Message-ID: <2cdbd8b216a37727c1996440f88229be05ee7efc.camel@imgtec.com>
Date: Thu, 7 Mar 2024 12:01:49 +0000
From: Frank Binns <Frank.Binns@...tec.com>
To: Matt Coster <Matt.Coster@...tec.com>,
        "geert@...ux-m68k.org"
	<geert@...ux-m68k.org>
CC: "robh@...nel.org" <robh@...nel.org>,
        "aford@...conembedded.com"
	<aford@...conembedded.com>,
        "krzysztof.kozlowski+dt@...aro.org"
	<krzysztof.kozlowski+dt@...aro.org>,
        "tzimmermann@...e.de"
	<tzimmermann@...e.de>,
        "mripard@...nel.org" <mripard@...nel.org>,
        "daniel@...ll.ch" <daniel@...ll.ch>,
        "linux-kernel@...r.kernel.org"
	<linux-kernel@...r.kernel.org>,
        "dri-devel@...ts.freedesktop.org"
	<dri-devel@...ts.freedesktop.org>,
        "conor+dt@...nel.org"
	<conor+dt@...nel.org>,
        "aford173@...il.com" <aford173@...il.com>,
        "devicetree@...r.kernel.org" <devicetree@...r.kernel.org>,
        "linux-renesas-soc@...r.kernel.org" <linux-renesas-soc@...r.kernel.org>,
        "maarten.lankhorst@...ux.intel.com" <maarten.lankhorst@...ux.intel.com>,
        "geert+renesas@...der.be" <geert+renesas@...der.be>,
        "magnus.damm@...il.com"
	<magnus.damm@...il.com>,
        "airlied@...il.com" <airlied@...il.com>
Subject: Re: [PATCH 2/6] arm64: dts: renesas: r8a774a1: Enable GPU

Hi Geert,

On Tue, 2024-02-27 at 12:04 +0100, Geert Uytterhoeven wrote:
> Hi Matt,
> 
> On Tue, Feb 27, 2024 at 10:31 AM Matt Coster <Matt.Coster@...tec.com> wrote:
> > Hi Adam,
> > 
> > Thanks for these patches! I'll just reply to this one patch, but my
> > comments apply to them all.
> > 
> > On 27/02/2024 03:45, Adam Ford wrote:
> > > The GPU on the RZ/G2M is a Rogue GX6250 which uses firmware
> > > rogue_4.45.2.58_v1.fw available from Imagination.
> > > 
> > > When enumerated, it appears as:
> > >   powervr fd000000.gpu: [drm] loaded firmware powervr/rogue_4.45.2.58_v1.fw
> > >   powervr fd000000.gpu: [drm] FW version v1.0 (build 6513336 OS)
> > 
> > These messages are printed after verifying the firmware blob’s headers,
> > *before* attempting to upload it to the device. Just because they appear
> > in dmesg does *not* imply the device is functional beyond the handful of
> > register reads in pvr_load_gpu_id().
> > 
> > Since Mesa does not yet have support for this GPU, there’s not a lot
> > that can be done to actually test these bindings.
> 
> OK.
> 
> > When we added upstream support for the first GPU (the AXE core in TI’s
> > AM62), we opted to wait until userspace was sufficiently progressed to
> > the point it could be used for testing. This thought process still
> > applies when adding new GPUs.
> > 
> > Our main concern is that adding bindings for GPUs implies a level of
> > support that cannot be tested. That in turn may make it challenging to
> > justify UAPI changes if/when they’re needed to actually make these GPUs
> > functional.
> 
> I guess that applies to "[PATCH 00/11] Device tree support for
> Imagination Series5 GPU", too, which has been in linux-next for about
> a month?
> https://lore.kernel.org/all/20240109171950.31010-1-afd@ti.com/
> 

The concern Matt has expressed is mostly right. However, adding DT bindings
doesn't prevent us from making UAPI changes. The thing that would prevent this
is adding the compatible(s) to the driver's `struct of_device_id` table, so this
is what we need to avoid doing until sufficient testing has been done.

