Re: [PATCH] drm/tegra: Enable cmu for Tegra186 and Tegra194

[Mikko Perttunen]

On Wednesday, January 28, 2026 2:57 AM Aaron Kling wrote:
> On Tue, Jan 27, 2026 at 4:32 AM Thierry Reding
> <thierry.reding@xxxxxxxxxx> wrote:
> >
> > On Tue, Jan 27, 2026 at 01:12:54PM +0900, Mikko Perttunen wrote:
> > > On Thursday, January 22, 2026 2:08 AM Kurt Kiefer wrote:
> > > >
> > > > > On Dec 8, 2025, at 8:23 PM, Aaron Kling <webgeek1234@xxxxxxxxx> wrote:
> > > > >
> > > > > On Wed, Nov 5, 2025 at 3:28 PM Jasper Korten <jja2000@xxxxxxxxx> wrote:
> > > > >>
> > > > >> Hi all,
> > > > >>
> > > > >> On 11/4/25 19:12, Aaron Kling wrote:
> > > > >>> On Tue, Nov 4, 2025 at 3:14 AM Thierry Reding <thierry.reding@xxxxxxxxx> wrote:
> > > > >>>> On Mon, Nov 03, 2025 at 12:39:57PM -0600, Aaron Kling wrote:
> > > > >>>>> On Mon, Nov 3, 2025 at 5:54 AM Thierry Reding <thierry.reding@xxxxxxxxx> wrote:
> > > > >>>>>> On Sat, Nov 01, 2025 at 06:15:17PM -0500, Aaron Kling via B4 Relay wrote:
> > > > >>>>>>> From: Aaron Kling <webgeek1234@xxxxxxxxx>
> > > > >>>>>>>
> > > > >>>>>>> Without the cmu, nvdisplay will display colors that are notably darker
> > > > >>>>>>> than intended. The vendor bootloader and the downstream display driver
> > > > >>>>>>> enable the cmu and sets a sRGB table. Loading that table here results in
> > > > >>>>>>> the intended colors.
> > > > >>>>>>>
> > > > >>>>>>> Signed-off-by: Aaron Kling <webgeek1234@xxxxxxxxx>
> > > > >>>>>>> ---
> > > > >>>>>>>  drivers/gpu/drm/tegra/dc.h  |  13 +++
> > > > >>>>>>>  drivers/gpu/drm/tegra/sor.c | 206 ++++++++++++++++++++++++++++++++++++++++++++
> > > > >>>>>>>  2 files changed, 219 insertions(+)
> > > > >>>>>> What does "darker than intended" mean? Who defines the intention? How do
> > > > >>>>>> we know what the intention is? What this patch ultimately seems to be
> > > > >>>>>> doing is define sRGB to be the default colorspace. Is that always the
> > > > >>>>>> right default choice? What if people want to specify a different
> > > > >>>>>> colorspace?
> > > > >>>>> I reported this issue almost a month ago. See kernel lore [0] and
> > > > >>>>> freedesktop issue [1]. The pictures in the latter show what nvdisplay
> > > > >>>>> looks like right now. It's nigh unusably dark. When booted into
> > > > >>>>> Android with a tv launcher that has a black background, as is default
> > > > >>>>> for LineageOS, it is really hard to read anything. Is it correct as a
> > > > >>>>> default? Well, cboot hardcodes this, so... presumably? It would be
> > > > >>>>> more ideal to expose this and csc to userspace, but I'm not sure if
> > > > >>>>> drm has a standardized interface for that or if tegra would have to
> > > > >>>>> make something vendor specific. I think that would be a separate
> > > > >>>>> change concept compared to setting this default, though.
> > > > >>>> The reason I'm asking is because I don't recall ever seeing "broken"
> > > > >>>> colors like you do. So I suspect that this may also be related to what
> > > > >>>> display is connected, or the mode that we're setting.
> > > > >> I have tried it on both a MacroSilicon HDMI capture card and an Arzopa
> > > > >> Z1FC 1080p portable monitor and run into the same darker colors. Both
> > > > >> have in common that they use HDMI which seems to line up with what Aaron
> > > > >> is reporting. I do not have an eDP display to test or another carrier
> > > > >> board with a different display out to test.
> > > > >>>> It could perhaps
> > > > >>>> also be related to what infoframes we're sending and how these are
> > > > >>>> supported/interpreted by the attached display.
> > > > >>>>
> > > > >>>> All of that is to say that maybe this looks broken on the particular
> > > > >>>> setup that you have but may works fine on other setups. Changing the
> > > > >>>> default may fix your setup and break others.
> > > > >>> Do you have a device set up so you can check? Or does the regression
> > > > >>> test bench have a display that can be forwarded?
> > > > >>>
> > > > >>> My current setup is a rack of units plugged via hdmi to a kvm which is
> > > > >>> then plugged to a pikvm. I also observed this issue before I had this
> > > > >>> setup, plugged directly to a 1080p monitor. I have not checked
> > > > >>> displayport. I can cycle through a couple other displays without this
