Anton Gorenko

I suspect that only hipcc package is needed because it contains some cmake stuff.

Yes, seems so. Considering that it worked in my case I suggest one change to your commit: keep the `host: - hipcc`. I hope this is enough to make it...

Good idea, I'll benchmark with DisablePmeStream = true. But I can say the in my old profiling results (I profile with DisablePmeStream = true, so kernels are not running concurrently)...

`column -t -s, results.hip_stats.csv` or even `column -t -s, -R 2,3,4 results.hip_stats.csv` can be used for better formatting, but I agree that human-readable output of rocprof would be great.

> The HIP platform works in both of them. Did you remove cached kernels to trigger recompilation? (`rm /tmp/openmm-hip-*`)

@peastman > Do we have any idea of what the solution is? I don't have any (yet). Can you replace `libhiprtc.so.6.0.32831-58582f0` in the current conda environment with a "clean" file...