An optimization for paths is to only create a texture for the original native
Path object, and have all copies of that object use that texture. This works in
most cases, but sometimes that original path object may get destroyed (when the
SDK path object is finalized) while we are still referencing and using that object
in the DisplayList code. This causes undefined errors such as crashes and hanging
as we iterate through the operations of a destroyed (and garbage-filled) path object.
The fix is to use the existing ResourceCache to refcount the original path until
we are done with it.
Issue #6414050 Analytics Dogfood App crashes reliably on Jellybean
Change-Id: I5dbec5c069f7d6a1e68c13424f454976a7d188e9
There was a bug in the DisplayList code for layers where we would
set the alpha value for a layer only if the alpha were non-1. This works
most of the time (since the value is usually 1 and doesn't need to
be set at all, and if the value is non-1, it is set correctly). But when
the value has been set to a non-1 value, setting alpha back to 1 cannot happen
due to this logic. This caused launcher to have some invisible pages when
returning to Home because those pages previously had an alpha value of 0, and
setting the alpha to 1 had no effect due to this DisplayList code.
The fix is to simply remove the check for non-1 values; we should
always set the alpha value of layers to handle all cases.
Issue #6413892 alpha value is messed up after AlphaAnimation
Change-Id: Ia51acb2eaaf0609ea7189998ed449bdd9ea7e05f
This flag was still hanging around pending any need to disable
DisplayList properties. But things seem stable, so it's time to clean up
and simplify the code.
At the same time, I reduced redundance in DisplayList dimensions. We
used to call drawDisplayList() with width/height parameters that were
used to do a clip reject. This is redundant with the DisplayList properties
that set the bounds of the DisplayList; the left/right and top/bottom properties
represent the same width/height properties formerly used in drawDisplayList().
The new approach is to not pass dimensions to drawDisplayList(), but to
instead pull those dimensions directly from the DisplayList when needed.
Change-Id: I8871beff03b1d4be95f7c6e079c31a71d31e0c56
Cached layers were sometimes retaining an obsolete blending
value, causing artifacts where translucent objects were disappearing
completely. Also, added extra tracing info for DisplayLists.
Issue #6303668 Flickering views during SwipeHelper drag
Change-Id: I66ce158652c4a3ed316040585b40b1744e1fad0c
Some views (such as ImageView and TextView) handle non-opaque alpha
values directly. This was originally an optimization, but we can handle it faster
in many cases without this optimization when DisplayList properties are enabled.
Basically, if a view has non-overlapping rendering, we set the alpha value directly
on the renderer (the equivalent of setting it on the Paint object) and draw each
primitive with that alpha value. Doing it this way avoids re-creating DisplayLists
while getting the same speedup that onSetAlpha() used to get pre-DisplayList properties.
Change-Id: I0f7827f075d3b35093a882d4adbb300a1063c288
Re-enabling DisplayList properties last week caused some app
errors due to the way that some transforms were being handled (specifically,
those coming from the old Animations and ViewGroup's childStaticTransformation
field). This change pushes *all* transform/alpha data from View.draw() into
the view's DisplayList, making DisplayLists more encapsulated (and correct).
Change-Id: Ia702c6aae050784bb3ed505aa87553113f8a1938
Several issues came up after DisplayList properties were enabled,
so they were disabled pending fixes. Those issues have been fixed, so
DisplayList properties are once again being enabled by default. This
CL both re-enables these properties (in View.java and DisplayListRenderer.h)
and fixes the various issues that enabling them caused the first time around.
Related issues (all currently marked as Fixed, though that was simply because
DL properties were disabled - this CL provides the real fixes now that
DL properties are enabled by default):
Issue #6198276 Text input broken
Issue #6198472 Native crash at pc 00076428 in many different apps in JRM80
Issue #6204173 Date/time picker isn't rendering all parts of UI
Issue #6203941 All Apps overscroll effect is rendered weirdly/has flickering
Issue #6200058 CAB rendering issue - not drawing items?
