Merge pull request #50501 from reduz/fix-forward-mobile

Fixes to forward mobile

5 files changed, 172 insertions, 155 deletions

diff --git a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
index 03419ec8f5..4985fd1687 100644
--- a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
+++ b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
@@ -35,6 +35,34 @@
 
 using namespace RendererSceneRenderImplementation;
 
+RenderForwardMobile::ForwardID RenderForwardMobile::_allocate_forward_id(ForwardIDType p_type) {
+	int32_t index = -1;
+	for (uint32_t i = 0; i < forward_id_allocators[p_type].allocations.size(); i++) {
+		if (forward_id_allocators[p_type].allocations[i] == false) {
+			index = i;
+			break;
+		}
+	}
+
+	if (index == -1) {
+		index = forward_id_allocators[p_type].allocations.size();
+		forward_id_allocators[p_type].allocations.push_back(true);
+		forward_id_allocators[p_type].map.push_back(0xFF);
+	} else {
+		forward_id_allocators[p_type].allocations[index] = true;
+	}
+
+	return index;
+}
+void RenderForwardMobile::_free_forward_id(ForwardIDType p_type, ForwardID p_id) {
+	ERR_FAIL_INDEX(p_id, (ForwardID)forward_id_allocators[p_type].allocations.size());
+	forward_id_allocators[p_type].allocations[p_id] = false;
+}
+
+void RenderForwardMobile::_map_forward_id(ForwardIDType p_type, ForwardID p_id, uint32_t p_index) {
+	forward_id_allocators[p_type].map[p_id] = p_index;
+}
+
 /* Render buffer */
 
 void RenderForwardMobile::RenderBufferDataForwardMobile::clear() {
@@ -68,7 +96,7 @@ void RenderForwardMobile::RenderBufferDataForwardMobile::configure(RID p_color_b
 	RD::DataFormat color_format = RenderForwardMobile::singleton->_render_buffers_get_color_format();
 
 	if (p_msaa == RS::VIEWPORT_MSAA_DISABLED) {
-		if (color_format == RD::DATA_FORMAT_B10G11R11_UFLOAT_PACK32) {
+		if (color_format == RD::DATA_FORMAT_A2B10G10R10_UNORM_PACK32) {
 			// @TODO add a second color buffer for alpha as this format is RGB only
 		}
 
@@ -139,7 +167,14 @@ RD::DataFormat RenderForwardMobile::_render_buffers_get_color_format() {
 	// Using 32bit buffers enables AFBC on mobile devices which should have a definate performance improvement (MALI G710 and newer support this on 64bit RTs)
 	// NO ALPHA and unsigned float.
 	// @TODO No alpha is an issue, recommendation here is to add a second RT for alpha
-	return RD::DATA_FORMAT_B10G11R11_UFLOAT_PACK32;
+	return RD::DATA_FORMAT_A2B10G10R10_UNORM_PACK32;
+}
+
+bool RenderForwardMobile::_render_buffers_can_be_storage() {
+	// Using 32bit buffers enables AFBC on mobile devices which should have a definate performance improvement (MALI G710 and newer support this on 64bit RTs)
+	// NO ALPHA and unsigned float.
+	// @TODO No alpha is an issue, recommendation here is to add a second RT for alpha
+	return false;
 }
 
 RID RenderForwardMobile::_setup_render_pass_uniform_set(RenderListType p_render_list, const RenderDataRD *p_render_data, RID p_radiance_texture, bool p_use_directional_shadow_atlas, int p_index) {
@@ -1403,6 +1438,44 @@ void RenderForwardMobile::_render_list_with_threads(RenderListParameters *p_para
 	}
 }
 
