1437556 - Canvas performance on Mac is worse with WR than Layers

Status: RESOLVED WORKSFORME

(Core :: Graphics: WebRender, enhancement, P2)

Description

[Jeff Muizelaar [:jrmuizel]]

We're not using SkiaGL, but also it seems like we could be doing better even in software.

https://jrmuizel.github.io/webrender/perf/canvas-perf.html

https://perfht.ml/2o0kpH7

Comment 1

[Milan Sreckovic [:milan] (needinfo for best results)]

This is WebRender specific?  As in, it got worse from Skia (not SkiaGL) without WR?

Comment 2

[Sotaro Ikeda [:sotaro]]

In current implementation, SkiaGL is used only with LayersBackend::LAYERS_OPENGL.

Comment 3

[Sotaro Ikeda [:sotaro]]

Created Bug 1438079 to enable SkiaGL on mac.

Comment 4

[Sotaro Ikeda [:sotaro]]

SkiaGL was disabled by Bug 1468801.

Comment 5

[Kelsey Gilbert [:jgilbert]]

Can we quantify what we mean here? The linked benchmark is not representative of real content, and would be the pessimal case for software canvas.

Can we either pick a different benchmark (or set of benchmarks), or alternatively narrow the scope of this bug to 'large canvas2d clears are slow' and decide whether to WONTFIX that specific case?

Comment 6

[Jeff Muizelaar [:jrmuizel]]

This bug is just meant to ensure that we're not regressing canvas2d performance on Mac with WebRender.

Comment 7

[Kelsey Gilbert [:jgilbert]]

Can you re-test now that we've dropped SkiaGL, since WR lacking SkiaGL was the likely cause here?

Comment 8

[Jeff Muizelaar [:jrmuizel]]

Dropping the n-i as this is important right now.

Comment 9

[Trevor Rowbotham [:rowbot]]

According to [1], canvas performance is still an issue on mac with webrender enabled.

[1] https://www.reddit.com/r/firefox/comments/brw6c7/post_your_webrender_results/eoo3nnz?utm_source=share&utm_medium=web2x

Comment 10

[Jessie [:jbonisteel] pls NI]

Miko, when you are done your current task, can you help investigate this to see if it is indeed a problem?

Comment 11

[Bert Peers [:bpeers]]

On 80.0a1 I'm seeing the expected results on the Chalkboard benchmark (13" MBP):

With WebRender: 7.03 seconds / 6.97 seconds / 6.96 seconds

Without WebRender (webrender.force-disable true): 22.89 seconds / 23.14 seconds / 22.87 seconds

With WebRender, the attached test case runs at a steady 32 FPS (via webrender.debug.profiler).

Without WebRender it maxes out at 26 FPS (via layers.acceleration.draw-fps).

Comment 12

[Bert Peers [:bpeers]]

Looking for more tests, I did find some differences on CanvasMark 2013 between WR on/off:
http://www.kevs3d.co.uk/dev/canvasmark/
About half the tests are slower with WR enabled.

Focusing just on the Plasma test, I wonder if possibly some of this is due to lower clock speeds, where having WR enabled allows running more efficiently. When the CPU clock with WR-on boosts to the same value as WR-off, it achieves the same benchmark score on the plasma test; the problem is that after the first spike, it stops boosting, whereas with WR-off it never reaches that level of efficiency.

Just a thought.

Comment 13

[Bert Peers [:bpeers]]

Attachment: 718AE232-1527-4599-A793-40E84F022EB0_4_5005_c.jpeg

Plasma test with WR off.
The ~2Ghz line is the idle "press space bar to start". The two peaks are two runs, both hitting ~3.9Ghz.

Comment 14

[Bert Peers [:bpeers]]

Attachment: B39D6B01-2E9C-4FE3-97C9-2FB7DDE40DC0_4_5005_c.jpeg

Plasma test with WR-on. The first peak (at about the midpoint of the graph) is the first run, it gets a similar score. The 2nd peak (at about 75% on the X-axis) got a score 33% lower.

Comment 15

[Miko Mynttinen]

(In reply to Bert Peers [:bpeers] from comment #11)

On 80.0a1 I'm seeing the expected results on the Chalkboard benchmark (13" MBP):

With WebRender: 7.03 seconds / 6.97 seconds / 6.96 seconds

Without WebRender (webrender.force-disable true): 22.89 seconds / 23.14 seconds / 22.87 seconds

With WebRender, the attached test case runs at a steady 32 FPS (via webrender.debug.profiler).

Without WebRender it maxes out at 26 FPS (via layers.acceleration.draw-fps).

I am getting similar numbers. On my 15" MBP WebRender is a lot faster, especially with a 4K monitor.

Jeff's testcase: Non-WR: 32 FPS, WR: 75 FPS.
Chalkboard: Non-WR: 47 seconds , WR: 2.7 seconds (Yes, these are correct numbers that were measured multiple times)

Comment 16

[Miko Mynttinen]

It seems that Jeff's testcase was improved by client storage patches of bug 1536515. These took the test from 13 FPS to 60 FPS.

Comment 17

[Jeff Muizelaar [:jrmuizel]]

If https://testdrive-archive.azurewebsites.net/performance/chalkboard/ is the Chalkboard benchmark that is being referred to, I'm pretty sure that doesn't use canvas.

Comment 18

[Jessie [:jbonisteel] pls NI]

After Miko re-lands client texture stuff, we will do further testing to see how we are doing.

Comment 19

[Miko Mynttinen]

(In reply to Jessie [:jbonisteel] pls NI from comment #18)

After Miko re-lands client texture stuff, we will do further testing to see how we are doing.

Client storage was re-enabled in today's Nightly, and the testcase still looks good.

I also benchmarked WebRender on/off in MotionMark canvas tests, and it seems that WebRender is doing better with both integrated and discrete GPUs on my 2019 MBP. Interestingly, it seems that Intel with the laptop screen benefits more from WebRender than AMD with an external 4K monitor.

WR on AMD: 987, 963
WR off AMD: 935, 905

WR on Intel: 801, 784
WR off Intel: 690, 672