OSDN Git Service
Lingfeng Yang [Fri, 22 Jul 2016 22:21:58 +0000 (15:21 -0700)]
More efficient swapBuffers
Previously, we were using glFinish() in the host to make sure
that a buffer that we are presenting has been completely written to.
That is bad for performance, because it usually means to block the CPU
until the GPU's command queue is completely empty and all operations
have completed.
This CL uses eglCreateSyncKHR with a native fence FD to guarantee
the draw buffer is finished before presenting. The idea is that
after rcFlushWindowColorBufferAsync has queued up GPU commands to
finish up the draw buffer, eglCreateSyncKHR follows that up with a fence
command.
The fence command in turn is represented by a FD (by providing
EGL_SYNC_NATIVE_FENCE_ANDROID as the sync object type). Giving that FD
to queueBuffer tells SurfaceFlinger/BufferQueue not to do anything
that depends on the contents of that buffer until the fence command
has completed.
Theoretically, we are supposed to wait on the FD coming from dequeueBuffer
as well, to make sure no one else is using it, but in practice,
it matters much less; we will just overwrite the contents anyway, and it
has not proved to be any problem in terms of visual artifacts.
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570 <- needed
https://googleplex-android-review.git.corp.google.com/
1219571 <- this CL
Change-Id: I3dd1b69c77665ac0143bd5a302e6a5a04736c9ea
Lingfeng Yang [Thu, 14 Jul 2016 00:44:03 +0000 (17:44 -0700)]
Implement EGL_KHR_fence_sync/EGL_ANDROID_native_fence_sync for emulator
The core purpose of these two extensions is to provide a way for one thread
to know when an EGL command has finished processing in another thread,
so that it is safe to do work that touches common memory regions,
such as frame buffers in a concurrent draw/present setup.
eglCreateSyncKHR inserts a fence command into the current context's
command queue, and returns a handle to that command. eglClientWaitSyncKHR
takes such a handle as input and blocks the CPU until that fence command
is completed.
For the emulator, we implement these commands by wrapping underlying
host implementations. That is enough for EGL_KHR_fence_sync, but we also
include an interface to the Goldfish Sync driver in order to have
EGL_ANDROID_native_fence_sync. Goldfish Sync driver adds the ability for
these fence commands and their completed/not completed state
to be represented as FDs, which the rest of the Android platform
uses for doing graphics asynchronously.
Related specs:
https://www.khronos.org/registry/vg/extensions/KHR/EGL_KHR_fence_sync.txt
https://www.khronos.org/registry/egl/extensions/ANDROID/EGL_ANDROID_native_fence_sync.txt
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570 <- uses this
All changes above, including host-side, are needed for this to work.
The following CL uses the functionality in this one:
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: I9bf0ba0994b2e8689a6bed9c5eeee206a5e21f45
Lingfeng Yang [Thu, 14 Jul 2016 00:24:59 +0000 (17:24 -0700)]
Add backing implementation of EGLSyncKHR objects
For EGL_KHR_fence_sync and related extensions in the emulator,
we'll need an object that can track host-side sync objects
along with any Android sync framework fence fd's that
may be produced, and be able to maintain the type and status
of sync objects (generally, EGL fence vs. EGL fence + fence fd,
and signaled vs. unsignaled).
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539 <- this CL
https://googleplex-android-review.git.corp.google.com/
1219570 <- uses this
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: Ic20d11500f3a6e7164c1a5919937b26a885b1987
Lingfeng Yang [Thu, 14 Jul 2016 00:11:47 +0000 (17:11 -0700)]
Add DEBUG_EGL / DPRINT macros for debug logging EGL.
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538 <- this CL
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570 <- uses this
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: I3041732beffb3bbc92196a082dc73d3dd74a0c4e
Lingfeng Yang [Thu, 14 Jul 2016 00:05:09 +0000 (17:05 -0700)]
Goldfish sync driver user-space interface
goldfish_sync.h implements the ioctls that interface
with the goldfish sync driver and the goldfish sync
virtual device on the host.
goldfish_sync_open() tells the kernel driver that
there is a new context in which sync work
may be required. This is represented by a user-space
FD. Note that this is not a timeline FD
as with SW_SYNC_USER.
The intention is to associate each EGL context
with a sync context, reflected in getGoldfishSyncFd()
in the EGLContext_t class.
The IOC_QUEUE_WORK ioctl may return a valid fence FD
that can be used by the system to do work while
waiting for that fence FD to be signaled.
In order to avoid bad behavior, timelines and sync fence objects
are still completely maintained and signaled by the
kernel driver and underlying host system.
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
In particular, the ioctl() implementation is listed in
https://android-review.googlesource.com/238399
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537 <- this CL
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570 <- uses this
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: I74b92415acd95b4bf2ec8aa3a6ff5392d86cf0b3
Lingfeng Yang [Wed, 13 Jul 2016 23:27:43 +0000 (16:27 -0700)]
Add encoder entries for EGL_KHR_fence_sync
- also update encoder templates to latest version
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535
https://googleplex-android-review.git.corp.google.com/
1219536 <- this CL
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570 <- uses this
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: Ief436ccf8947d35c4b7719dc9ef6def4488cb3f0
Lingfeng Yang [Wed, 13 Jul 2016 23:26:10 +0000 (16:26 -0700)]
ANDROID_EMU_ASYNC_SWAP extension searching + setting
In order for the async frame swapping feature to work
with all combinations of new/old emulators and system images,
one way to do it is to introduce an emulator-specific GL
extension. This is the same way the checksum calculation
support works.
