On Thu, Apr 04, 2024 at 09:59:03AM -0400, Harry Wentland wrote: > Hi all, > > On 2024-04-04 06:24, Pekka Paalanen wrote: > > On Wed, 3 Apr 2024 17:32:46 -0400 > > Leo Li <sunpeng.li@xxxxxxx> wrote: > > > >> On 2024-03-28 10:33, Pekka Paalanen wrote: > >>> On Fri, 15 Mar 2024 13:09:56 -0400 > >>> <sunpeng.li@xxxxxxx> wrote: > >>> > >>>> From: Leo Li <sunpeng.li@xxxxxxx> > >>>> > >>>> These patches aim to make the amdgpgu KMS driver play nicer with compositors > >>>> when building multi-plane scanout configurations. They do so by: > >>>> > >>>> 1. Making cursor behavior more sensible. > >>>> 2. Allowing placement of DRM OVERLAY planes underneath the PRIMARY plane for > >>>> 'underlay' configurations (perhaps more of a RFC, see below). > >>>> > >>>> Please see the commit messages for details. > >>>> > >>>> > >>>> For #2, the simplest way to accomplish this was to increase the value of the > >>>> immutable zpos property for the PRIMARY plane. This allowed OVERLAY planes with > >>>> a mutable zpos range of (0-254) to be positioned underneath the PRIMARY for an > >>>> underlay scanout configuration. > >>>> > >>>> Technically speaking, DCN hardware does not have a concept of primary or overlay > >>>> planes - there are simply 4 general purpose hardware pipes that can be maped in > >>>> any configuration. So the immutable zpos restriction on the PRIMARY plane is > >>>> kind of arbitrary; it can have a mutable range of (0-254) just like the > >>>> OVERLAYs. The distinction between PRIMARY and OVERLAY planes is also somewhat > >>>> arbitrary. We can interpret PRIMARY as the first plane that should be enabled on > >>>> a CRTC, but beyond that, it doesn't mean much for amdgpu. > >>>> > >>>> Therefore, I'm curious about how compositors devs understand KMS planes and > >>>> their zpos properties, and how we would like to use them. It isn't clear to me > >>>> how compositors wish to interpret and use the DRM zpos property, or > >>>> differentiate between OVERLAY and PRIMARY planes, when it comes to setting up > >>>> multi-plane scanout. > >>> > >>> You already quoted me on the Weston link, so I don't think I have > >>> anything to add. Sounds fine to me, and we don't have a standard plane > >>> arrangement algorithm that the kernel could optimize zpos ranges > >>> against, yet. > >>> > >>>> Ultimately, what I'd like to answer is "What can we do on the KMS driver and DRM > >>>> plane API side, that can make building multi-plane scanout configurations easier > >>>> for compositors?" I'm hoping we can converge on something, whether that be > >>>> updating the existing documentation to better define the usage, or update the > >>>> API to provide support for something that is lacking. > >>> > >>> I think there probably should be a standardised plane arrangement > >>> algorithm in userspace, because the search space suffers from > >>> permutational explosion. Either there needs to be very few planes (max > >>> 4 or 5 at-all-possible per CRTC, including shareable ones) for an > >>> exhaustive search to be feasible, or all planes should be more or less > >>> equal in capabilities and userspace employs some simplified or > >>> heuristic search. > >>> > >>> If the search algorithm is fixed, then drivers could optimize zpos > >>> ranges to have the algorithm find a solution faster. > >>> > >>> My worry is that userspace already has heuristic search algorithms that > >>> may start failing if drivers later change their zpos ranges to be more > >>> optimal for another algorithm. > >>> > >>> OTOH, as long as exhaustive search is feasible, then it does not matter > >>> how DRM drivers set up the zpos ranges. > >>> > >>> In any case, the zpos ranges should try to allow all possible plane > >>> arrangements while minimizing the number of arrangements that won't > >>> work. The absolute values of zpos are pretty much irrelevant, so I > >>> think setting one plane to have an immutable zpos is a good idea, even > >>> if it's not necessary by the driver. That is one less moving part, and > >>> only the relative ordering between the planes matters. > >>> > >>> > >>> Thanks, > >>> pq > >> > >> Right, thanks for your