Add a supersampling option to LightmapGI

This provides increased lightmap quality with less noise, smoother
shadows and better small-scale shadow detail. The downside is that
this significantly increases bake times and memory usage while baking
lightmaps, so this option is disabled by default.

Co-authored-by: Hugo Locurcio <[email protected]>
Co-authored-by: landervr <[email protected]>

Author: 3 people

doc/classes/LightmapGI.xml

@@ -39,10 +39,10 @@
 			If [code]true[/code], bakes lightmaps to contain directional information as spherical harmonics. This results in more realistic lighting appearance, especially with normal mapped materials and for lights that have their direct light baked ([member Light3D.light_bake_mode] set to [constant Light3D.BAKE_STATIC] and with [member Light3D.editor_only] set to [code]false[/code]). The directional information is also used to provide rough reflections for static and dynamic objects. This has a small run-time performance cost as the shader has to perform more work to interpret the direction information from the lightmap. Directional lightmaps also take longer to bake and result in larger file sizes.
 			[b]Note:[/b] The property's name has no relationship with [DirectionalLight3D]. [member directional] works with all light types.
 		</member>
-		<member name="environment_custom_color" type="Color" setter="set_environment_custom_color" getter="get_environment_custom_color">
+		<member name="environment_custom_color" type="Color" setter="set_environment_custom_color" getter="get_environment_custom_color" default="Color(1, 1, 1, 1)">
 			The color to use for environment lighting. Only effective if [member environment_mode] is [constant ENVIRONMENT_MODE_CUSTOM_COLOR].
 		</member>
-		<member name="environment_custom_energy" type="float" setter="set_environment_custom_energy" getter="get_environment_custom_energy">
+		<member name="environment_custom_energy" type="float" setter="set_environment_custom_energy" getter="get_environment_custom_energy" default="1.0">
 			The color multiplier to use for environment lighting. Only effective if [member environment_mode] is [constant ENVIRONMENT_MODE_CUSTOM_COLOR].
 		</member>
 		<member name="environment_custom_sky" type="Sky" setter="set_environment_custom_sky" getter="get_environment_custom_sky">
@@ -67,13 +67,23 @@
 		</member>
 		<member name="quality" type="int" setter="set_bake_quality" getter="get_bake_quality" enum="LightmapGI.BakeQuality" default="1">
 			The quality preset to use when baking lightmaps. This affects bake times, but output file sizes remain mostly identical across quality levels.
-			To further speed up bake times, decrease [member bounces], disable [member use_denoiser] and increase the lightmap texel size on 3D scenes in the Import dock.
+			To further speed up bake times, decrease [member bounces], disable [member use_denoiser] and/or decrease [member texel_scale].
+			To further increase quality, enable [member supersampling] and/or increase [member texel_scale].
 		</member>
 		<member name="shadowmask_mode" type="int" setter="set_shadowmask_mode" getter="get_shadowmask_mode" enum="LightmapGIData.ShadowmaskMode" default="0" experimental="">
 			The shadowmasking policy to use for directional shadows on static objects that are baked with this [LightmapGI] instance.
 			Shadowmasking allows [DirectionalLight3D] nodes to cast shadows even outside the range defined by their [member DirectionalLight3D.directional_shadow_max_distance] property. This is done by baking a texture that contains a shadowmap for the directional light, then using this texture according to the current shadowmask mode.
 			[b]Note:[/b] The shadowmask texture is only created if [member shadowmask_mode] is not [constant LightmapGIData.SHADOWMASK_MODE_NONE]. To see a difference, you need to bake lightmaps again after switching from [constant LightmapGIData.SHADOWMASK_MODE_NONE] to any other mode.
 		</member>
+		<member name="supersampling" type="bool" setter="set_supersampling_enabled" getter="is_supersampling_enabled" default="false">
+			If [code]true[/code], lightmaps are baked with the texel scale multiplied with [member supersampling_factor] and downsampled before saving the lightmap (so the effective texel density is identical to having supersampling disabled).
+			Supersampling provides increased lightmap quality with less noise, smoother shadows and better shadowing of small-scale features in objects. However, it may result in significantly increased bake times and memory usage while baking lightmaps. Padding is automatically adjusted to avoid increasing light leaking.
+		</member>
+		<member name="supersampling_factor" type="float" setter="set_supersampling_factor" getter="get_supersampling_factor" default="2.0">
+			The factor by which the texel density is multiplied for supersampling. For best results, use an integer value. While fractional values are allowed, they can result in increased light leaking and a blurry lightmap.
+			Higher values may result in better quality, but also increase bake times and memory usage while baking.
+			See [member supersampling] for more information.
+		</member>
 		<member name="texel_scale" type="float" setter="set_texel_scale" getter="get_texel_scale" default="1.0">
 			Scales the lightmap texel density of all meshes for the current bake. This is a multiplier that builds upon the existing lightmap texel size defined in each imported 3D scene, along with the per-mesh density multiplier (which is designed to be used when the same mesh is used at different scales). Lower values will result in faster bake times.
 			For example, doubling [member texel_scale] doubles the lightmap texture resolution for all objects [i]on each axis[/i], so it will [i]quadruple[/i] the texel count.

