Add partial and optional support for OCIO to MaterialX

[JGamache-autodesk]

The default color management system in MaterialX has support for a very limited set of color spaces. Some commonly used spaces such as linear Rec.2020 or sRGB-encoded AP1 are not available. MaterialX is often used in applications that already use OCIO for color management and so artists are sometimes confused as to why the color spaces in a menu that is available for textures in another part of the application are not available for textures in MaterialX. In some cases, a color space may even be mathematically the same as one in the MaterialX set, but it is simply named differently. This may lead to the artist picking the wrong color space and wasting time. In situations like this, the application is able to offer a better user experience by offloading the interpretation of color space names to OCIO. This allows artists to use the same color spaces with all textures rather than having a different set for MaterialX.
However, not all OCIO color spaces are appropriate for use in a MaterialX document and this PR does not offer support for full OCIO. Supported color spaces include any of those that are mathematically compatible with one of the default MaterialX spaces.
This PR does not change where the implicit color conversions happen in a MaterialX shader, and it does not provide any additional functionality to creators of MaterialX documents, other than expanding the set of allowed color space names. The MaterialX API already allows a client to override the default mx::ColorManagementSystem, this PR illustrates how that may be done in order to take advantage of a subset of OCIO.
It should be noted that only minimal changes would be needed to adapt this PR for use with either OCIO or the nanoColor library currently under development.
This integrates the OCIO shadergen into the MaterialX shadergen to provide on-the-fly color management for OpenGL, Vulkan, Metal, and OSL.
Targeting 1.39.1. Definitely not 1.39.0.
Tested with GLSL and OSL on Linux.

[ld-kerley]

This looks like really great work - when I get some cycles I'd be happy to kick the tires on the MSL side (might be next week at this point). I left a few observations below.

[ld-kerley]

Would be nice if we could keep this version in sync somehow with

const string MslShaderGenerator::VERSION = "2.3";

from here

[ld-kerley]

You should be able to use ${TARGET_NAME} to collect the correct monolithic/non-monolithic target to use.

It might also be cleaner to append the OCIO library to the target that way too, so then the entire OCIO addition here could just be added inside one guard

if(MATERIALX_BUILD_OCIO)
    find_package(OpenColorIO REQUIRED)
    target_link_libraries(${TARGET_NAME} OpenColorIO::OpenColorIO)
    target_compile_definitions(${TARGET_NAME} PUBLIC MATERIALX_BUILD_OCIO)
endif()

[ld-kerley]

I'm not an OCIO expert - but does this pull some pre-baked config from inside OCIO. Will people need to provide custom OCIO configs at some point?

[doug-walker]

@ld-kerley , the PR supports use of arbitrary OCIO configs, but to keep things simple for the purposes of this unit test, Jerry is using one of the configs that's built into the OCIO library itself.

In the real world, apps could create the config using OCIO::Config::CreateFromFile(path) rather than CreateFromBuiltinConfig.

[JGamache-autodesk]

Well, OSL was not as trivial as expected, but still manageable. Results are good:

And we can see the OCIO code being injected in GLSL:

vec4 ocio_color_conversion_8182d4da8df3d64c153c5e4252dbbf(vec4 inPixel)
{
  vec4 outColor = inPixel;
  
  // Add Gamma 'monCurveFwd' processing
  
  {
    vec4 breakPnt = vec4(0.0392857157, 0.0392857157, 0.0392857157, 1.);
    vec4 slope = vec4(0.077380158, 0.077380158, 0.077380158, 1.);
    vec4 scale = vec4(0.947867274, 0.947867274, 0.947867274, 0.999998987);
    vec4 offset = vec4(0.0521326996, 0.0521326996, 0.0521326996, 9.99998974e-07);
    vec4 gamma = vec4(2.4000001, 2.4000001, 2.4000001, 1.00000095);
    vec4 isAboveBreak = vec4(greaterThan( outColor, breakPnt));
    vec4 linSeg = outColor * slope;
    vec4 powSeg = pow( max( vec4(0., 0., 0., 0.), scale * outColor + offset), gamma);
    vec4 res = isAboveBreak * powSeg + ( vec4(1., 1., 1., 1.) - isAboveBreak ) * linSeg;
    outColor.rgb = vec3(res.x, res.y, res.z);
    outColor.a = res.w;
  }

  return outColor;
}

and in the OSL code:

color4 ocio_color_conversion_8182d4da8df3d64c153c5e4252dbbf(color4 inPixel)
{
  color4 outColor = inPixel;
  
  // Add Gamma 'monCurveFwd' processing
  
  {
    vector4 breakPnt = vector4(0.0392857157, 0.0392857157, 0.0392857157, 1.);
    vector4 slope = vector4(0.077380158, 0.077380158, 0.077380158, 1.);
    vector4 scale = vector4(0.947867274, 0.947867274, 0.947867274, 0.999998987);
    vector4 offset = vector4(0.0521326996, 0.0521326996, 0.0521326996, 9.99998974e-07);
    vector4 gamma = vector4(2.4000001, 2.4000001, 2.4000001, 1.00000095);
    vector4 isAboveBreak = vector4((outColor.rgb.r > breakPnt.x) ? 1.0 : 0.0, (outColor.rgb.g > breakPnt.y) ? 1.0 : 0.0, (outColor.rgb.b > breakPnt.z) ? 1.0 : 0.0, (outColor.a > breakPnt.w) ? 1.0 : 0.0);
    vector4 linSeg = outColor * slope;
    vector4 powSeg = pow( max( vector4(0., 0., 0., 0.), scale * outColor + offset), gamma);
    vector4 res = isAboveBreak * powSeg + ( vector4(1., 1., 1., 1.) - isAboveBreak ) * linSeg;
    outColor.rgb = vector(res.x, res.y, res.z);
    outColor.a = res.w;
  }

  return outColor;
}

[kwokcb]

It's quite possible that not all shading languages and code generation variants will be supported as this restricted to code generation built into OCIO and does not produce a shading language agnostic nodegraph. (*)(**)

A suggestion is to inherit OCIO from the base color transform and test for OCIO first and if it fails use the built in one (or check built in support first maybe better?).

Basically generation from an "OCIO" config will be called if it's not a supported transform in the "default". This would match more closely how targets and fallbacks work for code generation. I think this would also match now nanocolor is being proposed to be integrated for MaterialX -- i.e. if not supported by nanocolor then full OCIO can be used.

(*) The "default" color transforms in OCIO 2.0 are a superset of what's in MaterialX currently, but MaterialX has a super set of shading languages pre-supported with other variants possible.)
(**) I'm not sure whatever became of the discussion of OCIO providing operators which could map to MaterialX nodes but this would avoid "chasing" shading language support.

Just my 2 cents. :).