bevy/crates/bevy_pbr/src/render/light.rs

1752 lines
66 KiB
Rust
Raw Normal View History

use crate::{
directional_light_order, point_light_order, AlphaMode, AmbientLight, Cascade,
CascadeShadowConfig, Cascades, CascadesVisibleEntities, Clusters, CubemapVisibleEntities,
DirectionalLight, DirectionalLightShadowMap, DrawPrepass, EnvironmentMapLight,
GlobalVisiblePointLights, Material, MaterialPipelineKey, MeshPipeline, MeshPipelineKey,
Reorder render sets, refactor bevy_sprite to take advantage (#9236) This is a continuation of this PR: #8062 # Objective - Reorder render schedule sets to allow data preparation when phase item order is known to support improved batching - Part of the batching/instancing etc plan from here: https://github.com/bevyengine/bevy/issues/89#issuecomment-1379249074 - The original idea came from @inodentry and proved to be a good one. Thanks! - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the new ordering ## Solution - Move `Prepare` and `PrepareFlush` after `PhaseSortFlush` - Add a `PrepareAssets` set that runs in parallel with other systems and sets in the render schedule. - Put prepare_assets systems in the `PrepareAssets` set - If explicit dependencies are needed on Mesh or Material RenderAssets then depend on the appropriate system. - Add `ManageViews` and `ManageViewsFlush` sets between `ExtractCommands` and Queue - Move `queue_mesh*_bind_group` to the Prepare stage - Rename them to `prepare_` - Put systems that prepare resources (buffers, textures, etc.) into a `PrepareResources` set inside `Prepare` - Put the `prepare_..._bind_group` systems into a `PrepareBindGroup` set after `PrepareResources` - Move `prepare_lights` to the `ManageViews` set - `prepare_lights` creates views and this must happen before `Queue` - This system needs refactoring to stop handling all responsibilities - Gather lights, sort, and create shadow map views. Store sorted light entities in a resource - Remove `BatchedPhaseItem` - Replace `batch_range` with `batch_size` representing how many items to skip after rendering the item or to skip the item entirely if `batch_size` is 0. - `queue_sprites` has been split into `queue_sprites` for queueing phase items and `prepare_sprites` for batching after the `PhaseSort` - `PhaseItem`s are still inserted in `queue_sprites` - After sorting adjacent compatible sprite phase items are accumulated into `SpriteBatch` components on the first entity of each batch, containing a range of vertex indices. The associated `PhaseItem`'s `batch_size` is updated appropriately. - `SpriteBatch` items are then drawn skipping over the other items in the batch based on the value in `batch_size` - A very similar refactor was performed on `bevy_ui` --- ## Changelog Changed: - Reordered and reworked render app schedule sets. The main change is that data is extracted, queued, sorted, and then prepared when the order of data is known. - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the reordering. ## Migration Guide - Assets such as materials and meshes should now be created in `PrepareAssets` e.g. `prepare_assets<Mesh>` - Queueing entities to `RenderPhase`s continues to be done in `Queue` e.g. `queue_sprites` - Preparing resources (textures, buffers, etc.) should now be done in `PrepareResources`, e.g. `prepare_prepass_textures`, `prepare_mesh_uniforms` - Prepare bind groups should now be done in `PrepareBindGroups` e.g. `prepare_mesh_bind_group` - Any batching or instancing can now be done in `Prepare` where the order of the phase items is known e.g. `prepare_sprites` ## Next Steps - Introduce some generic mechanism to ensure items that can be batched are grouped in the phase item order, currently you could easily have `[sprite at z 0, mesh at z 0, sprite at z 0]` preventing batching. - Investigate improved orderings for building the MeshUniform buffer - Implementing batching across the rest of bevy --------- Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
2023-08-27 14:33:49 +00:00
NotShadowCaster, PointLight, PointLightShadowMap, PrepassPipeline, RenderMaterials, SpotLight,
VisiblePointLights,
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
};
use bevy_asset::Handle;
Camera Driven Rendering (#4745) This adds "high level camera driven rendering" to Bevy. The goal is to give users more control over what gets rendered (and where) without needing to deal with render logic. This will make scenarios like "render to texture", "multiple windows", "split screen", "2d on 3d", "3d on 2d", "pass layering", and more significantly easier. Here is an [example of a 2d render sandwiched between two 3d renders (each from a different perspective)](https://gist.github.com/cart/4fe56874b2e53bc5594a182fc76f4915): ![image](https://user-images.githubusercontent.com/2694663/168411086-af13dec8-0093-4a84-bdd4-d4362d850ffa.png) Users can now spawn a camera, point it at a RenderTarget (a texture or a window), and it will "just work". Rendering to a second window is as simple as spawning a second camera and assigning it to a specific window id: ```rust // main camera (main window) commands.spawn_bundle(Camera2dBundle::default()); // second camera (other window) commands.spawn_bundle(Camera2dBundle { camera: Camera { target: RenderTarget::Window(window_id), ..default() }, ..default() }); ``` Rendering to a texture is as simple as pointing the camera at a texture: ```rust commands.spawn_bundle(Camera2dBundle { camera: Camera { target: RenderTarget::Texture(image_handle), ..default() }, ..default() }); ``` Cameras now have a "render priority", which controls the order they are drawn in. If you want to use a camera's output texture as a texture in the main pass, just set the priority to a number lower than the main pass camera (which defaults to `0`). ```rust // main pass camera with a default priority of 0 commands.spawn_bundle(Camera2dBundle::default()); commands.spawn_bundle(Camera2dBundle { camera: Camera { target: RenderTarget::Texture(image_handle.clone()), priority: -1, ..default() }, ..default() }); commands.spawn_bundle(SpriteBundle { texture: image_handle, ..default() }) ``` Priority can also be used to layer to cameras on top of each other for the same RenderTarget. This is what "2d on top of 3d" looks like in the new system: ```rust commands.spawn_bundle(Camera3dBundle::default()); commands.spawn_bundle(Camera2dBundle { camera: Camera { // this will render 2d entities "on top" of the default 3d camera's render priority: 1, ..default() }, ..default() }); ``` There is no longer the concept of a global "active camera". Resources like `ActiveCamera<Camera2d>` and `ActiveCamera<Camera3d>` have been replaced with the camera-specific `Camera::is_active` field. This does put the onus on users to manage which cameras should be active. Cameras are now assigned a single render graph as an "entry point", which is configured on each camera entity using the new `CameraRenderGraph` component. The old `PerspectiveCameraBundle` and `OrthographicCameraBundle` (generic on camera marker components like Camera2d and Camera3d) have been replaced by `Camera3dBundle` and `Camera2dBundle`, which set 3d and 2d default values for the `CameraRenderGraph` and projections. ```rust // old 3d perspective camera commands.spawn_bundle(PerspectiveCameraBundle::default()) // new 3d perspective camera commands.spawn_bundle(Camera3dBundle::default()) ``` ```rust // old 2d orthographic camera commands.spawn_bundle(OrthographicCameraBundle::new_2d()) // new 2d orthographic camera commands.spawn_bundle(Camera2dBundle::default()) ``` ```rust // old 3d orthographic camera commands.spawn_bundle(OrthographicCameraBundle::new_3d()) // new 3d orthographic camera commands.spawn_bundle(Camera3dBundle { projection: OrthographicProjection { scale: 3.0, scaling_mode: