Correct stencil upload bugs

[unknownbrackets]

GLES was missing the semantic definitions. I guess it was just using the last ones on desktop or something. Now it works fine on mobile again.

Vulkan definitely looks like a driver bug, and it's a bug that could easily be affecting regular rendering too. I didn't test extensively but at least this case is affected:

• Stencil testing is enabled, thus stencil is being written.
• Alpha or color testing is enabled, and discard; is used in the shader.

I suppose we could statically write to depth (Adreno only?) whenever we also generate a discard. (EDIT by hrydgard: a commit for this has now been added)

If we bother detecting that case, we might put layout(early_fragment_tests) in; for GL4.2+/Vulkan/GLES3.1+? It might clarify some case with hardware tessellation or direct depal, etc.

Fixes #10634.

-[Unknown]

[unknownbrackets]

Seeing that it could be interesting to use ARM_shader_framebuffer_fetch_depth_stencil as an alternative to dual source blending (using in-shader blending replacing destColor.a), although that's probably impacted by early depth tests too.

-[Unknown]

[hrydgard]

I see you added a workaround by writing to depth - did layout(early_fragment_tests) do anything?

[hrydgard]

Good catch with the semantics by the way ... oops.

[unknownbrackets]

layout(early_fragment_tests) is the exact opposite of what we want in this case (we need the tests to be late so it will write our stencil only when not discarded.)

I meant we could add that to the regular drawing fragment shaders to explicitly allow early tests in cases where we don't discard. And regular drawing is probably impacted by this driver bug.

-[Unknown]

[unknownbrackets]

Posted the issue on their forum at least:
https://developer.qualcomm.com/forum/qdn-forums/software/adreno-gpu-sdk/35916

-[Unknown]

[hrydgard]

Oh right, forgot that it's indeed a hint the other way :)

They're usually not nearly as responsive as say ARM in their forum, but good to have it reported.