Re: [PATCH v2 0/5] better LUT segmentation for EOTFs

[Matthew Schwartz <matthew.schwartz@linux.dev>]

On 5/6/26 12:11 PM, Melissa Wen wrote:
> Hi,
> 
> With an external HDR monitor, we can see gradient banding around the sun
> in the intro of Ori and the Will of the Wisps game on steamOS/Gamescope.
> Gamescope uses AMD predefined transfer functions for degamma,
> shaper/pre-3D-LUT and blend/post-3D-LUT plus CRTC regamma, however, only
> degamma block has hardware curves. Shaper, blend, regamma predefined TFs
> are software-computed by AMD color module into PWL LUTs. In addition, we
> cannot use hardware curves on PRE_DEGAM with subsampled format, so that,
> predefined TFs are also translated to LUTs in this situation, using
> GAMCOR block instead. For this translation, the driver originally used
> the same helper for EOTFs and inverse EOTFs, even though they differ in
> input domain, number of regions and number of TF points per region.
> 
> Baring this in mind, patch 1 maps degamma predefined curves as LUT using
> GAMCOR block for AMD driver-specific property that are still in use by
> current gamescope. This was inspired by a similar patch from Harry for
> colorop [1]. Patch 2 reverts commit 8b89acc0b2ba ("drm/amd/display:
> Remove unused cm3_helper_translate_curve_to_degamma_hw_format") to
> reintroduce cm3_helper_translate_curve_to_degamma_hw_format() and patch
> 3 wire it up for encoded -> linear-light LUTs (degamma/blend). With 16
> samples per region across 12 regions for blend LUT (where hardware
> fixed-function curves are not available and predefined TFs are
> software-computed into LUTs), banding becomes almost imperceptible.
> 
> Patch 4 and 5 increase precision in the brightest half, where PQ/SRGB
> EOTFs are steeper, by enabling up to 256 samples per region and halving
> the per-region point count across 9 regions (128 in [0.5, 1], 64 in
> [0.25, 0.5], …). This better matches the shape of PQ/SRGB EOTFs.
> Although patches 4 and 5 seem conceptually correct to me, I couldn't see
> clear improvement in the bright end with or without them.
> 
> This series targets DCN3+ hw families. With this series:
> - degamma and blend LUTs use
>   cm3_helper_translate_curve_to_degamma_hw_format(): encoded input,
>   non-zero end slope, up to 256 points linearly interpolated between
>   adjacent TF pts, fitting [0,1] encoded input range.
> - shaper and regamma LUTs continue using
>   cm3_helper_translate_curve_to_hw_format(): linear-light input, zero
>   end slope, 16 points per region across 32 regions.
> 
> [1] https://lore.kernel.org/dri-devel/20260330153451.99472-8-harry.wentland@amd.com/
> 
> [v1] https://lore.kernel.org/dri-devel/20260414220237.184289-1-mwen@igalia.com/
> Changes:
> - new patch for GAMCOR usage in case of degamma predefined TF with subsampled formats
> - fix misleading information regarding degamma hw curves (Kruno)
> - clarify LUT segmentation choice using 8-bit sRGB as a reference (Kruno)
> 
> Best Regards,
> 
> Melissa

I tested this on a DCN35 device with an internal HDR panel that was affected by the gradient issue. I cannot see any banding present with this series applied on top of amd-staging-drm-next with AMD private color properties enabled.

Tested-by: Matthew Schwartz <matthew.schwartz@linux.dev>
for the series.

Thanks,
Matt

> 
> Melissa Wen (5):
>   drm/amd/display: use GAMCOR for degamma private props in subsampled
>     format
>   Revert "drm/amd/display: Remove unused
>     cm3_helper_translate_curve_to_degamma_hw_format"
>   drm/amd/display: use a separate helper to translate degamma curves
>   drm/amd/display: support up to 256 samples per region in degamma/blend
>     LUT
>   drm/amd/display: use halving distribution for PQ/sRGB linearizing LUT
> 
>  .../amd/display/amdgpu_dm/amdgpu_dm_color.c   |  16 +-
>  .../amd/display/dc/dcn30/dcn30_cm_common.c    | 184 ++++++++++++++++++
>  .../display/dc/dwb/dcn30/dcn30_cm_common.h    |   4 +
>  .../amd/display/dc/hwss/dcn32/dcn32_hwseq.c   |  10 +-
>  4 files changed, 204 insertions(+), 10 deletions(-)
>