RadeonSI Disables SDMA For Polaris To Fix Corruption Bugs

[agd5f]

SDMA isn't necessarily useful for all cases. It's designed for paging memory and a few other tasks related to GPUVM management (which is what the kernel drivers use it for). For UMDs, it's mainly only useful for moving stuff to/from vram. It's designed to saturate the pcie link between the card and the system. Using gfx or compute has more bandwidth for doing things in vram directly for example (e.g. moving stuff within vram or doing fills, etc.). SDMA does have the advantage of running asynchronously, but it's also not cache coherent with gfx (at least on pre-navi) so sometimes the overhead of synchronization outweighs the advantages.

[Space Heater]

agd5f do you think the kernel and/or mesa teams will have enough free cycles to look at this, or is this something that is relatively low priority given the age of the cards affected?

[agd5f]

Yes, we are looking at it.

[abott]

I had SDMA issues with my R9 270X that crashed me for a year randomly before you or Marek pushed SDMA disabling for that model card which fixed it. Might be worth messing with that hardware if you think it might be an architecture issue. I no longer have that card in use, though.

[pal666]

Disabling and never fixing seems like a trend for AMD. Intel GPUs can't come soon enough...

intel has plenty of gpus for long time, why aren't you enjoying them yet? what about their sdma state btw? do you even know that does it mean?

[nuetzel]

What performance loss could we expect from the 2 sdma disabling?

I don't 'see' any on Polaris 20 (8 GB version) so far.
Even better then ever (with latest optimizations).
Look here: #2

[geearf]

I don't 'see' any on Polaris 20 (8 GB version) so far.
Even better then ever (with latest optimizations).
Look here: #2

Good enough for me, thank you!

[Marc Driftmeyer]

Patchwork appears to be already showing fixes on SDMA.

[PuckPoltergeist]

Disabling and never fixing seems like a trend for AMD. Intel GPUs can't come soon enough...

How long did it take until RC6 worked? At least for one generation of GPUs?

[brent]

Intel GPUs are not a magic land of stability either like some here tend to imply, by the way. I have a total of three issues right now:

* Panel self-refresh is enabled by i915 by default, but it does not work well (flickering). I have to disable it manually with a kernel flag. The downside is significantly increased power usage.
* Recent kernels suffer from broken RC6 sleep state, resulting in extremely (!) increased power usage
* Kernel 5.4 also suffers from random GPU hangs on many systems, mine included

I don't know why, but Intel developers do not manage to backport workarounds/fixes to stable kernels. This is really frustrating.