This is a port to device/generic/goldfish-opengl, based on
https://googleplex-android-review.git.corp.google.com/#/c/
1065345/
This is part of a sequential, multi-CL change. There is also
a corresponding multi-CL change on the host side:
https://android-review.googlesource.com/#/q/topic:emu-glsync-host
The changes in the system image are as follows:
platform/build:
https://googleplex-android-review.git.corp.google.com/
1024926
device/generic/goldfish:
https://googleplex-android-review.git.corp.google.com/
1230942
device/generic/goldfish-opengl:
https://googleplex-android-review.git.corp.google.com/
1219535 <- this CL
https://googleplex-android-review.git.corp.google.com/
1219536
https://googleplex-android-review.git.corp.google.com/
1219537
https://googleplex-android-review.git.corp.google.com/
1219538
https://googleplex-android-review.git.corp.google.com/
1219539
https://googleplex-android-review.git.corp.google.com/
1219570
https://googleplex-android-review.git.corp.google.com/
1219571
Change-Id: Id07a72e5cb3397ff9d0b55788b8ad763c15050f3
Yahan Zhou [Fri, 22 Jul 2016 20:51:14 +0000 (13:51 -0700)]
Fix gralloc pipe and add flush to rcCloseColorBufferPuid
There is a typo in gralloc pipe. Also rcCloseColorBufferPuid needs to be
flushed to function correctly.
Change-Id: Ic006578f560dede0bdc073d6ab862e588b71e452
Colin Cross [Thu, 21 Jul 2016 04:47:57 +0000 (04:47 +0000)]
Merge "Fix master build"
Christopher Ferris [Thu, 21 Jul 2016 02:09:47 +0000 (02:09 +0000)]
Merge "Revert "Use the new <system/qemu_pipe.h>.""
Christopher Ferris [Thu, 21 Jul 2016 02:08:53 +0000 (02:08 +0000)]
Revert "Use the new <system/qemu_pipe.h>."
This reverts commit
a4b8bbbcc63dd03305bcdb6acf4d916da2889206.
Change-Id: I3bd5b8b88c9462f7af744c6ec626ff7fe0432dc2
Christopher Ferris [Thu, 21 Jul 2016 01:45:04 +0000 (01:45 +0000)]
Merge "Use the new <system/qemu_pipe.h>."
Christopher Ferris [Thu, 21 Jul 2016 01:42:55 +0000 (18:42 -0700)]
Use the new <system/qemu_pipe.h>.
The old qemu_pipe.h was moved so change to use the new location of the
header.
Change-Id: I032a9444a430f3426ec430fff27f6b117443b5ca
Yahan Zhou [Thu, 21 Jul 2016 00:15:15 +0000 (17:15 -0700)]
Fix master build
Change-Id: Ib81144748f520a9812c98c3cd23f114ba490d297
Yahan Zhou [Wed, 20 Jul 2016 23:26:20 +0000 (23:26 +0000)]
Merge "Fix jb build"
Yahan Zhou [Wed, 20 Jul 2016 23:12:02 +0000 (16:12 -0700)]
Fix jb build
Change-Id: I31b9e3e44523819f54da362e688b2c5098ec27bd
Bo Hu [Wed, 20 Jul 2016 21:37:20 +0000 (21:37 +0000)]
Merge "Fix jb arm boot"
Yahan Zhou [Wed, 20 Jul 2016 21:31:33 +0000 (14:31 -0700)]
Fix jb arm boot
Change-Id: Ibf124424f703cc30ee7b5db7c3589b69728a4c32
Yahan Zhou [Tue, 19 Jul 2016 23:27:32 +0000 (23:27 +0000)]
Merge "Misc fix in gralloc pipe"
Yahan Zhou [Mon, 18 Jul 2016 23:14:21 +0000 (16:14 -0700)]
Misc fix in gralloc pipe
Fix size and have a confirmation message in gralloc pipe. Also fix the
naming of puid.
Change-Id: I0b87b9e6bcf88172792705ac9643e150b227be96
Yahan Zhou [Mon, 18 Jul 2016 23:03:25 +0000 (23:03 +0000)]
Merge "Create pre-process gralloc pipe on the guest"
Yahan Zhou [Mon, 18 Jul 2016 23:03:08 +0000 (23:03 +0000)]
Merge "Make goldfish-opengl compatible with master"
Yahan Zhou [Mon, 11 Jul 2016 18:51:43 +0000 (11:51 -0700)]
Create pre-process gralloc pipe on the guest
https://buganizer.corp.google.com/issues/
29457657
This patch create a pre-process pipe in gralloc. It tells the host which
process holds which color buffer, so that on process exit the host can
correctly reduce the reference counters of all gralloc color buffers.
