On 2021-11-10 06:55, Ville Syrjälä wrote: > On Wed, Nov 10, 2021 at 10:49:24AM +0200, Pekka Paalanen wrote: >> On Wed, 10 Nov 2021 00:02:16 +0200 >> Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx> wrote: >> >>> On Tue, Nov 09, 2021 at 03:47:58PM -0500, Harry Wentland wrote: >>>> On 2021-11-08 04:54, Pekka Paalanen wrote: >>>>> On Thu, 4 Nov 2021 12:27:56 -0400 >>>>> Harry Wentland <harry.wentland@xxxxxxx> wrote: >>>>> >>>>>> On 2021-11-04 04:38, Pekka Paalanen wrote: >>>>>>> On Wed, 3 Nov 2021 11:08:13 -0400 >>>>>>> Harry Wentland <harry.wentland@xxxxxxx> wrote: >>>>>>> >>>>>>>> On 2021-09-06 17:38, Uma Shankar wrote: >>>>>>>>> Existing LUT precision structure is having only 16 bit >>>>>>>>> precision. This is not enough for upcoming enhanced hardwares >>>>>>>>> and advance usecases like HDR processing. Hence added a new >>>>>>>>> structure with 32 bit precision values. >>>>>>>>> >>>>>>>>> This also defines a new structure to define color lut ranges, >>>>>>>>> along with related macro definitions and enums. This will help >>>>>>>>> describe multi segmented lut ranges in the hardware. >>>>>>>>> >>>>>>>>> Signed-off-by: Uma Shankar <uma.shankar@xxxxxxxxx> >>>>>>>>> --- >>>>>>>>> include/uapi/drm/drm_mode.h | 58 +++++++++++++++++++++++++++++++++++++ >>>>>>>>> 1 file changed, 58 insertions(+) >> >> ... >> >>>>>> If the framebuffer is not in FP16 the question then becomes how >>>>>> the integer pixel values relate to LUT addressing. >>>>> >>>>> Traditionally, and in any API I've seen (GL, Vulkan), a usual mapping >>>>> is to match minimum unsigned integer value to 0.0, and unsigned maximum >>>>> integer value to 1.0. This is how things work on the cable too, right? >>>>> (Also taking full vs. limited range video signal into account. And >>>>> conversion to cable-YUV if that happens.) >>>>> >>>>> If you want integer format FB values to map to something else, then you >>>>> have to tag the FB with that range information, somehow. New UAPI. >>>>> >>>> >>>> On the cable we send integer values, not floating point. AMD HW uses >>>> floating point internally, though, and the PWL API defines floating >>>> point entries, so on some level we need to be clear what the floating >>>> point entries mean. Either we document that to be [0.0, 1.0] or we >>>> have some UAPI to define it. I'm leaning toward the latter but have >>>> to think about it some more. >>> >>> As for Intel hw if you have an integer pixel value of 0xff... (with >>> however many bits you have with a specific pixel format) it will get >>> extended to 0.fff... (to whatever precision the pipe has internally). >>> So if we go by that a fixed point 1.0 value in the proposed >>> drm_color_lut_range would be considered just outside the gamut. And >>> pretty sure fp16 input of 1.0 should also result in a 0.fff... internal >>> value as well [1]. I think that definition pretty much matches how GL >>> UNORM<->float conversion works as well. >> >> Does it work that way in GL though? >> >> I've always thought that with GL_UNSIGNED_BYTE, 0xff maps to 1.0, not >> 255.0/256.0. >> >> Taking a random spec: OpenGL ES 2.0.25 >> >> Section 2.1.2 Data Conversions says: >> >> Normalized unsigned integers represent numbers in the range >> [0, 1]. The conversion from a normalized unsigned integer c to >> the corresponding floating-point f is defined as >> f = c / (2^b - 1) >> >> Note how the divisor has -1. > > That seems to match what I said, or at least tried to say (~0 <-> 1.0 in > float). drm_color_lut_range being fixed point would follow the ~0 side of > that. Or at least that interpretation would very easily map to our hw. > If I understand you right Intel HW represents 0xff (assuming 8 bpc) as the largest (representable) float that is less than 1.0. That float would be bigger than 255.0/256.0 but smaller than 256.0/256.0. Harry