thoughts! I agree that there should be a common plane > >> arrangement algorithm. I think libliftoff is the most obvious candidate here. It > >> only handles overlay arrangements currently, but mixed-mode arrangements is > >> something I've been trying to look at. > >> > >> Taking the driver's reported zpos into account could narrow down the search > >> space for mixed arrangements. We could tell whether underlay, or overlay, or > >> both, is supported by looking at the allowed zpos ranges. > >> > >> I also wonder if it'll make underlay assignments easier. libliftoff has an > >> assumption that the PRIMARY plane has the lowest zpos (which now I realize, is > >> not always true). Therefore, the underlay buffer has to be placed on the > >> PRIMARY, with the render buffer on a higher OVERLAY. Swapping buffers between > >> planes when testing mixed-arrangements is kind of awkward, and simply setting > >> the OVERLAY's zpos to be lower or higher than the PRIMARY's sounds simpler. > >> > >> Currently only gamescope makes use of libliftoff, but I'm curious if patches > >> hooking it up to Weston would be welcomed? If there are other ways to have a > >> common arrangement algorithm, I'd be happy to hear that as well. > > > > A natural thing would be to document such an algorithm with the KMS > > UAPI. > > > > I don't know libliftoff well enough to say how welcome it would be in > > Weston. I have no fundamental or policy reason to keep an independent > > implementation in Weston though, so it's plausible at least. > > > > It would need investigation, and perhaps also extending Weston test > > suite a lot more towards VKMS to verify plane assignments. Currently > > all plane assignment testing is manual on real hardware. > > > > It looks like VKMS doesn't have explicit zpos yet, so someone would > probably need to add that. > > https://drmdb.emersion.fr/properties/4008636142/zpos Yes. If we look into adding that, maybe it should be done using with ConfigFS: https://patchwork.freedesktop.org/series/122618/ With that in and with zpos support, we could then run a batch of tests that can dynamically exercise on-the-fly all possible combinations. > > Harry > > >> Note that libliftoff's algorithm is more complex than weston, since it searches > >> harder, and suffers from that permutational explosion. But it solves that by > >> trying high benefit arrangements first (offloading surfaces that update > >> frequently), and bailing out once the search reaches a hard-coded deadline. > >> Since it's currently overlay-only, the goal could be to "simply" have no > >> regressions. > > > > Ensuring no regressions would indeed need to be taken care of by > > extending the VKMS-based automated testing. > > > > > > Thanks, > > pq > > > >>> > >>>> Some links to provide context and details: > >>>> * What is underlay?: https://gitlab.freedesktop.org/emersion/libliftoff/-/issues/76 > >>>> * Discussion on how to implement underlay on Weston: https://gitlab.freedesktop.org/wayland/weston/-/merge_requests/1258#note_2325164 > >>>> > >>>> Cc: Joshua Ashton <joshua@xxxxxxxxx> > >>>> Cc: Michel Dänzer <mdaenzer@xxxxxxxxxx> > >>>> Cc: Chao Guo <chao.guo@xxxxxxx> > >>>> Cc: Xaver Hugl <xaver.hugl@xxxxxxxxx> > >>>> Cc: Vikas Korjani <Vikas.Korjani@xxxxxxx> > >>>> Cc: Robert Mader <robert.mader@xxxxxxxxx> > >>>> Cc: Pekka Paalanen <pekka.paalanen@xxxxxxxxxxxxx> > >>>> Cc: Sean Paul <sean@xxxxxxxxxx> > >>>> Cc: Simon Ser <contact@xxxxxxxxxxx> > >>>> Cc: Shashank Sharma <shashank.sharma@xxxxxxx> > >>>> Cc: Harry Wentland <harry.wentland@xxxxxxx> > >>>> Cc: Sebastian Wick <sebastian.wick@xxxxxxxxxx> > >>>> > >>>> Leo Li (2): > >>>> drm/amd/display: Introduce overlay cursor mode > >>>> drm/amd/display: Move PRIMARY plane zpos higher > >>>> > >>>> .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 405 ++++++++++++++++-- > >>>> .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.h | 7 + > >>>> .../amd/display/amdgpu_dm/amdgpu_dm_crtc.c | 1 + > >>>> .../amd/display/amdgpu_dm/amdgpu_dm_plane.c | 28 +- > >>>> 4 files changed, 391 insertions(+), 50 deletions(-) > >>>> > >>> > > >
Attachment:
signature.asc
Description: PGP signature