Kokkos multispecies acceleration with type-sorting and matrix-matrix products #310
Conversation
There was a problem hiding this comment.
Looks great! I left two comments (which might take a bit time)
And here are another two minor ones:
- you can update the version number to 1.2.0 in
_version.py(the current development branch is 1.1.2) - you can add a trillion simulation image to our gallery: https://github.com/mir-group/flare/blob/gemm/docs/source/related.rst
| @@ -39,4 +39,21 @@ target_sources(lammps PRIVATE | |||
| ${LAMMPS_SOURCE_DIR}/radial.cpp | |||
| ${LAMMPS_SOURCE_DIR}/y_grad.cpp | |||
| ) | |||
There was a problem hiding this comment.
In line 17 (somehow I can not comment there...), we skipped the pair mgp. I remember it is because some syntax is not updated with the latest lammps. Could you check and fix that?
And correspondingly the unit tests in test_lammps.py (line 38 and 146) skip the mgp-kokkos tests, which after the fix we should resume.
There was a problem hiding this comment.
I think mgp/kk has had issues longer, I can look at that separately.
There was a problem hiding this comment.
ok, that also works. I thought it would be just simple syntax fixing. But we can do it separately
| Kokkos::parallel_for("FLARE: B2 norm2 evdwl w", | ||
| Kokkos::TeamPolicy<DeviceType, TagNorm2>(batch_size, TEAM_SIZE, vector_length), | ||
| *this | ||
| ); |
There was a problem hiding this comment.
Do we need to add team::barrier to the code? It is worth testing the large scale long time behavior, to ensure this update will not produce light-speed hydrogens.
There was a problem hiding this comment.
I hope not, without justification other than running this for 8 hours on 8xV100 and not seeing any issues
There was a problem hiding this comment.
for now that is sufficient
|
I updated the version and added a reference (hoping the RST will work) |
you can merge after GitHub actions pass |
This PR accelerates the LAMMPS plugin on GPUs by sorting the atoms by their type, allowing for the atom-wise matrix-vector products to be replaced with per-type matrix-matrix products, for which we now use KokkosKernels.
KokkosKernels will automatically be downloaded using FetchContent, the user should hopefully not notice any change to the installation procedure.
The neighborlist requests are now compatible with the newest version of LAMMPS, which requires a small ASE patch (added to the GitHub Action) and to use
halfneighbor lists when running the Tersoff potential with Kokkos (changed in unit test).Some whitespace/formatting happened.