This was originally added for b/14445579. An in-development app was
attempting to render to a window as an EGLSurface, then tear that
down, change some window properties, and create a new EGLSurface. The
second eglCreateWindowSurface failed because the window was already
connected. This change went in, but it turned out the real problem was
that the app still (unintentionally) had the surface current. After
the app bug was fixed, nobody revisited whether this change was
actually needed.
Turns out it wasn't needed. After an EGLSurface is both destroyed
*AND* not current (basically refcount==0), we were already
disconnecting the window in ~egl_surface_t().
Apart from being unnecessary and redundant, disconnecting the window
here is wrong for two reasons:
(a) The surface may still be in use after eglDestroySurface, if it was
still current. Rendering is undefined in that case, but disconnecting
the window leads to more catastrophic results than necessary.
(b) It's being called before calling the driver's eglDestroySurface.
The driver will almost definitely have a buffer dequeued that it needs
to cancel, and by disconnecting first we turn that into an error that
they don't have a graceful way to deal with.
Bug: 24524053
Change-Id: Ib063134413d25d3526f794aafb5e333e3417ea42
Previously the parsing found the next space and then added the the difference
between the current position and space to the set of tokens. This improperly
generated empty strings if there were consecutive spaces or if spaces existed at
the beginning or end of strings. To fix this, the parse is modified to use
simple stringstream parsing.
Bug: 22709246
Change-Id: I9e32c07bbf984eadccdadf1dc34437fa0c46088b
The product of width and height should be less than UINT32_MAX (in practice
smaller). Adding the checks prevents overflows when allocating buffers.
Bug: 20726612
Change-Id: I9769edf0688a9bfe69906d49fa0540cadf4c49b0
The Android EGL shim injects GL_EXT_debug_marker into the ES driver
EXTENSIONS string for the OpenGL ES 1.x and 2.0/3.0/3.1 drivers if
the extension is not already provided. This feature is used by
GLES_trace.
In Open GL ES 3.0 it became possible to query an indexed version of
the EXTENSIONS string via GetStringi(). NUM_EXTENSIONS Gets were
also added to the specification (taken from Open GL).
If the shim does not have to inject the extension, then there is no
problem, as glGetString() and glGetStringi() / NUM_EXTENSIONS will
be consistent. However, if the Android EGL shim injects the
extension, NUM_EXTENSIONS and GetStringi() will report one less
extension than is really available.
Consistency between these methods is tested by the dEQP framework
with the dEQP-GLES3.functional.state_query.string.extensions test.
If the driver does not provide GL_EXT_debug_marker, this test
fails.
This change wraps all of the affected entry points so that the
wrapped driver extensions are never visible directly to dEQP,
eliminating the inconsistency.
Old version of unity uses custom linker to link
itself and libmono. The custom linker does not
support lookups with gnu-hash style. Switching
libraries needed by libunity.so to --hash-style=both
fixes this problem.
Bug: http://b/20093774
Change-Id: I2a4887cf7a5ee80a8976ecaad54b52dc62518e55
This adds EGL wrapper functions for the following EGL extensions:
EGL_EXT_buffer_age
EGL_KHR_partial_update
EGL_KHR_swap_buffers_with_damage
Change-Id: I407acda1e0310f7f01a5efe9c915721a941138a4
- Wire up new dataSpace parameter through buffer queue stack
- Update tests to include the parameter
- Switch eglApi to using dataSpace to indicate sRGB gamma/linear
difference
- Remove RAW_SENSOR in favor of RAW16
- Remove use of sRGB format enums
- Add default dataspace to buffer queue core
- Add query for default dataspace
Cherry pick of I070bd2e7c56506055c419004c29e2e3feac725df
Change-Id: I461952389c18051176c6b75e664f20ad369f5760
Some time before kitkat, a workaround was put in place to keep tuna
working with obsolete PowerVR GPU drivers. The tuna product was dropped
in AOSP for kitkat, but the workaround remained.
There are still a number of devices shipping with OMAP4 processors, and
this workaround continues to be applied on those products even when it
is not required. The effect is that virtual display CTS will fail on
those devices.
Fix regression from 6701fbe5f0 .
The rdhwr instruction is unsupported by mips32r1 cpus, but is
supported via kernel emulation on those cpus.
Change-Id: Ieb3343fff0f0da54003044f451c90e1c761208eb
Replace the global lock in terminate for the ref count check with atomics and a
local lock.
Bug: 19072064
Change-Id: I0cfd6564e227a75b8387a8788b16381d5bc2cc88
Avoid ambiguity by using register numbers. Register names are
interpreted depending on the ABI. Fixes the issue with mips64 gcc,
where registers t0-t4 are translated into $8-$11 (o32 ABI) instead
of $12-$15 (o64 ABI).
Also, use t9 ($25) for storing function address
(in getProcAddress.cpp).
Change-Id: I12ef5cd98bf14dab3c01f2756889cc02856c1cd8
Combines 4 patches by Dragoslav and Douglas:
Use t9 to hold the address of the called function.
Add support for gl binding.
Fix inline assembly code to be R6 compatible.
Fix some more inline assembly code to be R6 compatible.
Change-Id: I0b899813790b2ba4ca538bb21692014cc1933f66
