1421199 - Use EGL_EXPERIMENTAL_PRESENT_PATH_FAST_ANGLE for WebRender

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, enhancement, P1)

Description

[Sotaro Ikeda [:sotaro]]

In Bug 1419876 Comment 1, there is an info to make fast ANGLE.

https://github.com/Microsoft/angle/wiki/Getting-Good-Performance-From-ANGLE#--use-our-render-to-backbuffer-flags

With the info, I looked into ANGLE source and found EGL_EXPERIMENTAL_PRESENT_PATH_FAST_ANGLE. It is used in Renderer11 as a flag for enabling fast path.

Comment 1

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use EGL_EXPERIMENTAL_PRESENT_PATH_FAST_ANGLE for WebRender

Comment 2

[Sotaro Ikeda [:sotaro]]

Attachment: wip patch - Use EGL_EXPERIMENTAL_PRESENT_PATH_FAST_ANGLE for WebRender

Comment 3

[Sotaro Ikeda [:sotaro]]

With attachment 8932376 [details] [diff] [review], peak fps of http://learningwebgl.com/lessons/lesson03/index.html on P50(Win10) was changed from 50fps to 60fps.

Comment 4

[Sotaro Ikeda [:sotaro]]

Attachment: patch part 1 - Fix GLES3 and "present path fast" validation issue

This change is from https://github.com/Microsoft/angle/commit/fb8a709e46fcf4385bf0793cac8962942b7b8f28

Without the change, ANGLE could not create GLES3 with "present path fast".

Comment 5

[Sotaro Ikeda [:sotaro]]

Attachment: patch part 2 - Change GLDefs.h for adding EGL_ANGLE_experimental_present_path

Comment 6

[Sotaro Ikeda [:sotaro]]

Attachment: patch part 3 - Add GetAndInitDisplayForWebRender()

Comment 7

[Kelsey Gilbert [:jgilbert]]

Comment on attachment 8932411 [details] [diff] [review]
patch part 3 - Add GetAndInitDisplayForWebRender()

Review of attachment 8932411 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/gl/GLLibraryEGL.cpp
@@ +148,5 @@
>  
> +static EGLDisplay
> +GetAndInitDisplayForWebRender(GLLibraryEGL& egl, void* displayType)
> +{
> +    EGLint attrib_list[] = {  LOCAL_EGL_PLATFORM_ANGLE_DEVICE_TYPE_ANGLE,

const-qualify const lists, please!

Comment 8

[Sotaro Ikeda [:sotaro]]

Attachment: patch part 3 - Add GetAndInitDisplayForWebRender()

Apply the comment.

Comment 9

[Sotaro Ikeda [:sotaro]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=b69d8dab58dbf95404c38d628f7f4dc6ef7b32e4

Comment 10

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/9e30a05d64c6
Use EGL_EXPERIMENTAL_PRESENT_PATH_FAST_ANGLE for WebRender r=jgilbert

Comment 11

[Dorel Luca [:dluca]]

https://hg.mozilla.org/mozilla-central/rev/9e30a05d64c6

Comment 12

[Jeff Muizelaar [:jrmuizel]]

It seems like Chrome doesn't use this. It would be interesting to know why.

Comment 13

[Dzmitry Malyshau [:kvark]]

Jeff, I'm trying to figure this out.
Filed https://github.com/baldurk/renderdoc/issues/810

Comment 14

[Dzmitry Malyshau [:kvark]]

Ok, Chrome must be getting some other way to achieve the same effect. I verified that the GPU capture of MotionMark running in Chrome doesn't do a final blit on Present() and instead just renders into the swap chain right away.

Comment 15

[Jeff Muizelaar [:jrmuizel]]

I talked to Geoff from ANGLE and he pointed out the Chrome is using EGL_SURFACE_ORIENTATION_INVERT_Y_ANGLE and just draws things upside down.

See https://bugs.chromium.org/p/angleproject/issues/detail?id=1262 and https://bugs.chromium.org/p/chromium/issues/detail?id=525379

Comment 16

[Dzmitry Malyshau [:kvark]]

Thanks Jeff, this is interesting. The discussion doesn't mention any comparisons between those two ways around the full-screen blit. I'd assume that the difference is small. We can investigate at some point in the future.

Comment 17

[Kelsey Gilbert [:jgilbert]]

I'll check if this needs to be upstreamed.