Adsk Contrib - Issues #1968 (mirrored) and #1992 (LUT-free builtins)

[cozdas]

For issue #1992 - "Make LUT-free implementations of certain built-in transforms"

• Adding fixed-function FIXED_FUNCTION_LIN_TO_PQ. This will be used to handle following built-in transforms if the LUT support is turned off:
  -- CURVE - ST-2084_to_LINEAR
  -- CURVE - LINEAR_to_ST-2084
  -- DISPLAY - CIE-XYZ-D65_to_REC.2100-PQ
  -- DISPLAY - CIE-XYZ-D65_to_ST2084-P3-D65

• Adding fixed function FIXED_FUNCTION_LIN_TO_GAMMA_LOG. This is a parametrized gamma + log segment curve with adjustable mirroring point. This will be used to handle following built-in transform if the LUT support is turned off:
  --CURVE - HLG-OETF
  --CURVE - HLG-OETF-INVERSE
  -- DISPLAY - CIE-XYZ-D65_to_REC.2100-HLG-1000nit
  -- CURVE - APPLE_LOG_to_LINEAR
  -- APPLE_LOG_to_ACES2065-1

• Adding fixed function FIXED_FUNCTION_LIN_TO_DOUBLE_LOG which has two log-affine segments with an optional linear segment in-between. This will be used to handle following built-in transform if the LUT support is turned off:
  -- CANON_CLOG2-CGAMUT_to_ACES2065-1
  -- CURVE - CANON_CLOG2_to_LINEAR
  -- CANON_CLOG3-CGAMUT_to_ACES2065-1
  -- CURVE - CANON_CLOG3_to_LINEAR

• Implemented CPU renderers for scalar, SSE2 (with fastPower) and SSE2 (with Intel SVML) intrinsics targets for the new PQ fixed-function. The other two fixed-functions are implemented in the basic form and will need more optimization in the future when we introduce the option to turn the LUT support off.

• Implemented GPU shader generator for all the new fixed-functions,

For issue #1968 - "Make display EOTFs in built-in transforms mirrored and unclamped to preserve sub-black and super-white"

• Both the existing LUT-based and the new LUT-free implementations using PQ or HLG curves are now unclamped at both ends and mirrored around the origin.
• Added new built-in transforms:
  --CURVE - HLG-OETF
  --CURVE - HLG-OETF-INVERSE

Aux Changes

• GetFixedFunctionCPURenderer() now takes a new bool parameter fastLogExpPow.
• Added a stand-in OCIO_LUT_SUPPORT preprocessor macro in preparation for ocio-lite where the lut support can be turned off.
• Added util functions to Config_tests.cpp to help creating configs with arbitrary version.
• Added ability to test fixed-function cpu renderers with and without fastLogExpPow.
• Added capability to specify custom extended ranges in the GPU unit tests. Default value is [-1.0,2.0] same as the previously hard-coded range.
• Fixed a bug in the GPU unit tests where the computed domain values would overshoot.
• In BuiltinTransform_tests it's now possible (and necessary) to specify the error threshold for each built-in transform. This was needed as the hard-coded 1e-6 threshold was too tight for the PQ curve which uses nested power functions and thus numerical stability was depending on the implementation.
• Added GpuShaderText::floatXGreaterThanEqual() functions

[cozdas]

adding ability to specify different thresholds for different built-ins as some are more sensitive than others to the underlying implementation change (i.e. lut-based vs scalar vs SSE implementation). Increased the threshold for PQ curve which has nested power() invocation thus numerically less stable. Also adding extended range values to make sure that all implementations can handle wider than [0,1] domain. Target values are computed with double-precision math.

[remia]

Great work @cozdas and @doug-walker!

Do we need to update the CTF minimum version as well?

[doug-walker]

@remia , yes absolutely right about the CTF version. I saw you bumped the version in your ACES PR, so my plan is to merge that first to update the CTF version to 2.4 and then we will add the corresponding tests to this PR. There will be some integration work as well in FixedFunction.

[doug-walker]

@remia , the merge of the ACES2 PR into this branch was fairly complicated, so I would appreciate it if you could make another pass over this one. And the CTF version check was added.

[remia]

@remia , the merge of the ACES2 PR into this branch was fairly complicated, so I would appreciate it if you could make another pass over this one. And the CTF version check was added.

Looks ok to me!

[zachlewis]

Man, this is a lot of work... Cuneyt, you're a maniac. This is awesome.

Looks like you've taken care of #1705 here too, that's cool to see.

I think these "Double Log" and "Gamma Log" FixedFunctions and Builtins are excellent additions, and are quite reasonably named. Having LUT-free alternatives for some of these builtins makes it a lot easier to sling representations of processors around, both in memory and on printed paper.

For the future, I think we need to figure out a better way of documenting FixedFunction parameters.

Fantastic work.