+void RenderForwardMobile::_fill_push_constant_instance_indices(GeometryInstanceForwardMobile::PushConstant *p_push_constant, const GeometryInstanceForwardMobile *p_instance) {
+	// first zero out our indices
+
+	p_push_constant->omni_lights[0] = 0xFFFF;
+	p_push_constant->omni_lights[1] = 0xFFFF;
+
+	p_push_constant->spot_lights[0] = 0xFFFF;
+	p_push_constant->spot_lights[1] = 0xFFFF;
+
+	p_push_constant->decals[0] = 0xFFFF;
+	p_push_constant->decals[1] = 0xFFFF;
+
+	p_push_constant->reflection_probes[0] = 0xFFFF;
+	p_push_constant->reflection_probes[1] = 0xFFFF;
+
+	for (uint32_t i = 0; i < MAX_RDL_CULL; i++) {
+		uint32_t ofs = i < 4 ? 0 : 1;
+		uint32_t shift = (i & 0x3) << 3;
+		uint32_t mask = ~(0xFF << shift);
+		if (i < p_instance->omni_light_count) {
+			p_push_constant->omni_lights[ofs] &= mask;
+			p_push_constant->omni_lights[ofs] |= uint32_t(forward_id_allocators[FORWARD_ID_TYPE_OMNI_LIGHT].map[p_instance->omni_lights[i]]) << shift;
+		}
+		if (i < p_instance->spot_light_count) {
+			p_push_constant->spot_lights[ofs] &= mask;
+			p_push_constant->spot_lights[ofs] |= uint32_t(forward_id_allocators[FORWARD_ID_TYPE_SPOT_LIGHT].map[p_instance->spot_lights[i]]) << shift;
+		}
+		if (i < p_instance->decals_count) {
+			p_push_constant->decals[ofs] &= mask;
+			p_push_constant->decals[ofs] |= uint32_t(forward_id_allocators[FORWARD_ID_TYPE_DECAL].map[p_instance->decals[i]]) << shift;
+		}
+		if (i < p_instance->reflection_probe_count) {
+			p_push_constant->reflection_probes[ofs] &= mask;
+			p_push_constant->reflection_probes[ofs] |= uint32_t(forward_id_allocators[FORWARD_ID_TYPE_REFLECTION_PROBE].map[p_instance->reflection_probes[i]]) << shift;
+		}
+	}
+}
+
 template <RenderForwardMobile::PassMode p_pass_mode>
 void RenderForwardMobile::_render_list_template(RenderingDevice::DrawListID p_draw_list, RenderingDevice::FramebufferFormatID p_framebuffer_Format, RenderListParameters *p_params, uint32_t p_from_element, uint32_t p_to_element) {
 	RD::DrawListID draw_list = p_draw_list;
@@ -1452,8 +1525,6 @@ void RenderForwardMobile::_render_list_template(RenderingDevice::DrawListID p_dr
 			push_constant.lightmap_uv_scale[3] = inst->lightmap_uv_scale.size.y;
 		};
 
-		_fill_instance_indices(inst->omni_lights, inst->omni_light_count, push_constant.omni_lights, inst->spot_lights, inst->spot_light_count, push_constant.spot_lights, inst->reflection_probes, inst->reflection_probe_count, push_constant.reflection_probes, inst->decals, inst->decals_count, push_constant.decals, push_constant.layer_mask);
-
 		RID material_uniform_set;
 		SceneShaderForwardMobile::ShaderData *shader;
 		void *mesh_surface;
@@ -1464,6 +1535,8 @@ void RenderForwardMobile::_render_list_template(RenderingDevice::DrawListID p_dr
 			mesh_surface = surf->surface_shadow;
 
 		} else {
+			_fill_push_constant_instance_indices(&push_constant, inst);
+
 #ifdef DEBUG_ENABLED
 			if (unlikely(get_debug_draw_mode() == RS::VIEWPORT_DEBUG_DRAW_LIGHTING)) {
 				material_uniform_set = scene_shader.default_material_uniform_set;
@@ -1772,13 +1845,13 @@ void RenderForwardMobile::geometry_instance_pair_light_instances(GeometryInstanc
 		switch (type) {
 			case RS::LIGHT_OMNI: {
 				if (ginstance->omni_light_count < (uint32_t)MAX_RDL_CULL) {
-					ginstance->omni_lights[ginstance->omni_light_count] = p_light_instances[i];
+					ginstance->omni_lights[ginstance->omni_light_count] = light_instance_get_forward_id(p_light_instances[i]);
 					ginstance->omni_light_count++;
 				}
 			} break;
 			case RS::LIGHT_SPOT: {
 				if (ginstance->spot_light_count < (uint32_t)MAX_RDL_CULL) {
-					ginstance->spot_lights[ginstance->spot_light_count] = p_light_instances[i];
+					ginstance->spot_lights[ginstance->spot_light_count] = light_instance_get_forward_id(p_light_instances[i]);
 					ginstance->spot_light_count++;
 				}
 			} break;
@@ -1794,7 +1867,7 @@ void RenderForwardMobile::geometry_instance_pair_reflection_probe_instances(Geom
 
