[AArch64] Neoverse V3/V3AE schedule model is not implemented

[fioraking]

Hello guys,
I noticed that the Neoverse V3/V3AE scheduling model is being used in the Neoverse N2 model in llvm/lib/Target/AArch64/AArch64Processors.td.

Is there a plan to implement it?

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: kindles (fioraking)

Hello guys, I noticed that the Neoverse V3/V3AE scheduling model is being used in the Neoverse N2 model in llvm/lib/Target/AArch64/AArch64Processors.td.

Is there a plan to implement it?

[DavidSpickett]

@davemgreen probably knows about these things.

[davemgreen]

Hi - I don't know of anyone adding one at the moment. It certainly would be good to have one.

I was just giving #131525 a try to see if it could show the differences required between the existing V2 scheduling model and Neoverse V3. It seems to be capable of showing the diffs, if that would still require some manual work at the moment to plug it into a V3 model.

[fioraking]

@DavidSpickett @davemgreen
Okay, if I finish it, I will try to push it to LLVM.

There is currently an confused problem with the scheduling model details. If you're familiar with it, could you please help me?

Here is link: https://discourse.llvm.org/t/confusion-for-arm-neoverse-v2-cpu-scheduling-model/85796

[davemgreen]

Yes - The script is trying too hard to match the throughput numbers from the limited info in the optimization guides. It is probably at this point better at discovering the differences between cpus, that can be used to see how V2 and V3 are different to see how llvm/lib/Target/AArch64/AArch64SchedNeoverseV2.td can be adjusted too bring it up to the V3.

There is also work under way by @lakshayk-nv and @sjoerdmeijer to see if llvm-exegesis can be used to see how well it can do at measuring the latencies/throughput on real devices. It would be nice if eventually it was all auto-generated in a way that could find the quirks of the machine and automatically generate a decent scheduling model for them, but we will have to see how well that goes and currently the process is a little manual to make sure the values are sensible.

[c-rhodes]

FYI @simonwallis2 is currently working on this.