Core26 released to full FAH ?

[muziqaz]

Yeah, I said the same last year around this time. That did not go very well

[arisu]

AMD has HIP out. It is now down to us to implement it. Initially core25 was supposed to be hip. Due to issues with other stuff, core26 was born, which fixed some issues, but had other issues, again HIP integration to backseat. Core27 then was thought to have HIP. But it was pushed without hip to make sure that all the issues with core26 were fixed on 27. It seems we have majority of things fixed now, so I hope core28 or revision of core27 will have hip, as excuses and reasoning is running out now

AMD will listen if we, as in FAH, have something to show and tell them that things can be improved, etc. I cannot tell them anything right now because we have nothing to work on yet.

Is it as simple as rebuilding one of the cores with OpenMM 8.2.0 with HIP enabled and doing tests? Or are there more blockers?

[muziqaz]

Not that simple. If it was, this sh*t would have been built early last year
Since I'm trying not to pressure the Devs, I keep the question to once a month or two.

[Joe_H]

At the least a core has to be built, go through validation testing to show whether or not the HIP implementation gives the same results as OpenCL or CUDA. Then they have to figure out the packaging with the necessary libraries, etc. Then test to see if those install properly over at least a couple versions of Windows and several Linux distros.

[calxalot]

The next client is also needed, because the arguments for the core are different for hip.

[arisu]

The next client is also needed, because the arguments for the core are different for hip.

I think that code is already written for 8.4.10 and can be released at any time.

[calxalot]

The client is not even buildable right now because of changes in cbang that have not been applied to the client.

Nobody has tested the code you are talking about. But yes, it seems to have been committed.

[arisu]

It's buildable, you just have to checkout the 8.4.10 tag. I've tested the code with a dummy core on 8.4.10 and it passes the correct arguments when enabled. Of course the dummy core does nothing but print out the arguments it was given, it can't actually fold.

But at least the client part seems to be working.

[calxalot]

I’m not seeing any 8.4.10 tags in the repo.

[arisu]

Huh, I forget how I did it then. Maybe I just used 8.4.9 for cbang when building the 8.4.10 client? Either way I was able to build 8.4.10 a few months ago, or whenever the hip commit was made (it was shortly after that).