Unity RubysAdventure demo does not render properly if WebGL2 is enabled.

The Unity emunittest demo RubysAdventure does not render in Safari STP if WebGL 2 experimental feature is enabled. It does work properly in WebGL 1 mode.

RubysAdventure build can be downloaded from http://clb.confined.space/emunittest_unity/RubysAdventure_20190722_201255_wasm_release_profiling.zip

The source Unity project can be downloaded from https://assetstore.unity.com/packages/essentials/tutorial-projects/2d-beginner-complete-project-140253 . The above build has been exported from this Unity project unmodified.

[Radar WebKit Bug Importer]

<rdar://problem/67817161>

[Kenneth Russell]

https://tiny.vision/ also renders all black in Safari Technology Preview Release 114 (Safari 14.1, WebKit 15611.1.1.3) . It was broken in Release 113 too.

jdarpinian@ said he's actively investigating this, so reassigning.

Raising to P1, Major - this is key content we've been aiming to enable with the WebGL 2.0 upgrade in Safari.

[James Darpinian]

So far I haven't found anything wrong with the rendering itself. What I have found using spector.js is that Unity is setting a bunch of uniforms to 0, which are set to 1 normally in Chrome. I'm looking at one of the first draw calls and there are uniforms _Color, _Flip, and _RendererColor which have all their values set to 0, when in Chrome they are all 1. At least, that's what spector.js sees.

[James Darpinian]

I haven't yet ruled out a bug in WebKit that's causing those uniforms to be set to 0 incorrectly, but it's also possible that Unity is setting them wrong. I will investigate more on Monday.

[Brendan Duncan]

Digging into the GL functions Unity is using, I found the WebGL2 variant of glUniform4fv, gl.uniform4fv(location, data, srcOffset, srcLength), is failing on Safari. When I changed that call to the WebGL1 variant, the test started working on Safari. I haven't verified yet if other functions are failing as well.

[Brendan Duncan]

Looks like there are problems with the WebGL2 variants for all of the gl.uniform*v functions. When I disabled the WebGL2 variants for those functions in the RubysAdventure build Jukka provided, it started working.

I did this by editing the framework.unityweb file in the Build folder, and replacing the code `if(GL.currentContext.supportsWebGL2EntryPoints)` in the `_glUniform*v` functions with `if(false)`, forcing those functions to use the WebGL1 variants.

[Brendan Duncan]

I built WebKit from source to figure out why the glUniform* functions were failing. Focusing on glUniform4fv, I found ANGLE was generating the error in ValidateUniformCommonBase (ANGLE/src/libANGLE/validationES.cpp). Specifically, count > 1 && !uniform.isArray(). 

The uniform it doesn't think was an array is `uniform vec4 _MainTex_ST`. This is the translate/scale transform for the screen blit UVs, so being considered invalid and not getting set, the transform has a 0 scale, and hence the black screen. 

The real bug then is why ANGLE doesn't think that vec4 uniform is an array.

[Brendan Duncan]

The WebGL1 variant calls validate with a count of 1, that's why it didn't have an issue with isArray being false. So the real bug is with Webkit's validateUniformParameters using a count size of 4 instead of 1 for the vec4 uniform.

[Brendan Duncan]

Here's the bug:

Source/WebCore/html/canvas/WebGL2RenderingContext.cpp,
WebGL2RenderingContext::uniform4fv:
m_context->uniform4fv(location->location(), data.data(), srcOffset, srcLength ? srcLength : (data.length() - srcOffset) / 4);

should be:
m_context->uniform4fv(location->location(), data.data(), srcOffset, srcLength ? srcLength / 4 : (data.length() - srcOffset) / 4);

And likewise for the other unform*v functions.

[James Darpinian]

Thanks for the help! This is definitely the issue. Confirmed that this fixes *all* of the emunittest demos on http://clb.confined.space/emunittest/ including RubysAdventure.

I'm currently investigating whether one of the known test failures covers this or whether we just lack test coverage. If the latter, I will write a new WebGL conformance test.

[James Darpinian]

Created attachment 411179 [details]
Patch

[James Darpinian]

Looks like the fix doesn't affect any conformance tests :( I will write a new one tomorrow.

[Kenneth Russell]

Comment on attachment 411179 [details]
Patch

Thank you Brendan for tracking down the root problem and James for fixing this! Yes, please do add a new WebGL 2.0 conformance test for all of these entry points. Looking at Chrome's implementation, I'm surprised it doesn't have the same bug. r+

[James Darpinian]

Created attachment 411381 [details]
Patch

[James Darpinian]

I found that there is actually a test covering this, but it is not enabled in the 2.0.0 webgl conformance snapshot. I guess in addition to passing 2.0.0, before release we should try a 2.0.1 run and examine the results to see if there are any other serious issues hiding in there.

[Kenneth Russell]

Comment on attachment 411381 [details]
Patch

Looks good to me again!

[James Darpinian]

Created attachment 411387 [details]
Patch

[EWS]

Committed r268502: <https://trac.webkit.org/changeset/268502>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 411387 [details].

[Kenneth Russell]

*** Bug 218327 has been marked as a duplicate of this bug. ***