[Mesa-dev] introducing radv - proof of concept vulkan driver for AMD VI chipsets

[Dave Airlie]

I was waiting for an open source driver to appear when I realised I
should really just write one myself, some talking with Bas later, and
we decided to see where we could get.

This is the point at which we were willing to show it to others, it's
not really a vulkan driver yet, so far it's a vulkan triangle demos
driver.

It renders the tri and cube demos from the vulkan loader,
and the triangle demo from Sascha Willems demos
and the Vulkan CTS smoke tests (all 4 of them one of which draws a triangle).

There is a lot of work to do, and it's at the stage where we are
seeing if anyone else wants to join in at the start, before we make
too many serious design decisions or take a path we really don't want
to.

So far it's only been run on Tonga and Fiji chips I think, we are
hoping to support radeon kernel driver for SI/CIK at some point, but I
think we need to get things a bit further on VI chips first.

The code is currently here:
https://github.com/airlied/mesa/tree/semi-interesting

There is a not-interesting branch which contains all the pre-history
which might be useful for someone else bringing up a vulkan driver on
other hardware.

The code is pretty much based on the Intel anv driver, with the winsys
ported from gallium driver,
and most of the state setup from there. Bas wrote the code to connect
NIR<->LLVM IR so we could reuse it in the future for SPIR-V in GL if
required. It also copies AMD addrlib over, (this should be shared).

Also we don't do SPIR-V->LLVM direct. We use NIR as it has the best
chance for inter shader stage optimisations (vertex/fragment combined)
which neither SPIR-V or LLVM handles for us, (nir doesn't do it yet
but it can).

If you want to submit bug reports, they will only be taken seriously
if accompanied by working patches at this stage, and we've no plans to
merge to master yet, but open to discussion on when we could do that
and what would be required.

Dave.