821450 - Add dithering support for opaque images

Status: NEW

(Core :: Graphics: ImageLib, defect, P5)

Description

[Jeff Muizelaar [:jrmuizel]]

This shouldn't be too hard and might be worth doing. There will of course be a performance impact, but might not be too bad if we do during conversion to 16bit surfaces.

Comment 1

[Brad Lassey [:blassey] (use needinfo?)]

Jeff, can you comment as to how much this will help and if we should track it for release?

Comment 2

[Jeff Muizelaar [:jrmuizel]]

(In reply to Brad Lassey [:blassey] from comment #1)
> Jeff, can you comment as to how much this will help and if we should track
> it for release?

It's hard to say how much it will help, without specific problems that we want to address. I would suggest not tracking it for release.

Comment 3

[Jeff Muizelaar [:jrmuizel]]

Attachment: Sketch of what this could look like

Comment 4

[Jeff Muizelaar [:jrmuizel]]

Attachment: Working version

Comment 5

[Brad Lassey [:blassey] (use needinfo?)]

qawanted to compare the results from this with the problem images they've identified

Comment 6

[Kevin Brosnan [Ex-Mozilla]]

Pushed this as https://tbpl.mozilla.org/?tree=Try&rev=68560043a297 now to find a few representative images.

Comment 7

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

We have a really badly banded image in prime real estate in the b2g UI.  I'll grab before/after screenshots as soon as I can.

(We're separately working around this with <canvas> for v1.)

Comment 8

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

Attachment: left: without, right: with

I don't see a difference.  Maybe we have to stick this somewhere else in the pipeline too?

Comment 9

[Jeff Muizelaar [:jrmuizel]]

(In reply to Chris Jones [:cjones] [:warhammer] from comment #8)
> Created attachment 703764 [details]
> left: without, right: with
> 
> I don't see a difference.  Maybe we have to stick this somewhere else in the
> pipeline too?

The banding on this comes from compositing the gradient mask and the wallpaper together. To fix that, I think we'd need a fully dithered pipeline which is unlikely to be performant enough.

Comment 10

[Jeff Muizelaar [:jrmuizel]]

(In reply to Chris Jones [:cjones] [:warhammer] from comment #8)
> Created attachment 703764 [details]
> left: without, right: with
> 
> I don't see a difference.  Maybe we have to stick this somewhere else in the
> pipeline too?

I was also talking with Patryk and Peter about this particular case and it sounds like we may just solve it by using a solid color dim instead of a gradient mask.

Comment 11

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

We have a <canvas> hack that gets the job done.

Comment 12

[Jeff Muizelaar [:jrmuizel]]

(In reply to Chris Jones [:cjones] [:warhammer] from comment #11)
> We have a <canvas> hack that gets the job done.

That comes at the cost of having to blend the 'opaque' canvas with what is underneath it. Probably not a big a deal, but things like this add up.

Comment 13

[Kevin Brosnan [Ex-Mozilla]]

Pushed this as https://tbpl.mozilla.org/?tree=Try&rev=68560043a297 now to find a few representative images.

Comment 14

[Chris Jones [:cjones] inactive; ni?/f?/r? if you need me]

Important use case for content; content doesn't seem to generally ship dithered resources so they look bad on low-end screens.

Comment 15

[Paul Silaghi, QA [:pauly]]

Dropping qawanted since the request was accomplished in comment 8.

Comment 16

[The 8472]

Attachment: gradients.png

Would this also fix the rendering of this testcase? The middle column should look like the right one.

Comment 17

[Jeff Muizelaar [:jrmuizel]]

(In reply to The 8472 from comment #16)
> Created attachment 8503355 [details]
> gradients.png
> 
> Would this also fix the rendering of this testcase? The middle column should
> look like the right one.

What is the test case from?

Comment 18

[The 8472]

I generated it with imagemagick. It's a 16bit PNG with gradients between shades of green.

The first column is with color steps truncated to 8bit precision, the second column with full 16bit precision and the 3rd column with the same 16bit gradient dithered to 8bit.

Assuming an 8bit output device and no dithering the 2nd column will look like the 1st column, i.e. banded
But ideally the 2nd column should look like the 3rd column.

Comment 19

[Brad Lassey [:blassey] (use needinfo?)]

filter on [mass-p5]