[3.x] GLES2 & GLES3 - Use gl_FragColor temporary

[lawnjelly]

On some hardware, modifying gl_FragColor multiple times (or more likely reading from gl_FragColor) can cause large performance drops. This PR writes to a standard temporary variable instead, and copies across to gl_FragColor once only at the end of the fragment shader.

This could potentially lead to large gains in performance on affected hardware (so far I've measured from 2x increase in fps (GLES2) to 4x (GLES3)).

Temporary variable is called frag_color in GLES2, and frag_color in GLES3 (In GLES3 the final output is renamed to frag_color_final).

Partially addresses #84526.

Notes

• The problem shows most in fill rate limited circumstances, e.g. with MSAA, and where frag_color is read / modified more than once.
• Should work with custom shaders as they write to e.g. ALBEDO rather than frag_color directly.
• No changes needed for 2D as far as I can see.
• May also be worth doing in 4.x, especially GLES3.
• See Write / reading frag_color multiple times in shader seems to cause performance problem #84526 (comment) for tests on different hardware.
• Although I've only measured increases in performance through scene.glsl, I've included the fix for other potentially affected shaders, as there is no evidence of a drop in performance and potential for gain, especially if these shaders are later modified.

Example Project

(Simple box with fog and emission)
Before 51 fps
After 104 fps

glFragColor.zip

[Calinou]

Tested locally, it works as expected (no rendering issues). I can't notice any performance difference on a RTX 4090 (even at 8K) but it's not exactly the kind of GPU targeted by this optimization 🙂

[lawnjelly]

Tested locally, it works as expected (no rendering issues). I can't notice any performance difference on a RTX 4090 (even at 8K) but it's not exactly the kind of GPU targeted by this optimization 🙂

It may not be a bottleneck on your super GPU, especially with such a simple test project (or not have the vulnerability). I bumped up the MSAA to make the difference show up more in the test project. It may appear more when the GPU is stressed by fill rate.

[clayjohn]

I haven't tested locally, but I trust your testing on this. Issues would be immediately apparent if they appear.

[akien-mga]

Thanks!