> > > > >>> patch to see if I get any other result. I am fairly certain I have
> > > > >>> consistently seen this issue since I started trying to work with
> > > > >>> tegra-drm on kernel 6.1 or maybe even 5.15. I've never seen it work to
> > > > >>> allow for a bisect.
> > > > >>>
> > > > >>> I am in contact with one other person with a tx2 devkit, who
> > > > >>> replicated the issue when I asked. Who plans to reply to this thread
> > > > >>> with setup info later.
> > > > >>
> > > > >> For reference, I am said person. I have a Jetson TX2 Devkit that uses
> > > > >> the P2771 Device Tree. I'm running a Fedora distrokernel with no
> > > > >> additional patches applied by myself. I have personally noticed the
> > > > >> issue to at least be present on 6.14.5 and 6.17.4.
> > > > >>
> > > > >>
> > > > >> I'm currently not at home to take screenshots with and without the
> > > > >> submitted patch, but will be able to do it tomorrownight or friday.
> > > > >
> > > > > Any further thoughts from the maintainers on this patch? As far as I
> > > > > know, this is an issue for all users, at the very least on hdmi.
> > > > >
> > > > > Aaron
> > > > >
> > > >
> > > > I can confirm that I have the same issue on a DisplayPort output of t194.
> > > > IMO, this patch will need to be reworked a bit to enable the CMU for this
> > > > output as well. I hacked this change in for DisplayPort, and then it
> > > > functioned as intended there as well.
> > > >
> > > > I've traced back to the reason this is necessary. The DC hub driver is
> > > > applying an sRGB degamma for every RGB plane (presumably for blending),
> > > > and then nothing reapplies the EOTF later on. Without gamma correction
> > > > in places where it is expected, images are going to look "too dark".
> > > >
> > > > Which does raise the point that there is an alternative implementation
> > > > where we do not degamma RGB planes in the first place. But this may have
> > > > unintended consequences when it comes to composition.
> > > >
> > > > The SOR does not appear to handle YCbCr outputs at this time, so enabling
> > > > the CMU assuming an sRGB EOTF seems like a reasonable path here, to me.
> > > >
> > > > Kurt
> > >
> > > I tested this patch locally and did some investigation. Can confirm
> > > that on my Jetson AGX Xavier, this patch (or disabling degamma) fixes
> > > the color output.
> > >
> > > The colorspace the display expects from the incoming data is specified
> > > in the AVI infoframe. This is generated in
> > > tegra_sor_hdmi_setup_avi_infoframe, which calls into
> > > drm_hdmi_avi_infoframe_from_display_mode, which leaves a lot of fields
> > > set at the default.
> > >
> > > Currently we're advertising:
> > > * colorimetry = no data -> for HD resolutions, use Rec. 709 primaries.
> > > These are the same as sRGB.
> > > * itc = false -> NOT IT content.
> > >
> > > My understanding (based on some LLM research and otherwise) is that
> > > this is likely to result in the display expecting Rec. 709 colors with
> > > Rec. 709 gamma. sRGB gamma is slightly different, and setting itc =
> > > true would hint the display towards using sRGB gamma.
> > >
> > > However, what seems clear to me is that the display would be expecting
> > > nonlinear data, so enabling gamma conversion at the output LUT seems
> > > correct to me. So this patch would be a clear improvement (with the
> > > fixes already discussed).
> >
> > I think ideally we want to hook this up to the DRM color management
> > facilities, so that it can both be properly reported and configured
> > at runtime.
> 
> That would be ideal yes. But I don't have nearly the knowledge of DRM
> or nvdisplay to wire this up. Is there someone at Nvidia that can make
> time to do so?
> 
> > Obviously we also want to make sure that the output pixels match what
> > is advertised via the AVI infoframe. Looks like there's concensus that
> > enabling the output LUT is the correct way to do that.
> 
> If more fully featured CMU support can't be done in a reasonable
> timeframe, I can address the earlier review comments and coordinate
> with Kurt to replicate this for DP, then send a v2.
> 
> Aaron

The DRM color management stuff should be mostly orthogonal to this, so yes, we should go forward with this patch.

Mikko