Issue #6198578 Front camera shows black screen after taking picture.
Issue #6232010 Layers not recreated when children change (DisplayList properties)
Change-Id: I8b5f9ec342208ecb20d3e6a60d26cf7c6112ec8b
WebView needs more fine-grained control over the behavior of the
framework upon execution of the display lists. The new status_t
allows WebView to requests its functor to be re-executed directly
without causing a redraw of the entire hierarchy.
Change-Id: I97a8141dc5c6eeb6805b6024cc1e76fce07d24cc
DisplayList properties are still disabled default (flags in View.java
and DisplayListRenderer.h). When they are enabled, and when a View has
a DisplayList, invalidations due to property changes are now optimized
to avoid causing DisplayList recreation. This eliminates the drawing step
of invalidation (due to changes in these properties), only requiring
issuing the previously-created DisplayList to the GL renderer. Invalidation
is slightly faster (less overhead as we walk up the hierarchy), getDisplayList()
is potentially much faster (going down to ~0ms), depending on the complexity
of the View being redrawn and the size of the invalidated hierarchy.
Change-Id: I57587d5b810c3595bdd72a6c52349c2a3d1bdf25
Bug #6157792
Previously, DisplayListRenderer would compute the number of bytes
written after a drawing op by looking at the difference between
the start pointer of the command stream and the end pointer of
the command stream. The SkWriter class used to record the commands
stream allocates blocks of storage which would cause a crash when
a command spanned two blocks.
Change-Id: I4d79d3feeb6d72d9d4e6ab05ecebd72d004be56c
This is the real fix to issue 6158892. We currently delete
transform/camera structures at DisplayList destructor time, if these
structures are not NULL. We set the fields to NULL in an init() method
called (eventually) by the constructor. But it is possible for the object
to be destroyed before that init code is called, resulting in the deref
bug reported. The fi is to set these structures to NULL directly in the
constructor.
Issue 6158892i: Device runtime restarts frequently
Change-Id: Ibfa0f9314767eed6fd51f4ec7edc0d0edd5fdd0f
Basic functionality of handling View properties (transforms,
left/right/top/bottom, and alpha) at the native DisplayList level.
This logic is disabled for now (via compile-time flags in View.java and
DisplayListRenderer.h) as we continue work on it (there is no advantage
to the new approach until we optimize invalidation and rendering paths
to use the new code path).
Change-Id: I370c8d21fbd291be415f55515ab8dced6f6d51a3
glGetError() was invoked every time a display list was updated. This
was unnecessary since display list updates do not execute OpenGL code.
Change-Id: Ia0a75a65ccf2a4701fb3a828848cd0c42f256c6c
Bug #6073717
Bug #6065504
Bug #6026515
Bug #5971725
Prior to this patch, the destructor of DisplayList would always run
on the finalizer thread. This could cause a race condition if the UI
thread was busy rendering display lists at the same time leading to
various random native crashes.
Change-Id: Ie11108e3b1538d4b358a1a8b4cce1b2d33152d0c
This optimization allows us to quickly skip operations that lie
entirely outside of the known bounds of a display list. Because
of ViewGroup.setClipChildren, we must keep the operations recorded
in the display list. setClipChildren(false) is however a very
uncommon operation and we will therefore often benefit from this
new optimization.
Change-Id: I0942c864e55298e6dccd9977d15adefbce3ba3ad
These markers will be used to group the GL commands by View in the
OpenGL ES debugging tool. This will help correlate individual GL
calls to higher level components like Views.
Change-Id: I73607ba2e7224a80ac32527968261ee008f049c6
- it was a regression introduced into this CL: https://android-git.corp.google.com/g/#/c/154240/5
- basically needed to set the GlyphID encoding to the Skia Paint as we are now using glyphID resulting
from the Harfbuzz shaping
- also define GlyphID encoding as the default on the Paint class
Change-Id: Idb7c2c57ac67595425ce3be9421258962690fcdd
Bug #5706056
A newly introduced optimization relied on the display list renderer
to properly measure text to perform fast clipping. The paint used
to measure text needs to have AA and glyph id encoding set to return
the correct results. Unfortunately these properties were set by
the GL renderer and not by the display list renderer. This change
simply sets the properties in the display list renderer instead.
