/xmlf90/default --with-libpsml=$FB/libpsml/default \
  --with-bigdft=$FB/bigdft/default
make -j 64
src/98_main/abinit -b      # check the CPP options

8. Open points

• eb_foss still needs --with-linalg-flavor=easybuild+elpa. ABINIT’s own detection could look for BLACS inside libscalapack, as the fallbacks do.
• FCFLAGS_HINTS still cannot override the vendor hints (the order problem of section 4.2).
• Upstream reports.
  • AOCC flang’s preprocessor bug.
  • The three BigDFT 1.7.1.33 problems of section 3.3.
• FFTW variant at run time on Cray. By default the loader picks the x86_rome build of cray-fftw (section 5.5). It would be cleaner to embed the module’s library path in the binary, for example by linking with CRAY_ADD_RPATH=yes. That has not been tried yet.
• FFTW3 threads. OpenMP is off in these builds, so ABINIT uses fftw3, not fftw3-threads, although cray-fftw provides the threaded libraries.