This patch allows the OpenGLES rendering library to use Unix
sockets instead of TCP ones when communicating with its clients.
On certain benchmarks (e.g. 0xBench teapot), this provides a
noticeable improvement (x1.05 fps) without any other changes.
On practice, Unix sockets are faster than TCP sockets, even
local ones. Also, this introduces a moderate amount of
abstraction that will allow us to use Win32 named pipes
on Windows (where TCP sockets are much slower than they
are on Unix).
Note that by default, TCP streams are still used.
The client (emulator) must call the new API 'setStreamMode'
to change it to STREAM_MODE_UNIX between 'initLibrary' and
'startOpenglRenderer' calls.
+ Adjust callers / user appropriately.
Change-Id: I4105bbf07541f3146b50a58d1a5b51e8cf044fab
This re-enables the build of the gralloc GLES emulation library.
This is only for experimentation / debugging purpose. GLES emulation
is still disabled by default in the emulator. Even if you force-enable
it by using the '-gpu on' flag, this module will not properly yet
and nothing will be displayed.
Change-Id: I8f0ca8805ed99037b011365d08507ca08bff5e75
This fixes the definition of the emulator-specific egl.cfg file
used to indicate to our EGL system library which support modules
are available in the emulator.
We need to list two modules here:
'emulation' corresponding to GPU emulation
'android' corresponding to the software renderer
The indices before each name correspond to the 'display' and 'impl'
numbers that are expected by frameworks/base/opengl/libs/EGL/Loader.cpp
In a nutshell, 'display' should always be 0, and 'impl' should be 0
for the software renderer, and 1 for a hardware-based one. See the code
under framewors/base/opengl/libs/EGL/egl.cpp for mode details.
Change-Id: I52c898759200c2dfba9049ed00b31b18e8c37f69
The gralloc.goldfish module cannot build in master because
the gralloc interface evolved considerably since gingerbread
(which the current code is based on).
For now, disable it completely, this will make it impossible
to test GPU emulation in the internal master tree for now.
We'll have to port this module to the new world-order to solve
this.
Change-Id: I864217f9c41fd1234f0df497005499800af879e5
This adds an initLibrary() function to libOpenglRender.
It will be used by the emulator to fallback on software rendering
if the library cannot be initialized properly, e.g. if it is not
possible to load the host EGL/GLES libraries.
Change-Id: I41e8ad73a315166e4a15cbee1db72c2552370f46
This patch changes the filter used to scale the GL window
in the rendering library. Instead of nearest-neighbour, use
bi-linear, which will provide for much more pleasant results
when scaling the emulator UI window, or toggling to full-screen
mode.
Change-Id: If2d9e1300fa4878c25e1221eaedb565b9150431a
This patch modifies the gralloc.goldfish module to check
that the emulator does, indeed, support GPU emulation.
If this is not the case, it loads the framebuffer-based fallback
module (gralloc.default) explicitely and uses it instead.
This is necessary because sometimes the emulator will be started
in headless mode in restricted environment, i.e. without access
to the display subsystem or GL libraries, which make GPU emulation
impossible.
We check for ro.kernel.qemu.gles which will be either undefined
or set to 0 if the emulator doesn't support GPU emulation, or 1
otherwise.
Change-Id: Ib6b143e6dcdfb44ff2c5b889138d0fb4118bb461
This patch modifies GLDispatch.cpp in the translator libraries
to open the Darwin GL library as a framework. Using a hard-coded
system path prevents the requirement to modify LD_LIBRARY_PATH
or DYLD_LIBRARY_PATH to point to the directory that contains the
library.
In this specific case, this is equivalent to
/System/Frameworks/OpenGL.framework/Libraries/libGL.dylib
But is the "kosher" way to open it the library.
Change-Id: Iccd6568d36fc97200617c268f05d7b4ea85f918e
You can still re-enable them by uncommenting one line in
development/tools/emulator/opengl/Android.mk
Change-Id: Ie0a640b6f3359b437aba40b124da0892defb00ef
These libraries are only used by libOpenglRender, so make them
static so that their code is included in this shared library
instead. This reduces the shared library dependencies that
will be required from the emulator.
Change-Id: I24438932f0d6b40852b515140ea6950c04124fdd
This patch forces the default definition of BUILD_EMULATOR_OPENGL_DRIVER
to true by default, which is the common case where we will build the
goldfish-specific gralloc hardware module to allow SurfaceFlinger to
use GLES emulation exclusively.
The only reason to define thi to 'false' today is to debug specific
apps, but this requires displaying the GL output in a different window
than the emulator to be usable.
Change-Id: Iaf03a9c780449bb6b991928d324a088021a73203
It is possible to move the final installation path of a shared
library by using LOCAL_MODULE_PATH/LOCAL_UNSTRIPPED_PATH in its
module declaration.
We do this for example to put certain libraries under /system/lib/egl
or /system/lib/hw.
