WebVR compositor bypass via BrowserMain context + mailbox

WebVR compositor bypass via BrowserMain context + mailbox

Implement compositor bypass for WebVR via mailbox copy:

- Create a GPU command buffer GL context in the VrShellGl thread that's backed
  by a native Surface

- SubmitFrame creates a mailbox and sends it as part of the mojo call

- On SubmitFrame, the GPU command buffer GL context converts the mailbox to a
  texture, and draws the content to the native Surface

- VrShellGl picks up the frame from the corresponding SurfaceTexture and
  proceeds with rendering as usual

Also implements on-the-fly canvas resizing, the transfer surface and GVR
renderbuffer are resized as needed to match the source canvas.

Added frame number as an argument to various TRACE_EVENTs for timing analysis.

It includes two separate modes for frame scheduling, the more aggressive one
is conditional on chrome:://flags#enable-webvr-experimental-rendering since the
framerate seems a bit less stable.

BUG=698643
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.android:android_optional_gpu_tests_rel;master.tryserver.chromium.linux:closure_compilation;master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

patch from issue 2729523002 at patchset 390001 (http://crrev.com/2729523002#ps390001)

[dcheng, 2017-03-07 10:00:04]

dcheng@chromium.org changed reviewers:
+ klausw@chromium.org

[dcheng, 2017-03-07 10:00:04]

Dubious fix that gets the component build further. Now it fails with:

obj/chrome/browser/android/vr_shell/libvr_common.a(vr_shell_command_buffer_gl.o):vr_shell_command_buffer_gl.cc:function
vr_shell::VrShellCommandBufferGl::OnGpuChannelEstablished(scoped_refptr<gpu::GpuChannelHost>):
error: undefined reference to
ui::ContextProviderCommandBuffer::ContextProviderCommandBuffer(scoped_refptr<gpu::GpuChannelHost>,
int, gpu::GpuStreamPriority, int, GURL const&, bool, bool,
gpu::SharedMemoryLimits const&, gpu::gles2::ContextCreationAttribHelper const&,
ui::ContextProviderCommandBuffer*, ui::command_buffer_metrics::ContextType)'

Which I /think/ means all the Mojo stuff works now, as this is a completely
different build target and the build seems to have progressed further.

As mentioned on chat, there's some weird issue with deps. The missing function
is defined in mailbox_struct_traits.cc.

$ gn refs out/android-vr //gpu/ipc/common/mailbox_struct_traits.cc
//gpu/ipc/common:interfaces

$ gn refs out/android-vr //gpu/ipc/common:interfaces
//cc/ipc:interfaces
//cc/ipc:interfaces_blink
//content/common:mojo_bindings
//device/vr:mojo_bindings
//device/vr:vr
//gpu/ipc/common:test_interfaces
//media/mojo/interfaces:interfaces
//services/ui/gpu/interfaces:interfaces
//services/ui/public/interfaces:interfaces

Interestingly enough, there's one _blink target depending on
//gpu/ipc/common:interfaces... but not //device/vr:mojo_bindings_blink.

$ gn path out/android-vr //gpu/ipc/common:interfaces
//device/vr:mojo_bindings_blink
No non-data paths found between these two targets.

The hack in PS2 "fixes" this by changing the VR mojo bindings to depend on
//cc/ipc:interfaces... by doing so, we magically gain a path between the two:

$ gn path out/android-vr //gpu/ipc/common:interfaces
//device/vr:mojo_bindings_blink
//device/vr:mojo_bindings_blink --[public]-->
//cc/ipc:interfaces_blink --[public]-->
//gpu/ipc/common:interfaces

Showing one of 2 "interesting" non-data paths. 2 of them are public.

It looks pretty bizarre that we only get the dependency on interfaces if we
indirectly depend on it, so it might be a bug in the mojo rules.

[mthiesse, 2017-03-07 15:43:01]

Thanks dcheng!

@klaus This now compiles on component build when you add
"//services/ui/public/cpp/gpu" to chrome/browser/android/vr_shell:vr_common

[dcheng, 2017-03-07 19:32:53]

Here's a version that should be officially blessed and less hacky.

[klausw, 2017-03-07 20:38:59]

Thank you very much! Merged as of PS23 into my
https://codereview.chromium.org/2729523002, PTAL at that one for the
mojo/IPC changes. (I removed the FrameView.cpp change.)

On Tue, Mar 7, 2017 at 11:32 AM, <dcheng@chromium.org> wrote:

> Here's a version that should be officially blessed and less hacky.
>
> https://codereview.chromium.org/2738683002/
>

-- 
You received this message because you are subscribed to the Google Groups
"Chromium-reviews" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to chromium-reviews+unsubscribe@chromium.org.

[klausw, 2017-03-07 20:45:05]

Thank you very much! Merged as of PS23 into my
https://codereview.chromium.org/2729523002, PTAL at that one for the
mojo/IPC changes. (I removed the FrameView.cpp change.)

On Tue, Mar 7, 2017 at 11:32 AM, <dcheng@chromium.org> wrote:

> Here's a version that should be officially blessed and less hacky.
>
> https://codereview.chromium.org/2738683002/
>

-- 
You received this message because you are subscribed to the Google Groups "Blink
Reviews" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to blink-reviews+unsubscribe@chromium.org.

[dcheng, 2017-03-08 00:26:56]

Description was changed from

==========
WebVR compositor bypass via BrowserMain context + mailbox

Implement compositor bypass for WebVR via mailbox copy:

- Create a GPU command buffer GL context in the VrShellGl thread that's backed
  by a native Surface

- SubmitFrame creates a mailbox and sends it as part of the mojo call

- On SubmitFrame, the GPU command buffer GL context converts the mailbox to a
  texture, and draws the content to the native Surface

- VrShellGl picks up the frame from the corresponding SurfaceTexture and
  proceeds with rendering as usual

Also implements on-the-fly canvas resizing, the transfer surface and GVR
renderbuffer are resized as needed to match the source canvas.

Added frame number as an argument to various TRACE_EVENTs for timing analysis.

It includes two separate modes for frame scheduling, the more aggressive one
is conditional on chrome:://flags#enable-webvr-experimental-rendering since the
framerate seems a bit less stable.

BUG=698643
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.android:android_optional_gpu_tests_rel;master.tryserver.chromium.linux:closure_compilation;master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

patch from issue 2729523002 at patchset 390001
(http://crrev.com/2729523002#ps390001)
==========

to

==========
WebVR compositor bypass via BrowserMain context + mailbox

Implement compositor bypass for WebVR via mailbox copy:

- Create a GPU command buffer GL context in the VrShellGl thread that's backed
  by a native Surface

- SubmitFrame creates a mailbox and sends it as part of the mojo call

- On SubmitFrame, the GPU command buffer GL context converts the mailbox to a
  texture, and draws the content to the native Surface

- VrShellGl picks up the frame from the corresponding SurfaceTexture and
  proceeds with rendering as usual

Also implements on-the-fly canvas resizing, the transfer surface and GVR
renderbuffer are resized as needed to match the source canvas.

Added frame number as an argument to various TRACE_EVENTs for timing analysis.

It includes two separate modes for frame scheduling, the more aggressive one
is conditional on chrome:://flags#enable-webvr-experimental-rendering since the
framerate seems a bit less stable.

BUG=698643
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.android:android_optional_gpu_tests_rel;master.tryserver.chromium.linux:closure_compilation;master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

patch from issue 2729523002 at patchset 390001
(http://crrev.com/2729523002#ps390001)
==========