Re: [PATCH v3 1/2] drm/xe/mcr: Take vcs1/vecs1 into account for first media slice

On Tue, 28 Jul 2026 13:57:10 -0700, Matt Roper wrote:
>
> On Mon, Jul 27, 2026 at 10:27:11AM -0700, Ashutosh Dixit wrote:
> > vcs1 and vecs1 may also be present on the first media slice, so take
> > vcs1/vecs1 also into account when determining if the first media slice is
> > absent.
> >
> > Bspec: 67103, 77977
> >
> > Cc: Gustavo Sousa <[email protected]>
> > Cc: Matt Roper <[email protected]>
> > Signed-off-by: Ashutosh Dixit <[email protected]>
> > ---
> > v2: Bug fix
> > v3: Fix for BMG (Matt Roper)
> > ---
> >  drivers/gpu/drm/xe/xe_gt_mcr.c | 8 ++++++--
> >  1 file changed, 6 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_gt_mcr.c b/drivers/gpu/drm/xe/xe_gt_mcr.c
> > index 9799ffadf85d2..8beb4006cd891 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_mcr.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_mcr.c
> > @@ -452,11 +452,15 @@ static void init_steering_dss(struct xe_gt *gt)
> >
> >  static void init_steering_oaddrm(struct xe_gt *gt)
> >  {
> > +	u64 hwe_mask = XE_HW_ENGINE_VCS0 | XE_HW_ENGINE_VECS0;
> > +
> > +	if (MEDIA_VERx100(gt_to_xe(gt)) >= 3500)
>
> This will work as a short-term bandaid, but fundamentally the media
> configuration isn't really a case where "old platforms were always 1/1/1
> and new platforms will probably always be 2/2/2."  It's something that
> can easily vary platform-by-platform or even SKU-by-SKU (e.g., bspec
> 70819 still shows some what was originally planned in the past for
> different SKUs of a platform, even though that platform eventually
> settled on something simpler).  If I recall correctly, there were also
> configurations like 2/1/1 used on older platforms like DG2, but we
> didn't need to worry about SCMI or OADDRM steering for MCR purposes on
> those platforms so it just never came up in the code.
>
> I'm okay landing this patch as a short-term fix for NVL and CRI, but we
> should add a TODO comment or something indicating that we really should
> move the "VD per SCMI" and "VE per SCMI" values into our media
> descriptors so that we can easily pick the appropriate config on a
> per-release / per-SKU manner to ensure we think about it on a
> release-by-release basis when enabling new platforms and IP versions in
> the future.

Thanks Matt, I have added the TODO in v4, I prefer to merge this for now as
a short-term fix for NVL and CRI. Thanks.

>
> > +		hwe_mask |= XE_HW_ENGINE_VCS1 | XE_HW_ENGINE_VECS1;
> >	/*
> >	 * First instance is only terminated if the entire first media slice
> > -	 * is absent (i.e., no VCS0 or VECS0).
> > +	 * is absent (i.e., no engines in hwe_mask).
> >	 */
> > -	if (gt->info.engine_mask & (XE_HW_ENGINE_VCS0 | XE_HW_ENGINE_VECS0))
> > +	if (gt->info.engine_mask & hwe_mask)
> >		gt->steering[OADDRM].group_target = 0;
> >	else
> >		gt->steering[OADDRM].group_target = 1;
> > --
> > 2.54.0
> >
>
> --
> Matt Roper
> Graphics Software Engineer
> Linux GPU Platform Enablement
> Intel Corporation