AMD optimization: How to proceed?

[Mushoz]

On my 7900XTX I found that setting MMVQ_MAX_BATCH_SIZE in to 3 mmvq.cuh leads to a sizable performance increase for batchsizes 4 through 8. I would like to implement this optimization and create a PR for this, but I am not quite sure how to proceed:

1. I only know that the value 3 is ideal on the 7900XTX, but that does not necessarily translate to all AMD GPUs.
2. I cannot simply change the variable, as that would also impact Nvidia devices and the current number is chosen for a reason.

What would be the best way to implement this? Would setting this conditionally based on the device ID be acceptable? Should I apply this to just the 7900XTX, all RDNA3 cards or even all AMD cards? Looking forwards to some pointers on how to proceed. Thanks!

[Mushoz]

Pinging @0cc4m as our discussion led me to investigate this :)