[Mesa-dev] [PATCH 2/2] [RFC] radv: add scratch support for spilling.

[Dave Airlie]

On 10 October 2016 at 13:25, Dave Airlie <airlied at gmail.com> wrote:
> From: Dave Airlie <airlied at redhat.com>
>
> This is a bit of a hack due to how llvm currently handles
> spilling in it's shader ABI. Currently llvm amdgpu backend
> uses relocations to patch the shader with the address of
> the tmpring. The driver loads the shader and patches the
> relocations.
>
> However for vulkan this doesn't work so well for a few reasons
> a) when we build/load the shaders we aren't constructing the
> command stream yet, and the same shader could be using in multiple
> command streams.
>
> b) multiple command execution engines for compute shaders.
>
> So ideally we'd fix LLVM to understand the ABI convention, possibly
> we'd fix it so user sgpr 0,1 are used (this hack uses 10/11).
>
> This patch when it gets the shader back from llvm it patches
> the relocation dword to a nop, and patches to previous mov command
> to move from SGPR 10 and 11. This works usually as it seems the
> SGPR loading of the spill stuff is at the start of shaders always
> so the 10/11 user sgprs haven't been trashed yet. I'm not 100%
> sure this will work all the time, but for now this should allow
> us to pass a bunch more CTS tests and make the Sascha computeshader
> demo to work.

So I found a shader that this doesn't work so well with unfortunately, so
while I'd like this as a temporary solution I probably need to start
digging into llvm.

My current plan is to add a flag to llvm to denote ability to spill to
the address
in userdata sgpr 0/1, have llvm preload those and use them instead.

Now the other question I have is, should I be killing two user data sgprs for
this purpose, or should we define a better ABI, so that the first descriptor
in the buffer pointed to by these is the scratch buffer, and other things
could be queued after it, (like push constants and dynamic descriptor).

Bas, nha? not sure if Matt is on this list.

Dave.