However, the Android build system has a small bug where you cannot
depend on these shared libraries because it will complain it doesn't
find them under /system/lib, the default install location.
More precisely, consider libfoo defined as:
include $(CLEAR_VARS)
LOCAL_MODULE := libfoo
LOCAL_MODULE_PATH := $(TARGET_OUT_SHARED_LIBRARIES)/egl
...
include $(BUILD_SHARED_LIBRARY)
Its final binary will be installed to /system/lib/egl/libfoo.so
Now, let's write a module that depends on it, as:
include $(CLEAR_VARS)
LOCAL_MODULE := libbar
LOCAL_SHARED_LIBRARIES += libfoo
...
include $(BUILD_SHARED_LIBRARY)
The build system will complain there is no rule to define the target
/system/lib/libfoo.so, but the target should be /system/lib/egl/libfoo.so.
A work-around is to define libbar as follows instead:
include $(CLEAR_VARS)
LOCAL_MODULE := libbar
LOCAL_ADDITIONAL_DEPENDENCIES := $(TARGET_OUT_SHARED_LIBRARIES)/egl/libfoo.so
...
include $(BUILD_SHARED_LIBRARY)
This works if you don't need to link against the library when building your
shared library (which is fortunately the case for the GLES emulation
libraries).
That's essentially what this patch implements under common.mk. We update
emugl-set-shared-library-subpath to record that a module has been "moved",
and we avoid adding them to the LOCAL_SHARED_LIBRARIES variable of the
modules that export it.
+ Simplify three Android.mk files accordingly.
Change-Id: Id15bef5359b0daa8d82bfaa5abf9346f00190ab5
Do not redefine SDL_CONFIG to avoid conflicts with the emulator's
build system which depend on its definition. Note that this has
been fixed in another patch (https://review.source.android.com/25684)
but it is also cleaner to do this here.
Change-Id: I1a30d200e467ac7a91a1e6135fa1de34033e8df0
add drawValidate when calling glCheckFramebufferStatus
in the draw validate, a color attachement will be added to the
framebuffer if there is no target attached to COLOR_ATTACHMENT
of the framebuffer.
opengl does not allow framebuffer with no color attachement
so CheckFramebufferStatus would return INCOMPLETE status.
We already call drawValidate before any draw, the call was
missing at glCheckFramebufferStatus.
Change-Id: I94b930024c3076ac5fe0f8662d11884729c4f953
There is a flashing effect we see on Mac when switching between
drawables (pbuffer and a window) on the same context.
This change make the FrameBuffer object in the renderer to use
two different OpenGL contexts, one which is used while the pbuffer
drawable is bound and another which is used when the subwindow is
bounded during swap. The two contexts share textures so both contexts
have access to all allocated color buffers.
While this fixes the flashing effect on Mac, it should be more effective on
other platforms as well.
Change-Id: I63b49f0cdae8e2785765503750bbf5fbd4d963a1
Wrong share group was set up when two user contexts are sharing !
That means that shared context was not really worked before, now it does.
Change-Id: Ia4edbba0441e309902b2c48c44865342410bae3d
This make the EGL/GLESv1/GLESv2 libraries on the guest to use
the OPENGL bionic tls slot for faster tls access.
Note that we still setting the slower tls in order to have the
tls destructor which allow us to close the host connection
when the thread exits.
Gives slightly performance improvement, the avg score for the
teapot test in 0xBench goes from 34 to 37 fps. (This test is
near immediate mode since it renders a lot of small primitives
so it is a good candicate to be improved).
Change-Id: I9060c75cc29c2e28721fa11d3f318b438edb5da9
This change removes the 'blitFromPbuffer' path in the renderer
which uses binding pbuffer to texture. This path does not work
on all platforms and is not relavent any more. It simplifies
a lot the FBConfig filtering code.
Also, added a check to fail the renderer initialization if
the translator plugin does not support EGL_KHR_gl_texture_2d_image
and GL_OES_EGL_IMAGE (by both GLES implementations [1 and 2])
since our implementation requires it.
Change-Id: Ifd4c66943a8da56308cfb7e38914c962b4f5befe
That replaces the readback to host we do at the end of each frame
with two blits on the GPU, one to copy the pixels into a texture
and another to render the texture to another in order to flip
the image.
Change-Id: I7e0e10493d38944d0b613e245023f34236d3dfc4
This patch improves the performance of the TcpStream implementation
on the host by disabling the Nagle algorithm, thus improving the
bandwidth of small packets.
When used with the corresponding change in the emulator, this
significantly improves the speed of OpenGLES emulation for many
applications.
Change-Id: Ic09e51ecddf04bc7d667f46e1d260030d5fcad88
This is a workaround to a driver bug on Intel/Linux, when destroying
a subwindow do not close the X display connection, keep it for the
next call to createSubWindow. It happens that the intel/linux driver
caches the display connection for each window ID and use it during
swap or something similar. Using the same display connection for
all subwindows workaround the problem.
