NaN generated on HDR inverses #158
Description
In the image dataset there are a few generated images with NaN, I specifically identified these:
images-v2/cubes/inverse_on_input/cms/pq2000_InvOutput.Academy.P3-D65_2000nit_in_P3-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cms/pq4000_InvOutput.Academy.P3-D65_4000nit_in_P3-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq1000_InvOutput.Academy.Rec2100-D65_1000nit_in_Rec2100-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq2000_InvOutput.Academy.P3-D65_2000nit_in_P3-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq2000_InvOutput.Academy.P3-D65_2000nit_in_Rec2100-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq2000_InvOutput.Academy.Rec2100-D65_2000nit_in_Rec2100-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq4000_InvOutput.Academy.P3-D65_4000nit_in_P3-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq4000_InvOutput.Academy.P3-D65_4000nit_in_Rec2100-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq4000_InvOutput.Academy.Rec2100-D65_4000nit_in_Rec2100-D65_ST2084.exr
images-v2/cubes/inverse_on_input/cube_faces/pq500_InvOutput.Academy.Rec2100-D65_500nit_in_Rec2100-D65_ST2084.exr
This seem to be due to a division by zero happening in the inverse compression function, the way it's been dealt with in OCIO was to change the lim < 1.0001 to lim <= 1.0001 in both forward and inverse direction.
For debugging purposes, you should be able to trigger the issue using RGB [0.40898755, 0.0, 0.6531319] with InvOutput.Academy.P3-D65_4000nit_in_P3-D65_ST2084.ctl
This change also seem to significantly improve the HDR roundtrip, eg. on this case.