Add support for AMD GPUs (consumer/datacenter)

[amacbride]

I'm working on a patch to support AMD consumer and datacenter GPUs (gfx1102/gfx942). I'm building a regression test suite and will be testing on a few CDNA3/4 and RDNA3/4 cards.

If anyone has specific suggestions about what to include, please comment here and I will incorporate them.

[orionarcher]

Thanks for working on this! Per our discussion on slack it sounds like the main change is to the default neighbor list?

I'm building a regression test suite and will be testing on a few CDNA3/4 and RDNA3/4 cards.

Is this something that would be integrated into TorchSim CI? Curious what you have in mind.

[amacbride]

torchsim-rocm-support.patch

[amacbride]

Thanks for working on this! Per our discussion on slack it sounds like the main change is to the default neighbor list?

I'm building a regression test suite and will be testing on a few CDNA3/4 and RDNA3/4 cards.

Is this something that would be integrated into TorchSim CI? Curious what you have in mind.

Yes, that's probably the right place for it. I want to check both correctness and performance.

[amacbride]

I'm going to continue testing on more-exotic cards, but the test suite passed on the AMD Radeon RX 7600. There was a warning about one MACE test, but that was already a float64 model being run as float32, so a 6% difference isn't entirely surprising.

• Test: test_mace_consistency[ti_sim_state] - the titanium (Ti) test case fails (14/15 pass)
• Error: Force comparison off by small margin
  • Absolute difference: 1.065e-05 vs tolerance 1.0e-05 (6.5% over tolerance)
  • Relative difference: 1.2 (120%) vs tolerance 1e-05
  • 2 out of 6 force components mismatched for this one case

[abhijeetgangan]

This should be moved to discussion along with the AMD neighborlist kernel request. I have had other people request this as well.

The problem will always be acceleration of the models as they are most of the time the bottleneck. A intermediate solution is to have triton kernels.