Hibernating EGL takes a long time (>100 ms) and blocks all other
rendering. During window animations, the outgoing activity begins
hibernation before the animation stops, causing visible stutter.
Hibernation is still available by setting 'BOARD_ALLOW_EGL_HIBERNATION
:= true' in the devices BoardConfig.mk
Change-Id: Iab4e00723a1adcd97481e81b2efdc821b3e9712f
When eglTerminate() is called with a window surface still exists, a
deadlock would occur since egl_display_t::terminate() holds a lock
while destroying the window surface, which calls
onWindowSurfaceDestroyed() which attempts to take the same lock.
This change refactors the hibernation code and data into a separate
object with its own lock, separate from the egl_display_t lock. This
avoids the deadlock and better encapsulates the hibernation logic.
The change also fixes a bug discovered incidentally while debugging:
hibernating after calling eglTerminate() succeeds, but will cause
awakens from subsequent eglInitialize() to fail. We will no longer
hibernate a terminated display.
Change-Id: If55e5bb603d4f8953babc439ffc8d8a60af103d9
If the EGL implementation supports the EGL_IMG_hibernate_process
extension, use it to hibernate (and hopefully release memory or other
resources) when the process isn't actively using EGL or OpenGL ES. The
idleness heuristic used in this change is:
(a) Wake up when entering any EGL API call, and remain awake for the
duration of the call.
(b) Do not hibernate when any window surface exists; this means the
application is very likely in the foreground.
(c) Do not hibernate while any context is made current to a thread.
The app may be using a client API without the EGL layer knowing,
so it is not safe to hibernate.
(d) Only check these conditions and attempt to hibernate after a
window surface is destroyed or a thread's context is detached. By
not attempting to hibernate at the end of every EGL call, we avoid
some transient wakeups/hibernate cycles when the app is mostly idle,
or is starting to become active but hasn't created its window
surface yet.
On a Galaxy Nexus, hibernating frees 1567 VM pages from the process.
Both hibernating and waking can take anywhere from 30ms to over 100ms
-- measurements have been very inconsistent.
Change-Id: Ib555f5d9d069aefccca06e8173a89625b5f32d7e
This is in preparation for a change that will hibernate the underlying
EGL when idle. Instead of a bare egl_display_t*, get_display() now
returns a egl_display_ptr, which acts like a smart pointer. The
"wakecount" counter managed by the smart pointer isn't used for
anything in this change. It will be used to make sure we don't
hibernate when any thread is in an EGL call, without having to hold a
mutex for the duration of the call.
Change-Id: Iee52f3549a51162efc3800e1195d3f76bba2f2ce
This change adds a debug option to EGL to use an EGLSyncKHR each frame to
determine when the GPU finishes rendering the frame.
Change-Id: I09ce071db904b44f07ca814c586c291c8b59385a
This patch adds a function setGlDebugLevel() to libEGL to enable
GL tracing. This will be used by the Java layer to add an option
to "am start" that can enable tracing for a particular application.
Change-Id: Ie1dbdd550f502df8633553595cb33ee9d9ae44e1
this is possible now that we support only a single
EGL implementation. this allows a large code simplification.
Change-Id: I7a6b9db4c5d60f4407c6061e7a68729af63d5242
from now on, the system can only have one EGL
implementation. this means the software and h/w renderer
cannot be used at the same time on a device. Of course, the
h/w renderer is always prefered; in its absence we
default to the software renderer.
Change-Id: Ib579f58055dd0ce4c4a99144131efa11c16ca3d3
this was introduced in a recent change. eglMakeCurrent can
end up calling eglDestroyImageKHR via ANativewWindow::disconnect
when the consumer is in the same process.
we make sure we don't hold the lock while this is happening.
Change-Id: Id17fe4fd76eecf5f962cefb9aa32be41fc1b042d
it would happen when a context was made non-current, in this
case we would call the implementation's eglMakeCurrent() which
would succeed, if we're rescheduled at that point, another
eglMakeCurrent() could make that context current to another thread,
however, when we came back to the original call we would
overwrite egl_context_t internal state.
this is fixed by moving the critical section under
egl_display_t's lock.
Change-Id: I743c85696e13263d3a9570824940263df0caacdc
- don't advertise extensions that are not supported
by any implementation
- remove EGL_ANDROID_swap_rectangle which is not
implemented by anybody and confuses people
- add some comments about mandatory extensions
Bug: 5428001
Change-Id: Id8dc48116ac1d1eb79ec9ef55d03e29d4257c1f3
the code that validated EGL objects dereferenced the object
to access its EGLDisplay -- needed for validation (!).
This was wrong for two reasons, first we dereferenced the object
before validating it (potentially leading to a crash), secondly
we didn't validate that the object existed in the right EGLDisplay.
We now use the EGLDisplay passed by the user API.
Change-Id: I66f9e851d4f8507892a6b1fee3065f124c4e7138
This change makes the makes the stub EGL_ANDROID_blob_cache callbacks
actually use a BlobCache object.
Bug: 5474671
Change-Id: I5cbaae2dea3aad2fe306c9f57029c3f215a0863a
This change adds a stub cache implementation that gets passed to the
underlying EGL implementation at initialization time.
Change-Id: I14437c5b6f91b7a34a19bb02ad802e6e54f88d2a
as specified by the EGL specification, terminated objects's
handles become invalid, the objects themselves are destroyed
when they're not current to some thread.
Change-Id: Id3a4a5736a5bbc3926a9ae8385d43772edb88eeb
