I implemented CL[1]. It can submit composit frame with or without Mojo IPC.
With it, I got the performance data from GFXBench on minnie. I found out only
test the ALU 2 & Texturing have performance regression. Results of other tests
are almost same. See attachment for detail.

Anyone has idea what are ALU2 & Texturing? And why only those two have performance regression?


[1] https://chromium-review.googlesource.com/c/522782/1/ui/aura/local/window_port_local.cc#118

Deleted: DFCvsMojo.png 1.5 MB

	DFCvsMojo.png 1.5 MB View Download

+anicolao@

Thank you for looking into this Peng! :-)

It would be good to know what parts of the code are contributing to the regression with mojo. Maybe the sampling profiler could provide some insight?

I tried doing mojo message serializing and deserializing in
CompositorFrameSinkLocal::SubmitCompositorFrame()(see [1]). And I found out
the mojo serializing and deserializing only contribute about 10 points of
the regression.


1282 vs 1270 vs 1107. (Direct Function Call vs DFC + Serializing vs Mojo IPC)

[1] https://chromium-review.googlesource.com/c/522782/3/ui/aura/local/compositor_frame_sink_local.cc#117

Do we have a bug to track understanding and mitigating this regression?

I just implemented one CFS per ARC++ window with CL[1]. Everything works fine.
I measured the ARC++ performance again. Numbers are in the [2].

There is small regressions in Driver overhead 2 testes comparing to current
implementation. For one CFS with mojo IPC, the result is similar to the current
implementation with Mojo IPC.


[1] https://chromium-review.googlesource.com/c/527475/
[2] https://docs.google.com/spreadsheets/d/1FO1E1iW9yXDHquRzR35nas-xLWL40xXXAgeJ5Fk4ToQ/edit#gid=0

It's hard to evaluate these numbers without understanding where the cost comes from. Can you record some traces that highlight the difference?

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/059a6f74910c4fe167d1559b49db3158ef58cd5d

commit 059a6f74910c4fe167d1559b49db3158ef58cd5d
Author: Peng Huang <penghuang@google.com>
Date: Wed Sep 20 00:31:11 2017

exo: Implement exo performance test

Bug: 729028
Change-Id: Ibda083042d68a681cffd8eacb0494221b8c6a5be
Reviewed-on: https://chromium-review.googlesource.com/653037
Reviewed-by: Scott Violet <sky@chromium.org>
Reviewed-by: David Reveman <reveman@chromium.org>
Commit-Queue: Peng Huang <penghuang@chromium.org>
Cr-Commit-Position: refs/heads/master@{#503005}
[modify] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/BUILD.gn
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/perftests.cc
[modify] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/simple.cc
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/simple.h
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/simple_main.cc
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/test/DEPS
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/test/run_all_client_perftests.cc
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/test/wayland_client_test.cc
[add] https://crrev.com/059a6f74910c4fe167d1559b49db3158ef58cd5d/components/exo/wayland/clients/test/wayland_client_test.h

We got soem data, but it not very useful. Remove this issue from  issue 728695 .