Each process will get a 64bit identifier from the host, to help the host
identifying it.
It works with host patch:
https://android-review.googlesource.com/#/c/246823/
Change-Id: I8d9a512a1f151f4f29a3b318ccf17d62d8f46512
Yahan Zhou [Mon, 18 Jul 2016 19:20:04 +0000 (12:20 -0700)]
Make goldfish-opengl compatible with master
Change-Id: I103424ff9e2bfdad8181850a216b7a0ee0290267
Chih-Hung Hsieh [Thu, 9 Jun 2016 21:02:30 +0000 (14:02 -0700)]
[cherry-pick] Fix misc-macro-parentheses warnings in goldfish/opengl.
Add parentheses around macro arguments used beside operators.
Use NOLINT to suppress warning or clang-tidy can put parentheses at
wrong places.
Bug:
28705665
cherry-pick from https://android-review.googlesource.com/#/c/237790/
Change-Id: I40653d3e617f84c07b308bb9c2869219f6563706
Test: build with clang-tidy
Tina Zhang [Mon, 7 Sep 2015 02:47:07 +0000 (10:47 +0800)]
[cherry-pick] Delete some useless files
Currently, some files are not used by the guest part of
GL emulation any more.
cherry-pick from https://android-review.googlesource.com/#/c/169695/
Change-Id: I1fb87614aaba19be6863d6fb85e80937d0778384
Signed-off-by: Tina Zhang <tina.zhang@intel.com>
Christopher Wiley [Fri, 12 Feb 2016 20:37:09 +0000 (12:37 -0800)]
[cp] Build camera related goldfish HALs as brilloemulator for Brillo
Brillo runs a reduced subset of the graphics stack and a different
system configuration from goldfish. However, many of these HALs still
make sense for Brillo. Build the HALs with a board name that will allow
the camera and graphics subsystems to pick them up.
Bug:
26936651
Test: brilloemulator is able to pick up and use these HALs to take
fake pictures.
cherry-pick from:
https://android-review.googlesource.com/#/c/202700/
Change-Id: I79265ada33a5a97c6f670b79f9c239606465c9fc
Christopher Wiley [Fri, 12 Feb 2016 20:32:07 +0000 (12:32 -0800)]
[cherry-pick] Refactor gralloc HAL to not repeat defintions
Bug:
26936651
cherry-pick from:
https://android-review.googlesource.com/#/c/202619/
Change-Id: If11a2af0496cdaf86e3bc87ded37fc8d93b7fa16
Miroslav Tisma [Wed, 23 Sep 2015 10:54:56 +0000 (12:54 +0200)]
[cherry-pick] GPU: Build gralloc.ranchu.so
cherry-pick from:
https://android-review.googlesource.com/#/c/172935/
Change-Id: Ie449043b06b6fefa52bdbd68c243a1165ec95ee1
Yahan Zhou [Fri, 1 Jul 2016 22:47:21 +0000 (15:47 -0700)]
Make it compatible with ics-mr1
Change-Id: I8661fa4cc93111dfdb5fe9fa73f8a47b2820f615
Yahan Zhou [Thu, 30 Jun 2016 21:16:16 +0000 (21:16 +0000)]
Merge "Make opengl compatible with gingerbread"
Yahan Zhou [Thu, 30 Jun 2016 21:09:10 +0000 (21:09 +0000)]
Merge "Revert "Workaround: disable checksum for a problem in glTexSubImage2D""
Yahan Zhou [Thu, 23 Jun 2016 20:58:09 +0000 (13:58 -0700)]
Make opengl compatible with gingerbread
This patch makes mnc opengl compatible with gingerbread. With this
patch, you can copy-paste the opengl folder to gingerbread and it will
just build.
Notice that in gingerbread, the GL folder is in
development/tools/emulator
Change-Id: I99f060edd517a7afd425fd2103cbceaa61f0fc82
Yahan Zhou [Fri, 3 Jun 2016 00:24:03 +0000 (00:24 +0000)]
Revert "Workaround: disable checksum for a problem in glTexSubImage2D"
The checksum enable/disable should be control by feature control now.
This reverts commit
2a5a09abd50933d7cc4b1302769a19ffdf9d104b.
Change-Id: I8f33718f7b2112fd6abcfd51fadc336e9bd15a3a
Yahan Zhou [Mon, 20 Jun 2016 19:28:34 +0000 (12:28 -0700)]
Make GL build script compatible with jb (and mr1.1, mr2)
Notice that in jb, the GL encoder is located in sdk/emulator/opengl. In
jb-mr1.1, the encoder is in development/tools/emulator/opengl.
After this patch, it should just work if you copy the opengl directory
from mnc to jb.
Change-Id: I4b6ba6daeb64315ecf72b52b142acbbccdc4dbe0
Yahan Zhou [Tue, 28 Jun 2016 00:00:11 +0000 (00:00 +0000)]
Merge "Revert "Temporarily rename makefile to avoid building duplicate opengl libs.""