ScalingMode::FixedVertical, ..default() }.into(), ..default() }) ``` Note that `Camera3dBundle` now uses a new `Projection` enum instead of hard coding the projection into the type. There are a number of motivators for this change: the render graph is now a part of the bundle, the way "generic bundles" work in the rust type system prevents nice `..default()` syntax, and changing projections at runtime is much easier with an enum (ex for editor scenarios). I'm open to discussing this choice, but I'm relatively certain we will all come to the same conclusion here. Camera2dBundle and Camera3dBundle are much clearer than being generic on marker components / using non-default constructors. If you want to run a custom render graph on a camera, just set the `CameraRenderGraph` component: ```rust commands.spawn_bundle(Camera3dBundle { camera_render_graph: CameraRenderGraph::new(some_render_graph_name), ..default() }) ``` Just note that if the graph requires data from specific components to work (such as `Camera3d` config, which is provided in the `Camera3dBundle`), make sure the relevant components have been added. Speaking of using components to configure graphs / passes, there are a number of new configuration options: ```rust commands.spawn_bundle(Camera3dBundle { camera_3d: Camera3d { // overrides the default global clear color clear_color: ClearColorConfig::Custom(Color::RED), ..default() }, ..default() }) commands.spawn_bundle(Camera3dBundle { camera_3d: Camera3d { // disables clearing clear_color: ClearColorConfig::None, ..default() }, ..default() }) ``` Expect to see more of the "graph configuration Components on Cameras" pattern in the future. By popular demand, UI no longer requires a dedicated camera. `UiCameraBundle` has been removed. `Camera2dBundle` and `Camera3dBundle` now both default to rendering UI as part of their own render graphs. To disable UI rendering for a camera, disable it using the CameraUi component: ```rust commands .spawn_bundle(Camera3dBundle::default()) .insert(CameraUi { is_enabled: false, ..default() }) ``` ## Other Changes * The separate clear pass has been removed. We should revisit this for things like sky rendering, but I think this PR should "keep it simple" until we're ready to properly support that (for code complexity and performance reasons). We can come up with the right design for a modular clear pass in a followup pr. * I reorganized bevy_core_pipeline into Core2dPlugin and Core3dPlugin (and core_2d / core_3d modules). Everything is pretty much the same as before, just logically separate. I've moved relevant types (like Camera2d, Camera3d, Camera3dBundle, Camera2dBundle) into their relevant modules, which is what motivated this reorganization. * I adapted the `scene_viewer` example (which relied on the ActiveCameras behavior) to the new system. I also refactored bits and pieces to be a bit simpler. * All of the examples have been ported to the new camera approach. `render_to_texture` and `multiple_windows` are now _much_ simpler. I removed `two_passes` because it is less relevant with the new approach. If someone wants to add a new "layered custom pass with CameraRenderGraph" example, that might fill a similar niche. But I don't feel much pressure to add that in this pr. * Cameras now have `target_logical_size` and `target_physical_size` fields, which makes finding the size of a camera's render target _much_ simpler. As a result, the `Assets<Image>` and `Windows` parameters were removed from `Camera::world_to_screen`, making that operation much more ergonomic. * Render order ambiguities between cameras with the same target and the same priority now produce a warning. This accomplishes two goals: 1. Now that there is no "global" active camera, by default spawning two cameras will result in two renders (one covering the other). This would be a silent performance killer that would be hard to detect after the fact. By detecting ambiguities, we can provide a helpful warning when this occurs. 2. Render order ambiguities could result in unexpected / unpredictable render results. Resolving them makes sense. ## Follow Up Work * Per-Camera viewports, which will make it possible to render to a smaller area inside of a RenderTarget (great for something like splitscreen) * Camera-specific MSAA config (should use the same "overriding" pattern used for ClearColor) * Graph Based Camera Ordering: priorities are simple, but they make complicated ordering constraints harder to express. We should consider adopting a "graph based" camera ordering model with "before" and "after" relationships to other cameras (or build it "on top" of the priority system). * Consider allowing graphs to run subgraphs from any nest level (aka a global namespace for graphs). Right now the 2d and 3d graphs each need their own UI subgraph, which feels "fine" in the short term. But being able to share subgraphs between other subgraphs seems valuable. * Consider splitting `bevy_core_pipeline` into `bevy_core_2d` and `bevy_core_3d` packages. Theres a shared "clear color" dependency here, which would need a new home.
2022-06-02 00:12:17 +00:00
use bevy_core_pipeline::core_3d::Transparent3d;
use bevy_ecs::prelude::*;
use bevy_math::{Mat4, UVec3, UVec4, Vec2, Vec3, Vec3Swizzles, Vec4, Vec4Swizzles};
use bevy_render::{
camera::Camera,
2021-06-02 02:59:17 +00:00
color::Color,
mesh::Mesh,
render_asset::RenderAssets,
Make render graph slots optional for most cases (#8109) # Objective - Currently, the render graph slots are only used to pass the view_entity around. This introduces significant boilerplate for very little value. Instead of using slots for this, make the view_entity part of the `RenderGraphContext`. This also means we won't need to have `IN_VIEW` on every node and and we'll be able to use the default impl of `Node::input()`. ## Solution - Add `view_entity: Option<Entity>` to the `RenderGraphContext` - Update all nodes to use this instead of entity slot input --- ## Changelog - Add optional `view_entity` to `RenderGraphContext` ## Migration Guide You can now get the view_entity directly from the `RenderGraphContext`. When implementing the Node: ```rust // 0.10 struct FooNode; impl FooNode { const IN_VIEW: &'static str = "view"; } impl Node for FooNode { fn input(&self) -> Vec<SlotInfo> { vec![SlotInfo::new(Self::IN_VIEW, SlotType::Entity)] } fn run( &self, graph: &mut RenderGraphContext, // ... ) -> Result<(), NodeRunError> { let view_entity = graph.get_input_entity(Self::IN_VIEW)?; // ... Ok(()) } } // 0.11 struct FooNode; impl Node for FooNode { fn run( &self, graph: &mut RenderGraphContext, // ... ) -> Result<(), NodeRunError> { let view_entity = graph.view_entity(); // ... Ok(()) } } ``` When adding the node to the graph, you don't need to specify a slot_edge for the view_entity. ```rust // 0.10 let mut graph = RenderGraph::default(); graph.add_node(FooNode::NAME, node); let input_node_id = draw_2d_graph.set_input(vec![SlotInfo::new( graph::input::VIEW_ENTITY, SlotType::Entity, )]); graph.add_slot_edge( input_node_id, graph::input::VIEW_ENTITY, FooNode::NAME, FooNode::IN_VIEW, ); // add_node_edge ... // 0.11 let mut graph = RenderGraph::default(); graph.add_node(FooNode::NAME, node); // add_node_edge ... ``` ## Notes This PR paired with #8007 will help reduce a lot of annoying boilerplate with the render nodes. Depending on which one gets merged first. It will require a bit of clean up work to make both compatible. I tagged this as a breaking change, because using the old system to get the view_entity will break things because it's not a node input slot anymore. ## Notes for reviewers A lot of the diffs are just removing the slots in every nodes and graph creation. The important part is mostly in the graph_runner/CameraDriverNode.
