Rewrite culling to be more cache/thread friendly.

-Uses a single array with all data
-Massive performance improvement
-Does not support threads yet, but code is now thread friendly

4 files changed, 20 insertions, 13 deletions

diff --git a/servers/rendering/renderer_rd/renderer_scene_render_forward.cpp b/servers/rendering/renderer_rd/renderer_scene_render_forward.cpp
index c6b2fa6dc0..983fbc9b64 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_forward.cpp
+++ b/servers/rendering/renderer_rd/renderer_scene_render_forward.cpp
@@ -1532,7 +1532,7 @@ void RendererSceneRenderForward::_add_geometry_with_material(InstanceBase *p_ins
 	}
 }
 
-void RendererSceneRenderForward::_fill_render_list(const PagedArray<InstanceBase *> &p_instances, PassMode p_pass_mode, bool p_using_sdfgi) {
+void RendererSceneRenderForward::_fill_render_list(const PagedArray<InstanceBase *> &p_instances, PassMode p_pass_mode, const CameraMatrix &p_cam_projection, const Transform &p_cam_transform, bool p_using_sdfgi) {
 	scene_state.current_shader_index = 0;
 	scene_state.current_material_index = 0;
 	scene_state.used_sss = false;
@@ -1540,6 +1540,10 @@ void RendererSceneRenderForward::_fill_render_list(const PagedArray<InstanceBase
 	scene_state.used_normal_texture = false;
 	scene_state.used_depth_texture = false;
 
+	Plane near_plane(p_cam_transform.origin, -p_cam_transform.basis.get_axis(Vector3::AXIS_Z));
+	near_plane.d += p_cam_projection.get_z_near();
+	float z_max = p_cam_projection.get_z_far() - p_cam_projection.get_z_near();
+
 	uint32_t geometry_index = 0;
 
 	//fill list
@@ -1547,6 +1551,9 @@ void RendererSceneRenderForward::_fill_render_list(const PagedArray<InstanceBase
 	for (int i = 0; i < (int)p_instances.size(); i++) {
 		InstanceBase *inst = p_instances[i];
 
+		inst->depth = near_plane.distance_to(inst->transform.origin);
+		inst->depth_layer = CLAMP(int(inst->depth * 16 / z_max), 0, 15);
+
 		//add geometry for drawing
 		switch (inst->base_type) {
 			case RS::INSTANCE_MESH: {
@@ -1782,7 +1789,7 @@ void RendererSceneRenderForward::_render_scene(RID p_render_buffer, const Transf
 	_update_render_base_uniform_set(); //may have changed due to the above (light buffer enlarged, as an example)
 
 	render_list.clear();
-	_fill_render_list(p_instances, PASS_MODE_COLOR, using_sdfgi);
+	_fill_render_list(p_instances, PASS_MODE_COLOR, p_cam_projection, p_cam_transform, using_sdfgi);
 
 	bool using_sss = !low_end && render_buffer && scene_state.used_sss && sub_surface_scattering_get_quality() != RS::SUB_SURFACE_SCATTERING_QUALITY_DISABLED;
 
@@ -2046,7 +2053,7 @@ void RendererSceneRenderForward::_render_shadow(RID p_framebuffer, const PagedAr
 
 	PassMode pass_mode = p_use_dp ? PASS_MODE_SHADOW_DP : PASS_MODE_SHADOW;
 
-	_fill_render_list(p_instances, pass_mode);
+	_fill_render_list(p_instances, pass_mode, p_projection, p_transform);
 
 	RID rp_uniform_set = _setup_render_pass_uniform_set(RID(), RID(), RID(), RID(), PagedArray<RID>());
 
@@ -2079,7 +2086,7 @@ void RendererSceneRenderForward::_render_particle_collider_heightfield(RID p_fb,
 
 	PassMode pass_mode = PASS_MODE_SHADOW;
 
-	_fill_render_list(p_instances, pass_mode);
+	_fill_render_list(p_instances, pass_mode, p_cam_projection, p_cam_transform);
 
 	RID rp_uniform_set = _setup_render_pass_uniform_set(RID(), RID(), RID(), RID(), PagedArray<RID>());
 
@@ -2112,7 +2119,7 @@ void RendererSceneRenderForward::_render_material(const Transform &p_cam_transfo
 	render_list.clear();
 
 	PassMode pass_mode = PASS_MODE_DEPTH_MATERIAL;
-	_fill_render_list(p_instances, pass_mode);
+	_fill_render_list(p_instances, pass_mode, p_cam_projection, p_cam_transform);
 
 	RID rp_uniform_set = _setup_render_pass_uniform_set(RID(), RID(), RID(), RID(), PagedArray<RID>());
 
@@ -2151,7 +2158,7 @@ void RendererSceneRenderForward::_render_uv2(const PagedArray<InstanceBase *> &p
 	render_list.clear();
 
 	PassMode pass_mode = PASS_MODE_DEPTH_MATERIAL;
-	_fill_render_list(p_instances, pass_mode);
+	_fill_render_list(p_instances, pass_mode, CameraMatrix(), Transform());
 
 	RID rp_uniform_set = _setup_render_pass_uniform_set(RID(), RID(), RID(), RID(), PagedArray<RID>());
 
@@ -2209,7 +2216,7 @@ void RendererSceneRenderForward::_render_sdfgi(RID p_render_buffers, const Vecto
 	render_list.clear();
 