Yahan Zhou [Mon, 27 Jun 2016 22:48:39 +0000 (22:48 +0000)]
Revert "Temporarily rename makefile to avoid building duplicate opengl libs."
This reverts commit
8b2652824b14889092d41d263762b9f3bdf6f86c.
Change-Id: I93ceaea56f8ef4ed90a73af5b04f7a1fed0083e0
Yahan Zhou [Mon, 27 Jun 2016 22:16:40 +0000 (22:16 +0000)]
Merge "Fix 3Dmark crash - preserve more host connections"
Lingfeng Yang [Fri, 24 Jun 2016 02:06:28 +0000 (19:06 -0700)]
Fix 3Dmark crash - preserve more host connections
The CL
https://googleplex-android-review.git.corp.google.com/#/c/
1131769/
for killing host connections was too aggressive.
It incorrectly assumed that if eglInitialize was not called in the
current thread, it'd be treated as a gralloc-only connection.
It's perfectly possible to create a new thread,
not call eglInitialize, and call eglCreateContext instead,
which is within proper usage since after there is supposed
to be only one display per process.
This situation resulted in 3DMark crashing, because once the
connection is killed on a thread that is not actually gralloc-only,
all the state goes away and segfault is just a matter of time.
This CL makes it so that eglCreateContext / eglMakeCurrent
mark the connection as not gralloc only, preventing the
connection from being erroneously killed.
The behavior where no extra render threads are created
when opening recents is still preserved.
(Cherry-pick from
ebff5363850076de2623b6f6a0ff035ea3e823d5)
Change-Id: I20a204aefca91e5ac8f6b097d9fcd799f6b02583
Prathmesh Prabhu [Mon, 27 Jun 2016 20:30:13 +0000 (13:30 -0700)]
Temporarily rename makefile to avoid building duplicate opengl libs.
We've yet to drop the opengl code from the old location. Turns out that
simply including this new source location leads the Android build system
to build the lib from both the old and new location.
Renamed the Android.mk file here to avoid building from this location
until the migration is complete.
Change-Id: I66b5f4201f99b8a17c814918e5b42489d686d7c5
Surendra Vispute [Fri, 24 Jun 2016 00:24:14 +0000 (17:24 -0700)]
Merge opengl/ from device/generic/goldfish to /
Yahan Zhou [Wed, 22 Jun 2016 23:13:17 +0000 (23:13 +0000)]
Merge "Fix glTexSubImage2D" into mnc-emu-dev
Yahan Zhou [Wed, 22 Jun 2016 22:36:04 +0000 (15:36 -0700)]
Fix glTexSubImage2D
This is to port
https://googleplex-android-review.git.corp.google.com/#/c/
1145109
into mnc
Change-Id: Ic53c1181ab1c69fb33c5758034c0c4ec131b010e
Yahan Zhou [Wed, 22 Jun 2016 22:26:08 +0000 (15:26 -0700)]
Fix build scripts for klp
The previous build script makes klp fail to boot. This patch fixes it.
Change-Id: I091fdd29ef3f37d10d5f56c81e9ba17618214ebf
Yahan Zhou [Sat, 18 Jun 2016 00:40:14 +0000 (17:40 -0700)]
Make build script compatible with klp and up
It gets rid of c++11 and makes build script compatible with klp and up.
It also includes GL headers here, because headers in framework/native
might be different across API levels.
Change-Id: Ida425416a392cef878256965d69b54afe42b7263
Bill Rassieur [Mon, 20 Jun 2016 22:30:41 +0000 (15:30 -0700)]
Initial empty repository
Yahan Zhou [Mon, 20 Jun 2016 19:58:05 +0000 (19:58 +0000)]
Merge "Fix EGL cleanup code for CTS android.nativeopengl" into mnc-emu-dev
Yahan Zhou [Fri, 10 Jun 2016 00:54:34 +0000 (17:54 -0700)]
Fix EGL cleanup code for CTS android.nativeopengl
In CTS, android.nativeopengl.EGLCleanupTest tries to repeatatively
create and destroy egl context for 1000 times to check for memory leaks.
This patch releases resources when destroying contexts.
It turns android.nativeopengl.EGLCleanupTest from failure to flaky,
because of two other issue:
(1) There is a 1 min timeout in the test:
cts/tools/tradefed-host/src/com/android/cts/tradefed/testtype/
WrappedGTest.java
Our emulator can barely finish it within 1 min.
(2) There is another bug that may trigger UI rendering events after the
test begun, which crashes randomly because this test releases rendering
resources. It needs to be fixed in another patch.
Change-Id: Id285f526149638c63eb78c167c7d2daef7bd2412
Yahan Zhou [Mon, 20 Jun 2016 16:58:21 +0000 (16:58 +0000)]
Merge "Regenerate gl encoder from qemu" into mnc-emu-dev
Yahan Zhou [Fri, 17 Jun 2016 22:06:37 +0000 (15:06 -0700)]
Regenerate gl encoder from qemu
Regenerated from:
https://android-review.googlesource.com/#/c/237292/
Change-Id: I7972c4fddfb9765af39856ccd8aa29ac67558301
Lingfeng Yang [Mon, 13 Jun 2016 15:39:28 +0000 (08:39 -0700)]
Don't leave gralloc-only host connections hanging.
bug:
29318455
Memory leaks are becoming more of a problem, especially
with host machines that don't have much RAM and when
we use the emulator for long periods of time.
