ui/gfx: optimize UnPreMultiply operation by checking alpha value

ui/gfx: optimize UnPreMultiply operation by checking alpha value

Don't need to do Un-premultiplying operation,
if alpha value equals to 0 or 255.

BUG=none
Test(DONE)=gfx_unittests --gtest_filter="ColorAnalysisTest*"

Committed: https://src.chromium.org/viewvc/chrome?view=rev&revision=284166

[hyunki, 2014-07-17 10:05:41]

Many cases alpha value would be 0 or 255. It needs to be optimized.
Please take a look. Thanks. :)

[hyunki, 2014-07-17 16:51:03]

hello sky@: gentle ping. PTAL. thank you!

[Alexei Svitkine (slow), 2014-07-17 18:32:43]

If this is truly a performance win, shouldn't this be done internally by
SkUnPreMultiply::PMColorToColor(). +reed for his thoughts.

By the way, is there a benchmark or something that shows that this helps
performance or is this speculative?

[hyunki, 2014-07-17 18:59:50]

Thanks for the comment.

> 
> By the way, is there a benchmark or something that shows that this helps
> performance or is this speculative?

I've found 3-4 similar alpha checks.
https://code.google.com/p/chromium/codesearch#search/&q=%22alpha%20!=%200%20&...

It depends on alpha value of given bitmap. I'll investigate it by checking
websites such as wsj.com and nytimes.com.

[reed1, 2014-07-17 20:33:25]

This sort of change highlights a slightly fragile assumption, that the swizzle
order of SkPMColor and SkColor are the same... which is not always true. The
swizzle order if in control of chrome (or whom ever controls SkUserConfig.h) but
just a note that this decision would impact this sort of code.

[hyunki, 2014-07-18 12:32:53]

For 10000 times unittest for UnPreMultiply() on my desktop Intel(R) Core(TM)
i5-4670 CPU @ 3.40GHz when I set alpha to 255 to check fast path,
results are as belows.

- without patch: 6.4ms
- with patch: 2ms
- with patch but not taking fast path: 6.5ms

In summary, 
When alpha value equals to 0 or 255, this idea is 3.2 times faster
and even when alpha value is not 0 and 255, the performance is almost same.

Thus, I think this idea can be beneficial.
Any thoughts?

[hyunki, 2014-07-18 16:00:48]

Alexei Svitkine@: Gentle ping, PTAL. Thanks.

[Alexei Svitkine (slow), 2014-07-18 16:25:12]

lgtm

[hyunki, 2014-07-18 16:30:40]

thanks for the review.

[hyunki, 2014-07-18 16:30:46]

The CQ bit was checked by hyunki.baik@samsung.com

[commit-bot: I haz the power, 2014-07-18 16:31:38]

CQ is trying da patch. Follow status at
 https://chromium-status.appspot.com/cq/hyunki.baik@samsung.com/399053002/1

[commit-bot: I haz the power, 2014-07-18 18:44:36]

Change committed as 284166

[reed1, 2015-05-12 13:25:46]

fwiw -- the new version is potentially fragile, as the swizzle format for
SkPMColor (which is what is stored in the bitmap) is not required to match the
swizzle format of SkColor (the output of PMUnPremultiply).