Fix voxel GI issues (2)

[RandomShaper]

Follow up of #76437, in the spirit of what I suggested here: #76437 (comment).

Fixes additional manifestations of #76243 when other effects are toggled (I've experienced, and confirmed no longer an issue with this PR, with SSAO). It may also fix other still not caught bugs.

[clayjohn]

Looks good!

[Calinou]

It may also fix other still not caught bugs.

It might be worth checking whether this resolves #55334. (The issue can be reproduced in https://github.com/Calinou/godot-reflection/tree/voxelgi if you want to check there.)

[jcostello]

It may also fix other still not caught bugs.

It might be worth checking whether this resolves #55334. (The issue can be reproduced in https://github.com/Calinou/godot-reflection/tree/voxelgi if you want to check there.)

Does it make sense to combine both?

[Calinou]

Does it make sense to combine both?

In some situations, yes. For instance, if you want fully real-time GI in a specific area of the world, but want to use SDFGI for general open world GI. (SDFGI isn't fully real-time as it lacks support for dynamic occluders and emissive objects.)

[RandomShaper]

I've included another fix in the commit; namely, initializers added to some fields of LightInstance.

Without this extra change, I was being able to reproduce #76243.* And, considering the kind of fix, @Calinou's hypothesis about optimization levels playing a role makes full sense. Without initializers, garbage data was reaching the renderer and, if certain values happened to be Inf or Nan, related computations delivered wrong values. Specifically, the computation of pssm_coord here was affected:

godot/servers/rendering/renderer_rd/shaders/environment/volumetric_fog_process.glsl

Lines 385 to 428 in 2a0aef5

	for (uint i = 0; i < params.directional_light_count; i++) {
	if (directional_lights.data[i].volumetric_fog_energy > 0.001) {
	vec3 shadow_attenuation = vec3(1.0);
	if (directional_lights.data[i].shadow_opacity > 0.001) {
	float depth_z = -view_pos.z;
	vec4 pssm_coord;
	vec3 light_dir = directional_lights.data[i].direction;
	vec4 v = vec4(view_pos, 1.0);
	float z_range;
	if (depth_z < directional_lights.data[i].shadow_split_offsets.x) {
	pssm_coord = (directional_lights.data[i].shadow_matrix1 * v);
	pssm_coord /= pssm_coord.w;
	z_range = directional_lights.data[i].shadow_z_range.x;
	} else if (depth_z < directional_lights.data[i].shadow_split_offsets.y) {
	pssm_coord = (directional_lights.data[i].shadow_matrix2 * v);
	pssm_coord /= pssm_coord.w;
	z_range = directional_lights.data[i].shadow_z_range.y;
	} else if (depth_z < directional_lights.data[i].shadow_split_offsets.z) {
	pssm_coord = (directional_lights.data[i].shadow_matrix3 * v);
	pssm_coord /= pssm_coord.w;
	z_range = directional_lights.data[i].shadow_z_range.z;
	} else {
	pssm_coord = (directional_lights.data[i].shadow_matrix4 * v);
	pssm_coord /= pssm_coord.w;
	z_range = directional_lights.data[i].shadow_z_range.w;
	}
	float depth = texture(sampler2D(directional_shadow_atlas, linear_sampler), pssm_coord.xy).r;
	float shadow = exp(min(0.0, (depth - pssm_coord.z)) * z_range * INV_FOG_FADE);
	shadow = mix(shadow, 1.0, smoothstep(directional_lights.data[i].fade_from, directional_lights.data[i].fade_to, view_pos.z)); //done with negative values for performance
	shadow_attenuation = mix(vec3(1.0 - directional_lights.data[i].shadow_opacity), vec3(1.0), shadow);
	}
	total_light += shadow_attenuation * directional_lights.data[i].color * directional_lights.data[i].energy * henyey_greenstein(dot(normalize(view_pos), normalize(directional_lights.data[i].direction)), params.phase_g) * directional_lights.data[i].volumetric_fog_energy;
	}
	}

*: Interestingly enough, I could only see this specific issue in the D3D12 renderer. That made me think it was a D3D12-specific issue, but "happily" it turned out to be a menace for the whole rendering infrastructure.

[Calinou]

Specifically, the computation of pssm_coord here was affected:

I think this is a bug I introduced when adding either Volumetric Fog Energy or Shadow Opacity properties to Light3D. Sorry for the inconvenience 🙁

[mhilbrunner]

Thanks! 🎉 Merged.

[akien-mga]

Cherry-picked for 4.0.3.