Add another interesting profile to the test corpus

This one has three different tables, and each table has a "shoulder" at
the high end. The blue table is much shallower than red or green (which
are both fairly close to gamma 2.2).

Our current approximation code produces a good curve for green, but red
and blue are quite bad (inverse curves top out at 0.83 and 0.87). The
"best" single curve ends up being the red curve, so we end up darkening
all colors substantially.

My upcoming change to iteration/fitting strategy should improve these
results, but this really points to a failure in our current testing:
We don't have enough information about the quality of the "best single
curve" against the other two tables in our .txt files, nor about
objective quality measures of the curves (abs(f(1) - 1)). Improving the
testing should be straightforward.

We also may want to revisit separate curves per channel?

Bug: chromium:1017960
Change-Id: I6e662612d2d7d9a8b1bae9cdcc0b38eb60e87a3f
Reviewed-by: Mike Klein <>
Commit-Queue: Brian Osman <>
3 files changed