Platform/GFX/2012-January-23
From MozillaWiki
- Every Monday at 2:30 PM Pacific/5:30 PM Eastern/22:30 UTC/Tuesday at 11:30 AM NZDT
- +1 650 903 0800 x92 Conf# 99366
- +1 416 848 3114 x92 Conf# 99366
- +1 800 707 2533 (pin 369) Conf# 99366 (toll free, Skype)
- Video (Vidyo) link: https://v.mozilla.com/flex.html?roomdirect.html&key=vu1FKlkBlT29
- Vidyo room 9366 (if you have LDAP and can log in at https://v.mozilla.com)
Agenda:
- OMTC, Java compositor
- benwa has the java compositor thread being created and initializing, but needs to make some changes to hook it in to off-main-thread usage
- benwa had a discussion with pcwalton about having a temporary java compositor (which renders chrome and a screenshot of the previous page) while the regular compositor (which is maintained inside gecko) is being initialized. There'll then be a transition from the Java compositor and the Gecko compositor, but it has to be transparent to the user.
- jpr: Will the Java compositor always be around?
- pcwalton: yes. We need to draw scrollbars, etc in Java, but it won't be driving GL. It will be small & probably won't use GL.
- jpr: What about unaccelerated devices?
- pcwalton: We should always have acceleration. Any device that doesn't support GL layers is a bug.
- benwa: gw280 worries about using NPOT textures. Specifically from a performance standpoint. We need to benchmark to find out what the cost is (there is definitely a perf hit on IMG hardware; can't speak for Adreno)
- bjacob: Many devices definitely don't support NPOT pbuffers, including Maguru. jgilbert and he are working on removing usage of pbuffers on EGL, in favour of using FBOs.
- BenWa thinks he'll have at least some part of OMTC working on-device by the end of the week, if he doesn't hit any bugs.
- ajuma has started looking at asynchronous scrolling on the new compositor vs the Java compositor. There's quite some work to be done to implement it in the IPDL PLayers.
- dougt: How should we characterize winning from this?
- ajuma: We should have perceived performance wins.
- pcwalton: We should have improvements in checkerboarding because Layers is smarter. We need to be careful to be sure we don't regress panning/zooming perf when a finger is down. Java path for adjusting scroll and zoom is very simple; Layers is less simple.
- benwa: WebGL should be better too.
- dougt: Timeline?
- jrmuizel/joe/benwa: post-12. We don't want to uplift this; too great a risk of regression. Threads are right in the name.
- 280.13 Driver Bugs
- bjacob: XRender, probably. We would not have these problems if we used pixman or skia.
- We're not sure whether this is blocking the new test slaves.
- joe to talk to bhearsum to find out how urgent these are.
- B2G/Mobile (jpr)
- joe to coordinate b2g/gfx work
- bjacob can not debug. mwu says you can just flash the phone, which is in direct contradiction to what cjones told us.
- Issues on maguru seem largely explained by lack of support for NPOT PBuffers.
- We're switching from PBuffers to FBOs altogether on EGL, which also means Windows/ANGLE. Raise your hand if you disagree. bug 720467
- Profiler on Mac (benwa)
- https://developer.mozilla.org/en/Performance/Profiling_with_the_Built-in_Profiler#Running_the_profiler
- Switch to the profiling nightly permenantly, install the extension. Then simply hit APPLE+SHIFT+O whenever!
- Give BenWa some feedback
- getImage/putImage with Azure bug 720428 (jrmuizel)
- bas and jrmuizel discussed