2023-03-21 20:11:13 +00:00
render_graph::{Node, NodeRunError, RenderGraphContext},
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
render_phase::{
CachedRenderPipelinePhaseItem, DrawFunctionId, DrawFunctions, PhaseItem, RenderPhase,
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
},
render_resource::*,
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
renderer::{RenderContext, RenderDevice, RenderQueue},
2021-06-02 02:59:17 +00:00
texture::*,
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
view::{ExtractedView, ViewVisibility, VisibleEntities},
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
Extract,
2021-06-02 02:59:17 +00:00
};
use bevy_transform::{components::GlobalTransform, prelude::Transform};
Mesh vertex buffer layouts (#3959) This PR makes a number of changes to how meshes and vertex attributes are handled, which the goal of enabling easy and flexible custom vertex attributes: * Reworks the `Mesh` type to use the newly added `VertexAttribute` internally * `VertexAttribute` defines the name, a unique `VertexAttributeId`, and a `VertexFormat` * `VertexAttributeId` is used to produce consistent sort orders for vertex buffer generation, replacing the more expensive and often surprising "name based sorting" * Meshes can be used to generate a `MeshVertexBufferLayout`, which defines the layout of the gpu buffer produced by the mesh. `MeshVertexBufferLayouts` can then be used to generate actual `VertexBufferLayouts` according to the requirements of a specific pipeline. This decoupling of "mesh layout" vs "pipeline vertex buffer layout" is what enables custom attributes. We don't need to standardize _mesh layouts_ or contort meshes to meet the needs of a specific pipeline. As long as the mesh has what the pipeline needs, it will work transparently. * Mesh-based pipelines now specialize on `&MeshVertexBufferLayout` via the new `SpecializedMeshPipeline` trait (which behaves like `SpecializedPipeline`, but adds `&MeshVertexBufferLayout`). The integrity of the pipeline cache is maintained because the `MeshVertexBufferLayout` is treated as part of the key (which is fully abstracted from implementers of the trait ... no need to add any additional info to the specialization key). * Hashing `MeshVertexBufferLayout` is too expensive to do for every entity, every frame. To make this scalable, I added a generalized "pre-hashing" solution to `bevy_utils`: `Hashed<T>` keys and `PreHashMap<K, V>` (which uses `Hashed<T>` internally) . Why didn't I just do the quick and dirty in-place "pre-compute hash and use that u64 as a key in a hashmap" that we've done in the past? Because its wrong! Hashes by themselves aren't enough because two different values can produce the same hash. Re-hashing a hash is even worse! I decided to build a generalized solution because this pattern has come up in the past and we've chosen to do the wrong thing. Now we can do the right thing! This did unfortunately require pulling in `hashbrown` and using that in `bevy_utils`, because avoiding re-hashes requires the `raw_entry_mut` api, which isn't stabilized yet (and may never be ... `entry_ref` has favor now, but also isn't available yet). If std's HashMap ever provides the tools we need, we can move back to that. Note that adding `hashbrown` doesn't increase our dependency count because it was already in our tree. I will probably break these changes out into their own PR. * Specializing on `MeshVertexBufferLayout` has one non-obvious behavior: it can produce identical pipelines for two different MeshVertexBufferLayouts. To optimize the number of active pipelines / reduce re-binds while drawing, I de-duplicate pipelines post-specialization using the final `VertexBufferLayout` as the key. For example, consider a pipeline that needs the layout `(position, normal)` and is specialized using two meshes: `(position, normal, uv)` and `(position, normal, other_vec2)`. If both of these meshes result in `(position, normal)` specializations, we can use the same pipeline! Now we do. Cool! To briefly illustrate, this is what the relevant section of `MeshPipeline`'s specialization code looks like now: ```rust impl SpecializedMeshPipeline for MeshPipeline { type Key = MeshPipelineKey; fn specialize( &self, key: Self::Key, layout: &MeshVertexBufferLayout, ) -> RenderPipelineDescriptor { let mut vertex_attributes = vec![ Mesh::ATTRIBUTE_POSITION.at_shader_location(0), Mesh::ATTRIBUTE_NORMAL.at_shader_location(1), Mesh::ATTRIBUTE_UV_0.at_shader_location(2), ]; let mut shader_defs = Vec::new(); if layout.contains(Mesh::ATTRIBUTE_TANGENT) { shader_defs.push(String::from("VERTEX_TANGENTS")); vertex_attributes.push(Mesh::ATTRIBUTE_TANGENT.at_shader_location(3)); } let vertex_buffer_layout = layout .get_layout(&vertex_attributes) .expect("Mesh is missing a vertex attribute"); ``` Notice that this is _much_ simpler than it was before. And now any mesh with any layout can be used with this pipeline, provided it has vertex postions, normals, and uvs. We even got to remove `HAS_TANGENTS` from MeshPipelineKey and `has_tangents` from `GpuMesh`, because that information is redundant with `MeshVertexBufferLayout`. This is still a draft because I still need to: * Add more docs * Experiment with adding error handling to mesh pipeline specialization (which would print errors at runtime when a mesh is missing a vertex attribute required by a pipeline). If it doesn't tank perf, we'll keep it. * Consider breaking out the PreHash / hashbrown changes into a separate PR. * Add an example illustrating this change * Verify that the "mesh-specialized pipeline de-duplication code" works properly Please dont yell at me for not doing these things yet :) Just trying to get this in peoples' hands asap. Alternative to #3120 Fixes #3030 Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-02-23 23:21:13 +00:00
use bevy_utils::{
tracing::{error, warn},
HashMap,
};
use std::{hash::Hash, num::NonZeroU64};
2021-06-02 02:59:17 +00:00
2021-11-22 23:16:36 +00:00
#[derive(Component)]
2021-06-02 02:59:17 +00:00
pub struct ExtractedPointLight {
color: Color,
bevy_pbr2: Improve lighting units and documentation (#2704) # Objective A question was raised on Discord about the units of the `PointLight` `intensity` member. After digging around in the bevy_pbr2 source code and [Google Filament documentation](https://google.github.io/filament/Filament.html#mjx-eqn-pointLightLuminousPower) I discovered that the intention by Filament was that the 'intensity' value for point lights would be in lumens. This makes a lot of sense as these are quite relatable units given basically all light bulbs I've seen sold over the past years are rated in lumens as people move away from thinking about how bright a bulb is relative to a non-halogen incandescent bulb. However, it seems that the derivation of the conversion between luminous power (lumens, denoted `Φ` in the Filament formulae) and luminous intensity (lumens per steradian, `I` in the Filament formulae) was missed and I can see why as it is tucked right under equation 58 at the link above. As such, while the formula states that for a point light, `I = Φ / 4 π` we have been using `intensity` as if it were luminous intensity `I`. Before this PR, the intensity field is luminous intensity in lumens per steradian. After this PR, the intensity field is luminous power in lumens, [as suggested by Filament](https://google.github.io/filament/Filament.html#table_lighttypesunits) (unfortunately the link jumps to the table's caption so scroll up to see the actual table). I appreciate that it may be confusing to call this an intensity, but I think this is intended as more of a non-scientific, human-relatable general term with a bit of hand waving so that most light types can just have an intensity field and for most of them it works in the same way or at least with some relatable value. I'm inclined to think this is reasonable rather than throwing terms like luminous power, luminous intensity, blah at users. ## Solution - Documented the `PointLight` `intensity` member as 'luminous power' in units of lumens. - Added a table of examples relating from various types of household lighting to lumen values. - Added in the mapping from luminous power to luminous intensity when premultiplying the intensity into the colour before it is made into a graphics uniform. - Updated the documentation in `pbr.wgsl` to clarify the earlier confusion about the missing `/ 4 π`. - Bumped the intensity of the point lights in `3d_scene_pipelined` to 1600 lumens. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-08-23 23:48:11 +00:00
/// luminous intensity in lumens per steradian
2021-06-02 02:59:17 +00:00
intensity: f32,
range: f32,
radius: f32,
transform: GlobalTransform,
shadows_enabled: bool,
shadow_depth_bias: f32,
shadow_normal_bias: f32,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
spot_light_angles: Option<(f32, f32)>,
}
#[derive(Component, Debug)]
pub struct ExtractedDirectionalLight {
color: Color,
illuminance: f32,
Take DirectionalLight's GlobalTransform into account when calculating shadow map volume (not just direction) (#6384) # Objective This PR fixes #5789, by enabling movable (and scalable) directional light shadow volumes. ## Solution This PR changes `ExtractedDirectionalLight` to hold a copy of the `DirectionalLight` entity's `GlobalTransform`, instead of just a `direction` vector. This allows the shadow map volume (as defined by the light's `shadow_projection` field) to be transformed honoring translation _and_ scale transforms, and not just rotation. It also augments the texel size calculation (used to determine the `shadow_normal_bias`) so that it now takes into account the upper bound of the x/y/z scale of the `GlobalTransform`. This change makes the directional light extraction code more consistent with point and spot lights (that already use `transform`), and allows easily moving and scaling the shadow volume along with a player entity based on camera distance/angle, immediately enabling more real world use cases until we have a more sophisticated adaptive implementation, such as the one described in #3629. **Note:** While it was previously possible to update the projection achieving a similar effect, depending on the light direction and distance to the origin, the fact that the shadow map camera was always positioned at the origin with a hardcoded `Vec3::Y` up value meant you would get sub-optimal or inconsistent/incorrect results. --- ## Changelog ### Changed - `DirectionalLight` shadow volumes now honor translation and scale transforms ## Migration Guide - If your directional lights were positioned at the origin and not scaled (the default, most common scenario) no changes are needed on your part; it just works as before; - If you previously had a system for dynamically updating directional light shadow projections, you might now be able to simplify your code by updating the directional light entity's transform instead; - In the unlikely scenario that a scene with directional lights that previously rendered shadows correctly has missing shadows, make sure your directional lights are positioned at (0, 0, 0) and are not scaled to a size that's too large or too small.
