{phys}[nvofbf/2024.9] VASP v6.4.3 w/ hdf5

[sassy-crick]

(created using eb --new-pr)

Requires:

• PR {data,tools}[GCCcore/13.3.0,NVHPC/24.9-CUDA-12.6.0] Szip v2.1.1, HDF5 v1.14.5 w/ serial #21632
• PR {numlib}[NVHPC/24.9-CUDA-12.6.0,nvompi/2024.9] FFTW v3.3.10 #21634
• PR {lib,mpi}[GCCcore/13.3.0,NVHPC/24.9] NCCL v2.22.3, UCC-CUDA v1.3.0, OpenMPI v5.0.3 w/ CUDA 12.6.0 #21546
• PR {toolchain}[system/system] nvofbf v2024.9, nvompi v2024.9 #21639

Currently the patches have the CUDA version hard-coded in it (see comment in the EC files) and a better way needs to be found here.
Help from @bedroge and @Thyre for helpful discussions is much appreciated!

[tanmoy1989]

Hi, this is just to add-in that while building VASP/6.4.3 using this PR we noticed that it only works for L40 GPUs because of the use of only cc89 parameter against -gpu flag in the makefile.include coming from vasp-6.4.3_nv-flags.patch. To make the build more robust if cc80 is added then that seems to be working well on both A100s and L40s. So I propose the following modification in lines 15-17 in vasp-6.4.3_nv-flags.patch

+CC          = mpicc  -acc -gpu=cc80,cc89,cuda12.6 -mp
+FC          = mpif90 -acc -gpu=cc80,cc89,cuda12.6 -mp
+FCL         = mpif90 -acc -gpu=cc80,cc89,cuda12.6 -mp -c++libs

[sassy-crick]

There is another PR open for VASP-6.5.1: PR #23188
I would suggest to leave that one open until the NVHPC toolchain has settled a bit and then look into it again, and either close it or improve it and merge it.