852578 - Minesweeper benchmark incredibly slow

Status: RESOLVED WONTFIX

(Core :: Graphics, defect)

Description

[Scoobidiver (away)]

This table is explicit. It's at least seven times worse than IE 10.
                      IE 10      Checker     Island
Nightly/20130318:   17.2 sec    112.1 sec   29.6 sec
Chrome 25:           5.8 sec     29.3 sec    9.8 sec
IE 10:               2.5 sec      5.5 sec    4.2 sec

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Can we get a profile of this?

Comment 2

[(mostly gone) XtC4UaLL [:xtc4uall]]

I get different Results:

IE 10:
Trunk ~ Chrome 27 = 5s

Checker:
Trunk ~ Chrome 27 = 26s

Island:
Trunk ~ Chrome 27 = 7,6s

on a ATI Radeon HD 4850 with default HWA Settings on Win 7 with Platform Update.
Btw, [parity-xy] is used for Features, not Perf.

Comment 3

[Scoobidiver (away)]

My graphics section of about:support is:
        Adapter Description
        Mobile Intel(R) 4 Series Express Chipset Family
        Adapter Drivers
        igdumd64 igd10umd64 igdumdx32 igd10umd32
        Adapter RAM
        Unknown
        ClearType Parameters
        Gamma: 2200 Pixel Structure: RGB ClearType Level: 100 Enhanced Contrast: 100
        Device ID
        0x2a42
        Direct2D Enabled
        true
        DirectWrite Enabled
        true (6.2.9200.16492)
        Driver Date
        10-4-2012
        Driver Version
        8.15.10.2869
        GPU #2 Active
        false
        GPU Accelerated Windows
        1/1 Direct3D 10
        Vendor ID
        0x8086
        WebGL Renderer
        Google Inc. -- ANGLE (Mobile Intel(R) 4 Series Express Chipset Family)
        AzureCanvasBackend
        direct2d
        AzureContentBackend
        direct2d
        AzureFallbackCanvasBackend
        cairo

With HW acceleration turn off, it's worse with a ratio of 1.5 (28 sec, 181 sec, 44 sec).

Comment 4

[Scoobidiver (away)]

Attachment: Profile IE10

Comment 5

[Scoobidiver (away)]

Attachment: Profile Checker

Comment 6

[Scoobidiver (away)]

Attachment: Profile Island

Comment 7

[Jeff Muizelaar [:jrmuizel]]

Can you profile using the Gecko profiler add-on and upload the result using the share button on the bottom right?

Comment 8

[Scoobidiver (away)]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #7)
> Can you profile using the Gecko profiler add-on and upload the result using
> the share button on the bottom right?
The add-on doesn't show any toolbar or button so I can't use it.

Comment 9

[Jeff Muizelaar [:jrmuizel]]

Did you get it from here?
https://github.com/bgirard/Gecko-Profiler-Addon/raw/master/geckoprofiler.xpi

Comment 10

[Scoobidiver (away)]

The one from AMO doesn't work.

With the Github's one:
* IE 10: http://people.mozilla.com/~bgirard/cleopatra/#report=0f8dbe347a5898d96381af3f5d0a99f860a5dfc0
* Checker: http://people.mozilla.com/~bgirard/cleopatra/#report=bb0e6e149202c5b79f10e3e65c2f53074b2dbc88
* Island: http://people.mozilla.com/~bgirard/cleopatra/#report=c5d230a33af6dae433d6eb9906c2f1111f12517b

Comment 11

[Jeff Muizelaar [:jrmuizel]]

It looks like we're being hurt a lot by GDI interop.

It would be worth trying this build to see if it helps:
https://tbpl.mozilla.org/?tree=Try&rev=d1921bda6a6a

Comment 12

[Scoobidiver (away)]

(In reply to Jeff Muizelaar [:jrmuizel] from comment #11)
> It would be worth trying this build to see if it helps:
> https://tbpl.mozilla.org/?tree=Try&rev=d1921bda6a6a
I don't find it in ftp://ftp.mozilla.org/pub/firefox/try-builds/

Comment 13

[Jeff Muizelaar [:jrmuizel]]

http://ftp.mozilla.org/pub/mozilla.org/firefox/try-builds/[e-mail]-d1921bda6a6a/try-win32/

Comment 14

[Scoobidiver (away)]

With the try build, it's slightly better:
* IE 10: 15.2 seconds http://people.mozilla.com/~bgirard/cleopatra/#report=7dfa07a47f5b28981957c24ad7ec50aba140d2b9
* Checker: 94.2 seconds http://people.mozilla.com/~bgirard/cleopatra/#report=91487277b8dc4ee2fe7cd7fc36a7048b00c44489
* Island: 23.3 seconds http://people.mozilla.com/~bgirard/cleopatra/#report=66fd640abed543a92d023deef1192fcfcd58de26

Comment 15

[Jeff Muizelaar [:jrmuizel]]

Hmm.. we're still getting hurt by the texture creation locks. I think we probably need to be creating less textures.

Comment 16

[Guilherme Lima]

Windows 7 with HWA, on the IE10 map
Nightly 27 - 6.50s
Chrome 29 - 6.60s
IE 10 - 4.70s

On Checker
Nightly 27 - 34.10s
Chrome 29 - 39.00s
IE 10 - 13.20s

Comment 17

[Guilherme Lima]

On the same hardware from years ago:

Huge Checker
Firefox 54 - 33.20s
Nightly (Advanced Layers on) - 28.30s
Chrome 59 - 32.90s
IE 11 - 4.30s (!!)

Comment 18

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

(In reply to Guilherme Lima from comment #17)
> On the same hardware from years ago:
> 
> Huge Checker
> Firefox 54 - 33.20s
> Nightly (Advanced Layers on) - 28.30s
> Chrome 59 - 32.90s
> IE 11 - 4.30s (!!)

Do you have the numbers for nightly with advanced layers off?  layers.mlgpu.dev-enabled and layers.mlgpu.enable-on-windows7 set to false in about:config and Firefox restarted.

Comment 19

[Guilherme Lima]

Tested again today:
Nightly (Advanced Layers on) - 28.30s (exactly same number from yesterday, nice!)
Nightly (Advanced Layers off) - 30.60s

Comment 20

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

David, let's make sure this is expected regression.

Comment 21

[Guilherme Lima]

It's an improvement... it takes less time to complete.

Comment 22

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

I ran it locally, got the same improvement, and still read it as a regression.  And it's still early in the day :)

Comment 23

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

Although, the "Island" one does regress for me, ~10% or so...

Comment 24

[David Anderson [:dvander] - inactive, e-mail if emergency]

I see a slight improvement on all 3 on my machine, but I can't tell what this benchmark is actually measuring. It doesn't look like anything. Maybe it was supposed to measure Canvas 2D and now instead is testing whether setTimeout is implemented the same as Edge.