[Intel-gfx] [RFC 00/33] Add Support for Plane Color Pipeline
Sebastian Wick
sebastian.wick at redhat.com
Tue Sep 5 12:57:12 UTC 2023
On Tue, Sep 05, 2023 at 02:33:04PM +0200, Sebastian Wick wrote:
> On Tue, Sep 05, 2023 at 02:33:26PM +0300, Pekka Paalanen wrote:
> > On Mon, 4 Sep 2023 14:29:56 +0000
> > "Shankar, Uma" <uma.shankar at intel.com> wrote:
> >
> > > > -----Original Message-----
> > > > From: Sebastian Wick <sebastian.wick at redhat.com>
> > > > Sent: Thursday, August 31, 2023 2:46 AM
> > > > To: Shankar, Uma <uma.shankar at intel.com>
> > > > Cc: Harry Wentland <harry.wentland at amd.com>; intel-
> > > > gfx at lists.freedesktop.org; dri-devel at lists.freedesktop.org; wayland-
> > > > devel at lists.freedesktop.org; Ville Syrjala <ville.syrjala at linux.intel.com>; Pekka
> > > > Paalanen <pekka.paalanen at collabora.com>; Simon Ser <contact at emersion.fr>;
> > > > Melissa Wen <mwen at igalia.com>; Jonas Ådahl <jadahl at redhat.com>; Shashank
> > > > Sharma <shashank.sharma at amd.com>; Alexander Goins <agoins at nvidia.com>;
> > > > Naseer Ahmed <quic_naseer at quicinc.com>; Christopher Braga
> > > > <quic_cbraga at quicinc.com>
> > > > Subject: Re: [RFC 00/33] Add Support for Plane Color Pipeline
> > > >
> > > > On Wed, Aug 30, 2023 at 08:47:37AM +0000, Shankar, Uma wrote:
> > > > >
> > > > >
> > > > > > -----Original Message-----
> > > > > > From: Harry Wentland <harry.wentland at amd.com>
> > > > > > Sent: Wednesday, August 30, 2023 12:56 AM
> > > > > > To: Shankar, Uma <uma.shankar at intel.com>;
> > > > > > intel-gfx at lists.freedesktop.org; dri- devel at lists.freedesktop.org
> > > > > > Cc: wayland-devel at lists.freedesktop.org; Ville Syrjala
> > > > > > <ville.syrjala at linux.intel.com>; Pekka Paalanen
> > > > > > <pekka.paalanen at collabora.com>; Simon Ser <contact at emersion.fr>;
> > > > > > Melissa Wen <mwen at igalia.com>; Jonas Ådahl <jadahl at redhat.com>;
> > > > > > Sebastian Wick <sebastian.wick at redhat.com>; Shashank Sharma
> > > > > > <shashank.sharma at amd.com>; Alexander Goins <agoins at nvidia.com>;
> > > > > > Naseer Ahmed <quic_naseer at quicinc.com>; Christopher Braga
> > > > > > <quic_cbraga at quicinc.com>
> > > > > > Subject: Re: [RFC 00/33] Add Support for Plane Color Pipeline
> > > > > >
> > > > > > +CC Naseer and Chris, FYI
> > > > > >
> > > > > > See https://patchwork.freedesktop.org/series/123024/ for whole series.
> > > > > >
> > > > > > On 2023-08-29 12:03, Uma Shankar wrote:
> > > > > > > Introduction
> > > > > > > ============
> > > > > > >
> > > > > > > Modern hardwares have various color processing capabilities both
> > > > > > > at pre-blending and post-blending phases in the color pipeline.
> > > > > > > The current drm implementation exposes only the post-blending
> > > > > > > color hardware blocks. Support for pre-blending hardware is missing.
> > > > > > > There are multiple use cases where pre-blending color hardware
> > > > > > > will be
> > > > > > > useful:
> > > > > > > a) Linearization of input buffers encoded in various transfer
> > > > > > > functions.
> > > > > > > b) Color Space conversion
> > > > > > > c) Tone mapping
> > > > > > > d) Frame buffer format conversion
> > > > > > > e) Non-linearization of buffer(apply transfer function)
> > > > > > > f) 3D Luts
> > > > > > >
> > > > > > > and other miscellaneous color operations.
> > > > > > >
> > > > > > > Hence, there is a need to expose the color capabilities of the
> > > > > > > hardware to user-space. This will help userspace/middleware to use
> > > > > > > display hardware for color processing and blending instead of
> > > > > > > doing it through GPU shaders.
