1798929 - Large percentage of time spent in ScopedResolveTexturesForDraw in Proxx-Tables-Canvas benchmark

Status: RESOLVED FIXED

(Core :: Graphics: Canvas2D, enhancement, P3)

Description

[Markus Stange [:mstange]]

Profile: https://share.firefox.dev/3hbJr5H

WebGLContext::DrawArraysInstanced seems to spend a fair amount of time allocating and freeing things via ScopedResolveTexturesForDraw.

I found this on the Proxx-Tables-Canvas benchmark (bug 1798923): https://grandprixbench.netlify.app/?suites=Proxx-Tables-Canvas

Comment 1

[Brad Werth [:bradwerth] | Out until Feb 2025]

Kelsey, you've been in this code quite a bit. Are these memory allocations avoidable, or can the number of texUnits be known up-front? ScopedResolveTexturesForDraw::ScopedResolveTexturesForDraw is also profiling in a weird way: the profiler seems to show that the function invokes itself and also can allocate multiple maps in the same call -- perhaps that's just showing that the reserve amount has been exceeded.

Comment 2

[Kelsey Gilbert [:jgilbert]]

We should just reuse the allocations somehow.
The max size is known and small, I'm mostly using map for ergonomics.

Comment 3

[Kelsey Gilbert [:jgilbert]]

I'm tempted to just have this be a static thread_local, but we can just tag it off of the this object.

Comment 4

[Kelsey Gilbert [:jgilbert]]

Also showing up kinda hot there is overhead around these four related commands:

• SetEnabled
• Clear
• ClearColor
• Scissor

Feels like we're ending up doing something like:

for () {
  Enable(SCISSOR_TEST)
  Scissor(x,y,...)
  ClearColor(r,g,b,a)
  Clear(COLOR)
  Disable(SCISSOR_TEST)
}

You can imagine us e.g. lazily combining ClearColor+Clear into Clear(flags, r,g,b,a, d,s), and/or eliding unused Enables/Disables.
Deser overhead is cheap but not free, and it does show up here weighing things down.

Comment 5

[Kelsey Gilbert [:jgilbert]]

Attachment: Bug 1798929 - Reuse samplerByTexUnit capacity to avoid (de)allocs.

Comment 6

[Kelsey Gilbert [:jgilbert]]

While that profile link does make it look like we're spending 15% in this func, that's because we're dropping samples that are waiting on events, so this is closer to a 2% win, all else equal.

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/25b53dd457af
Reuse samplerByTexUnit capacity to avoid (de)allocs. r=gfx-reviewers,lsalzman

Comment 8

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/25b53dd457af

Comment 9

[Andrej (:aglavic)]

== Change summary for alert #36053 (as of Sun, 13 Nov 2022 01:24:52 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
8%	motionmark_webgl 3DGraphics-WebGL	windows10-64-shippable-qr	e10s fission stylo webrender	10.59 -> 11.46

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=36053