 	ginstance->reflection_probe_count = p_reflection_probe_instance_count < (uint32_t)MAX_RDL_CULL ? p_reflection_probe_instance_count : (uint32_t)MAX_RDL_CULL;
 	for (uint32_t i = 0; i < ginstance->reflection_probe_count; i++) {
-		ginstance->reflection_probes[i] = p_reflection_probe_instances[i];
+		ginstance->reflection_probes[i] = reflection_probe_instance_get_forward_id(p_reflection_probe_instances[i]);
 	}
 }
 
@@ -1804,7 +1877,7 @@ void RenderForwardMobile::geometry_instance_pair_decal_instances(GeometryInstanc
 
 	ginstance->decals_count = p_decal_instance_count < (uint32_t)MAX_RDL_CULL ? p_decal_instance_count : (uint32_t)MAX_RDL_CULL;
 	for (uint32_t i = 0; i < ginstance->decals_count; i++) {
-		ginstance->decals[i] = p_decal_instances[i];
+		ginstance->decals[i] = decal_instance_get_forward_id(p_decal_instances[i]);
 	}
 }
 
diff --git a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h
index bf3aa7ed5c..5d28cdb5d3 100644
--- a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h
+++ b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h
@@ -42,6 +42,18 @@ namespace RendererSceneRenderImplementation {
 class RenderForwardMobile : public RendererSceneRenderRD {
 	friend SceneShaderForwardMobile;
 
+	struct ForwardIDAllocator {
+		LocalVector<bool> allocations;
+		LocalVector<uint8_t> map;
+	};
+
+	ForwardIDAllocator forward_id_allocators[FORWARD_ID_MAX];
+
+	virtual ForwardID _allocate_forward_id(ForwardIDType p_type) override;
+	virtual void _free_forward_id(ForwardIDType p_type, ForwardID p_id) override;
+	virtual void _map_forward_id(ForwardIDType p_type, ForwardID p_id, uint32_t p_index) override;
+	virtual bool _uses_forward_ids() const override { return true; }
+
 protected:
 	/* Scene Shader */
 
@@ -152,6 +164,7 @@ protected:
 	};
 
 	virtual RD::DataFormat _render_buffers_get_color_format() override;
+	virtual bool _render_buffers_can_be_storage() override;
 
 	RID _setup_render_pass_uniform_set(RenderListType p_render_list, const RenderDataRD *p_render_data, RID p_radiance_texture, bool p_use_directional_shadow_atlas = false, int p_index = 0);
 	virtual void _render_scene(RenderDataRD *p_render_data, const Color &p_default_bg_color) override;
@@ -515,14 +528,14 @@ protected:
 		GeometryInstanceLightmapSH *lightmap_sh = nullptr;
 
 		// culled light info
-		uint32_t reflection_probe_count;
-		RID reflection_probes[MAX_RDL_CULL];
-		uint32_t omni_light_count;
-		RID omni_lights[MAX_RDL_CULL];
-		uint32_t spot_light_count;
-		RID spot_lights[MAX_RDL_CULL];
-		uint32_t decals_count;
-		RID decals[MAX_RDL_CULL];
+		uint32_t reflection_probe_count = 0;
+		ForwardID reflection_probes[MAX_RDL_CULL];
+		uint32_t omni_light_count = 0;
+		ForwardID omni_lights[MAX_RDL_CULL];
+		uint32_t spot_light_count = 0;
+		ForwardID spot_lights[MAX_RDL_CULL];
+		uint32_t decals_count = 0;
+		ForwardID decals[MAX_RDL_CULL];
 