This CL fixes the problem detailed in the bug link,
which is that the number of render threads
(and thus thread-specific storage) skyrockets when
repeatedly pressing the Recents button while an app is active.
The thing is that there are lots of threads
that never use the EGL display, just gralloc, in order
to do gralloc things like update the screenshot display
of apps in recents.
These threads are not cleaned up because a thread pool
is often used to manage these sorts of tasks.
If a new or different thread suddenly wants to update
the recents screenshots through gralloc, a new connection
will be established.
This CL uses gralloc_unregister_buffer as a signal
for when a thread might be done with gralloc in general,
exiting the host connection immediately.
Of course, there may be other things that the thread wants
to do, but then it can just go ahead and create a new connection.
We also use the heuristic that we do not exit on gralloc_unregister_buffer
when the thread seems to be a bona fide rendering thread
(eglInitialize has been called).
There are still more leaks going on, this fixes just
one source of them.
After this CL, the render thread count should be held constant
when repeatedly starting apps and going to recents.
Change-Id: I5f57539429f149e3a938467e3b82fa911ab118c3
Bjoern Johansson [Mon, 13 Jun 2016 23:09:20 +0000 (16:09 -0700)]
Use correct format when reading color buffers
When reading back data from color buffers on the host make sure gralloc
uses the correct format and type parameters. Previously it was hardcoded
based on RGBA at 8 bits per component which would cause incorrect reads
for other pixel formats. This could cause QEMU pipe reads to call abort
because the reads went past a page boundary if the calculated size
exceeded the actual buffer size.
Change-Id: Ifa61f0a97d79577744b899b1bbb4120a0b9170eb
Yahan Zhou [Mon, 6 Jun 2016 19:42:17 +0000 (19:42 +0000)]
Merge "Move stream->flush to non-generated code" into mnc-emu-dev
Yahan Zhou [Mon, 6 Jun 2016 19:30:39 +0000 (12:30 -0700)]
Move stream->flush to non-generated code
Move flush command out of auto-generated code.
Change-Id: I3f3c8ca49344ea4c5357d7d6953364ce1015ea99
Yahan Zhou [Mon, 6 Jun 2016 17:51:45 +0000 (17:51 +0000)]
Merge "[cherry-pick] Lose HAVE_ANDROID_OS in goldfish." into mnc-emu-dev
Yahan Zhou [Thu, 2 Jun 2016 00:24:28 +0000 (17:24 -0700)]
Check vertex buffer size before glDraw*
This patch verifies if a vertex buffer is big enough before calling
glDraw*. It fixes crashes when trying to index outside of vertex buffer.
It should make it less crashy in glDrawElements and glDrawArrays:
https://code.google.com/p/android/issues/detail?id=207284
Change-Id: I36f0f568d96b576f0e0975435d2d7fd41cd9e732
Elliott Hughes [Wed, 12 Aug 2015 21:30:38 +0000 (14:30 -0700)]
[cherry-pick] Lose HAVE_ANDROID_OS in goldfish.
cherry pick from:
https://android-review.googlesource.com/#/c/166427
Change-Id: Id626aee8d65e0ded543ac12416c37f938d196676
Yahan Zhou [Wed, 1 Jun 2016 19:49:39 +0000 (12:49 -0700)]
Fix mnc build
Missing a bracket when cherry-pick.
Change-Id: I5202bf4dc5c64cb4663fdcadcbf6704272708b42
Lizhe Liu [Wed, 27 May 2015 03:13:23 +0000 (11:13 +0800)]
[cherry-pick] Optimize glGetIntegerv to reduce the cost of translating
Hardware GLES emulation in the Android platform is implemented
with a mix of components. Translator module is one of the most
important components in emulator. It translate the function
calls into calls to the appropriate desktop APIs and cost lots
of reference to complete it. So optimization must be beneficial
to us. GL_MAX_CUBE_MAP_TEXTURE_SIZE, GL_MAX_RENDERBUFFER_SIZE
and GL_MAX_TEXTURE_SIZE are the rough estimate value of the
largest texture which the OPENGL can handle. All of them are
constant value. So there is no need for us to translate them
from guest os to host os every time. We can cache the value
after first time and reuse it next time.
Cherry-pick from:
https://android-review.googlesource.com/#/c/151899
Change-Id: I1a0352c40b667bafd6ec1fb66a8060d7fe4a415d
Signed-off-by: Lizhe Liu <lizhe.liu@intel.com>
Yahan Zhou [Wed, 1 Jun 2016 17:38:02 +0000 (17:38 +0000)]
Merge "[cherry-pick] opengl: GLEncoder: Fix potential array buffer misbinding" into mnc-emu-dev
Yu Ning [Wed, 19 Aug 2015 07:22:13 +0000 (15:22 +0800)]
[cherry-pick] opengl: GLEncoder: Fix potential array buffer misbinding
When encoding a glDrawElements() or glDrawArrays() call, the encoder
needs to figure out whether each array (e.g. vertex, color, etc.) uses
a bound buffer or not. It is possible that some arrays use bound buffers
while others immediate buffers. In this case, there is a chance that the
encoder fails to preserve the semantics of the input GLES calls,
resulting in misinterpretation by the translator that all arrays use
bound buffers.