Also added a wait loop to wait until the window becomes mapped
instead of just calling XSync after calling XMapWindow. This
is more accurate and make sure we do not miss the first window
refresh.
Change-Id: I63c69d736df433666a8c79d71127afe72776e89f
on eglCreateContext the translator created a window for the
specified config with a matching pixelformat and saved it's
device context for future use.
however when the thread exited the window was destroyed by windows
and the handle saved became invalid.
the next call to wglCreateContext with the handle saved failed.
fixed the problem by creating windows per thread and not in a
global pool.
each thread will create its own window for each config when asked
the window will be destroyed when the thread exits.
Change-Id: I63668a20e8c30c42c4cfcaca664938683600b28e
our framebuffer implementation in gralloc does not support
setUpdateRect but we erronously publish as we do. Reset
the setUpdateRect so that flinger will not try to use it.
Change-Id: I2e4f2ef7e5701aed69caa194426945bafd5ac4c9
Now that we are not running the renderer as a seperate process
we need to have the XInitThreads call we put in the renderer process
inside emulator_test_renderer since the render server runs inside that
thread.
NOTE FOR INTEGRATION: this will need to be added to the emulator as
well.
Change-Id: I70592ad976d08c86e2c9e75b11b6ad57d6067606
Changes the render_api to support re-creating the display
sub-window and the ability to rotate the displayed framebuffer
image.
That currently works only when the renderer runs as a thread
and not as seperate process. Therefore we setup the library
to run the renderer as thread(s) inside the calling process
on *all* platforms for now.
Change-Id: Ifd009db903759042a7edcf89866d3c3fe076cae9
On Mac it happens that when a context is bound to a pbuffer
and you want to bind it to a window (NSView) instead you must
release it from the pbuffer before binding the window by calling
clearDrawable handle of NSOpenGLContext.
This change added an override of NSOpenGLContext in order to track
to which drawable type the context was previously bound and
call clearDrawable when necessary.
Change-Id: Iece5ab16a46aa0d107ccb773986a6b280d09d181
Workaround to intel/linux driver issue.
Desktop OpenGL does not support precision qulifiers in shaders,
so we remove them by defining "precision" "mediump" "lowp" and
"highp" to blank.
That translates the line "precision mediump float;" to "float;" which
is a valid command according to the GLSL spec however it fails
to compile using the intel/linux driver.
This change adds a parser which removes completely the
precision command from the shader source.
Change-Id: I29a16bb78ac0000d624aa2f83e93ef5c6ad60345
Fixed the location shift WAR to work for arrays that starts at
location 0. Fixed conformance regressions made by this workaround.
The conversion should be smarter than just shifting 16-bits back
and forth, it should take into account if array element is being
accessed.
Change-Id: Icb746c67e16edfacb8264a1e687fd24ac6e868e7
That fixes the event injection code to build correctly on
Windows.
Also renamed SDL_{CFLAGS,LDFLAGS} used in the Android.mk
to have a LOCAL_ prefix.
Change-Id: Ie82838ffb931880e700322116c3190fbb754d889
on some platforms info log length of programs and shaders was
not returned correctly, we now cache the info log for these objects
and calculate info log length intenrally to be both correct and
compliant to the spec and conformance tests.
Change-Id: I017932482ba36f9e2e1858fe372683f7ac3a98ca
In GLES, a vertex shader attribute can be at location 0 and have
a current value. In OpenGL, the spec is not clear, resulting in
absurdities like the ATI driver binding an attribute to location
0, does not give an erro when you set it, but gives an error when
you try to get it back. And it doesn't actually set the value in
the shader.
So, in this patch we:
1. Track attribute 0 value internally, setting and getting it
as necessary.
2. Upon glDrawArrays and glDrawElements, if attribute 0 is not
"array enabled" (that is, it should use a current value) we
create a dummy array, fill it with the intended current value,
and attach and enable it. After the draw, we disable it.
Change-Id: I35f3e8a924e6fba236f4f4d85423b04ae448dad4
adding more params validations to some GL functions
which the OpenGL layer below us should check , but using Intel
Graphic driver those checks are ignored
Change-Id: I3ca2a891f1d9f988a7b662d6bc32b81e12cc333d
fixin a bug when we declared that our implementation supports
OES_standard_derivatives without checking that the
openGL layer below supports this extention
Change-Id: Ic74f18eb753f0cfe067e23c3bf83905e40b17665
add special handling of locations smaller or equal to 0
do not translate location -1 (this location means the
uniform is not found)
location 0 does not need translation, it has no effect
Change-Id: Idbde51f08433ed70a5a1a2cf1ede51043f3cca76
fix glGetAttachedShaders
NULL is allowed in count parameter
fix problem with nvidia driver which returned invalid count
Change-Id: Icfaa5688b04325de2890486eb87b163babaddf30