 		GeometryInstanceSurfaceDataCache *surface_caches = nullptr;
 
@@ -554,6 +567,8 @@ protected:
 				dirty_list_element(this) {}
 	};
 
+	_FORCE_INLINE_ void _fill_push_constant_instance_indices(GeometryInstanceForwardMobile::PushConstant *p_push_constant, const GeometryInstanceForwardMobile *p_instance);
+
 public:
 	static void _geometry_instance_dependency_changed(RendererStorage::DependencyChangedNotification p_notification, RendererStorage::DependencyTracker *p_tracker);
 	static void _geometry_instance_dependency_deleted(const RID &p_dependency, RendererStorage::DependencyTracker *p_tracker);
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
index 9b32971e30..7b5f448c18 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
@@ -567,6 +567,8 @@ int RendererSceneRenderRD::reflection_atlas_get_size(RID p_ref_atlas) const {
 RID RendererSceneRenderRD::reflection_probe_instance_create(RID p_probe) {
 	ReflectionProbeInstance rpi;
 	rpi.probe = p_probe;
+	rpi.forward_id = _allocate_forward_id(FORWARD_ID_TYPE_REFLECTION_PROBE);
+
 	return reflection_probe_instance_owner.make_rid(rpi);
 }
 
@@ -659,7 +661,7 @@ bool RendererSceneRenderRD::reflection_probe_instance_begin_render(RID p_instanc
 			tf.mipmaps = mipmaps;
 			tf.width = atlas->size;
 			tf.height = atlas->size;
-			tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
+			tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT | (_render_buffers_can_be_storage() ? RD::TEXTURE_USAGE_STORAGE_BIT : 0);
 
 			atlas->reflection = RD::get_singleton()->texture_create(tf, RD::TextureView());
 		}
@@ -1233,6 +1235,9 @@ RID RendererSceneRenderRD::light_instance_create(RID p_light) {
 	light_instance->self = li;
 	light_instance->light = p_light;
 	light_instance->light_type = storage->light_get_type(p_light);
+	if (light_instance->light_type != RS::LIGHT_DIRECTIONAL) {
+		light_instance->forward_id = _allocate_forward_id(light_instance->light_type == RS::LIGHT_OMNI ? FORWARD_ID_TYPE_OMNI_LIGHT : FORWARD_ID_TYPE_SPOT_LIGHT);
+	}
 
 	return li;
 }
@@ -1306,6 +1311,7 @@ RendererSceneRenderRD::ShadowCubemap *RendererSceneRenderRD::_get_shadow_cubemap
 RID RendererSceneRenderRD::decal_instance_create(RID p_decal) {
 	DecalInstance di;
 	di.decal = p_decal;
+	di.forward_id = _allocate_forward_id(FORWARD_ID_TYPE_DECAL);
 	return decal_instance_owner.make_rid(di);
 }
 
@@ -2122,6 +2128,10 @@ RD::DataFormat RendererSceneRenderRD::_render_buffers_get_color_format() {
 	return RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
 }
 
+bool RendererSceneRenderRD::_render_buffers_can_be_storage() {
+	return true;
+}
+
 void RendererSceneRenderRD::render_buffers_configure(RID p_render_buffers, RID p_render_target, int p_width, int p_height, RS::ViewportMSAA p_msaa, RenderingServer::ViewportScreenSpaceAA p_screen_space_aa, bool p_use_debanding, uint32_t p_view_count) {
 	ERR_FAIL_COND_MSG(p_view_count == 0, "Must have atleast 1 view");
 