Fix this by encoding an unbind operation before any glXXXPointer() call
that does not mean to use a bound buffer. For GLESv2, that is already
done:
40a674e Fix emulator crashes on glDrawElements command
This CL does the same for GLESv1.
Cherry-pick from:
https://android-review.googlesource.com/#/c/167443/
Change-Id: Ifd17321b124a4f4b1037a43d5a54094541d367f1
Signed-off-by: Yu Ning <yu.ning@intel.com>
Lingfeng Yang [Wed, 25 May 2016 15:31:06 +0000 (08:31 -0700)]
Don't optimize out rcUpdateColorBuffer
bug:
28406160
If we want to use rcUpdateColorBuffer to
synchronize on the host in order to remove
out of order webcam frames, we need to call it
if (cb->hostHandle), not just when
the width and height are nonzero.
Otherwise, we get deadlocks.
Change-Id: Ib98774b935e8786e307dd38711a0f2f9ffea440e
Lingfeng Yang [Fri, 27 May 2016 00:32:31 +0000 (00:32 +0000)]
Merge "Save a few glBindBuffer calls" into mnc-emu-dev
Lingfeng Yang [Thu, 26 May 2016 21:18:33 +0000 (14:18 -0700)]
Save a few glBindBuffer calls
When sending vertex attributes, we only need to
bind the context-wide GLClientState's VBO
if a particular vertex attrib state is enabled.
This CL makes it so we call glBindBuffer only if
the currently bound vbo changes, in many cases
completely eliminating extra glBindBuffer calls
in sendVertexAttributes. Each extra glBindBuffer
call takes a few microseconds.
Change-Id: I04de3be0114ec6df10e67bc801b6d089bf5096db
Lingfeng Yang [Mon, 23 May 2016 19:16:04 +0000 (12:16 -0700)]
Refactor GL extension getting part in checksum setup.
Change-Id: Ieb26ce5682a87821935a5b639e5242942d13ac21
Adam Buchbinder [Fri, 20 May 2016 22:25:59 +0000 (15:25 -0700)]
Fix memory-handling issues in the OpenGL code.
This fixes the folloiwng cppcheck reports:
[device/generic/goldfish/opengl/system/GLESv2_enc/GL2Encoder.cpp:789]:
(error) Mismatching allocation and deallocation: str
[device/generic/goldfish/opengl/system/GLESv2_enc/GL2Encoder.cpp:795]:
(error) Mismatching allocation and deallocation: str
[device/generic/goldfish/opengl/tests/gles_android_wrapper/egl.cpp:492]:
(error) Mismatching allocation and deallocation: attrib
Change-Id: Iaf494eca1b927b2fa4b91f01ba340746b9da9bd3
(cherry picked from commit
3375e5e21450b78daf109225707d574976b7ec1c)
Yahan Zhou [Tue, 17 May 2016 16:56:04 +0000 (16:56 +0000)]
Merge "Sends tex coord only when the texture is enabled in gles1" into mnc-emu-dev
Yahan Zhou [Mon, 16 May 2016 23:23:22 +0000 (16:23 -0700)]
Sends tex coord only when the texture is enabled in gles1
Only send tex coord if the corresponding texture unit is enabled.
This is to prevent a crash in An3DBench:
https://b.corp.google.com/issues/
9631994
Change-Id: I23ae030adba3a3fa05e00f63c96040b439b86baf
Bjoern Johansson [Thu, 12 May 2016 17:38:27 +0000 (10:38 -0700)]
Create and read color buffer when expecting reads
Make sure color buffers are created when gralloc is allocating buffers
that indicate that software reads are going to happen. In this case
software will require the rendered data to be available from the host at
a later point and that means that rendering has to be stored in a color
buffer. Also make sure that when gralloc locks a buffer for reading it
will actually read the data from the host if the buffer is marked for
reading.
Change-Id: Ief4b7fe34401727cabec9fdd5c2b9812a6c49612
Yahan Zhou [Wed, 11 May 2016 23:04:02 +0000 (23:04 +0000)]
Merge "Cleanup debugging code in a previous CL" into mnc-emu-dev
Yahan Zhou [Wed, 11 May 2016 22:54:56 +0000 (15:54 -0700)]
Cleanup debugging code in a previous CL
Clean up my logs.
Change-Id: Ia447994af03358964bb39330824e8da8eeebf08b
Yahan Zhou [Wed, 11 May 2016 22:34:26 +0000 (22:34 +0000)]
Merge "Add framebuffer checks to pass dEQP" into mnc-emu-dev
Yahan Zhou [Wed, 11 May 2016 22:06:34 +0000 (15:06 -0700)]
Add framebuffer checks to pass dEQP
They are defined in GLES3 compatible context, but used in dEQP-GLES2
tests. So we need to enable them.
