Bug 160520
| Summary: | Add support for ColorSpaces and deep color to Color.h | ||
|---|---|---|---|
| Product: | WebKit | Reporter: | Dean Jackson <dino> |
| Component: | Layout and Rendering | Assignee: | Nobody <webkit-unassigned> |
| Status: | NEW | ||
| Severity: | Normal | CC: | simon.fraser |
| Priority: | P2 | Keywords: | InRadar |
| Version: | WebKit Nightly Build | ||
| Hardware: | Unspecified | ||
| OS: | Unspecified | ||
| Bug Depends on: | 161853, 161855, 161856 | ||
| Bug Blocks: | |||
Dean Jackson
For extended and deep colors, we'll need to update Color.h.
We'll need a reference to a ColorSpace and probably floating point component values.
| Attachments | ||
|---|---|---|
| Add attachment proposed patch, testcase, etc. |
Dean Jackson
<rdar://problem/27667095>
Dean Jackson
8/2/16, 4:17 PM Dean Jackson:
Doing some allocation tracking it seems an average page (macrumors daringfireball) has on the order of 3000 - 6000 colors.
8/2/16, 4:19 PM Dean Jackson:
Lea's amazing http://leaverou.github.io/conic-gradient/ got to about 8200 live color instances.
8/2/16, 4:25 PM Dean Jackson:
Although, that number would probably be smaller if we implemented conic-gradient :)
The single page HTML5 got to about 300,000 colors at its highest point, so an extra 64bit pointer would be about 2.25 Mb of extra memory (double what the current Color is).
8/3/16, 11:14 AM Dean Jackson:
If we start counting unique colors....
daringfireball.net 22
html5 spec 106 (out of 350k creations)
macrumors 76
apple.com/iphone 26
BBC 68
etc
8/3/16, 1:18 PM Dean Jackson:
With discussion from Sam:
Current
- 64 bits
(32 for inline color, 1 bit for validity)
- Inline access
Double Size
- 128 bits
(32 for inline color, 1 bit for validity, 64 bits for pointer to extended data)
- Direct access
Con:
- Double the size
~2MB extra on HTML spec test
Atomic Color (the color of tomorrow)
- 64 bits
(64 bits for pointer to uniqued color)
- Indirect access
Con:
- Indirect access might be a perf hit
Tagged pointer
- 64 bits
(same as current, except you have a bit which tells you if this is a pointer to the extended data)
- Direct access after branch
Con:
- Adds a branch to all color access
- Might not work with current tools
8/3/16, 1:18 PM Dean Jackson:
Dean Jackson
1. Color will get rid of m_isValid and m_color, and use RefPtr<ColorImpl>.
2. isValid() becomes m_colorImpl != nullptr
3. ColorImpl will start with just the RGBA struct (add space and deeper colors later)
4. Table will live in Color.cpp
Simon Fraser (smfr)
Isn't this mostly done?