This change also improves the error message printed out when the
application attempts to use a bitmap larger than the max texture
size.
Change-Id: I4d84e1c7d194aed9ad476f69434eaa2c8f3836a8
Bug: 5062011
Previously, each GLES20DisplayList would hold onto an instance of
GLES20RecordingCanvas. In turn, each GLES20RecordingCanvas
held onto an SkWriter with a 16Kb buffer along with several other
objects. With one display list per view and hundreds of views,
the overhead could add up to a few megabytes.
Ensured that the GLES20RecordingCanvas is reset as soon as
the display list has been constructed, thereby promptly freeing
the 16Kb buffer.
Changed GLES20DisplayList so that it acquires a GLES20RecordingCanvas
from a pool as needed and recycles it when done.
Removed some dead code and cruft related to the construction of
GLES20Canvas objects in general. Some code was written with the
assumption that the underlying renderer object could change
behind the scenes or might be lazily constructed, but that isn't
actually the case so we can simplify things.
Removed an unnecessary weak reference from GLES20DisplayList
to the View. It isn't actually used anywhere.
Fixed a bug in GLES20DisplayList where isValid() would return
true while the display list was being recorded. This is incorrect
because the native display list might not actually exist. Worse,
even if the native display list does exist, it is stale and
potentially refers to old Bitmaps that have been GC'd (because the
mBitmaps list was cleared when recording started).
Change-Id: Ib12d5483688cb253478edeb0156d34c476c2566b
Clicking on a node in hierarchyviewer1 and hierarchyviewer2 and then
clicking the new "Dump DisplayList" button will cause the display
list for the selected node (including its children) to be output into
logcat.
Change-Id: Iad05f5f6cca0f8b465dccd962b501dc18fe6e053
Clicking on a node in hierarchyviewer1 and hierarchyviewer2 and then
clicking the new "Dump DisplayList" button will cause the display
list for the selected node (including its children) to be output into
logcat.
Change-Id: Id32f62569ad1ab4d533bc62987f3a7390c1bb4e6
Bug #4092053
The problem always existed but was made visible by partial invalidation.
When saving a layer, the renderer would try to postpone glClear()
operations until the next drawing command. This however does not work
since the clip might have changed. The fix is rather simple and
simply gets rid of this "optimization" (that turned out to be
usless anyway given how View issues saveLayer() calls.)
This change also fixes an issue with gradients (color stops where
not properly computed when using a null stops array) and optimizes
display lists rendering (quickly rejects larger portions of the
tree to avoid executing unnecessary code.)
Change-Id: I0f5b5f6e1220d41a09cc2fa84c212b0b4afd9c46
Bug #3461349
This change also fixes two bugs that prevented partial invalidates
from working with other views. Both bugs were in our EGL implementation:
they were preventing the caller from comparing the current context/surface
with another context/surface. This was causing HardwareRenderer to always
redraw the entire screen.
Change-Id: I33e096b304d4a0b7e6c8f92930f71d2ece9bebf5
A recent change to optimize path rendering didn't account for
the destruction of native objects by the VM finalizer. We may be
done with the Java level version before we're done with the native
structure that's used by the display list. For example, a drawing
method on a View that creates a temporary path to render into the
canvas will implicitly create a native structure that is put onto
the GL display list. That temporary path may go away, but the native
version should stick around as long as the display list does.
The fix is to refcount the original native version of the path
and only delete it when the refcoutn reaches zero (which means that
it is no longer needed by any display list). This is a similar mechanism
used for bitmaps and shaders.
Change-Id: I4de1047415066d425d1c689aa60827f97729b470