Should be used with this CL in the host:
https://android-review.googlesource.com/#/c/229251
dEQP-GLES2.functional.fbo.completeness.renderable failures 21->0
Change-Id: Ieacbbcc4afa688711c4d090971036505a74cc1ab
Yahan Zhou [Wed, 11 May 2016 17:54:59 +0000 (17:54 +0000)]
Merge "Allow rgb10_a2 internal format in renderbuffer" into mnc-emu-dev
Yahan Zhou [Wed, 11 May 2016 17:45:31 +0000 (10:45 -0700)]
Allow rgb10_a2 internal format in renderbuffer
dEQP-GLES2.functional.fbo.completeness.renderable.renderbuffer failures
3->0.
Change-Id: If5b94f5b1edc182ec5ccf2bd3321306a826e757e
Yahan Zhou [Mon, 9 May 2016 16:48:23 +0000 (16:48 +0000)]
Merge "eglCreateContext returns error code when taking wrong attributes" into mnc-emu-dev
Yahan Zhou [Thu, 5 May 2016 01:10:46 +0000 (18:10 -0700)]
eglCreateContext returns error code when taking wrong attributes
CTS.android.nativeopengl.EGLCreateContextTest fail -> pass
Change-Id: Iec95f568b65cd60cd78cc5eb80ce3ff0d1ae1731
Lingfeng Yang [Wed, 4 May 2016 16:12:53 +0000 (16:12 +0000)]
Merge "Flush draw calls" into mnc-emu-dev
Lingfeng Yang [Mon, 2 May 2016 17:47:00 +0000 (10:47 -0700)]
Flush draw calls
Increases Antutu Garden performance by 2-4 FPS,
depending on if host renderthread uses lock
(faster if no lock)
The idea is that if the draw calls are not flushed,
they may be called too late on the host,
resulting in lower performance:
the GPU is just sitting idly between pipe flushes
for more instructions.
Change-Id: I78e278fc9b30b4a135b48f56d912b6aba8aadea7
Evan Birenbaum [Thu, 28 Apr 2016 21:07:02 +0000 (21:07 +0000)]
Merge "FBO completeness check in glDrawElements." into mnc-emu-dev
Evan Birenbaum [Thu, 28 Apr 2016 19:47:57 +0000 (12:47 -0700)]
FBO completeness check in glDrawElements.
dEQP-GLES2.functional.negative_api.vertex_array.*:
3 Fail -> 0 Fail
The host framebuffer (not the guest state) is not complete in these
tests, so a call to the host glCheckFramebufferStatus is needed to
confirm that. The cached state is unreliable, since things may have
changed since the last guest call to glCheckFramebufferStatus.
Change-Id: I2cf026129c871df68b328949e9c79b3f5df590eb
Yahan Zhou [Thu, 28 Apr 2016 18:19:12 +0000 (18:19 +0000)]
Merge "glActiveTexture should set GL_INVALID_ENUM when taking wrong input" into mnc-emu-dev
Lingfeng Yang [Mon, 25 Apr 2016 16:39:43 +0000 (16:39 +0000)]
Merge "Stricter FBO completeness checks" into mnc-emu-dev
Lingfeng Yang [Mon, 25 Apr 2016 16:39:36 +0000 (16:39 +0000)]
Merge "Validate presence of framebuffer attachment's object" into mnc-emu-dev
Yahan Zhou [Wed, 20 Apr 2016 23:02:10 +0000 (16:02 -0700)]
Support glDrawElements with GL_UNSIGNED_INT again
This is defined in GLES2 extension.
Change-Id: Id6ac22bb207a88ebb13282c443c0bc990f25cdf3
Yahan Zhou [Tue, 12 Apr 2016 17:02:38 +0000 (10:02 -0700)]
Workaround: disable checksum for a problem in glTexSubImage2D
This is a temporary workaround for a problem in glTexSubImage2D. It
disables the GL checksum to avoid crashes. Please revert this CL when
the root cause is solved.
Change-Id: I8401449e42da87f580f119982a1bcb1b6b9a4c8b
(cherry picked from commit
c1a6ab7b5ce1e23cf65a8e936860bcea879b6eef)
Lingfeng Yang [Tue, 12 Apr 2016 16:29:11 +0000 (09:29 -0700)]
Stricter FBO completeness checks
dEQP-GLES2.functional.fbo.completeness.*:
39 Fail -> 0 Fail
Based on similar logic in Swiftshader's libGLESv2/libGLESv2.cpp
Also cache the result of glCheckFramebufferStatus,
if already called from somewhere else.
Change-Id: I3fb42dd6902cfce4043d08fd459a7069ef491f7e
Lingfeng Yang [Fri, 8 Apr 2016 21:42:34 +0000 (14:42 -0700)]
Validate presence of framebuffer attachment's object
dEQP-GLES2.functional.negative_api.state.get_framebuffer_attachment_parameteriv
Fail -> Pass
dEQP-GLES2.functional.fbo.api.attachment_query_empty_fbo
Fail -> Pass
Track texture and renderbuffer object associations to
framebuffer attachment points.