modules/lightmapper_rd/lightmapper_rd.cpp

@@ -234,14 +234,14 @@ void LightmapperRD::_sort_triangle_clusters(uint32_t p_cluster_size, uint32_t p_
 	}
 }
 
-Lightmapper::BakeError LightmapperRD::_blit_meshes_into_atlas(int p_max_texture_size, int p_denoiser_range, Vector<Ref<Image>> &albedo_images, Vector<Ref<Image>> &emission_images, AABB &bounds, Size2i &atlas_size, int &atlas_slices, BakeStepFunc p_step_function, void *p_bake_userdata) {
+Lightmapper::BakeError LightmapperRD::_blit_meshes_into_atlas(int p_max_texture_size, int p_denoiser_range, Vector<Ref<Image>> &albedo_images, Vector<Ref<Image>> &emission_images, AABB &bounds, Size2i &atlas_size, int &atlas_slices, float p_supersampling_factor, BakeStepFunc p_step_function, void *p_bake_userdata) {
 	Vector<Size2i> sizes;
 
 	for (int m_i = 0; m_i < mesh_instances.size(); m_i++) {
 		MeshInstance &mi = mesh_instances.write[m_i];
 		Size2i s = Size2i(mi.data.albedo_on_uv2->get_width(), mi.data.albedo_on_uv2->get_height());
 		sizes.push_back(s);
-		atlas_size = atlas_size.max(s + Size2i(2, 2).maxi(p_denoiser_range));
+		atlas_size = atlas_size.max(s + Size2i(2, 2).maxi(p_denoiser_range) * p_supersampling_factor);
 	}
 
 	int max = nearest_power_of_2_templated(atlas_size.width);
@@ -271,7 +271,10 @@ Lightmapper::BakeError LightmapperRD::_blit_meshes_into_atlas(int p_max_texture_
 		source_sizes.resize(sizes.size());
 		source_indices.resize(sizes.size());
 		for (int i = 0; i < source_indices.size(); i++) {
-			source_sizes.write[i] = sizes[i] + Vector2i(2, 2).maxi(p_denoiser_range); // Add padding between lightmaps.
+			// Add padding between lightmaps.
+			// Scale the padding if the lightmap will be downsampled at the end of the baking process
+			// Otherwise the padding would be insufficient.
+			source_sizes.write[i] = sizes[i] + Vector2i(2, 2).maxi(p_denoiser_range) * p_supersampling_factor;
 			source_indices.write[i] = i;
 		}
 		Vector<Vector3i> atlas_offsets;
@@ -1041,7 +1044,7 @@ LightmapperRD::BakeError LightmapperRD::_denoise(RenderingDevice *p_rd, Ref<RDSh
 	return BAKE_OK;
 }
 
-LightmapperRD::BakeError LightmapperRD::bake(BakeQuality p_quality, bool p_use_denoiser, float p_denoiser_strength, int p_denoiser_range, int p_bounces, float p_bounce_indirect_energy, float p_bias, int p_max_texture_size, bool p_bake_sh, bool p_bake_shadowmask, bool p_texture_for_bounces, GenerateProbes p_generate_probes, const Ref<Image> &p_environment_panorama, const Basis &p_environment_transform, BakeStepFunc p_step_function, void *p_bake_userdata, float p_exposure_normalization) {
+LightmapperRD::BakeError LightmapperRD::bake(BakeQuality p_quality, bool p_use_denoiser, float p_denoiser_strength, int p_denoiser_range, int p_bounces, float p_bounce_indirect_energy, float p_bias, int p_max_texture_size, bool p_bake_sh, bool p_bake_shadowmask, bool p_texture_for_bounces, GenerateProbes p_generate_probes, const Ref<Image> &p_environment_panorama, const Basis &p_environment_transform, BakeStepFunc p_step_function, void *p_bake_userdata, float p_exposure_normalization, float p_supersampling_factor) {
 	int denoiser = GLOBAL_GET("rendering/lightmapping/denoising/denoiser");
 	String oidn_path = EDITOR_GET("filesystem/tools/oidn/oidn_denoise_path");
 
@@ -1074,7 +1077,7 @@ LightmapperRD::BakeError LightmapperRD::bake(BakeQuality p_quality, bool p_use_d
 	Vector<Ref<Image>> albedo_images;
 	Vector<Ref<Image>> emission_images;
 
