293353 - Canvas - alpha transparency renders wrong colors

Status: RESOLVED FIXED

(Core :: Graphics: Canvas2D, defect)

Description

[Peter van der Woude [:Peter6]]

Mozilla/5.0 (Windows; U; Windows NT 5.0; en-US; rv:1.8b2) Gecko/20050508
Firefox/1.0+

See testcase
left to right is red - green - blue
top to bottom transparency 1 to 0 in 0.1 steps

related to bug 293304

testcase next

Comment 1

[Peter van der Woude [:Peter6]]

Attachment: testcase

Comment 2

[Nickolay_Ponomarev]

Another testcase I posted on IRC:
http://mozilla.doslash.org/stuff/canvas/test.xhtml
http://mozilla.doslash.org/stuff/canvas/test.png

* rgba(0, 0, 255, 0.5) is like rgba(127, 0, 127, 1), only lighter [because of
the white bgr]
* same for rgba(0, 255, 0, 0.5) / rgba(127, 127, 0, 1)
* rgba(255, 0, 0, 0.5) == rgba(0, 255, 0, 0.5)

Comment 3

[Nickolay_Ponomarev]

So the root problem is that the code above on black bgr draws bright red square.
I.e. if fille style is partially transparent, 255 is added to the actual (BGR)
color before blending(?). This explains all of the above problems.
 ctx.fillStyle = "rgba(0, 0, 0, 0.1)";
 ctx.fillRect(0, 0, 100, 100);

Comment 4

[Peter van der Woude [:Peter6]]

Attachment: screensot before/after Bug 294134 landing

Bug 294134 causes the transparent colors to render worse

Comment 5

[Stuart Parmenter]

Attachment: What it should look like...

This is what it should look like and what it does look like for me on windows
xp...

Comment 6

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

There is certainly something odd going on; what platform and what builds are
seeing the wrong colors?  If windows and a nightly build, can you post your
display settings? (video card, driver version, display depth)  It's a long shot,
but both stuart and I are having a hard time reproducing this...

Comment 7

[Peter van der Woude [:Peter6]]

(In reply to comment #6)
> There is certainly something odd going on; what platform and what builds are
> seeing the wrong colors?  If windows and a nightly build, can you post your
> display settings? (video card, driver version, display depth)  It's a long shot,
> but both stuart and I are having a hard time reproducing this...

Win2k (3 different machines)
Beast hourly builds
Geforce G200 32-bit color and Geforce G400 32-bit color
default Windows driver/Geforce driver/Nvidia driver (all tried)

Comment 8

[Steve England [:stevee]]

Mozilla/5.0 (Windows; U; Windows NT 5.0; en-US; rv:1.8b2) Gecko/20050514
Firefox/1.0+ ID:2005051410

For me, the testcase renders the same as the right hand side of 'screensot
before/after Bug 294134 landing'

GeForce4 Ti4400, 1280x1024x32 @ 85Hz, Forceware v76.41

Comment 9

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Ah ha, I can reproduce this now with the beast build.  Will take a look tonight.

Comment 10

[Dave Townsend [:mossop]]

I see the result from the right of the first screenshot attachment.

nVidia GeForce4 MX.

Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv:1.8b2) Gecko/20050514
Firefox/1.0+ ID:2005051409

Comment 11

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

I just did a local build with the same buildconfig that the nightlies/beast
build uses, and I don't see the wrong colors with this build.  I've zipped it up
at http://www.vlad1.com/~vladimir/misc/firefox-1.0+-vlad.zip -- can those of you
who are seeing the problem try this build and let me know what the test looks like?

The only difference that I'm aware of between my build and the nightly builds is
that mine is built with VC7.1, whereas the nightlies are VC6... that scares me a
little.