@@ -2152,9 +2162,9 @@ void RendererSceneRenderRD::render_buffers_configure(RID p_render_buffers, RID p
 		tf.width = rb->width;
 		tf.height = rb->height;
 		tf.array_layers = rb->view_count; // create a layer for every view
-		tf.usage_bits = RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
+		tf.usage_bits = RD::TEXTURE_USAGE_SAMPLING_BIT | (_render_buffers_can_be_storage() ? RD::TEXTURE_USAGE_STORAGE_BIT : 0);
 		if (rb->msaa != RS::VIEWPORT_MSAA_DISABLED) {
-			tf.usage_bits |= RD::TEXTURE_USAGE_CAN_COPY_TO_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
+			tf.usage_bits |= RD::TEXTURE_USAGE_CAN_COPY_TO_BIT | (_render_buffers_can_be_storage() ? RD::TEXTURE_USAGE_STORAGE_BIT : 0);
 		} else {
 			tf.usage_bits |= RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT;
 		}
@@ -2328,10 +2338,13 @@ void RendererSceneRenderRD::_setup_reflections(const PagedArray<RID> &p_reflecti
 		sort_array.sort(cluster.reflection_sort, cluster.reflection_count);
 	}
 
+	bool using_forward_ids = _uses_forward_ids();
 	for (uint32_t i = 0; i < cluster.reflection_count; i++) {
 		ReflectionProbeInstance *rpi = cluster.reflection_sort[i].instance;
 
-		rpi->render_index = i;
+		if (using_forward_ids) {
+			_map_forward_id(FORWARD_ID_TYPE_REFLECTION_PROBE, rpi->forward_id, i);
+		}
 
 		RID base_probe = rpi->probe;
 
@@ -2628,6 +2641,8 @@ void RendererSceneRenderRD::_setup_lights(const PagedArray<RID> &p_lights, const
 		shadow_atlas = shadow_atlas_owner.getornull(p_shadow_atlas);
 	}
 
+	bool using_forward_ids = _uses_forward_ids();
+
 	for (uint32_t i = 0; i < (cluster.omni_light_count + cluster.spot_light_count); i++) {
 		uint32_t index = (i < cluster.omni_light_count) ? i : i - (cluster.omni_light_count);
 		Cluster::LightData &light_data = (i < cluster.omni_light_count) ? cluster.omni_lights[index] : cluster.spot_lights[index];
@@ -2635,6 +2650,10 @@ void RendererSceneRenderRD::_setup_lights(const PagedArray<RID> &p_lights, const
 		LightInstance *li = (i < cluster.omni_light_count) ? cluster.omni_light_sort[index].instance : cluster.spot_light_sort[index].instance;
 		RID base = li->light;
 
+		if (using_forward_ids) {
+			_map_forward_id(type == RS::LIGHT_OMNI ? FORWARD_ID_TYPE_OMNI_LIGHT : FORWARD_ID_TYPE_SPOT_LIGHT, li->forward_id, index);
+		}
+
 		Transform3D light_transform = li->transform;
 
 		float sign = storage->light_is_negative(base) ? -1 : 1;
@@ -2767,7 +2786,6 @@ void RendererSceneRenderRD::_setup_lights(const PagedArray<RID> &p_lights, const
 			light_data.shadow_enabled = false;
 		}
 
-		li->light_index = index;
 		li->cull_mask = storage->light_get_cull_mask(base);
 
 		if (current_cluster_builder != nullptr) {
@@ -2836,11 +2854,15 @@ void RendererSceneRenderRD::_setup_decals(const PagedArray<RID> &p_decals, const
 		sort_array.sort(cluster.decal_sort, cluster.decal_count);
 	}
 
+	bool using_forward_ids = _uses_forward_ids();
 	for (uint32_t i = 0; i < cluster.decal_count; i++) {
 		DecalInstance *di = cluster.decal_sort[i].instance;
 		RID decal = di->decal;
 
-		di->render_index = i;
+		if (using_forward_ids) {
+			_map_forward_id(FORWARD_ID_TYPE_DECAL, di->forward_id, i);
+		}
+
 		di->cull_mask = storage->decal_get_cull_mask(decal);
 
 		Transform3D xform = di->transform;
@@ -2957,116 +2979,6 @@ void RendererSceneRenderRD::_setup_decals(const PagedArray<RID> &p_decals, const
 	}
 }
 