> > > > > > >
> > > > > >
> > > > > > Thanks, Uma, for sending this. I've been working on something
> > > > > > similar but you beat me to it. :)
> > > > >
> > > > > Thanks Harry for the useful feedback and overall collaboration on this so far.
> > > > >
> > > > > > >
> > > > > > > Work done so far and relevant references
> > > > > > > ========================================
> > > > > > >
> > > > > > > Some implementation is done by Intel and AMD/Igalia to address the same.
> > > > > > > Broad consensus is there that we need a generic API at drm core to
> > > > > > > suffice the use case of various HW vendors. Below are the links
> > > > > > > capturing the discussion so far.
> > > > > > >
> > > > > > > Intel's Plane Color Implementation:
> > > > > > > https://patchwork.freedesktop.org/series/90825/
> > > > > > > AMD's Plane Color Implementation:
> > > > > > > https://patchwork.freedesktop.org/series/116862/
> > > > > > >
> > > > > > >
> > > > > > > Hackfest conclusions
> > > > > > > ====================
> > > > > > >
> > > > > > > HDR/Color Hackfest was organised by Redhat to bring all the
> > > > > > > industry stakeholders together and converge on a common uapi
> > > > expectations.
> > > > > > > Participants from Intel, AMD, Nvidia, Collabora, Redhat, Igalia
> > > > > > > and other prominent user-space developers and maintainers.
> > > > > > >
> > > > > > > Discussions happened on the uapi expectations, opens, nature of
> > > > > > > hardware of multiple hardware vendors, challenges in generalizing
> > > > > > > the same and the path forward. Consensus was made that drm core
> > > > > > > should implement descriptive APIs and not go with prescriptive
> > > > > > > APIs. DRM core should just expose the hardware capabilities;
> > > > > > > enabling, customizing and programming the same should be done by
> > > > > > > the user-space. Driver should just
> > > > > > honor the user space request without doing any operations internally.
> > > > > > >
> > > > > > > Thanks to Simon Ser, for nicely documenting the design consensus
> > > > > > > and an UAPI RFC which can be referred to here:
> > > > > > >
> > > > > > > https://lore.kernel.org/dri-devel/QMers3awXvNCQlyhWdTtsPwkp5ie9bze
> > > > > > > _hD5
> > > > > > >
> > > > > >
> > > > nAccFW7a_RXlWjYB7MoUW_8CKLT2bSQwIXVi5H6VULYIxCdgvryZoAoJnC5lZgyK1
> > > > Q
> > > > > > Wn48
> > > > > > > 8=@emersion.fr/
> > > > > > >
> > > > > > >
> > > > > > > Design considerations
> > > > > > > =====================
> > > > > > >
> > > > > > > Following are the important aspects taken into account while
> > > > > > > designing the current RFC
> > > > > > > proposal:
> > > > > > >
> > > > > > > 1. Individual HW blocks can be muxed. (e.g. out of two HW blocks
> > > > > > > only one
> > > > > > can be used)
> > > > > > > 2. Position of the HW block in the pipeline can be programmable
> > > > > > > 3. LUTs can be one dimentional or three dimentional
> > > > > > > 4. Number of LUT entries can vary across platforms
> > > > > > > 5. Precision of LUT entries can vary across platforms
> > > > > > > 6. Distribution of LUT entries may vary. e.g Mutli-segmented,
> > > > Logarithmic,
> > > > > > > Piece-Wise Linear(PWL) etc
> > > > > > > 7. There can be parameterized/non-parameterized fixed function HW
> > > > > > blocks.
> > > > > > > e.g. Just a hardware bit, to convert from one color space to another.
> > > > > > > 8. Custom non-standard HW implementation.
> > > > > > > 9. Leaving scope for some vendor defined pescriptive
> > > > > > > implementation if
> > > > > > required.
> > > > > > > 10.Scope to handle any modification in hardware as technology
> > > > > > > evolves
> > > > > > >
> > > > > > > The current proposal takes into account the above considerations
> > > > > > > while keeping the implementation as generic as possible leaving
> > > > > > > scope for future
> > > > > > additions or modifications.
> > > > > > >
> > > > > > > This proposal is also in line to the details mentioned by Simon's
> > > > > > > RFC covering all the aspects discussed in hackfest.
> > > > > > >
> > > > > > >
> > > > > > > Outline of the implementation
> > > > > > > ============================
> > > > > > >
> > > > > > > Each Color Hardware block will be represented by a data structure
> > > > drm_color_op.
> > > > > > > These color operations will form the building blocks of a color
> > > > > > > pipeline which best represents the underlying Hardware. Color
> > > > > > > operations can be re-arranged, substracted or added to create
> > > > > > > distinct color pipelines to accurately describe the Hardware
> > > > > > > blocks present in the display
> > > > > > engine.
