[libre-riscv-dev] porting AMDVLK to the Libre RISC-V 3D GPU: NLNet EUR 50, 000 Grant application

[Michael Pham]

On Wed, Sep 25, 2019 at 10:14 PM Jacob Lifshay <programmerjake at gmail.com> wrote:
>
> On Wed, Sep 25, 2019 at 9:44 AM Luke Kenneth Casson Leighton
> <lkcl at lkcl.net> wrote:
> >
> > On Wed, Sep 25, 2019 at 4:03 PM Michael Pham <pham.michael.98 at gmail.com> wrote:
> >
> > > In the end though, since you are turning the AMD driver into a Libre
> > > RISC-V version, it might not even matter. But the point of me bringing
> > > this up is that instead of porting AMDVLK, RADV might be a better base
> > > to start from. Since Jacob is a graphics expert though, he'll have
> > > better comments than mine (once he gets back).
> >
> > yehyeh.  it's not about the performance of RADV or AMDVLK, at the
> > assembly-level or what-goes-on-the-screen level, it's about whether
> > the SPIR-V to LLVM-IR code preserves the intrinsics and vectorisation
> > semantics that we need for the SV Vector Engine.
>
> Most of my concerns were addressed in the previous emails, though I
> did want to point out that the part we would be doing the vast
> majority of the work on is LLVM itself.
> I am not completely sure about AMDVLK, but I know that RADV uses
> upstream LLVM as the shader compiler backend.
>
> If we were to base it off of one of AMDVLK or RADV, I'd pick RADV,
> because we can then benefit from all the work that Mesa (not just AMD)
> has put into NIR, where they do graphics-specific optimizations before
> handing off the code to LLVM.
>

This is actually what I was trying to tell Luke as well (although I
said it more clumsily and not as succinct).

-- Michael