2022-11-04 20:12:26 +00:00
transform: GlobalTransform,
shadows_enabled: bool,
shadow_depth_bias: f32,
shadow_normal_bias: f32,
cascade_shadow_config: CascadeShadowConfig,
cascades: HashMap<Entity, Vec<Cascade>>,
2021-06-02 02:59:17 +00:00
}
#[derive(Copy, Clone, ShaderType, Default, Debug)]
pub struct GpuPointLight {
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// For point lights: the lower-right 2x2 values of the projection matrix [2][2] [2][3] [3][2] [3][3]
// For spot lights: 2 components of the direction (x,z), spot_scale and spot_offset
light_custom_data: Vec4,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
color_inverse_square_range: Vec4,
position_radius: Vec4,
flags: u32,
shadow_depth_bias: f32,
shadow_normal_bias: f32,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
spot_light_tan_angle: f32,
}
#[derive(ShaderType)]
pub struct GpuPointLightsUniform {
data: Box<[GpuPointLight; MAX_UNIFORM_BUFFER_POINT_LIGHTS]>,
}
impl Default for GpuPointLightsUniform {
fn default() -> Self {
Self {
data: Box::new([GpuPointLight::default(); MAX_UNIFORM_BUFFER_POINT_LIGHTS]),
}
}
}
#[derive(ShaderType, Default)]
pub struct GpuPointLightsStorage {
#[size(runtime)]
data: Vec<GpuPointLight>,
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
pub enum GpuPointLights {
Uniform(UniformBuffer<GpuPointLightsUniform>),
Storage(StorageBuffer<GpuPointLightsStorage>),
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
impl GpuPointLights {
fn new(buffer_binding_type: BufferBindingType) -> Self {
match buffer_binding_type {
BufferBindingType::Storage { .. } => Self::storage(),
BufferBindingType::Uniform => Self::uniform(),
}
}
fn uniform() -> Self {
Self::Uniform(UniformBuffer::default())
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
fn storage() -> Self {
Self::Storage(StorageBuffer::default())
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
fn set(&mut self, mut lights: Vec<GpuPointLight>) {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
match self {
GpuPointLights::Uniform(buffer) => {
let len = lights.len().min(MAX_UNIFORM_BUFFER_POINT_LIGHTS);
let src = &lights[..len];
let dst = &mut buffer.get_mut().data[..len];
dst.copy_from_slice(src);
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
GpuPointLights::Storage(buffer) => {
buffer.get_mut().data.clear();
buffer.get_mut().data.append(&mut lights);
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
}
fn write_buffer(&mut self, render_device: &RenderDevice, render_queue: &RenderQueue) {
match self {
GpuPointLights::Uniform(buffer) => buffer.write_buffer(render_device, render_queue),
GpuPointLights::Storage(buffer) => buffer.write_buffer(render_device, render_queue),
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
pub fn binding(&self) -> Option<BindingResource> {
match self {
GpuPointLights::Uniform(buffer) => buffer.binding(),
GpuPointLights::Storage(buffer) => buffer.binding(),
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
pub fn min_size(buffer_binding_type: BufferBindingType) -> NonZeroU64 {
match buffer_binding_type {
BufferBindingType::Storage { .. } => GpuPointLightsStorage::min_size(),
BufferBindingType::Uniform => GpuPointLightsUniform::min_size(),
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
// NOTE: These must match the bit flags in bevy_pbr/src/render/mesh_view_types.wgsl!
bitflags::bitflags! {
#[repr(transparent)]
struct PointLightFlags: u32 {
const SHADOWS_ENABLED = (1 << 0);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
const SPOT_LIGHT_Y_NEGATIVE = (1 << 1);
const NONE = 0;
const UNINITIALIZED = 0xFFFF;
}
}
#[derive(Copy, Clone, ShaderType, Default, Debug)]
pub struct GpuDirectionalCascade {
view_projection: Mat4,
texel_size: f32,
far_bound: f32,
}
#[derive(Copy, Clone, ShaderType, Default, Debug)]
pub struct GpuDirectionalLight {
cascades: [GpuDirectionalCascade; MAX_CASCADES_PER_LIGHT],
color: Vec4,
dir_to_light: Vec3,
flags: u32,
shadow_depth_bias: f32,
shadow_normal_bias: f32,
num_cascades: u32,
cascades_overlap_proportion: f32,
depth_texture_base_index: u32,
2021-06-02 02:59:17 +00:00
}
// NOTE: These must match the bit flags in bevy_pbr/src/render/mesh_view_types.wgsl!
bitflags::bitflags! {
#[repr(transparent)]
struct DirectionalLightFlags: u32 {
const SHADOWS_ENABLED = (1 << 0);
const NONE = 0;
const UNINITIALIZED = 0xFFFF;
}
}
#[derive(Copy, Clone, Debug, ShaderType)]
2021-06-02 02:59:17 +00:00
pub struct GpuLights {
directional_lights: [GpuDirectionalLight; MAX_DIRECTIONAL_LIGHTS],
bevy_pbr2: Add support for most of the StandardMaterial textures (#4) * bevy_pbr2: Add support for most of the StandardMaterial textures Normal maps are not included here as they require tangents in a vertex attribute. * bevy_pbr2: Ensure RenderCommandQueue is ready for PbrShaders init * texture_pipelined: Add a light to the scene so we can see stuff * WIP bevy_pbr2: back to front sorting hack * bevy_pbr2: Uniform control flow for texture sampling in pbr.frag From 'fintelia' on the Bevy Render Rework Round 2 discussion: "My understanding is that GPUs these days never use the "execute both branches and select the result" strategy. Rather, what they do is evaluate the branch condition on all threads of a warp, and jump over it if all of them evaluate to false. If even a single thread needs to execute the if statement body, however, then the remaining threads are paused until that is completed." * bevy_pbr2: Simplify texture and sampler names The StandardMaterial_ prefix is no longer needed * bevy_pbr2: Match default 'AmbientColor' of current bevy_pbr for now * bevy_pbr2: Convert from non-linear to linear sRGB for the color uniform * bevy_pbr2: Add pbr_pipelined example * Fix view vector in pbr frag to work in ortho * bevy_pbr2: Use a 90 degree y fov and light range projection for lights * bevy_pbr2: Add AmbientLight resource * bevy_pbr2: Convert PointLight color to linear sRGB for use in fragment shader * bevy_pbr2: pbr.frag: Rename PointLight.projection to view_projection The uniform contains the view_projection matrix so this was incorrect. * bevy_pbr2: PointLight is an OmniLight as it has a radius * bevy_pbr2: Factoring out duplicated code * bevy_pbr2: Implement RenderAsset for StandardMaterial * Remove unnecessary texture and sampler clones * fix comment formatting * remove redundant Buffer:from * Don't extract meshes when their material textures aren't ready * make missing textures in the queue step an error Co-authored-by: Aevyrie <aevyrie@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2021-06-27 23:10:23 +00:00
ambient_color: Vec4,
// xyz are x/y/z cluster dimensions and w is the number of clusters
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
cluster_dimensions: UVec4,
// xy are vec2<f32>(cluster_dimensions.xy) / vec2<f32>(view.width, view.height)
// z is cluster_dimensions.z / log(far / near)
// w is cluster_dimensions.z * log(near) / log(far / near)
cluster_factors: Vec4,
n_directional_lights: u32,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// offset from spot light's light index to spot light's shadow map index
spot_light_shadowmap_offset: i32,
environment_map_smallest_specular_mip_level: u32,
2021-06-02 02:59:17 +00:00
}
// NOTE: this must be kept in sync with the same constants in pbr.frag
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
pub const MAX_UNIFORM_BUFFER_POINT_LIGHTS: usize = 256;
//NOTE: When running bevy on Adreno GPU chipsets in WebGL, any value above 1 will result in a crash
// when loading the wgsl "pbr_functions.wgsl" in the function apply_fog.