Comment 12

[Steve England [:stevee]]

With vlad's build from comment 11, the testcase now renders properly, in that it
looks like https://bugzilla.mozilla.org/attachment.cgi?id=183591

Comment 13

[Peter van der Woude [:Peter6]]

(In reply to comment #11)
> I just did a local build with the same buildconfig that the nightlies/beast
> build uses, and I don't see the wrong colors with this build.  I've zipped it up
> at http://www.vlad1.com/~vladimir/misc/firefox-1.0+-vlad.zip -- can those of you
> who are seeing the problem try this build and let me know what the test looks
like?
> 
> The only difference that I'm aware of between my build and the nightly builds is
> that mine is built with VC7.1, whereas the nightlies are VC6... that scares me a
> little.
> 

your build WFM Vlad

Comment 14

[Peter van der Woude [:Peter6]]

I asked on forums if all windows user sees the wrong rendering.
Answer, yes, confirmed problem on ME/98/2k/XP

Comment 15

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Ohboy, this will be exciting.  I'll talk to pav and chase tomorrow, see if we
can figure out what's going on.

Comment 16

[Mike Shaver (:shaver emeritus)]

Attachment: screenshot from OS X nightly (may 15)

So much is wrong here!

Comment 17

[Peter van der Woude [:Peter6]]

Vlad,

Bug 293226
Bug 293259
Bug 293304
Bug 293382

are also the result of this

Comment 18

[Peter van der Woude [:Peter6]]

pressed enter to soon.
The 4 bugs above look fine in your build but bad on beast builds

Comment 19

[Mike Shaver (:shaver emeritus)]

This has gotta block 1.1, at least the complete disaster that is the OS X result.

Comment 20

[Peter van der Woude [:Peter6]]

(In reply to comment #16)
> Created an attachment (id=183660) [edit]
> screenshot from OS X nightly (may 15)
> 
> So much is wrong here!

this may well have a different cause / need a different bug (Mac specific)

Comment 21

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Mac bug is a different issue (I've got a patch for that).

Comment 22

[Sjoerd Visscher]

What's the status on this bug? It this going to be fixed in time for Firefox 1.1?

Comment 23

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

It will be fixed for 1.1; I just confirmed that the issue is still there since
the 0.5.0 update, so I'll be trying to figure out what's going on with VC6 this
week.

Comment 24

[(not currently active) Ted Mielczarek]

*** Bug 293304 has been marked as a duplicate of this bug. ***

Comment 25

[(not currently active) Ted Mielczarek]

*** Bug 296252 has been marked as a duplicate of this bug. ***

Comment 26

[(not currently active) Ted Mielczarek]

*** Bug 293226 has been marked as a duplicate of this bug. ***

Comment 27

[(not currently active) Ted Mielczarek]

*** Bug 293382 has been marked as a duplicate of this bug. ***

Comment 28

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

So; VC6 miscompiles more than a few bits in pixman, in both debug and opt
builds.  There are a lot of macros in use of the form `q = Foo(x,y,a,t);', which
expand out to something like `q = ((t) = ...., (t) = ... t ..., .... t ....;',
with a good amount of shifts and multiple-type casts (e.g. `(uint16_t)
(uint8_t)') happening.  I was able to fix IcOver() by using a few extra
temporary variables, but fixing that just revealed other problems.

I'm not really sure what a solution here is for 1.1/1.8b3; a sufficiently gross
but perhaps workable solution would be to use VC7.1 to generate assembly, and
only assemble with VC6.  The pixman code itself has no external dependencies, so
this shouldn't be difficult, if that ends up being the right way to go.

Comment 29

[Steve England [:stevee]]

Mozilla/5.0 (Windows; U; Windows NT 5.0; en-US; rv:1.8b2) Gecko/20050706
Firefox/1.0+ ID:2005070602
This now WFM; since the landing of bug 299407

Comment 30

[Anne (:annevk)]

Per comment 28 and comment 29 this bug is fixed by the patch for bug 299407.

Comment 31

[(not currently active) Ted Mielczarek]

*** Bug 306755 has been marked as a duplicate of this bug. ***

Comment 32

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

*** Bug 306755 has been marked as a duplicate of this bug. ***