-void RendererSceneRenderRD::_fill_instance_indices(const RID *p_omni_light_instances, uint32_t p_omni_light_instance_count, uint32_t *p_omni_light_indices, const RID *p_spot_light_instances, uint32_t p_spot_light_instance_count, uint32_t *p_spot_light_indices, const RID *p_reflection_probe_instances, uint32_t p_reflection_probe_instance_count, uint32_t *p_reflection_probe_indices, const RID *p_decal_instances, uint32_t p_decal_instance_count, uint32_t *p_decal_instance_indices, uint32_t p_layer_mask, uint32_t p_max_dst_words) {
-	// first zero out our indices
-	for (uint32_t i = 0; i < p_max_dst_words; i++) {
-		p_omni_light_indices[i] = 0;
-		p_spot_light_indices[i] = 0;
-		p_reflection_probe_indices[i] = 0;
-		p_decal_instance_indices[i] = 0;
-	}
-
-	{
-		// process omni lights
-		uint32_t dword = 0;
-		uint32_t shift = 0;
-
-		for (uint32_t i = 0; i < p_omni_light_instance_count && dword < p_max_dst_words; i++) {
-			LightInstance *li = light_instance_owner.getornull(p_omni_light_instances[i]);
-
-			if ((li->cull_mask & p_layer_mask) && (li->light_index < 255)) {
-				p_omni_light_indices[dword] += li->light_index << shift;
-				if (shift == 24) {
-					dword++;
-					shift = 0;
-				} else {
-					shift += 8;
-				}
-			}
-		}
-
-		if (dword < 2) {
-			// put in ending mark
-			p_omni_light_indices[dword] += 0xFF << shift;
-		}
-	}
-
-	{
-		// process spot lights
-		uint32_t dword = 0;
-		uint32_t shift = 0;
-
-		for (uint32_t i = 0; i < p_spot_light_instance_count && dword < p_max_dst_words; i++) {
-			LightInstance *li = light_instance_owner.getornull(p_spot_light_instances[i]);
-
-			if ((li->cull_mask & p_layer_mask) && (li->light_index < 255)) {
-				p_spot_light_indices[dword] += li->light_index << shift;
-				if (shift == 24) {
-					dword++;
-					shift = 0;
-				} else {
-					shift += 8;
-				}
-			}
-		}
-
-		if (dword < 2) {
-			// put in ending mark
-			p_spot_light_indices[dword] += 0xFF << shift;
-		}
-	}
-
-	{
-		// process reflection probes
-		uint32_t dword = 0;
-		uint32_t shift = 0;
-
-		for (uint32_t i = 0; i < p_reflection_probe_instance_count && dword < p_max_dst_words; i++) {
-			ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_reflection_probe_instances[i]);
-
-			if ((rpi->cull_mask & p_layer_mask) && (rpi->render_index < 255)) {
-				p_reflection_probe_indices[dword] += rpi->render_index << shift;
-				if (shift == 24) {
-					dword++;
-					shift = 0;
-				} else {
-					shift += 8;
-				}
-			}
-		}
-
-		if (dword < 2) {
-			// put in ending mark
-			p_reflection_probe_indices[dword] += 0xFF << shift;
-		}
-	}
-
-	{
-		// process decals
-		uint32_t dword = 0;
-		uint32_t shift = 0;
-
-		for (uint32_t i = 0; i < p_decal_instance_count && dword < p_max_dst_words; i++) {
-			DecalInstance *decal = decal_instance_owner.getornull(p_decal_instances[i]);
-
-			if ((decal->cull_mask & p_layer_mask) && (decal->render_index < 255)) {
-				p_decal_instance_indices[dword] += decal->render_index << shift;
-				if (shift == 24) {
-					dword++;
-					shift = 0;
-				} else {
-					shift += 8;
-				}
-			}
-		}
-
-		if (dword < 2) {
-			// put in ending mark
-			p_decal_instance_indices[dword] += 0xFF << shift;
-		}
-	}
-}
-
 void RendererSceneRenderRD::_volumetric_fog_erase(RenderBuffers *rb) {
 	ERR_FAIL_COND(!rb->volumetric_fog);
 
