RadeonSI Gets Some Tidying Ahead Of Navi/GFX10 Support (Radeon RX 5700 Series)

[phoronix]

Phoronix: RadeonSI Gets Some Tidying Ahead Of Navi/GFX10 Support (Radeon RX 5700 Series)

Well known open-source AMD developer Marek Olšák sent out a set of eight RadeonSI Gallium3D patches this morning that appear mostly mundane and namely come down to some minor code alterations. This work though is in stepping towards the actual Navi/GFX10 support we expect to be dropped incredibly soon...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-GFX10-Prepping

[shmerl]

Is there enough documentation for radv developers to work on it, or they are still waiting for information?

[bridgman]

Is there enough documentation for radv developers to work on it, or they are still waiting for information?

Still waiting AFAIK

[shmerl]

Still waiting AFAIK

I see, thanks. Then radv can take a while to be usable.

[smitty3268]

I see, thanks. Then radv can take a while to be usable.

We'll see. I suspect it may not be that difficult to get it working once the radeonsi changes get published. It will likely just perform slower than it should be for a while until they figure out all the details.

[smitty3268]

And there it is, Navi support in Mesa: https://gitlab.freedesktop.org/mesa/..._requests/1173

Major deltas from the driver point of view:

• The GS stage is used instead of VS. GS is now a merged ES-GS-VS and is more efficient. In this mode, API shaders VS, TES, and GS are said to be compiled as NGG.
• Transform feedback is completely emulated with the help of GDS ordered atomics. Transform feedback queries have to be emulated too. (it might still be faster than the previous hw)

Slightly bigger than I expected, but I still bet it only takes a few weeks for some level of radv support once they get the hardware.

[marek]

And there it is, Navi support in Mesa: https://gitlab.freedesktop.org/mesa/..._requests/1173



Slightly bigger than I expected, but I still bet it only takes a few weeks for some level of radv support once they get the hardware.

The biggest part is registers definitions and addrlib. The real code that is added is about 5K lines of code.

[smitty3268]

The biggest part is registers definitions and addrlib. The real code that is added is about 5K lines of code.

Yeah, I saw that.

NGG seems like something that could potentially be complicated, but given that radeonsi did it in relatively few lines I'm assuming radv probably will too. I know that the last generation was already moving that direction, though I don't know if radv is structured well enough to easily extend it to NGG or not.

The rest seems like it should be relatively simple, with most of the hard work being done by AMD in the shared code like the kernel, addrlib, llvm, etc.