AMD Publishes Micro Engine Scheduler "MES" Firmware Documentation

[agd5f]

Even if it was only used as a validation vehicle on RDNA2, I cannot see a downside of opening it up and letting the community work on it if there was interest in it? I refer to the talk of TinyCorp writing a Fuzzer, fixing some bugs for the MES for RDNA3 and potentially optimize it further for AI workloads which I assume could also trickle down or be extended in scope to RDNA2 if there were no fundamental differences on the firmware/hardware level, right? As I see it, opening up the MES firmware and letting the community work on MES for RDNA2 could be an opportunity to pull in some fixes for RDNA2, too.

There is no production MES firmware for RDNA2,as I said, it was for proof of concept and early validation for RDNA3. If someone wants to play with it, they can patch the code back into the driver, but otherwise, it's just dead, unused code to maintain. I doubt the current MES 10.1 code even works at all at this point.

[ms178]

There is no production MES firmware for RDNA2,as I said, it was for proof of concept and early validation for RDNA3. If someone wants to play with it, they can patch the code back into the driver, but otherwise, it's just dead, unused code to maintain. I doubt the current MES 10.1 code even works at all at this point.

As the decision for opening up MES firmware was already made for RDNA3, I cannot see any fundamental hurdles as to why that couldn't be the case for a pre-production firmware (with the usual strings for legal reasons attached) for RDNA2, too. And with that, the now unused code could be of use for people who like to tinker around with it (e.g. ISVs that see some value in bringing that up to production level). Sure, it could be brought back later or maintained out-of-tree, but I seriously hoped for seeing some firmware fixes and use of that functionality on Linux after the announcement of opening up the RDNA3 firmware.