1087530 - Retain container's intermediate surfaces for up to one frame

Status: RESOLVED FIXED

(Core :: Graphics, defect, P2)

Description

[Benoit Girard (:BenWa)]

+++ This bug was initially created as a clone of Bug #1037220 +++

Right now hitting group opacity is a large performance cliff. This bug is about avoiding alloc/free of a render target for each draw where a intermediate container layer is required.

Comment 1

[Benoit Girard (:BenWa)]

Attachment: WIP (missing d3d implementation)

Comment 2

[Benoit Girard (:BenWa)]

Attachment: WIP (missing d3d implementation)

Comment 3

[Benoit Girard (:BenWa)]

Profiling results on mac:
We spend less time flushing under glDeleteTexture.

I did more in depth profiling b2g:
Default: http://people.mozilla.org/~bgirard/cleopatra/#report=f4a7e7d25c39d7e7a3890737d818473c77d73cf5&select=3632,4320
With WIP: http://people.mozilla.org/~bgirard/cleopatra/#report=9059fc993257ce438b90614b3cc609c95d674ff5&select=93603,94456

Differences:
Total GPU time: 9ms -> 7.5ms (GPU profiling already coming in handy)
CPU Frame times: Bimodal frame times, half at 20ms and half near 28ms -> Most frames at 16.6ms, two frames at ~28ms

Risk:
Reusing a surface that's still in use from the previous frame might cause a hazard that some GPU drivers might fallback to flushing. We need to keep an eye out to make sure this doesn't regress on any platforms.

Comment 4

[Benoit Girard (:BenWa)]

Attachment: WIP

Needs testing on windows

Comment 5

[Benoit Girard (:BenWa)]

Attachment: patch v1

Tested d3d9/d3d11/ogl(mac,b2g)/basic(windows)

Comment 6

[Jeff Muizelaar [:jrmuizel]]

Comment on attachment 8509834 [details] [diff] [review]
patch v1

Review of attachment 8509834 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/layers/composite/TextureHost.h
@@ +671,5 @@
>  
> +  /**
> +   * Perform a clear when recycling a non opaque surface.
> +   * This must be call before binding.
> +   */

I don't like the semantics of this that much, but don't have suggestions for improvement.

@@ +680,3 @@
>  
>  private:
> +  gfx::IntRect mRect;

It seems like this will store the size twice. Let's not.

Comment 7

[Benoit Girard (:BenWa)]

Attachment: patch v2

Comment 8

[Benoit Girard (:BenWa)]

http://people.mozilla.org/~bgirard/cleopatra/#report=c4a756c160c0c35286b2b6db472df4337ceb6eda&search=delete

Comment 9

[Benoit Girard (:BenWa)]

This is a profile of a testcase that uses group opacity. Jeff wanted a profile showing deleting a texture being costly:

http://people.mozilla.org/~bgirard/cleopatra/#report=c4a756c160c0c35286b2b6db472df4337ceb6eda&search=delete

Comment 10

[Benoit Girard (:BenWa)]

Attachment: patch v3

windows fixed

Comment 11

[Benoit Girard (:BenWa)]

Attachment: part 1: Avoid needless RenderTarget switches (WIP)

Comment 12

[Benoit Girard (:BenWa)]

with this patch and bug 1092360 my testcase goes from ~40ms to ~23ms frame time on the flame:
http://people.mozilla.org/~bgirard/many_framebuffer.html

Comment 13

[Benoit Girard (:BenWa)]

Attachment: patch v4

Comment 14

[Benoit Girard (:BenWa)]

https://treeherder.mozilla.org/ui/#/jobs?repo=try&revision=7f61069549e7

Comment 15

[Benoit Girard (:BenWa)]

https://tbpl.mozilla.org/?tree=Try&rev=1038e56afa44

Comment 16

[Jeff Muizelaar [:jrmuizel]]

Comment on attachment 8516291 [details] [diff] [review]
patch v4

Review of attachment 8516291 [details] [diff] [review]:
-----------------------------------------------------------------

Sure.

Comment 17

[Benoit Girard (:BenWa)]

https://tbpl.mozilla.org/?tree=Try&rev=38852eb68ad5

Comment 18

[Benoit Girard (:BenWa)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/1608ca65d899

Comment 19

[Benoit Girard (:BenWa)]

Improvement: Mozilla-Inbound - Session Restore no Auto Restore Test - MacOSX 10.6 (rev4) - 2.25% decrease

Comment 20

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/1608ca65d899

Comment 21

[Benoit Girard (:BenWa)]

Attachment: Part 2: Fix OGL Scissor

Comment 22

[Benoit Girard (:BenWa)]

Regression follow up: https://hg.mozilla.org/integration/mozilla-inbound/rev/d757834c073b

Comment 23

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/d757834c073b