Bug 49330 – [chromium] Canvas getImageData/putImageData cycle should not change values for valid premultiplied colours.

• Component: New Bugs
• Status: RESOLVED FIXED

[Stephen White]

[chromium] Canvas getImageData/putImageData cycle should not change values for valid premultiplied colours.

[Stephen White]

Created attachment 73516 [details]
Patch

[James Robinson]

Comment on attachment 73516 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=73516&action=review

R=me, please consider comment.

> WebCore/platform/graphics/skia/ImageBufferSkia.cpp:241
> +// This function does the equivalent of (a * b + 254) / 255, without an integer divide.

Is there any reason to avoid an integer divide?  I think it'd be a lot easier to read that way and the compiler can figure out the best code to generate for a given platform.

[Stephen White]

(In reply to comment #2)
> (From update of attachment 73516 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=73516&action=review
> 
> R=me, please consider comment.
> 
> > WebCore/platform/graphics/skia/ImageBufferSkia.cpp:241
> > +// This function does the equivalent of (a * b + 254) / 255, without an integer divide.
> 
> Is there any reason to avoid an integer divide?  I think it'd be a lot easier to read that way and the compiler can figure out the best code to generate for a given platform.

Sadly, divides are still slow on modern CPUs.  And since this is integer, one can't prerecompute the reciprocal and do a multiply, as you can in FP.

This trick is used a lot in skia (see SkMulDiv255Round).  The compiler can't do this optimization, since it's only valid for a restricted set of inputs, which the compiler doesn't know.

[Stephen White]

Committed r71760: <http://trac.webkit.org/changeset/71760>

[WebKit Review Bot]

http://trac.webkit.org/changeset/71760 might have broken GTK Linux 32-bit Release