785848 - WebGL shader loop unroll issue via angle d3d with ps_3_0

Status: RESOLVED WONTFIX

(Core :: Graphics: CanvasWebGL, defect)

Description

[Florian Bösch]

User Agent: Mozilla/5.0 (X11; Linux i686) AppleWebKit/536.11 (KHTML, like Gecko) Chrome/20.0.1132.8 Safari/536.11

Steps to reproduce:

Run this code: http://codeflow.org/issues/unroll-problem/


Actual results:

Output indicated 2 minutes time to error and shown this error message:

"error X3511: unable to unroll loop, loop does not appear to terminate in a timely manner (9 iterations), use the [unroll(n)] attribute to force an exact higher number"

On the shader info log. CPU use was at 25% and the page/browser was non-responding.


Expected results:

Linux/OSX have no problem with this code and show an output of "Compile Time: 11ms".

Comment 1

[Dzmitry Malyshau [:kvark]]

Unable to repro on Linux x64 with FF 53.0a1, "Intel(R) HD Graphics 5500 (Broadwell GT2)"

Comment 2

[Florian Bösch]

You need to try on Windows since it's a Direct3D HLSL compiler related problem.

Comment 3

[Dzmitry Malyshau [:kvark]]

Florian, I tested on Windows 7 with Radeon Rx 480, no issues either.

Comment 4

[Florian Bösch]

If you've merged in some angle fixes or updated the d3d runtime to dx11, the bug could be fixed by those. If you want to know when you incidentially "fixed" it you'll have to bisect FF versions all the way back to 2012 since apparently only now somebody got around to look at the bug.

Comment 5

[Dzmitry Malyshau [:kvark]]

Ah, apologies, that was OSX (sharing the monitor and input got me confused).
Loading on actual Windows gets FF stuck. Investigating...

Comment 6

[Dzmitry Malyshau [:kvark]]

Looks like Angle puts `[unroll]` on top of the loop, and then FXC is having problems actually doing it.
Related - https://codereview.appspot.com/7139055/

Comment 7

[Dzmitry Malyshau [:kvark]]

Attachment: test.hlsl

Attaching HLSL produced by Angle

Comment 8

[Kelsey Gilbert [:jgilbert]]

I think we're also unrolling loops always, when we don't always need to.

Comment 9

[Dzmitry Malyshau [:kvark]]

Attachment: Output.txt

Attached the FXC output when compiling the Angle-produced HTML with targets ps_3_0, ps_4_0, ps_5_0:

ps_3_0 gives up after ~2 minutes, although reporting it stopped after 710 iterations (6 iterations per second?)
ps_4_0 and ps_5_0 succeed in 13 seconds

One possible solution is to submit a patch to Angle limiting the maximum loop unroll, similar to https://codereview.appspot.com/7139055/. Although, it doesn't seem straightforward that it will be merged, especially given the freely adjustable constant it would introduce.

Another solution is to force angle to work with D3D11 and HLSL 4.0 targets, if possible. Given that we are fading out WinXP, I don't see much value in supporting D3D9 anyway.

Comment 10

[Dzmitry Malyshau [:kvark]]

Florian, I accidentally had "webgl.angle.try-d3d11 == false" on that Windows machine. When I set it back to default (true), the page loads normally, reporting: "Compile Time: 2140ms". I suppose the issue only persists where d3d11 is not available or fails to activate. Could you check again and confirm if it's still a problem?

Comment 11

[Florian Bösch]

I currently can't test on a windows configuration, but I know that d3d11 helps some.

However, 2140ms is much too long to compile a shader. It takes 21ms to compile on Linux (and probably similar amounts on other OS'es).

Lengthy shader compile times have been a persistent problem for WebGL and should be eliminated.

Comment 12

[Dzmitry Malyshau [:kvark]]

Florian, I do agree that 2140ms is little outside of the comfort zone, and we should be doing something about it. However, this bug doesn't seem to be a proper medium, since the original report was about FXC running out of time after 2 minutes, which was a much more severe issue that is no longer a problem.

Jeff, do we have a general bug for tracking Angle's shader compiling performance?

Comment 13

[Kelsey Gilbert [:jgilbert]]

(In reply to Dzmitry Malyshau from comment #12)
> Florian, I do agree that 2140ms is little outside of the comfort zone, and
> we should be doing something about it. However, this bug doesn't seem to be
> a proper medium, since the original report was about FXC running out of time
> after 2 minutes, which was a much more severe issue that is no longer a
> problem.
> 
> Jeff, do we have a general bug for tracking Angle's shader compiling
> performance?

I don't think so. I made one: Bug 1326482

Comment 14

[Dzmitry Malyshau [:kvark]]

Ok, thanks Jeff! I consider this issue to be closed then.