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
pub const MAX_DIRECTIONAL_LIGHTS: usize = 1;
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
pub const MAX_DIRECTIONAL_LIGHTS: usize = 10;
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
pub const MAX_CASCADES_PER_LIGHT: usize = 4;
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
pub const MAX_CASCADES_PER_LIGHT: usize = 1;
2021-06-02 02:59:17 +00:00
pub const SHADOW_FORMAT: TextureFormat = TextureFormat::Depth32Float;
#[derive(Resource, Clone)]
pub struct ShadowSamplers {
pub point_light_sampler: Sampler,
pub directional_light_sampler: Sampler,
2021-06-02 02:59:17 +00:00
}
// TODO: this pattern for initializing the shaders / pipeline isn't ideal. this should be handled by the asset system
impl FromWorld for ShadowSamplers {
2021-06-02 02:59:17 +00:00
fn from_world(world: &mut World) -> Self {
let render_device = world.resource::<RenderDevice>();
2021-06-21 23:28:52 +00:00
ShadowSamplers {
point_light_sampler: render_device.create_sampler(&SamplerDescriptor {
address_mode_u: AddressMode::ClampToEdge,
address_mode_v: AddressMode::ClampToEdge,
address_mode_w: AddressMode::ClampToEdge,
mag_filter: FilterMode::Linear,
min_filter: FilterMode::Linear,
mipmap_filter: FilterMode::Nearest,
compare: Some(CompareFunction::GreaterEqual),
..Default::default()
}),
directional_light_sampler: render_device.create_sampler(&SamplerDescriptor {
address_mode_u: AddressMode::ClampToEdge,
address_mode_v: AddressMode::ClampToEdge,
address_mode_w: AddressMode::ClampToEdge,
mag_filter: FilterMode::Linear,
min_filter: FilterMode::Linear,
mipmap_filter: FilterMode::Nearest,
compare: Some(CompareFunction::GreaterEqual),
..Default::default()
}),
}
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
#[derive(Component)]
pub struct ExtractedClusterConfig {
/// Special near value for cluster calculations
near: f32,
Dynamic light clusters (#3968) # Objective provide some customisation for default cluster setup avoid "cluster index lists is full" in all cases (using a strategy outlined by @superdump) ## Solution Add ClusterConfig enum (which can be inserted into a view at any time) to allow specifying cluster setup with variants: - None (do not do any light assignment - for views which do not require light info, e.g. minimaps etc) - Single (one cluster) - XYZ (explicit cluster counts in each dimension) - FixedZ (most similar to current - specify Z-slices and total, then x and y counts are dynamically determined to give approximately square clusters based on current aspect ratio) Defaults to FixedZ { total: 4096, z: 24 } which is similar to the current setup. Per frame, estimate the number of indices that would be required for the current config and decrease the cluster counts / increase the cluster sizes in the x and y dimensions if the index list would be too small. notes: - I didn't put ClusterConfig in the camera bundles to avoid introducing a dependency from bevy_render to bevy_pbr. the ClusterConfig enum comes with a pbr-centric impl block so i didn't want to move that into bevy_render either. - ~Might want to add None variant to cluster config for views that don't care about lights?~ - Not well tested for orthographic - ~there's a cluster_muck branch on my repo which includes some diagnostics / a modified lighting example which may be useful for tyre-kicking~ (outdated, i will bring it up to date if required) anecdotal timings: FPS on the lighting demo is negligibly better (~5%), maybe due to a small optimisation constraining the light aabb to be in front of the camera FPS on the lighting demo with 100 extra lights added is ~33% faster, and also renders correctly as the cluster index count is no longer exceeded
2022-03-08 04:56:42 +00:00
far: f32,
/// Number of clusters in `X` / `Y` / `Z` in the view frustum
dimensions: UVec3,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
#[derive(Component)]
pub struct ExtractedClustersPointLights {
data: Vec<VisiblePointLights>,
}
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
pub fn extract_clusters(
mut commands: Commands,
views: Extract<Query<(Entity, &Clusters), With<Camera>>>,
) {
for (entity, clusters) in &views {
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
commands.get_or_spawn(entity).insert((
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
ExtractedClustersPointLights {
data: clusters.lights.clone(),
},
ExtractedClusterConfig {
near: clusters.near,
Dynamic light clusters (#3968) # Objective provide some customisation for default cluster setup avoid "cluster index lists is full" in all cases (using a strategy outlined by @superdump) ## Solution Add ClusterConfig enum (which can be inserted into a view at any time) to allow specifying cluster setup with variants: - None (do not do any light assignment - for views which do not require light info, e.g. minimaps etc) - Single (one cluster) - XYZ (explicit cluster counts in each dimension) - FixedZ (most similar to current - specify Z-slices and total, then x and y counts are dynamically determined to give approximately square clusters based on current aspect ratio) Defaults to FixedZ { total: 4096, z: 24 } which is similar to the current setup. Per frame, estimate the number of indices that would be required for the current config and decrease the cluster counts / increase the cluster sizes in the x and y dimensions if the index list would be too small. notes: - I didn't put ClusterConfig in the camera bundles to avoid introducing a dependency from bevy_render to bevy_pbr. the ClusterConfig enum comes with a pbr-centric impl block so i didn't want to move that into bevy_render either. - ~Might want to add None variant to cluster config for views that don't care about lights?~ - Not well tested for orthographic - ~there's a cluster_muck branch on my repo which includes some diagnostics / a modified lighting example which may be useful for tyre-kicking~ (outdated, i will bring it up to date if required) anecdotal timings: FPS on the lighting demo is negligibly better (~5%), maybe due to a small optimisation constraining the light aabb to be in front of the camera FPS on the lighting demo with 100 extra lights added is ~33% faster, and also renders correctly as the cluster index count is no longer exceeded
2022-03-08 04:56:42 +00:00
far: clusters.far,
dimensions: clusters.dimensions,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
},
));
}
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
#[allow(clippy::too_many_arguments)]
2021-06-02 02:59:17 +00:00
pub fn extract_lights(
mut commands: Commands,
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
point_light_shadow_map: Extract<Res<PointLightShadowMap>>,
directional_light_shadow_map: Extract<Res<DirectionalLightShadowMap>>,
global_point_lights: Extract<Res<GlobalVisiblePointLights>>,
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
point_lights: Extract<
Query<(
&PointLight,
&CubemapVisibleEntities,
&GlobalTransform,
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
&ViewVisibility,
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
)>,
>,
spot_lights: Extract<
Query<(
&SpotLight,
&VisibleEntities,
&GlobalTransform,
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
&ViewVisibility,
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
)>,
>,
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
directional_lights: Extract<
Query<
(
Entity,
&DirectionalLight,
&CascadesVisibleEntities,
&Cascades,
&CascadeShadowConfig,
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
&GlobalTransform,
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
&ViewVisibility,
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
),
Without<SpotLight>,
>,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
>,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
mut previous_point_lights_len: Local<usize>,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
mut previous_spot_lights_len: Local<usize>,
2021-06-02 02:59:17 +00:00
) {
// NOTE: These shadow map resources are extracted here as they are used here too so this avoids
// races between scheduling of ExtractResourceSystems and this system.