@@ -4042,11 +3954,13 @@ bool RendererSceneRenderRD::free(RID p_rid) {
 		}
 		reflection_atlas_owner.free(p_rid);
 	} else if (reflection_probe_instance_owner.owns(p_rid)) {
-		//not much to delete, just free it
-		//ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_rid);
+		ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_rid);
+		_free_forward_id(FORWARD_ID_TYPE_REFLECTION_PROBE, rpi->forward_id);
 		reflection_probe_release_atlas_index(p_rid);
 		reflection_probe_instance_owner.free(p_rid);
 	} else if (decal_instance_owner.owns(p_rid)) {
+		DecalInstance *di = decal_instance_owner.getornull(p_rid);
+		_free_forward_id(FORWARD_ID_TYPE_DECAL, di->forward_id);
 		decal_instance_owner.free(p_rid);
 	} else if (lightmap_instance_owner.owns(p_rid)) {
 		lightmap_instance_owner.free(p_rid);
@@ -4081,6 +3995,9 @@ bool RendererSceneRenderRD::free(RID p_rid) {
 			shadow_atlas->shadow_owners.erase(p_rid);
 		}
 
+		if (light_instance->light_type != RS::LIGHT_DIRECTIONAL) {
+			_free_forward_id(light_instance->light_type == RS::LIGHT_OMNI ? FORWARD_ID_TYPE_OMNI_LIGHT : FORWARD_ID_TYPE_SPOT_LIGHT, light_instance->forward_id);
+		}
 		light_instance_owner.free(p_rid);
 
 	} else if (shadow_atlas_owner.owns(p_rid)) {
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.h b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
index 58d9a765d0..cd2056d390 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.h
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
@@ -148,6 +148,23 @@ protected:
 		}
 	}
 
+	//used for mobile renderer mostly
+
+	typedef int32_t ForwardID;
+
+	enum ForwardIDType {
+		FORWARD_ID_TYPE_OMNI_LIGHT,
+		FORWARD_ID_TYPE_SPOT_LIGHT,
+		FORWARD_ID_TYPE_REFLECTION_PROBE,
+		FORWARD_ID_TYPE_DECAL,
+		FORWARD_ID_MAX,
+	};
+
+	virtual ForwardID _allocate_forward_id(ForwardIDType p_type) { return -1; }
+	virtual void _free_forward_id(ForwardIDType p_type, ForwardID p_id) {}
+	virtual void _map_forward_id(ForwardIDType p_type, ForwardID p_id, uint32_t p_index) {}
+	virtual bool _uses_forward_ids() const { return false; }
+
 private:
 	RS::ViewportDebugDraw debug_draw = RS::VIEWPORT_DEBUG_DRAW_DISABLED;
 	static RendererSceneRenderRD *singleton;
@@ -189,9 +206,10 @@ private:
 
 		uint32_t render_step = 0;
 		uint64_t last_pass = 0;
-		uint32_t render_index = 0;
 		uint32_t cull_mask = 0;
 
+		ForwardID forward_id = -1;
+
 		Transform3D transform;
 	};
 
@@ -202,8 +220,8 @@ private:
 	struct DecalInstance {
 		RID decal;
 		Transform3D transform;
-		uint32_t render_index;
 		uint32_t cull_mask;
+		ForwardID forward_id = -1;
 	};
 
 	mutable RID_Owner<DecalInstance> decal_instance_owner;
@@ -347,7 +365,6 @@ private:
 		uint64_t last_scene_pass = 0;
 		uint64_t last_scene_shadow_pass = 0;
 		uint64_t last_pass = 0;
-		uint32_t light_index = 0;
 		uint32_t cull_mask = 0;
 		uint32_t light_directional_index = 0;
 
@@ -359,6 +376,8 @@ private:
 
 		Set<RID> shadow_atlases; //shadow atlases where this light is registered
 
+		ForwardID forward_id = -1;
+
 		LightInstance() {}
 	};
 
