NormalMap Shader: fix Adreno work-around

[WestLangley]

We need a test to replace gl_FrontFacing. I believe this change is correct.

It appears to work correctly even if the geometry and/or UVs are mirrored -- that is, even if the winding order of the UVs is clockwise when the winding order of the vertices is counterclockwise.

[WestLangley]

Merging now so @sunag can make progress with his changes.

[mrdoob]

Thanks!

[WestLangley]

/ping @JordyvanDortmont

If you would like to extensively test, that would be great. :-)

[WestLangley]

@mrdoob

Back in the day, we used this pattern to avoid if branching:

mapN.xy *= ( float( gl_FrontFacing ) * 2.0 - 1.0 );

If that is still an issue, this PR could be rewritten as:

bool frontFacing = dot( cross( S, T ), N ) > 0.0;

mapN.xy *= ( float( frontFacing ) * 2.0 - 1.0 );

[JordyvanDortmont]

Works correctly for several material examples and also on an Adreno GPU. Thanks for fixing the fix!

I think preventing instruction divergence is still very relevant for GPU performance, so it's definitely an improvement if the branching is not the same for each "fragment".

[mrdoob]

@WestLangley

If that is still an issue, this PR could be rewritten as:

bool frontFacing = dot( cross( S, T ), N ) > 0.0;

mapN.xy *= ( float( frontFacing ) * 2.0 - 1.0 );

That looks good to me.

[mrdoob]

You almost had it. This should have been:

if ( dot( cross( q0, q1 ), N ) < 0.0 ) mapN.xy *= - 1.0;

[mrdoob]

But we no longer need the workaround: #21205