Change-Id: Iff646bc57ef0202105768123cfa99384f37b01a3
Evan Birenbaum [Fri, 15 Apr 2016 22:03:51 +0000 (22:03 +0000)]
Merge "Sanity checks in vertex array functions." into mnc-emu-dev
Evan Birenbaum [Wed, 13 Apr 2016 19:53:09 +0000 (12:53 -0700)]
Sanity checks in vertex array functions.
dEQP-GLES2.functional.negative_api.vertex_array.*:
6 Fail -> 3 Fail.
dEQP-GLES2.functional.rasterization.*:
1 Fail -> 0 Fail.
The remaining vertex array failures will be fixed using some of
Frank's framebuffer status code.
Change-Id: Id8623029180beb6e792a3250763f08c4a03c7499
Lingfeng Yang [Fri, 15 Apr 2016 19:23:55 +0000 (12:23 -0700)]
Send glDrawArrays even if there is no vertex array
The spec says that if there are no arrays,
vertex attributes are to be used next.
dEQP-GLES2.functional.draw.random.50
Fail -> Pass
Change-Id: I81ab0bcac1c483436f8930212159a5079fe7b725
Lingfeng Yang [Fri, 8 Apr 2016 19:46:42 +0000 (12:46 -0700)]
Validation for glGetVertexAttribPointerv
dEQP-GLES2.functional.negative_api.state.get_vertex_attrib_pointerv
Fail -> Pass
Based on Intel's OpenGL validation for goldfish opengl,
introduced in nyc-dev
Change-Id: If67192368935f21de20c0f531036db027b867b37
Lingfeng Yang [Fri, 8 Apr 2016 19:38:35 +0000 (12:38 -0700)]
Validation for glGetVertexAttrib i/f v
dEQP-GLES2.functional.negative_api.state.get_vertex_attrib i/f v:
Fail -> Pass
Based on Intel's OpenGL validation for goldfish opengl,
introduced in nyc-dev
Change-Id: I9b539024e88a5afb7c78f8543e7c9ffff4f837e0
Lingfeng Yang [Fri, 8 Apr 2016 19:21:13 +0000 (12:21 -0700)]
Validation for glGetUniform i/f v
dEQP-GLES2.functional.negative_api.state.get_string:
Fail -> Pass
Based on Intel's OpenGL validation for goldfish opengl,
introduced in nyc-dev
Change-Id: I9c3b820a5fa6b3b3e4b5e8590cca590cf9ef86d2
Lingfeng Yang [Fri, 8 Apr 2016 19:13:45 +0000 (12:13 -0700)]
Validation for glGetString
dEQP-GLES2.functional.negative_api.state.get_string:
Fail -> Pass
Based on Intel's OpenGL validation for goldfish opengl,
introduced in nyc-dev
Change-Id: I74c480cf38a01e9c13f35ee3f58168415f7b6ab3
Lingfeng Yang [Mon, 4 Apr 2016 16:26:19 +0000 (09:26 -0700)]
More gracefulness on dequeueBuffer_* fail on flip
BUG: https://code.google.com/p/android/issues/detail?id=204176
Features like Instant Run will do things like abandon
the current set of buffers SurfaceFlinger gave to the UI,
which makes functions like eglSwapBuffers fail,
since they are dequeueing a buffer that is considered 'invalid'
(as it is before the instant run).
If we set the EGL error to EGL_BAD_ALLOC, this interacts badly
with libs/hwui/renderthread/EGLManager.cpp, causing
a LOG_ALWAYS_FATAL(...) to be called, which aborts the current process.
This CL sets the EGL error to EGL_BAD_SURFACE instead,
which allows hwui to recover.
Change-Id: I7cb5bf3b3834b3696981746f2319804c1adf6a25
Yahan Zhou [Mon, 21 Mar 2016 23:17:53 +0000 (23:17 +0000)]
Merge "Fix a typo in GLES encoder headers" into mnc-emu-dev
Yahan Zhou [Mon, 21 Mar 2016 23:11:38 +0000 (16:11 -0700)]
Fix a typo in GLES encoder headers
Unimportant but let's fix it.
Change-Id: I360ba0dc0b676238adaeb5d4360c10a9ec8c2371
Yahan Zhou [Fri, 18 Mar 2016 21:25:44 +0000 (21:25 +0000)]
Merge "Exit when QemuPipe is corrupted by GL" into mnc-emu-dev
Yahan Zhou [Wed, 16 Mar 2016 19:41:08 +0000 (12:41 -0700)]
Checksum for GLES messages from host to guest
Add checksum for GLES messages from host to guest. This patch contains
generated code (gl_enc.cpp, gl2_enc.cpp, renderControl_enc.cpp) and
shared code with the host (ChecksumCalculator.cpp,
ChecksumCalculator.h).
Change-Id: I577a34cae5208d3f11a165f861a69fa0971cf864