-	BakeError bake_error = _blit_meshes_into_atlas(p_max_texture_size, p_denoiser_range, albedo_images, emission_images, bounds, atlas_size, atlas_slices, p_step_function, p_bake_userdata);
+	BakeError bake_error = _blit_meshes_into_atlas(p_max_texture_size, p_denoiser_range, albedo_images, emission_images, bounds, atlas_size, atlas_slices, p_supersampling_factor, p_step_function, p_bake_userdata);
 	if (bake_error != BAKE_OK) {
 		return bake_error;
 	}
@@ -1330,6 +1333,7 @@ LightmapperRD::BakeError LightmapperRD::bake(BakeQuality p_quality, bool p_use_d
 	bake_parameters.shadowmask_light_idx = shadowmask_light_idx;
 	// Same number of rays for transparency regardless of quality (it's more of a retry rather than shooting new ones).
 	bake_parameters.transparency_rays = GLOBAL_GET("rendering/lightmapping/bake_performance/max_transparency_rays");
+	bake_parameters.supersampling_factor = p_supersampling_factor;
 
 	bake_parameters_buffer = rd->uniform_buffer_create(sizeof(BakeParameters));
 	rd->buffer_update(bake_parameters_buffer, 0, sizeof(BakeParameters), &bake_parameters);

modules/lightmapper_rd/lightmapper_rd.h

@@ -59,7 +59,7 @@ class LightmapperRD : public Lightmapper {
 		float bounce_indirect_energy = 0.0f;
 		uint32_t shadowmask_light_idx = 0;
 		uint32_t transparency_rays = 0;
-		uint32_t pad[1] = {};
+		float supersampling_factor = 0.0f;
 	};
 
 	struct MeshInstance {
@@ -260,6 +260,11 @@ class LightmapperRD : public Lightmapper {
 	Vector<Ref<Image>> shadowmask_textures;
 	Vector<Color> probe_values;
 
+	struct DilateParams {
+		uint32_t radius;
+		uint32_t pad[3];
+	};
+
 	struct DenoiseParams {
 		float spatial_bandwidth;
 		float light_bandwidth;
@@ -271,7 +276,7 @@ class LightmapperRD : public Lightmapper {
 		float pad[2];
 	};
 
-	BakeError _blit_meshes_into_atlas(int p_max_texture_size, int p_denoiser_range, Vector<Ref<Image>> &albedo_images, Vector<Ref<Image>> &emission_images, AABB &bounds, Size2i &atlas_size, int &atlas_slices, BakeStepFunc p_step_function, void *p_bake_userdata);
+	BakeError _blit_meshes_into_atlas(int p_max_texture_size, int p_denoiser_range, Vector<Ref<Image>> &albedo_images, Vector<Ref<Image>> &emission_images, AABB &bounds, Size2i &atlas_size, int &atlas_slices, float p_supersampling_factor, BakeStepFunc p_step_function, void *p_bake_userdata);
 	void _create_acceleration_structures(RenderingDevice *rd, Size2i atlas_size, int atlas_slices, AABB &bounds, int grid_size, uint32_t p_cluster_size, Vector<Probe> &probe_positions, GenerateProbes p_generate_probes, Vector<int> &slice_triangle_count, Vector<int> &slice_seam_count, RID &vertex_buffer, RID &triangle_buffer, RID &lights_buffer, RID &r_triangle_indices_buffer, RID &r_cluster_indices_buffer, RID &r_cluster_aabbs_buffer, RID &probe_positions_buffer, RID &grid_texture, RID &seams_buffer, BakeStepFunc p_step_function, void *p_bake_userdata);
 	void _raster_geometry(RenderingDevice *rd, Size2i atlas_size, int atlas_slices, int grid_size, AABB bounds, float p_bias, Vector<int> slice_triangle_count, RID position_tex, RID unocclude_tex, RID normal_tex, RID raster_depth_buffer, RID rasterize_shader, RID raster_base_uniform);
 
