Open DanielWicz opened 5 months ago
The changes required for 8.1 are ported, it would be great to have more test results before merging them. See https://github.com/amd/openmm-hip/issues/11
I can test it on MI250. The only problem is that I can't compile it on the host server, so I need some beta conda package or something like this.
What it ROCm version there?
What it ROCm version there?
5.5.1
Update, I just got access to 5.7.1 and 6.0.0
Great! I can't promise but I'll try to find time this weekend to create the conda package (and I also need to test on 6.0)
Great! I can't promise but I'll try to find time this weekend to create the conda package (and I also need to test on 6.0)
Thanks! I tried to use the old versiom for 8.0 with 6.0.0, but it gives error:
openmm.OpenMMException: Error launching HIP compiler: 256
clang: error: cannot find HIP runtime; provide its path via '--rocm-path', or pass '-nogpuinc' to build without HIP runtime
But I guess the old version is not prepared for the 6.0.0. There was a change of compiler directory in the 6.0.0
Edit: 5.7.1 gives in fact the same error. The ROCM_PATH is set and valid in both cases.
That's really strange, I think 5.7 should work (the error has nothing to do with binary compatibility which I would expect).
Can you check with export OPENMM_USE_HIPRTC=1
? See https://github.com/amd/openmm-hip#the-kernel-compilation-hipcc-and-hiprtc
Perhaps, now it's time to make hipRTC enabled by default (or even the only way to compile), but this requires benchmarking as previously hipRTC generated slower code than the standalone compiler. Likely it has changed in recent version of ROCm.
Please check #14, I've prepared a beta package built on ROCm 6.0.0.
export OPENMM_USE_HIPRTC=1
I checked with "export OPENMM_USE_HIPRTC=1" and it makes it work. There is a performance drop of 1-2% for my system, for the version 5.7.1.
Please check #14, I've prepared a beta package built on ROCm 6.0.0.
Thank! I will check this week and give a feedback
Is the support for OpenMM 8.1 planned ?