Feature Request: HDR support!

[AnyOldName3]

Quote:finally asked someone knowledgeable

I'm going to go with them being wrong, too. It can happen. Regarding tone mapping, all of that does apply, and you're right about it potentially making banding worse. It's only really good when you've got a wide bit depth (for example you've just rendered a thing to a floating point render target so have 32-bits of data per channel, which is what I mentioned several posts ago) so it won't make Dolphin look better. The steps are closer together in most HDR standards, though, and while some might exist that directly include the 2^24 SRGB colours as a contiguous region, it isn't a requirement.

A big wall of text will follow and some of it may not be especially well-written, so in case you want to avoid having to read it, I'm going to include a link to an article that also agrees with me but probably makes more sense: https://www.tomshardware.co.uk/what-is-10-bit-color,news-58382.html

The parts of the Wikipedia page I was referring to include, but are not limited to:

Quote:Rec. 2020 defines a bit depth of either 10-bits per sample or 12-bits per sample.

i.e. it doesn't define a specific bit depth that you have to use, instead letting you choose between two options. The paragraphs immediately following this don't say that 12-bit covers a wider colour space, so I read it to mean that they both cover the same colour space.

Quote:In coverage of the CIE 1931 color space the Rec. 2020 color space covers 75.8%, [...] and the Rec. 709 color space covers 35.9%.[6]

Two things here:
1. We don't see different CIE 1931 coverage values for 10-bit and 12-bit representations, but we would if 12-bit had a wider range.
2. Less importantly, and technically this doesn't prove my point, but it does suggest it: CIE colour spaces tend to aim to be perceptually uniform (and then turn out not to be), and SRGB also aims to be perceptually uniform (but isn't), and it would be very silly for Rec 2020 implementations to not be roughly perceptually uniform and all the different attempts at perceptual uniformity turn this into very approximate napkin maths. Anyway, if Rec 2020 covers 75.8% of the volume* of CIE 1931 and SRGB/Rec 709 35.9%, assuming everything was perceptually uniform, the doubling of the volume covered would require just one extra bit per sample to keep the same step width. 12-bit Rec 2020 instead adds four bits, multiplying the potential volume coverage by 16x. Either Rec 2020 implementations are putting all those new colours in the new regions, allowing tiny steps between colours that burn your eyes out or that you can't see in a well-lit room but traditional chunky steps between normal colours, or the step size between Rec 709 colours has shrunk.

* It's volume, even though the diagrams you usually see are 2D pictures of an area because usually brightness is left out.

Quote:The NHK measured contrast sensitivity for the Rec. 2020 color space [...] 11-bits per sample for the Rec. 2020 color space is below the visual modulation threshold [...] for the entire luminance range.

This bit doesn't even make sense if you don't interpret it as the NHK trying to work out how many bits they need to make the step size so small you can't see it as your interpretation has the colour space size and bit depth being dependent on each other.

Quote:Transfer characteristics [...] [MATHS]

The maths for the transfer function is the same for 10-bit and 12-bit (although they allow different minimum precisions for the constants) and the values are normalised relative to the camera-input light intensity, so range from 0 to 1 regardless of whether there are 1024 values in the range or 4096. This is literally the maths that defines the step size and distribution and it doesn't have any special handling for ensuring that the SRGB colours remain contiguous. If Rec 2020 had two different colour spaces explicitly declared for 10-bit and 12-bit representations, this could be made to work as you describe, but then this would be obvious from the rest of the page.



One final non-Wikipedia point. The Adobe RGB colour space is by many definitions an HDR standard because it offers a wider range than Rec 709. However, it still only uses 8 bits per sample, so has wider steps than SRGB which uses the same number of steps to cover a smaller range. That right there is an example of an HDR standard which is actually used for things and has a different step with in the SRGB range than SRGB does.