if point_light_shadow_map.is_changed() {
commands.insert_resource(point_light_shadow_map.clone());
}
if directional_light_shadow_map.is_changed() {
commands.insert_resource(directional_light_shadow_map.clone());
}
Scale normal bias by texel size (#26) * 3d_scene_pipelined: Use a shallower directional light angle to provoke acne * cornell_box_pipelined: Remove bias tweaks * bevy_pbr2: Simplify shadow biases by moving them to linear depth * bevy_pbr2: Do not use DepthBiasState * bevy_pbr2: Do not use bilinear filtering for sampling depth textures * pbr.wgsl: Remove unnecessary comment * bevy_pbr2: Do manual shadow map depth comparisons for more flexibility * examples: Add shadow_biases_pipelined example This is useful for stress testing biases. * bevy_pbr2: Scale the point light normal bias by the shadow map texel size This allows the normal bias to be small close to the light source where the shadow map texel to screen texel ratio is high, but is appropriately large further away from the light source where the shadow map texel can easily cover multiple screen texels. * shadow_biases_pipelined: Add support for toggling directional / point light * shadow_biases_pipelined: Cleanup * bevy_pbr2: Scale the directional light normal bias by the shadow map texel size * shadow_biases_pipelined: Fit the orthographic projection around the scene * bevy_pbr2: Directional lights should have no shadows outside their projection Before this change, sampling a fragment position from outside the ndc volume would result in the return sample being clamped to the edge in x,y or possibly always casting a shadow for fragment positions past the orthographic projection's far plane. * bevy_pbr2: Fix the default directional light normal bias * Revert "bevy_pbr2: Do manual shadow map depth comparisons for more flexibility" This reverts commit 7df1bab38a42d8a33bc50ca583d4be37bd9c9f0d. * shadow_biases_pipelined: Adjust directional light normal bias in 0.1 increments * pbr.wgsl: Add a couple of clarifying comments * Revert "bevy_pbr2: Do not use bilinear filtering for sampling depth textures" This reverts commit f53baab0232ce218866a45cad6902b470f4cf2c4. * shadow_biases_pipelined: Print usage to terminal
2021-07-19 19:20:59 +00:00
// This is the point light shadow map texel size for one face of the cube as a distance of 1.0
// world unit from the light.
// point_light_texel_size = 2.0 * 1.0 * tan(PI / 4.0) / cube face width in texels
// PI / 4.0 is half the cube face fov, tan(PI / 4.0) = 1.0, so this simplifies to:
// point_light_texel_size = 2.0 / cube face width in texels
// NOTE: When using various PCF kernel sizes, this will need to be adjusted, according to:
// https://catlikecoding.com/unity/tutorials/custom-srp/point-and-spot-shadows/
let point_light_texel_size = 2.0 / point_light_shadow_map.size as f32;
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
let mut point_lights_values = Vec::with_capacity(*previous_point_lights_len);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
for entity in global_point_lights.iter().copied() {
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
if let Ok((point_light, cubemap_visible_entities, transform, view_visibility)) =
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
point_lights.get(entity)
{
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
if !view_visibility.get() {
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
continue;
}
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
// TODO: This is very much not ideal. We should be able to re-use the vector memory.
// However, since exclusive access to the main world in extract is ill-advised, we just clone here.
let render_cubemap_visible_entities = cubemap_visible_entities.clone();
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
point_lights_values.push((
entity,
(
ExtractedPointLight {
color: point_light.color,
// NOTE: Map from luminous power in lumens to luminous intensity in lumens per steradian
// for a point light. See https://google.github.io/filament/Filament.html#mjx-eqn-pointLightLuminousPower
// for details.
intensity: point_light.intensity / (4.0 * std::f32::consts::PI),
range: point_light.range,
radius: point_light.radius,
transform: *transform,
shadows_enabled: point_light.shadows_enabled,
shadow_depth_bias: point_light.shadow_depth_bias,
// The factor of SQRT_2 is for the worst-case diagonal offset
shadow_normal_bias: point_light.shadow_normal_bias
* point_light_texel_size
* std::f32::consts::SQRT_2,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
spot_light_angles: None,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
},
render_cubemap_visible_entities,
),
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
));
}
2021-06-02 02:59:17 +00:00
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
*previous_point_lights_len = point_lights_values.len();
commands.insert_or_spawn_batch(point_lights_values);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let mut spot_lights_values = Vec::with_capacity(*previous_spot_lights_len);
for entity in global_point_lights.iter().copied() {
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
if let Ok((spot_light, visible_entities, transform, view_visibility)) =
spot_lights.get(entity)
{
if !view_visibility.get() {
Visibilty Inheritance, universal ComputedVisibility and RenderLayers support (#5310) # Objective Fixes #4907. Fixes #838. Fixes #5089. Supersedes #5146. Supersedes #2087. Supersedes #865. Supersedes #5114 Visibility is currently entirely local. Set a parent entity to be invisible, and the children are still visible. This makes it hard for users to hide entire hierarchies of entities. Additionally, the semantics of `Visibility` vs `ComputedVisibility` are inconsistent across entity types. 3D meshes use `ComputedVisibility` as the "definitive" visibility component, with `Visibility` being just one data source. Sprites just use `Visibility`, which means they can't feed off of `ComputedVisibility` data, such as culling information, RenderLayers, and (added in this pr) visibility inheritance information. ## Solution Splits `ComputedVisibilty::is_visible` into `ComputedVisibilty::is_visible_in_view` and `ComputedVisibilty::is_visible_in_hierarchy`. For each visible entity, `is_visible_in_hierarchy` is computed by propagating visibility down the hierarchy. The `ComputedVisibility::is_visible()` function combines these two booleans for the canonical "is this entity visible" function. Additionally, all entities that have `Visibility` now also have `ComputedVisibility`. Sprites, Lights, and UI entities now use `ComputedVisibility` when appropriate. This means that in addition to visibility inheritance, everything using Visibility now also supports RenderLayers. Notably, Sprites (and other 2d objects) now support `RenderLayers` and work properly across multiple views. Also note that this does increase the amount of work done per sprite. Bevymark with 100,000 sprites on `main` runs in `0.017612` seconds and this runs in `0.01902`. That is certainly a gap, but I believe the api consistency and extra functionality this buys us is worth it. See [this thread](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for more info. Note that #5146 in combination with #5114 _are_ a viable alternative to this PR and _would_ perform better, but that comes at the cost of api inconsistencies and doing visibility calculations in the "wrong" place. The current visibility system does have potential for performance improvements. I would prefer to evolve that one system as a whole rather than doing custom hacks / different behaviors for each feature slice. Here is a "split screen" example where the left camera uses RenderLayers to filter out the blue sprite. ![image](https://user-images.githubusercontent.com/2694663/178814868-2e9a2173-bf8c-4c79-8815-633899d492c3.png) Note that this builds directly on #5146 and that @james7132 deserves the credit for the baseline visibility inheritance work. This pr moves the inherited visibility field into `ComputedVisibility`, then does the additional work of porting everything to `ComputedVisibility`. See my [comments here](https://github.com/bevyengine/bevy/pull/5146#issuecomment-1182783452) for rationale. ## Follow up work * Now that lights use ComputedVisibility, VisibleEntities now includes "visible lights" in the entity list. Functionally not a problem as we use queries to filter the list down in the desired context. But we should consider splitting this out into a separate`VisibleLights` collection for both clarity and performance reasons. And _maybe_ even consider scoping `VisibleEntities` down to `VisibleMeshes`?. * Investigate alternative sprite rendering impls (in combination with visibility system tweaks) that avoid re-generating a per-view fixedbitset of visible entities every frame, then checking each ExtractedEntity. This is where most of the performance overhead lives. Ex: we could generate ExtractedEntities per-view using the VisibleEntities list, avoiding the need for the bitset. * Should ComputedVisibility use bitflags under the hood? This would cut down on the size of the component, potentially speed up the `is_visible()` function, and allow us to cheaply expand ComputedVisibility with more data (ex: split out local visibility and parent visibility, add more culling classes, etc). --- ## Changelog * ComputedVisibility now takes hierarchy visibility into account. * 2D, UI and Light entities now use the ComputedVisibility component. ## Migration Guide If you were previously reading `Visibility::is_visible` as the "actual visibility" for sprites or lights, use `ComputedVisibilty::is_visible()` instead: ```rust // before (0.7) fn system(query: Query<&Visibility>) { for visibility in query.iter() { if visibility.is_visible { log!("found visible entity"); } } } // after (0.8) fn system(query: Query<&ComputedVisibility>) { for visibility in query.iter() { if visibility.is_visible() { log!("found visible entity"); } } } ``` Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-15 23:24:42 +00:00
continue;
}
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
// TODO: This is very much not ideal. We should be able to re-use the vector memory.
