The current code for this filter premuls, then clamps. We should be
clamping, then premulling.
If the matrix makes alpha greater than 1, these two orderings can result
in different color values. Alpha will clamp to 1 either way, but the
color channels are multiplied by that >1 alpha in one case, and by =1
in the other.
The left column of the gm imagefilterscropexpand demonstrates this.
Its matrix adds 32/255 to alpha and 255/255 to green. This produces
alpha ~= 1.12. That's then multiplied by the relatively small red
and blue values in the grey checkerboard, resulting in different
in-range values than the ones we would have gotten if we clamped
alpha first. Green wasn't affected because it was already fully
saturated. 255 * 1.12 == 255 no matter when we clamp.
Change-Id: I4b30bf64c30fe62526674ad5f32e9ca19ec84714
Reviewed-on: https://skia-review.googlesource.com/77902
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
