Bug: 18226391
The issue occurs as a result of a dispatchFrame itself
taking longer than 12ms, the alloted budget. The result
is that a vsync request (which occured at the end) would
miss the vsync that occured 1ms prior to the end of the frame.
As a result it would end up waiting for the following vsync,
essentially dropping to 30fps even though 60 could have been
sustained.
Fix this with a few tweaks.
First, adjust the UI thread's delay bias from (now + 4ms) to
(vsync + 4ms), this prevents RT animators from slowly drifting
if the vsync occurs mid-task.
Second, request a vsync preemptively prior to running callbacks.
This way if any callbacks needs the next vsync and it takes
"too long", we will catch that vsync.
Finally, fix an issue where the display event queue was always
drained & rejected at the end of a task loop. Instead, drain
and reject all stale vsyncs. This still prevents the issue of
both UI thread & RT thread trying to drive 2 frames in a single
pulse, but also allows RT to notice that it missed a vsync
pulse it needed and that it should speed-up a bit in response
Change-Id: I9d6be037737e9283297898cac2e3563453e797cd
Bug: 18317479
RenderNode::prepareSubTree calls prefetchAndMarkInUse
on every bitmapResoruce in the DisplayList. However,
this resulted in textures being uploaded for bitmaps
that would be drawn from the AssetAtlas instead.
To fix this we teach TextureCache about the AssetAtlas
so that calls to TextureCache return the Texture from
AssetAtlas if it exists. Thus usage of AssetAtlas
is now purely to allow for further optimizations via
draw merging instead of a requirement to get
any benefit at all.
Change-Id: I65282fa05bac46f4e93822b3467ffa0261ccf200
Bug: 18203577
The issue occurs as a result of performTraversals() both doing
a window relayout call *and* early-returning because it's not dirty.
To fix this pauseSurface() returns whether or not the RT-side is
"dirty" to force ViewRootImpl to do a draw even if mDirty is
otherwise empty.
Change-Id: I534f367e75d18d273ebf14df3927f5c464ef6bef
Bug: 18518580
When destroying CanvasContext, the surface can be invalid
state. So the surface should be updated to null prior to
destroying GL resources to ensure that GL functions
are not called with an invalid surface.
Some GL implementation makes an error if GL functions
are called with an invalid surface. (Adreno 3xx)
Cherry picked from AOSP: f76d36f96bf221672e98e440c9df7cbf0e02e84e
Change-Id: Ie6f6ea081ec931fc9df30b2c3ed066ec1ae9d294
When destroying CanvasContext, the surface can be invalid
state. So the surface should be updated to null prior to
destroying GL resources to ensure that GL functions
are not called with an invalid surface.
Some GL implementation makes an error if GL functions
are called with an invalid surface. (Adreno 3xx)
Change-Id: Ie6f6ea081ec931fc9df30b2c3ed066ec1ae9d294
Bug: 18518580
If CanvasContext is being destroyed() the Surface
is probably no longer valid as well, so make sure to
makeCurrent() to the pbuffer surface so that the
subsequent GL operations are not using an invalid
EGLSurface
Change-Id: Ica5d6a065841772c47e00ad65aa7894c7e27e043
Remove Clang cutout for unused parameters. Fix warnings.
Remove Clang cutout for deprecated Skia function usage. Has been
fixed in the L push.
Change-Id: I7ea073ff67127cc1e14e798b655e2c50615fe8e7
Bug: 17516789
This will force a relayout/reinitialize pass if the Surface
is lost mid-render instead of crashing on the next frame
Change-Id: If08bfa16f740728fa7c05904fa11e26f07b81e2e
Bug: 17208461
* Switch Layer to be VirtualLightRefBase instead of
Caches' side-channel ref-counting
* Include active layers in gfxinfo dump
* Run gfxinfo dump on the correct thread
* Dump gfxinfo on Layer creation failure
Change-Id: I28d195699e2334518e215ab28c7a17355aee9678
Bug: 18099195
Don't use EGL_SWAP_BUFFER_PRESERVED on surfaces that will
never benefit. Also clean up some confusing naming
Change-Id: I674ca64e0464a3282cff79e5ecd350d08f47c014
Graphics memory usually gets trimmed in applications when the
activity goes into the background. We use quite a lot of graphics
memory when the shade/lockscreen is open, and some of them never gets
freed unless the recents activity is closed, because we don't have
these activity-trimming-heuristics for the shade. This change
proactively trims the graphics memory when the shade gets closed or
when the lockscreen is hidden, to emulate the same heuristics as for
activities.
This change also adds trimMemory on RenderThread to systrace to
verify that no jank is introduced with this change.
This change immediately saves around 10-30 MB on an xxhdpi device
after the shade is closed.
Bug: 17581375
Change-Id: I4fb622efb51815fe08187be97ba15d012d4de5d4
Bug: 17765082
DeferredLayerUpdater had fallen behind RT updates. Re-snap to
latest expectations, ensuring to call requireGlContext() prior
to detachSurfaceTexture to avoid leaking SurfaceTextures
Change-Id: Ic65fb9831e5284f658866da8da9ad5af1d227699
bug:17463894
Fixes the setViewport method to immediately affect the return values
of getViewportWidth/Height methods.
Also works around tiling extension issues observed on first frame
after window resize by disabling tiling for that frame.
Change-Id: Ie172d572d20d74a1be9cc58ad389af2cffa0e4b6
Bug: 17372309
AnimationContext::startFrame() happens both with and without
the UI thread lock. Pass the TraversalMode into it so
that ThreadedRenderer's subclass can correctly decide
when it is safe to push over mPendingAnimatingRenderNodes, as doing
so outside of the lock is Very Bad.
Change-Id: Ife5dd3a2b46b0a207cd9234c159a674afdbf5efd
Bug: 17372309
Fixes a case where UI thread and RT thread both used the same method
which wasn't safe for either of them.
Adds additional assertions & logging in unusual circumstances to
try and track down where the issue is occuring from.
Change-Id: I93d31a6fd0c5927259b67bdf96a475944226eee6
Bug: 17208461
There's a potential race condition between HardwareRenderer.destroy()
being called (which calls destroyCanvasAndSurface()) and the renderer
being finalized (which is what calls freePrefetchedLayers), during which
time it's possible we get a TRIM_MEMORY_COMPLETE and destroy the EGL
context.
Fix this race condition by moving stopDrawing() and freePrefetchedLayers()
into destroyCanvasAndSurface() where they should have been in the first
place.
Also, if we hit the assertion failure, dump the current state of
Caches to try and provide more context for the failure.
Change-Id: Ife0ba3562041e8b08e87e3e13640472b3004eed6
bug:15860114
Savelayers and HW layers both now support shadow casting.
For save layers, the light source should always be correct, for HW
layers, the light source position is set when the layer is created,
and updated when it is resized.
Change-Id: Ie85567dd43c2bb0a0b08fd0bd4db41efa793ac2b
bug:15859361
Moves lighting info out of StatefulBaseRenderer, since it's not useful
at record time, and only used by OGLR.
Change-Id: I7ab065d02d9304afad1dc4c48597a7a621366f8e