> > > > > >
> > > > > > Who is doing the arranging of color operations? IMO a driver should
> > > > > > define one or more respective pipelines that can be selected by
> > > > > > userspace. This seems to be what you're talking about after (I
> > > > > > haven't reviewed the whole thing yet). Might be best to drop this sentence or
> > > > to add clarifications in order to avoid confusion.
> > > > >
> > > > > Yes it's the driver who will set the pipeline based on the underlying
> > > > > hardware arrangement and possible combinations. There can be multiple
> > > > > pipelines possible if hardware can be muxed or order can be re-arranged (all
> > > > viable combinations should be defined as a pipeline in driver).
> > > > > Yeah, I will re-phrase this to help clarify it and avoid any ambiguity.
> > > > >
> > > > > > >
> > > > > > > In this proposal, a color pipeline is represented as an array of
> > > > > > > struct drm_color_op. For individual color operation, we add blobs
> > > > > > > to advertise the capability of the particular Hardware block.
> > > > > > >
> > > > > > > This color pipeline is then packaged within a blob for the user
> > > > > > > space to retrieve it.
> > > > > > >
> > > > > >
> > > > > > Would it be better to expose the drm_color_ops directly, instead of
> > > > > > packing a array of drm_color_ops into a blob and then giving that to
> > > > userspace.
> > > > >
> > > > > Advantage we see in packing in blobs is that interface will be
> > > > > cleaner. There will be just one GET_COLOR_PIPELINE property to invoke by user
> > > > and then its just parsing the data.
> > > > > This way the entire underlying hardware blocks with pipeline will be available to
> > > > user.
> > > >
> > > > I don't see how parsing a blob is easier than requesting the color ops from the
> > > > kernel. User space is already equiped with getting KMS objects and the igt test
> > > > code from Harry shows that this is all pretty trivial plumbing.
> > >
> > > There are multiple color operations possible with unique lut distribution and
> > > capabilities. Also the order of hardware blocks and possibility of multiple pipelines.
> > > Having all the information with one query and property and also be able to set the
> > > same with just one property call using blobs is better than multiple calls to driver.
> > > This can be useful in high refresh rate cases where not much time is there to program
> > > the hardware state. Latency of multiple calls to driver can be avoided.
> >
> > Hi,
> >
> > querying all that information only happens once, at KMS client start-up.
> >
> > Setting up a color pipeline is already a single call: the atomic commit ioctl.
>
> Well, clients also issue a bunch of ioctl to set some properties to the
> desired state and then you might run through a whole bunch of
> configurations to find one that actually works, so there is a case to be
> made that there are a lot of ioctls involved.
pq pointed at that for atomic this is actually really just a single
ioctl so really there is no advantage here at all.
> I just don't think this is an issue right now. Nobody has been
> complaining about the ioctls being a limiting factor so why should we
> optimize for this? Especially because it brings with it a bunch of
> disadvantages.
>
> Anyway, I agree with the sentiment here: this is not something we should
> optimize for.
>
> >
> > >
> > > > > For a particular hardware block in pipeline, user can get the relevant
> > > > > details from blob representing that particular block. We have created
> > > > > IGT tests (links mentioned in cover-letter) to demonstrate how it can be done.
> > > > This is just to clarify the idea.
> > > >
> > > > The blob is also not introspectable with the usual tools whereas exposing them as
> > > > properties would be.
> > > >
> > > > It also would, like Pekka correctly noted, bring a whole bunch of issues about
> > > > compatibility and versioning that are well understood with objects
> > > > + properties.
> > >
> > > The blob should be standardized in the UAPI and structures to parse them should be fixed.
> > > With this compatibility issues can be prevented.
> >
> > I think that is short-sighted.
> >
> > > > > Also since info is passed via blobs we have the flexibility to even
> > > > > define segmented LUTs and PWL hardware blocks. Also we have left scope
> > > > > for custom private hardware blocks as well which driver can work with
> > > > respective HAL and get that implemented.
> > > >
> > > > When color ops are real KMS objects they still can have properties which are
> > > > blobs that can store LUTs and other data.
> > > >
> > > > And we should avoid private blocks at all cost. In fact, I don't think the KMS rules
> > > > have changed in that regard and it simply is not allowed.
> > >
> > > Private blocks are not standardized but are vendor specific. So generic userspace will
> > > ignore these. However vendor and its respective HAL can make use of this field and leaves
> > > a scope to cater to such hardware vendors need. This doesn't affect the expectation of the
> > > standardized color operations which will be defined as enum in UAPI.