> > > Signed-off-by: Adam Ford <aford173@...il.com>
> > > 
> > > diff --git a/arch/arm64/boot/dts/renesas/r8a774a1.dtsi b/arch/arm64/boot/dts/renesas/r8a774a1.dtsi
> > > index a8a44fe5e83b..8923d9624b39 100644
> > > --- a/arch/arm64/boot/dts/renesas/r8a774a1.dtsi
> > > +++ b/arch/arm64/boot/dts/renesas/r8a774a1.dtsi
> > > @@ -2352,6 +2352,16 @@ gic: interrupt-controller@...10000 {
> > >                       resets = <&cpg 408>;
> > >               };
> > > 
> > > +             gpu: gpu@...00000 {
> > > +                     compatible = "renesas,r8a774a1-gpu", "img,img-axe";
> > 
> > The GX6250 is *not* an AXE core - it shouldn’t be listed as compatible
> > with one. For prior art, see [1] where we added support for the MT8173
> > found in Elm Chromebooks R13 (also a Series6XT GPU).
> 
> IC. And the bindings in [2].
> 
> > > +                     reg = <0 0xfd000000 0 0x20000>;
> > > +                     clocks = <&cpg CPG_MOD 112>;
> > > +                     clock-names = "core";
> > 
> > Series6XT cores have three clocks (see [1] again). I don’t have a
> > Renesas TRM to hand – do you know if their docs go into detail on the
> > GPU integration?
> 
> Not really. The diagram in the Hardware User's Manual just shows the
> following clock inputs:
>   - Clock (ZGϕ) from CPG,
>   - Clock (S3D1ϕ) from CPG,
>   - MSTP (ST112) from CPG.
> 
> ZG is the main (programmable) 3DGE clock, running at up to 600 MHz.
> S3D1 is the fixed 266 MHz AXI bus clock.
> MSTP112 is the gateable module clock (part of the SYSC/CPG clock
> domain), and its parent is ZG.
> 
> According to the sources:
>   - "core" is the primary clock used by the entire GPU core, so we use
>     MSTP112 for that.
>   - "sys" is the optional system bus clock, so that could be S3D1,
>   - "mem" is the optional memory clock, no idea what that would map to.
> 

The sys and mem clocks are optional in the driver as AXE can be configured at
integration time to operate with a single clock ("core") or three clocks
("core", "sys" and "mem). Series6XT doesn't have this configurability and
expects all three clocks.

> But IMHO the two optional clocks do not matter at all (the driver
> doesn't care about their rates, and just enables them together with
> the core clock), and S3D1 is always-on, so I'd just limit clocks to
> a single item.
> 

They matter in that, if present, we may be able to make additional power
savings. The driver doesn't attempt to take advantage of this at present (and
can't for AM62x anyway), but this shouldn't influence the DT bindings.

> Just wondering: is the availability of 1 clock specific to AXE, or to
> the AXE integration on AM62x?
> 

This is something TI have configured as part of their integration for AM62x.

Thanks
Frank

> > +                     interrupts = <GIC_SPI 119 IRQ_TYPE_LEVEL_HIGH>;
> > +                     power-domains = <&sysc R8A774A1_PD_3DG_B>;
> > +                     resets = <&cpg 112>;
> > +             };
> > [1]: https://gitlab.freedesktop.org/imagination/linux/-/blob/b3506b8bc45ed6d4005eb32a994df0e33d6613f1/arch/arm64/boot/dts/mediatek/mt8173.dtsi#L993-1006
> 
> [2] https://gitlab.freedesktop.org/imagination/linux/-/blob/b3506b8bc45ed6d4005eb32a994df0e33d6613f1/Documentation/devicetree/bindings/gpu/img,powervr.yaml
> 
> 
> Gr{oetje,eeting}s,
> 
>                         Geert
> 
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@...ux-m68k.org
> 
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
>                                 -- Linus Torvalds

Powered by blists - more mailing lists