 	PassMode pass_mode = PASS_MODE_SDF;
-	_fill_render_list(p_instances, pass_mode);
+	_fill_render_list(p_instances, pass_mode, CameraMatrix(), Transform());
 	render_list.sort_by_key(false);
 	_fill_instances(render_list.elements, render_list.element_count, true);
 
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_forward.h b/servers/rendering/renderer_rd/renderer_scene_render_forward.h
index 5d77c13b43..eca83893c3 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_forward.h
+++ b/servers/rendering/renderer_rd/renderer_scene_render_forward.h
@@ -574,7 +574,7 @@ class RendererSceneRenderForward : public RendererSceneRenderRD {
 	_FORCE_INLINE_ void _add_geometry(InstanceBase *p_instance, uint32_t p_surface, RID p_material, PassMode p_pass_mode, uint32_t p_geometry_index, bool p_using_sdfgi = false);
 	_FORCE_INLINE_ void _add_geometry_with_material(InstanceBase *p_instance, uint32_t p_surface, MaterialData *p_material, RID p_material_rid, PassMode p_pass_mode, uint32_t p_geometry_index, bool p_using_sdfgi = false);
 
-	void _fill_render_list(const PagedArray<InstanceBase *> &p_instances, PassMode p_pass_mode, bool p_using_sdfgi = false);
+	void _fill_render_list(const PagedArray<InstanceBase *> &p_instances, PassMode p_pass_mode, const CameraMatrix &p_cam_projection, const Transform &p_cam_transform, bool p_using_sdfgi = false);
 
 	Map<Size2i, RID> sdfgi_framebuffer_size_cache;
 
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
index 02ec399f58..c9edaa3d37 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
@@ -1153,7 +1153,7 @@ void RendererSceneRenderRD::_sdfgi_update_cascades(RID p_render_buffers) {
 	RD::get_singleton()->buffer_update(rb->sdfgi->cascades_ubo, 0, sizeof(SDFGI::Cascade::UBO) * SDFGI::MAX_CASCADES, cascade_data, true);
 }
 
-void RendererSceneRenderRD::sdfgi_update_probes(RID p_render_buffers, RID p_environment, const PagedArray<RID> &p_directional_light_instances, const RID *p_positional_light_instances, uint32_t p_positional_light_count) {
+void RendererSceneRenderRD::sdfgi_update_probes(RID p_render_buffers, RID p_environment, const Vector<RID> &p_directional_lights, const RID *p_positional_light_instances, uint32_t p_positional_light_count) {
 	RenderBuffers *rb = render_buffers_owner.getornull(p_render_buffers);
 	ERR_FAIL_COND(rb == nullptr);
 	if (rb->sdfgi == nullptr) {
@@ -1179,12 +1179,12 @@ void RendererSceneRenderRD::sdfgi_update_probes(RID p_render_buffers, RID p_envi
 
 			SDGIShader::Light lights[SDFGI::MAX_DYNAMIC_LIGHTS];
 			uint32_t idx = 0;
-			for (uint32_t j = 0; j < (uint32_t)p_directional_light_instances.size(); j++) {
+			for (uint32_t j = 0; j < (uint32_t)p_directional_lights.size(); j++) {
 				if (idx == SDFGI::MAX_DYNAMIC_LIGHTS) {
 					break;
 				}
 
-				LightInstance *li = light_instance_owner.getornull(p_directional_light_instances[j]);
+				LightInstance *li = light_instance_owner.getornull(p_directional_lights[j]);
 				ERR_CONTINUE(!li);
 
 				if (storage->light_directional_is_sky_only(li->light)) {
@@ -8485,7 +8485,7 @@ RendererSceneRenderRD::RendererSceneRenderRD(RendererStorageRD *p_storage) {
 		cluster.lights_instances = memnew_arr(RID, cluster.max_lights);
 		cluster.lights_shadow_rect_cache = memnew_arr(Rect2i, cluster.max_lights);
 
-		cluster.max_directional_lights = 8;
+		cluster.max_directional_lights = MAX_DIRECTIONAL_LIGHTS;
 		uint32_t directional_light_buffer_size = cluster.max_directional_lights * sizeof(Cluster::DirectionalLightData);
 		cluster.directional_lights = memnew_arr(Cluster::DirectionalLightData, cluster.max_directional_lights);
 		cluster.directional_light_buffer = RD::get_singleton()->uniform_buffer_create(directional_light_buffer_size);
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.h b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
index ded6d99e47..e4dc98571e 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.h
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
@@ -1516,7 +1516,7 @@ public:
 	virtual int sdfgi_get_pending_region_count(RID p_render_buffers) const;
 	virtual AABB sdfgi_get_pending_region_bounds(RID p_render_buffers, int p_region) const;
 	virtual uint32_t sdfgi_get_pending_region_cascade(RID p_render_buffers, int p_region) const;
-	virtual void sdfgi_update_probes(RID p_render_buffers, RID p_environment, const PagedArray<RID> &p_directional_light_instances, const RID *p_positional_light_instances, uint32_t p_positional_light_count);
+	virtual void sdfgi_update_probes(RID p_render_buffers, RID p_environment, const Vector<RID> &p_directional_lights, const RID *p_positional_light_instances, uint32_t p_positional_light_count);
 	RID sdfgi_get_ubo() const { return gi.sdfgi_ubo; }
 	/* SKY API */