Simplify Color Space Conversion APIs

[JimBobSquarePants]

Prerequisites

• I have written a descriptive pull-request title
• I have verified that there are no overlapping pull-requests open
• I have verified that I am following the existing coding patterns and practice as demonstrated in the repository. These follow strict Stylecop rules 👮.
• I have provided test coverage for my change (where applicable)

Description

Note: I don't expect anyone to thoroughly review every file. It's far too big.

See #2531 for the original details. I've improved on the original design there by reducing the generic parameters required and removing the requirement for marker interfaces.

Replaces ColorSpaceConverter with ColorProfileConverter. The new generic API brings several advantages.

• Simplifies and normalizes usage across single and multiple input APIs.
• Makes the API extensible (No more N*) implementations.
• Fixes several conversion accuracy issues.
• Improves performance, especially around companding.

There's potential here now also for reusing the API to allow for converting between ICC profiles. We can use the information to create dynamically calculated working space inputs.

The new API is faster than the old implementation and alternative libraries.

BenchmarkDotNet v0.13.10, Windows 11 (10.0.22631.3593/23H2/2023Update/SunValley3)
11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
.NET SDK 8.0.300-preview.24203.14
  [Host]     : .NET 8.0.4 (8.0.424.16909), X64 RyuJIT AVX2
  DefaultJob : .NET 8.0.4 (8.0.424.16909), X64 RyuJIT AVX2

CieXyz => CieLab

Method	Mean	Error	StdDev	Ratio	RatioSD
'Colourful Convert'	50.85 ns	0.720 ns	0.638 ns	1.00	0.00
'ImageSharp Convert Old'	64.99 ns	0.875 ns	0.819 ns	1.28	0.02
'ImageSharp Convert'	43.62 ns	0.864 ns	1.061 ns	0.86	0.02

CieXyz => HunterLab

Method	Mean	Error	StdDev	Ratio	RatioSD
'Colourful Convert'	25.30 ns	0.517 ns	1.335 ns	1.00	0.00
'ImageSharp Convert'	13.51 ns	0.244 ns	0.217 ns	0.57	0.03

CieXyz => Lms

Method	Mean	Error	StdDev	Median	Ratio	RatioSD
'Colourful Convert'	28.653 ns	0.6023 ns	1.6487 ns	27.979 ns	1.00	0.00
'ImageSharp Convert'	8.215 ns	0.1884 ns	0.2821 ns	8.123 ns	0.29	0.02

Rgb Chromatic Adaption

Method	Mean	Error	StdDev	Ratio	RatioSD
'Colourful Adapt'	347.4 ns	6.95 ns	12.70 ns	1.00	0.00
'ImageSharp Adapt'	261.8 ns	3.35 ns	2.97 ns	0.77	0.03

[gfoidl]

No thorough review, just what jumped into my eye while looking at the first files shown in GH.

[gfoidl]

Maybe it's better to have this as properties like

Suggested change

	public static readonly CieXyz A = new(1.09850F, 1F, 0.35585F);
	public static CieXyz A => new(1.09850F, 1F, 0.35585F);

because (I didn't measure it here, so only theory / feeling):

• with the static readonly it's a memory load (potentially not from cache, but from RAM, so hard to measure in micro-benchmarks)
• with the property it's a simple assignment of three floats
• the JIT sees what's going on, so can potentially optimize the use of the three floats (which isn't that easy w/ loading the struct from memory)

Note: that would be a change in the public API.

[JimBobSquarePants]

I've changed the pattern to use the auto property with initializer. That's the pattern we've used elsewhere.

The types are immutable, and I think this approach should be best for more tightly constrained memory environments.

[gfoidl]

See the other comment about property / field. I won't comment on these later on.

[JimBobSquarePants]

I don't actually need this property. Was missed during the cleanup.

[gfoidl]

Here readonly isn't strictly needed, as the whole struct is readonly.

[gfoidl]

Suggested change

	this.L.Equals(other.L)
	&& this.A.Equals(other.A)
	&& this.B.Equals(other.B);
	new Vector3(this.L, this.A, this.B) == new Vector3(other.L, other.A, other.B);

I don't know if this is faster, but at least it's less machine code.

[JimBobSquarePants]

Yep. The machine code diff is much larger than expected and, looking at the instructions, SIMDified.

[gfoidl]

Another idea: see sharplab with the private AsVector4 property.
This overlays the vector with the struct (is OK for float, as there's no padding in between), so the creation of the vector can be avoided.

That should work for other places (Vector3) too.

[JimBobSquarePants]

Oh that's clever yeah!

[JimBobSquarePants]

Done!