On Tue, May 30, 2023 at 08:35:14AM -0700, Bjorn Andersson wrote:
>
> On Mon, May 29, 2023 at 02:16:14PM +0530, Manivannan Sadhasivam wrote:
> > On Mon, May 29, 2023 at 09:38:59AM +0200, Konrad Dybcio wrote:
> > > On 28.05.2023 19:07, Manivannan Sadhasivam wrote:
> > > > On Tue, May 23, 2023 at 09:59:53AM +0200, Konrad Dybcio wrote:
> > > >> On 23.05.2023 03:15, Bjorn Andersson wrote:
> > > >>> From: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx>
> [..]
> > > >>> diff --git a/arch/arm64/boot/dts/qcom/sc8280xp.dtsi b/arch/arm64/boot/dts/qcom/sc8280xp.dtsi
> [..]
> > > >>> + gmu: gmu@3d6a000 {
> [..]
> > > >>> + status = "disabled";
> > > >> I've recently discovered that - and I am not 100% sure - all GMUs are
> > > >> cache-coherent. Could you please ask somebody at qc about this?
> > > >>
> > > >
> > > > AFAIU, GMU's job is controlling the voltage and clock to the GPU.
> > > Not just that, it's only the limited functionality we've implemented
> > > upstream so far.
> > >
> >
> > Okay, good to know!
> >
> > > It doesn't do
> > > > any data transactions on its own.
> > > Of course it does. AP communication is done through MMIO writes and
> > > the GMU talks to RPMh via the GPU RSC directly. Apart from that, some
> > > of the GPU registers (that nota bene don't have anything to do with
> > > the GMU M3 core itself) lay within the GMU address space.
> > >
>
> But those aren't shared memory accesses.
>
> >
> > That doesn't justify the fact that cache coherency is needed, especially
> > MMIO writes, unless GMU could snoop the MMIO writes to AP caches.
> >
>
> In reviewing the downstream state again I noticed that the GPU smmu is
> marked dma-coherent, so I will adjust that in v3.
Bjorn,
Would you mind sharing a perf delta (preferrably manhattan offscreen)
you see with and without this dma-coherent property?
-Akhil.
>
> Regards,
> Bjorn
