791422 - Support more flexibility in qcms output format

Status: RESOLVED FIXED

(Core :: Graphics: Color Management, defect)

Description

[Jeff Muizelaar [:jrmuizel]]

This will help support Chrome and should also let us output directly to a cairo compatible format.

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Attachment: Support more flexibility in qcms output format

Comment 2

[Benoit Girard (:BenWa)]

Yay for merging :)

Did a quick read-through and it looks good. I'll do a full re-through Monday and make sure it's not missing any sites.

Comment 3

[Benoit Girard (:BenWa)]

I think this is O.K. in practice but supporting this feature will limit us when chaining transformation because we can only do swizzling at the end. We can get away with this patch now because we precache the entire transformation regardless of its size. I was hoping before enabling v4 support we would do some performance testing and only precache when transforming images large enough (~33k pixels for 33^3 precache) however if we take this change doing so will be harder.

I think we have 3 options here:
1) Make the swizzling work with a variable. This will likely make the generated code worse but by how much?
2) Make a swizzle+non swizzle copy as needed. This will yield code bloat.
3) Always precache chain transforms. This will give bad performance when decoding many small images with unique profiles.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

(In reply to Benoit Girard (:BenWa) from comment #3)
> I think this is O.K. in practice but supporting this feature will limit us
> when chaining transformation because we can only do swizzling at the end. We
> can get away with this patch now because we precache the entire
> transformation regardless of its size. I was hoping before enabling v4
> support we would do some performance testing and only precache when
> transforming images large enough (~33k pixels for 33^3 precache) however if
> we take this change doing so will be harder.
> 
> I think we have 3 options here:
> 1) Make the swizzling work with a variable. This will likely make the
> generated code worse but by how much?
> 2) Make a swizzle+non swizzle copy as needed. This will yield code bloat.
> 3) Always precache chain transforms. This will give bad performance when
> decoding many small images with unique profiles.

I don't think we'll need to duplicate much of the code because afaik all of the routines do full src->dest transformations and not the individual pieces that are needed for chaining.

Comment 5

[Benoit Girard (:BenWa)]

Right. I'm fine landing this patch now and dealing with the problem once, if, we ever want to change this code.

Comment 6

[Jeff Muizelaar [:jrmuizel]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/bd16a5d51273

Comment 7

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/bd16a5d51273

Comment 8

[noel gordon]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #0)
> This will help support Chrome and should also let us output directly to a cairo compatible format.

Jeff did you mean Google Chrome?