> >
> > Vendors can have and expose their own unique snowflake operations
> > without any "private" as well: pick an unused operation type code, and
> > document what it does. Advertise it in some pipelines.
> >
> > Vendor HALs can make use of it, but it also allows generic userspace to
> > make use of it at will, and it allows other vendors to implement the
> > same and benefit from it without needing to patch every userspace.
> >
> > Or does Intel not want other vendors to potentially make use of their
> > UAPIs?
> >
> > > > > We can even define prescriptive operations as a private entry and
> > > > > enable it if a certain driver and HAL agree.
> > > > >
> > > > > > > To advertise the available color pipelines, an immutable ENUM
> > > > > > > property "GET_COLOR_PIPELINE" is introduced. This enum property has
> > > > blob id's as values.
> > > > > > > With each blob id representing a distinct color pipeline based on
> > > > > > > underlying HW capabilities and their respective combinations.
> > > > > > >
> > > > > > > Once the user space decides on a color pipeline, it can set the
> > > > > > > pipeline and the corresponding data for the hardware blocks within
> > > > > > > the pipeline with the BLOB property "SET_COLOR_PIPELINE".
> > > > > > >
> > > > > >
> > > > > > When I discussed this at the hackfest with Simon he proposed a new
> > > > > > DRM object, (I've called it a drm_colorop) to represent a color operation.
> > > > > > Each drm_colorop has a "NEXT" pointer to another drm_colorop, or
> > > > > > NULL if its the last in the pipeline. You can then have a mutable
> > > > > > enum property on the plane to discover and select a color pipeline.
> > > > >
> > > > > Yes, the proposal is inspired by this idea. Sure, we can work together to enhance
> > > > the design.
> > > > > Personally I feel the one proposed in the current RFC will do the same
> > > > > thing envisioned by Simon and you Harry. Management of the pipeline,
> > > > > addition, deletion and flexibility to represent hardware is more with blobs with
> > > > the relevant structures agreed in UAPI.
> > > > >
> > > > > > This seems a bit more transparent than a blob. You can see my
> > > > > > changes (unfortunately very WIP, don't look too closely at
> > > > > > individual patches) at
> > > > > > https://gitlab.freedesktop.org/hwentland/linux/-/merge_requests/5/di
> > > > > > ffs
> > > > > >
> > > > > > libdrm changes:
> > > > > > https://gitlab.freedesktop.org/hwentland/drm/-/merge_requests/1/diff
> > > > > > s
> > > > >
> > > > > Sure, will check the same.
> > > > >
> > > > > > IGT changes:
> > > > > > https://gitlab.freedesktop.org/hwentland/igt-gpu-tools/-/merge_reque
> > > > > > sts/1/diffs
> > > > > >
> > > > > > I'll take time to review your whole series and will see whether we
> > > > > > can somehow keep the best parts of each.
> > > > >
> > > > > Thanks and agree. Let's work together and get this implemented in DRM.
> > > > >
> > > > > > Curious to hear other opinions on the blob vs new DRM object question as
> > > > well.
> > > >
> > > > I can see a few advantages with the blob approach.
> > > >
> > > > User space can store the entire state in one blob and just assign a new blob to
> > > > change to another pipeline configuration.
> > >
> > > Agree
> > >
> > > > However, I would argue that changing a lot of properties is already common
> > > > practice and works well. User space can deal with it and has abstractions to
> > > > make this managable.
> > >
> > > Blob gives a lot of flexibility and ability to define the hardware capabilities generically.
> >
> > The structure of the atomic commit ioctl and the KMS property system is
> > even more flexible.
> >
> > > Lut distribution, number of segments, samples in each segment, precision of luts etc.
> > > can be precisely defined and userspace will get a complete picture of the underlying
> > > hardware and its capabilities. Also this is being done with just 2 properties. Leaving
> > > scope for future addition of any standard color operation as well.
> >
> > The number of properties does not seem too useful to strictly minimise
> > over other aspects.
> >
> >
> > Thanks,
> > pq
> >
> > >
> > > I feel blob approach once properly documented is a bit more flexible and scalable.
> > > Maybe I am bit biased here 😊 but all ideas are welcome.
> > >
> > > We have implemented some IGT's as well to explain the design better. Link below:
> > > https://patchwork.freedesktop.org/series/123018/
> > >
> > > Thanks Sebastian for the feedback.
> > >
> > > Regards,
> > > Uma Shankar
>
>
More information about the Intel-gfx
mailing list