207146 - Use GDI+ on Win32 to speed up -moz-opacity

Status: RESOLVED WONTFIX

(Core Graveyard :: GFX: Win32, defect)

Description

[Kevin McCluskey (gone)]

Comment 1

[Kevin McCluskey (gone)]

Attachment: patch which uses GDI+ to render -moz-opacity for images

This patch optionally implements -moz-opacity using GDI+ for images. (Which is
the hard part :) ). 

It needs additional work to support opacity on other elements.	This shouldn't
be too difficult since GDI+ supports alpha when defining colors used for
brushes.

Comment 2

[Kevin McCluskey (gone)]

Comment on attachment 124221 [details] [diff] [review]
patch which uses GDI+ to render -moz-opacity for images 

sorry!, Wrong patch!

Comment 3

[Kevin McCluskey (gone)]

Attachment: patch that optionally using GDI+ to implement -moz-opacity for images

Comment 4

[Kevin McCluskey (gone)]

Attachment: gif with mask - speed test

Comment 5

[Kevin McCluskey (gone)]

Attachment: test-image - png with alpha (jellyheart)

Comment 6

[Kevin McCluskey (gone)]

Attachment: test image - jpeg (raptor)

Comment 7

[Kevin McCluskey (gone)]

Attachment: speed test 1 - png with alpha

Comment 8

[Kevin McCluskey (gone)]

Attachment: speed test 2 - jpeg

Comment 9

[Kevin McCluskey (gone)]

Attachment: speed test 3 - gif

Comment 10

[Kevin McCluskey (gone)]

Attachment: speed test 4 - png with alpha (Without setting moz-opacity)

Comment 11

[Kevin McCluskey (gone)]

speed test 1 - (pgn with alpha) + global opacity 
------------------------------------------------
original                 10264 (ms)
using AlphaBlend          4767 (ms)
using GDI+                2764 (ms)   3.7X faster


speed test 2 - jpeg + global opacity
------------------------------------
original                 8752 (ms)
using AlphaBlend         5148 (ms)
Using GDI+               3475 (ms)   2.5X faster


speed test 3 - (gif)
--------------------
(I don't have recent numbers for this, but should be faster as well)

speed test 4 -(pgn with alpha) *without* global opacity
--------------------------------------------------------
original               18757 (ms)
AlphaBlend              4456 (ms)
using GDI+              3195 (ms)    5.8X faster


summary: The patch which I attached can use use either the AlphaBlend call or 
GDI+ for alpha compositing of images. When using GDI+ it gives approx
2.5X - 5.8X speed improvement over the original.  Using Alpha blend gives a nice
speed improvement but there isn't an easy way to implement the full moz-opacity
for elements besides images. Using GDI+ on the otherhand will make it easy to
implement -moz-opacity on other elemnts since GDI+ supports alpha when defining
colors for brushes.

There are two aspects of the patch which speeds up -moz-opacity for images. 

(1) Using GDI+ bitmaps to perform alpha compositing instead of using the
existing cross platform global -moz-opacity implementation in the viewmanager.
(Note: There seems to be significant memory overhead when using GDI+. (5-6Mb
overhead just for initializing GDI+)

(2) Creating the backbuffer as a DIB instead of DDB.  This speeds up rendering
of  png's regardless of whether GDI+ is used. (Note: This will increase memory
footprint since the DIB can not be kept in video card memory and it is cached to
avoid the overhead of re-creating the DIB on each NS_PAINT)

Comment 12

[Benjamin Smedberg]

Is this going to affect the SVG branch, which has a generic system in place to
do opacity, and already uses GDI+ to do opacity on SVG elements?

Comment 13

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Kevin, I don't think this approach is compatible with the group opacity model.

Comment 14

[david avery]

the other issue is that doing this ( SVG branch has the same issue) requires we
distribute a copy of the MS gdiplus.dll with mozilla , since it only comes with
WinXp by default

Comment 15

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Kevin, we should talk about this, both about why you want to do this (what
exactly is the workload you're trying to optimize for here) and where is the
right place to do it.

The approach you're trying will work when the view with -moz-opacity consists of
exactly one image and no other visible stuff (no borders or backgrounds etc) but
it won't give the right results (as per CSS) in other cases.

Comment 16

[Zibi Braniecki [:zbraniecki][:gandalf]]

I wanted to apply patch into local build, but i was unable to. nsImageWin.cpp is
completly different in patch and in cvs. Kevin, could You make a patch for
recent trunk?

Comment 17

[Kevin McCluskey (gone)]

roc: The issue I was looking at was the slow resize performance of images in
mail and composer when the image is made large.

I see your point, the group opacity is incorrect with my approach since it will
be compositing each element as it is drawn, rather than compositing the elements
as a group.

Comment 18

[Colin S]

If GDI+ had been used before Microsoft patched the JPEG problem with it, would
Mozilla have been susceptible to this exploit? If so, I would be wary of adding
GDI+ to Mozilla, despite the performance gain. If not, go GDI+!

Comment 19

[Robert O'Callahan (:roc) (email my personal email if necessary)]

No, the exploit only affected JPEG decoding, which would not be used here.

Comment 20

[Hixie (not reading bugmail)]

And even if it did, we're already linked to millions of lines of Microsoft code 
in Windows builds, purely by virtue of using the Win32 API. If you are worried 
about holes in Microsoft code affecting Mozilla, don't use a Microsoft OS.

Comment 21

[Colin S]

Attachment: speed test 3 - gif

Obsoletes 124228 (incorrect mime-type and marked as patch when it is a
testcase), but of course I don't have appropriate permissions to do so :)

Comment 22

[Gordon P. Hemsley [:GPHemsley]]

This bug has been buried in the graveyard and has not been updated in over 5 years. It is probably safe to assume that it will never be fixed, so resolving as WONTFIX.

[Mass-change filter: graveyard-wontfix-2014-09-24]