test cp2k build

[simonpintarelli]

No description provided.

[simonpintarelli]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2051.1

[simonpintarelli]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2051.1

[mtaillefumier]

I wonder if we should only consider 2025.2 and leave 2025.1 alone as I doubt any normal user will be on beverin. I did all my performances tests with cp2k master on beverin as well.

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2051.1

The recipe itself is 2025.2 btw. Let's see if this goes through, I have not been able to go through a whole build so far due to insane hipblaslt building times.

[mtaillefumier]

we can safely add

-spglib

• libxc
• libint
• smm=libxsmm

[mtaillefumier]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2051.1

The recipe itself is 2025.2 btw. Let's see if this goes through, I have not been able to go through a whole build so far due to insane hipblaslt building times.

I was confused by the directory name that's why I was asking.

[mtaillefumier]

we have an error
[error] Environment: the recipe cp2k/2051.1 is not available for the mi300 target

[RMeli]

I wonder if we should only consider 2025.2

I think this is fair. We are also in the process of releasing 2025.2 on Alps (#259).

[mtaillefumier]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2051.1

[mtaillefumier]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

we can safely add

-spglib

* libxc

* libint

* smm=libxsmm

All of these are enables by default, except spglib.

[mtaillefumier]

we can safely add
-spglib

* libxc

* libint

* smm=libxsmm

All of these are enables by default, except spglib.

Good point. i keep forgetting what is enabled by default.

[mtaillefumier]

that's the spec I used for my performance tests

cp2k@master+cosma~cuda+dbm_gpu~deepmd+dftd4~dlaf~elpa~enable_regtests~greenx+grid_gpu+grpp+hdf5~ipo+libint~libvori+libxc+mpi~mpi_f08~nlcg~opencl+openmp~pexsi~plumed+pw_gpu~pytorch~quip+rocm+sirius~smeagol+spglib+spla+tblite+trexio~vcsqnm+vdwxc amdgpu_target:=gfx942 build_system=cmake build_type=Release commit=06bd950d54ef45179bf552cbcd37c24b2a2e34e3 generator=ninja lmax=5 patches:=178e62e smm=libxsmm arch=linux-sles15-zen3 %c,cxx,[email protected]

[mtaillefumier]

I am not allowed to trigger the build.

[mtaillefumier]

we may need custom recipe for cp2k and dbcsr because AMD introduced some patch in the package.yaml files that do not always work.

[simonpintarelli]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

spla is not found by CMake. And despite [email protected] +spla, I cannot find it in the concretization. I added the package explicitly.

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

It's still not clear why SPLA is not found by CMake. From CP2K's package.py, we have:

with when("+spla"):
        depends_on("spla+cuda+fortran", when="+cuda")
        depends_on("spla+rocm+fortran", when="+rocm")

And the concretization shows spla~fortran. I don't understand why it goes through nonetheless.

[abussy]

I am at a loss here. I have no idea what causes the spla issue, and therefore how to fix it.

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[abussy]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.1

[simonpintarelli]

cscs-ci run alps;system=beverin;uarch=mi300;uenv=cp2k:2025.2

[abussy]

I did some testing: everything looks fine to me. Numerical results are correct on the usual benchmarks, and compared to 2025.2 on GH200. Timing wise, I am close to what @mtaillefumier achieved with his manual build.

Notes: It is important to export HSA_ENABLE_SDMA=0 in order to run with more than 1 rank per GPU. Finally, DBCSR does not always behave well on GPUs, and I get better performance with export DBCSR_RUN_ON_GPU=0 in some cases.

[mtaillefumier]

that's good news.