1124060 - Premultiplication destroys some alpha images in web content use cases

Status: NEW

(Core :: Graphics: ImageLib, defect, P3)

Description

[noel gordon]

Attachment: dataURL-alpha-instability.html

User Agent: Mozilla/5.0 (Windows NT 6.1; WOW64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/39.0.2171.95 Safari/537.36

Steps to reproduce:

For example, encoding <canvas> content via toDataURL as image/png (lossless) into a new Image(), where the canvas contains alpha values near 0.5, and then drawing that Image() onto a new canvas with drawImage(), and repeating this process over and over, should be idempotent.  For some alpha values, the image is slowly destroyed.

Test case attached: for Firefox, set alpha=0.498 for example, and keep pressing the test button.


Actual results:

The canvas image slowly fades to gray with each test button press.

Other browsers:
  IE9+: works.
  Chrome 39 stable: broken.
  Firefox 33.0.3: broken.
  Safari 6: broken.
  Safari 7: works.



Expected results:

The canvas image should not fade away for any alpha in the neigbourhood of 0.5.

Comment 1

[noel gordon]

Attachment: firefox-33.0.3.png

Test-case result for alpha=0.498 on Firefox 33.0.3

Comment 2

[noel gordon]

Chrome bug https://code.google.com/p/chromium/issues/detail?id=450189

Comment 3

[:Gijs (he/him)]

Seth, can you help place/prioritize/find-a-someone for this? Not sure if this is canvas, imagelib, gfx, or somewhere in the middle. :-)

Comment 4

[Seth Fowler [:seth] [:s2h]]

This probably exists at the intersection of canvas and ImageLib, which is good news since I'm about to start focusing on canvas work once I get the recent ImageLib regressions cleaned up.

I'll come back to this once I switch over to canvas.

Comment 5

[noel gordon]

Fixed in chromium: https://code.google.com/p/chromium/issues/detail?id=450189

Comment 6

[noel gordon]

I believe that Gecko uses (or once used) a table-driven method to compute pixel alpha pre-multiplication. Perhaps the table data was computed from a truncating pre-multiply routine?

Perhaps try a rounding one.  That was the problem in Chrome a least - a truncating premultiply in the image decoders. I changed the decoders to use a rounding pre-multiply and the alpha channel instability was gone.

HTH.

Comment 7

[Bas Schouten (:bas.schouten)]

I actually suspect this is more likely to be in the Canvas code. We should probably just fix this.

Comment 8

[Jonathan Watt [:jwatt]]

Mass bug change to replace various 'parity' whiteboard flags with the new canonical keywords. (See bug 1443764 comment 13.)

Comment 9

[BugBot [:suhaib / :marco/ :calixte]]

Clear a needinfo that is pending on an inactive user.

Inactive users most likely will not respond; if the missing information is essential and cannot be collected another way, the bug maybe should be closed as INCOMPLETE.

For more information, please visit auto_nag documentation.