// However, since exclusive access to the main world in extract is ill-advised, we just clone here.
let render_visible_entities = visible_entities.clone();
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let texel_size =
2.0 * spot_light.outer_angle.tan() / directional_light_shadow_map.size as f32;
spot_lights_values.push((
entity,
(
ExtractedPointLight {
color: spot_light.color,
// NOTE: Map from luminous power in lumens to luminous intensity in lumens per steradian
// for a point light. See https://google.github.io/filament/Filament.html#mjx-eqn-pointLightLuminousPower
// for details.
// Note: Filament uses a divisor of PI for spot lights. We choose to use the same 4*PI divisor
// in both cases so that toggling between point light and spot light keeps lit areas lit equally,
// which seems least surprising for users
intensity: spot_light.intensity / (4.0 * std::f32::consts::PI),
range: spot_light.range,
radius: spot_light.radius,
transform: *transform,
shadows_enabled: spot_light.shadows_enabled,
shadow_depth_bias: spot_light.shadow_depth_bias,
// The factor of SQRT_2 is for the worst-case diagonal offset
shadow_normal_bias: spot_light.shadow_normal_bias
* texel_size
* std::f32::consts::SQRT_2,
spot_light_angles: Some((spot_light.inner_angle, spot_light.outer_angle)),
},
render_visible_entities,
),
));
}
}
*previous_spot_lights_len = spot_lights_values.len();
commands.insert_or_spawn_batch(spot_lights_values);
for (
entity,
directional_light,
visible_entities,
cascades,
cascade_config,
transform,
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
view_visibility,
) in &directional_lights
{
Split `ComputedVisibility` into two components to allow for accurate change detection and speed up visibility propagation (#9497) # Objective Fix #8267. Fixes half of #7840. The `ComputedVisibility` component contains two flags: hierarchy visibility, and view visibility (whether its visible to any cameras). Due to the modular and open-ended way that view visibility is computed, it triggers change detection every single frame, even when the value does not change. Since hierarchy visibility is stored in the same component as view visibility, this means that change detection for inherited visibility is completely broken. At the company I work for, this has become a real issue. We are using change detection to only re-render scenes when necessary. The broken state of change detection for computed visibility means that we have to to rely on the non-inherited `Visibility` component for now. This is workable in the early stages of our project, but since we will inevitably want to use the hierarchy, we will have to either: 1. Roll our own solution for computed visibility. 2. Fix the issue for everyone. ## Solution Split the `ComputedVisibility` component into two: `InheritedVisibilty` and `ViewVisibility`. This allows change detection to behave properly for `InheritedVisibility`. View visiblity is still erratic, although it is less useful to be able to detect changes for this flavor of visibility. Overall, this actually simplifies the API. Since the visibility system consists of self-explaining components, it is much easier to document the behavior and usage. This approach is more modular and "ECS-like" -- one could strip out the `ViewVisibility` component entirely if it's not needed, and rely only on inherited visibility. --- ## Changelog - `ComputedVisibility` has been removed in favor of: `InheritedVisibility` and `ViewVisiblity`. ## Migration Guide The `ComputedVisibilty` component has been split into `InheritedVisiblity` and `ViewVisibility`. Replace any usages of `ComputedVisibility::is_visible_in_hierarchy` with `InheritedVisibility::get`, and replace `ComputedVisibility::is_visible_in_view` with `ViewVisibility::get`. ```rust // Before: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, computed_visibility: ComputedVisibility::default(), }); // After: commands.spawn(VisibilityBundle { visibility: Visibility::Inherited, inherited_visibility: InheritedVisibility::default(), view_visibility: ViewVisibility::default(), }); ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_hierarchy() { // After: fn my_system(q: Query<&InheritedVisibility>) { for inherited_visibility in &q { if inherited_visibility.get() { ``` ```rust // Before: fn my_system(q: Query<&ComputedVisibilty>) { for vis in &q { if vis.is_visible_in_view() { // After: fn my_system(q: Query<&ViewVisibility>) { for view_visibility in &q { if view_visibility.get() { ``` ```rust // Before: fn my_system(mut q: Query<&mut ComputedVisibilty>) { for vis in &mut q { vis.set_visible_in_view(); // After: fn my_system(mut q: Query<&mut ViewVisibility>) { for view_visibility in &mut q { view_visibility.set(); ``` --------- Co-authored-by: Robert Swain <robert.swain@gmail.com>
2023-09-01 13:00:18 +00:00
if !view_visibility.get() {
continue;
}
Make `RenderStage::Extract` run on the render world (#4402) # Objective - Currently, the `Extract` `RenderStage` is executed on the main world, with the render world available as a resource. - However, when needing access to resources in the render world (e.g. to mutate them), the only way to do so was to get exclusive access to the whole `RenderWorld` resource. - This meant that effectively only one extract which wrote to resources could run at a time. - We didn't previously make `Extract`ing writing to the world a non-happy path, even though we want to discourage that. ## Solution - Move the extract stage to run on the render world. - Add the main world as a `MainWorld` resource. - Add an `Extract` `SystemParam` as a convenience to access a (read only) `SystemParam` in the main world during `Extract`. ## Future work It should be possible to avoid needing to use `get_or_spawn` for the render commands, since now the `Commands`' `Entities` matches up with the world being executed on. We need to determine how this interacts with https://github.com/bevyengine/bevy/pull/3519 It's theoretically possible to remove the need for the `value` method on `Extract`. However, that requires slightly changing the `SystemParam` interface, which would make it more complicated. That would probably mess up the `SystemState` api too. ## Todo I still need to add doc comments to `Extract`. --- ## Changelog ### Changed - The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. Resources on the render world can now be accessed using `ResMut` during extract. ### Removed - `Commands::spawn_and_forget`. Use `Commands::get_or_spawn(e).insert_bundle(bundle)` instead ## Migration Guide The `Extract` `RenderStage` now runs on the render world (instead of the main world as before). You must use the `Extract` `SystemParam` to access the main world during the extract phase. `Extract` takes a single type parameter, which is any system parameter (such as `Res`, `Query` etc.). It will extract this from the main world, and returns the result of this extraction when `value` is called on it. For example, if previously your extract system looked like: ```rust fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { for cloud in clouds.iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` the new version would be: ```rust fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` The diff is: ```diff --- a/src/clouds.rs +++ b/src/clouds.rs @@ -1,5 +1,5 @@ -fn extract_clouds(mut commands: Commands, clouds: Query<Entity, With<Cloud>>) { - for cloud in clouds.iter() { +fn extract_clouds(mut commands: Commands, mut clouds: Extract<Query<Entity, With<Cloud>>>) { + for cloud in clouds.value().iter() { commands.get_or_spawn(cloud).insert(Cloud); } } ``` You can now also access resources from the render world using the normal system parameters during `Extract`: ```rust fn extract_assets(mut render_assets: ResMut<MyAssets>, source_assets: Extract<Res<MyAssets>>) { *render_assets = source_assets.clone(); } ``` Please note that all existing extract systems need to be updated to match this new style; even if they currently compile they will not run as expected. A warning will be emitted on a best-effort basis if this is not met. Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-07-08 23:56:33 +00:00