@@ -289,7 +294,7 @@ class LightmapperRD : public Lightmapper {
 	virtual void add_omni_light(const String &p_name, bool p_static, const Vector3 &p_position, const Color &p_color, float p_energy, float p_indirect_energy, float p_range, float p_attenuation, float p_size, float p_shadow_blur) override;
 	virtual void add_spot_light(const String &p_name, bool p_static, const Vector3 &p_position, const Vector3 p_direction, const Color &p_color, float p_energy, float p_indirect_energy, float p_range, float p_attenuation, float p_spot_angle, float p_spot_attenuation, float p_size, float p_shadow_blur) override;
 	virtual void add_probe(const Vector3 &p_position) override;
-	virtual BakeError bake(BakeQuality p_quality, bool p_use_denoiser, float p_denoiser_strength, int p_denoiser_range, int p_bounces, float p_bounce_indirect_energy, float p_bias, int p_max_texture_size, bool p_bake_sh, bool p_bake_shadowmask, bool p_texture_for_bounces, GenerateProbes p_generate_probes, const Ref<Image> &p_environment_panorama, const Basis &p_environment_transform, BakeStepFunc p_step_function = nullptr, void *p_bake_userdata = nullptr, float p_exposure_normalization = 1.0) override;
+	virtual BakeError bake(BakeQuality p_quality, bool p_use_denoiser, float p_denoiser_strength, int p_denoiser_range, int p_bounces, float p_bounce_indirect_energy, float p_bias, int p_max_texture_size, bool p_bake_sh, bool p_bake_shadowmask, bool p_texture_for_bounces, GenerateProbes p_generate_probes, const Ref<Image> &p_environment_panorama, const Basis &p_environment_transform, BakeStepFunc p_step_function = nullptr, void *p_bake_userdata = nullptr, float p_exposure_normalization = 1.0, float p_supersampling_factor = 1.0f) override;
 
 	int get_bake_texture_count() const override;
 	Ref<Image> get_bake_texture(int p_index) const override;

modules/lightmapper_rd/lm_common_inc.glsl

@@ -18,7 +18,7 @@ layout(set = 0, binding = 0) uniform BakeParameters {
 	float bounce_indirect_energy;
 	int shadowmask_light_idx;
 	uint transparency_rays;
-	uint pad0;
+	float supersampling_factor;
 }
 bake_params;
 

modules/lightmapper_rd/lm_compute.glsl

@@ -820,7 +820,7 @@ void main() {
 		// Empty texel, try again.
 		neighbor_position.xyz = texelFetch(sampler2DArray(source_position, linear_sampler), ivec3(atlas_pos + ivec2(-1, 0), params.atlas_slice), 0).xyz;
 	}
-	float texel_size_world_space = distance(position, neighbor_position.xyz);
+	float texel_size_world_space = distance(position, neighbor_position.xyz) * bake_params.supersampling_factor;
 
 	vec3 light_for_texture = vec3(0.0);
 	vec3 light_for_bounces = vec3(0.0);
@@ -1059,43 +1059,35 @@ void main() {
 
 #endif
 
-#if defined(MODE_DILATE)
-
-	vec4 c = texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos, params.atlas_slice), 0);
-	//sides first, as they are closer
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-1, 0), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(0, 1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(1, 0), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(0, -1), params.atlas_slice), 0);
-	//endpoints second
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-1, -1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-1, 1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(1, -1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(1, 1), params.atlas_slice), 0);
-
-	//far sides third
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-2, 0), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(0, 2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(2, 0), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(0, -2), params.atlas_slice), 0);
-
-	//far-mid endpoints
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-2, -1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-2, 1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(2, -1), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(2, 1), params.atlas_slice), 0);
-
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-1, -2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-1, 2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(1, -2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(1, 2), params.atlas_slice), 0);
-	//far endpoints
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-2, -2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(-2, 2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(2, -2), params.atlas_slice), 0);
-	c = c.a > 0.5 ? c : texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos + ivec2(2, 2), params.atlas_slice), 0);
-
-	imageStore(dest_light, ivec3(atlas_pos, params.atlas_slice), c);
+#ifdef MODE_DILATE
+
+	const int max_radius = int(4.0 * bake_params.supersampling_factor);
+	const ivec2 directions[8] = ivec2[8](ivec2(-1, 0), ivec2(0, 1), ivec2(1, 0), ivec2(0, -1), ivec2(-1, -1), ivec2(-1, 1), ivec2(1, -1), ivec2(1, 1));
+
+	vec4 texel_color = texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(atlas_pos, params.atlas_slice), 0);
+
+	for (int radius = 1; radius <= max_radius; radius++) {
+		for (uint i = 0; i < 8; i++) {
+			const ivec2 sample_pos = atlas_pos + directions[i] * radius;
+			// Texture bounds check for robustness.
+			if (any(lessThan(sample_pos, ivec2(0))) ||
+					any(greaterThanEqual(sample_pos, textureSize(source_light, 0).xy))) {
+				continue;
+			}
+
+			vec4 neighbor_color = texelFetch(sampler2DArray(source_light, linear_sampler), ivec3(sample_pos, params.atlas_slice), 0);
+			if (neighbor_color.a > 0.5) {
+				texel_color = neighbor_color;
+				break;
+			}
+		}
+
+		if (texel_color.a > 0.5) {
+			break;
+		}
+	}
+
+	imageStore(dest_light, ivec3(atlas_pos, params.atlas_slice), texel_color);
 
 #endif