Add maximum roughness cutoff to SSR to improve performance

[Calinou]

master version of #56815.

In a test scene with mixed rough and non-rough materials, this saves upwards of 0.15 ms of GPU time with very little visual artifacting (GTX 1080, 2560×1440). This improves performance the most in scenes that have mostly rough materials. The optimization works regardless of the SSR roughness quality selected in the Project Settings.

The roughness cutoff was empirically chosen – it seems to strike a good balance between improving performance and avoiding obvious visual artifacts. The roughness cutoff could be exposed to a project setting if needed, but the Params push constant is already full (128 bytes).

Testing project: test_ssr_performance.zip

Preview

Some more comparisons: Trying to find a good roughness cutoff, High-quality settings

SSR Roughness Low quality (default)

Before

After

SSR Roughness Disabled

Before

After

[fire]

I made a pr in the same area of roughness #56822.

[clayjohn]

With the changes listed here this could be merged after #69828

[Calinou]

Benchmark with the new roughness cutoff of 0.6 in 3840×2160 on a AMD Radeon RX 6900 XT (Fedora 36, RADV), using the same test scene as before:

Before	After (this PR)
479 FPS (2.08 mspf)	551 FPS (1.81 mspf)

[akien-mga]

Thanks!

[Calinou]

Despite the comment in the shader, I noticed the roughness cutoff in master doesn't actually match #69828, since it's 0.6 instead of 0.7:

godot/servers/rendering/renderer_rd/shaders/effects/screen_space_reflection.glsl

Lines 71 to 72 in 950f4c3

	// The roughness cutoff of 0.6 is chosen to match the roughness fadeout from GH-69828.
	if (roughness > 0.6) {

godot/servers/rendering/renderer_rd/shaders/effects/screen_space_reflection.glsl

Lines 226 to 228 in 950f4c3

	// This is an ad-hoc term to fade out the SSR as roughness increases. Values used
	// are meant to match the visual appearance of a ReflectionProbe.
	float roughness_fade = smoothstep(0.4, 0.7, 1.0 - normal_roughness.w);

In practice, this appears to have no visual impact, but restoring the "correct" cutoff value of 0.7 makes performance slightly worse (+0.03 mspf in 3840×2160 on a GeForce RTX 4090):

0.6 (master)

Screenshot

0.7 (proposed change)

Screenshot

What should we do? For reference, I've made #56815 use the "correct" cutoff value of 0.7, but maybe we should make it use 0.6.