@@ -1006,14 +1025,9 @@ public:
 		return li->last_pass;
 	}
 
-	_FORCE_INLINE_ void light_instance_set_index(RID p_light_instance, uint32_t p_index) {
-		LightInstance *li = light_instance_owner.getornull(p_light_instance);
-		li->light_index = p_index;
-	}
-
-	_FORCE_INLINE_ uint32_t light_instance_get_index(RID p_light_instance) {
+	_FORCE_INLINE_ ForwardID light_instance_get_forward_id(RID p_light_instance) {
 		LightInstance *li = light_instance_owner.getornull(p_light_instance);
-		return li->light_index;
+		return li->forward_id;
 	}
 
 	_FORCE_INLINE_ RS::LightType light_instance_get_type(RID p_light_instance) {
@@ -1050,17 +1064,11 @@ public:
 		return rpi->probe;
 	}
 
-	_FORCE_INLINE_ void reflection_probe_instance_set_render_index(RID p_instance, uint32_t p_render_index) {
-		ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_instance);
-		ERR_FAIL_COND(!rpi);
-		rpi->render_index = p_render_index;
-	}
-
-	_FORCE_INLINE_ uint32_t reflection_probe_instance_get_render_index(RID p_instance) {
+	_FORCE_INLINE_ ForwardID reflection_probe_instance_get_forward_id(RID p_instance) {
 		ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_instance);
 		ERR_FAIL_COND_V(!rpi, 0);
 
-		return rpi->render_index;
+		return rpi->forward_id;
 	}
 
 	_FORCE_INLINE_ void reflection_probe_instance_set_render_pass(RID p_instance, uint32_t p_render_pass) {
@@ -1098,6 +1106,11 @@ public:
 		return decal->decal;
 	}
 
+	_FORCE_INLINE_ ForwardID decal_instance_get_forward_id(RID p_decal) const {
+		DecalInstance *decal = decal_instance_owner.getornull(p_decal);
+		return decal->forward_id;
+	}
+
 	_FORCE_INLINE_ Transform3D decal_instance_get_transform(RID p_decal) const {
 		DecalInstance *decal = decal_instance_owner.getornull(p_decal);
 		return decal->transform;
@@ -1118,8 +1131,6 @@ public:
 		return li->transform;
 	}
 
-	void _fill_instance_indices(const RID *p_omni_light_instances, uint32_t p_omni_light_instance_count, uint32_t *p_omni_light_indices, const RID *p_spot_light_instances, uint32_t p_spot_light_instance_count, uint32_t *p_spot_light_indices, const RID *p_reflection_probe_instances, uint32_t p_reflection_probe_instance_count, uint32_t *p_reflection_probe_indices, const RID *p_decal_instances, uint32_t p_decal_instance_count, uint32_t *p_decal_instance_indices, uint32_t p_layer_mask, uint32_t p_max_dst_words = 2);
-
 	/* gi light probes */
 
 	virtual RID voxel_gi_instance_create(RID p_base) override;
@@ -1131,6 +1142,7 @@ public:
 	/* render buffers */
 
 	virtual RD::DataFormat _render_buffers_get_color_format();
+	virtual bool _render_buffers_can_be_storage();
 	virtual RID render_buffers_create() override;
 	virtual void render_buffers_configure(RID p_render_buffers, RID p_render_target, int p_width, int p_height, RS::ViewportMSAA p_msaa, RS::ViewportScreenSpaceAA p_screen_space_aa, bool p_use_debanding, uint32_t p_view_count) override;
 	virtual void gi_set_use_half_resolution(bool p_enable) override;
diff --git a/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl b/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
index 30673745ca..c28493d9e3 100644
--- a/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
+++ b/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
@@ -1291,7 +1291,7 @@ void main() {
 
 			blur_shadow(shadow);
 
-			light_compute(normal, directional_lights.data[i].direction, normalize(view), directional_lights.data[i].color * directional_lights.data[i].energy, shadow, f0, orms, 1.0,
+			light_compute(normal, directional_lights.data[i].direction, normalize(view), 0.0, directional_lights.data[i].color * directional_lights.data[i].energy, shadow, f0, orms, 1.0,
 #ifdef LIGHT_BACKLIGHT_USED
 					backlight,
 #endif