// TODO: As above
let render_visible_entities = visible_entities.clone();
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
commands.get_or_spawn(entity).insert((
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
ExtractedDirectionalLight {
color: directional_light.color,
illuminance: directional_light.illuminance,
Take DirectionalLight's GlobalTransform into account when calculating shadow map volume (not just direction) (#6384) # Objective This PR fixes #5789, by enabling movable (and scalable) directional light shadow volumes. ## Solution This PR changes `ExtractedDirectionalLight` to hold a copy of the `DirectionalLight` entity's `GlobalTransform`, instead of just a `direction` vector. This allows the shadow map volume (as defined by the light's `shadow_projection` field) to be transformed honoring translation _and_ scale transforms, and not just rotation. It also augments the texel size calculation (used to determine the `shadow_normal_bias`) so that it now takes into account the upper bound of the x/y/z scale of the `GlobalTransform`. This change makes the directional light extraction code more consistent with point and spot lights (that already use `transform`), and allows easily moving and scaling the shadow volume along with a player entity based on camera distance/angle, immediately enabling more real world use cases until we have a more sophisticated adaptive implementation, such as the one described in #3629. **Note:** While it was previously possible to update the projection achieving a similar effect, depending on the light direction and distance to the origin, the fact that the shadow map camera was always positioned at the origin with a hardcoded `Vec3::Y` up value meant you would get sub-optimal or inconsistent/incorrect results. --- ## Changelog ### Changed - `DirectionalLight` shadow volumes now honor translation and scale transforms ## Migration Guide - If your directional lights were positioned at the origin and not scaled (the default, most common scenario) no changes are needed on your part; it just works as before; - If you previously had a system for dynamically updating directional light shadow projections, you might now be able to simplify your code by updating the directional light entity's transform instead; - In the unlikely scenario that a scene with directional lights that previously rendered shadows correctly has missing shadows, make sure your directional lights are positioned at (0, 0, 0) and are not scaled to a size that's too large or too small.
2022-11-04 20:12:26 +00:00
transform: *transform,
shadows_enabled: directional_light.shadows_enabled,
shadow_depth_bias: directional_light.shadow_depth_bias,
// The factor of SQRT_2 is for the worst-case diagonal offset
shadow_normal_bias: directional_light.shadow_normal_bias * std::f32::consts::SQRT_2,
cascade_shadow_config: cascade_config.clone(),
cascades: cascades.cascades.clone(),
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
},
render_visible_entities,
));
}
2021-06-02 02:59:17 +00:00
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
pub(crate) const POINT_LIGHT_NEAR_Z: f32 = 0.1f32;
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
pub(crate) struct CubeMapFace {
pub(crate) target: Vec3,
pub(crate) up: Vec3,
2021-07-01 23:48:55 +00:00
}
// Cubemap faces are [+X, -X, +Y, -Y, +Z, -Z], per https://www.w3.org/TR/webgpu/#texture-view-creation
// Note: Cubemap coordinates are left-handed y-up, unlike the rest of Bevy.
// See https://registry.khronos.org/vulkan/specs/1.2/html/chap16.html#_cube_map_face_selection
//
// For each cubemap face, we take care to specify the appropriate target/up axis such that the rendered
// texture using Bevy's right-handed y-up coordinate space matches the expected cubemap face in
// left-handed y-up cubemap coordinates.
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
pub(crate) const CUBE_MAP_FACES: [CubeMapFace; 6] = [
// +X
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::X,
up: Vec3::Y,
2021-07-01 23:48:55 +00:00
},
// -X
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::NEG_X,
up: Vec3::Y,
2021-07-01 23:48:55 +00:00
},
// +Y
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::Y,
2021-07-01 23:48:55 +00:00
up: Vec3::Z,
},
// -Y
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::NEG_Y,
up: Vec3::NEG_Z,
2021-07-01 23:48:55 +00:00
},
// +Z (with left-handed conventions, pointing forwards)
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::NEG_Z,
up: Vec3::Y,
2021-07-01 23:48:55 +00:00
},
// -Z (with left-handed conventions, pointing backwards)
2021-07-01 23:48:55 +00:00
CubeMapFace {
target: Vec3::Z,
up: Vec3::Y,
2021-07-01 23:48:55 +00:00
},
];
fn face_index_to_name(face_index: usize) -> &'static str {
match face_index {
0 => "+x",
1 => "-x",
2 => "+y",
3 => "-y",
4 => "+z",
5 => "-z",
_ => "invalid",
}
}
2021-11-22 23:16:36 +00:00
#[derive(Component)]
pub struct ShadowView {
2021-07-01 23:48:55 +00:00
pub depth_texture_view: TextureView,
pub pass_name: String,
2021-06-02 02:59:17 +00:00
}
2021-11-22 23:16:36 +00:00
#[derive(Component)]
pub struct ViewShadowBindings {
pub point_light_depth_texture: Texture,
pub point_light_depth_texture_view: TextureView,
pub directional_light_depth_texture: Texture,
pub directional_light_depth_texture_view: TextureView,
}
2021-11-22 23:16:36 +00:00
#[derive(Component)]
pub struct ViewLightEntities {
2021-06-02 02:59:17 +00:00
pub lights: Vec<Entity>,
}
2021-11-22 23:16:36 +00:00
#[derive(Component)]
pub struct ViewLightsUniformOffset {
pub offset: u32,
2021-06-02 02:59:17 +00:00
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
// NOTE: Clustered-forward rendering requires 3 storage buffer bindings so check that
// at least that many are supported using this constant and SupportedBindingType::from_device()
pub const CLUSTERED_FORWARD_STORAGE_BUFFER_COUNT: u32 = 3;
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
#[derive(Resource)]
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
pub struct GlobalLightMeta {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
pub gpu_point_lights: GpuPointLights,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
pub entity_to_index: HashMap<Entity, usize>,
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
impl FromWorld for GlobalLightMeta {
fn from_world(world: &mut World) -> Self {
Self::new(
world
.resource::<RenderDevice>()
.get_supported_read_only_binding_type(CLUSTERED_FORWARD_STORAGE_BUFFER_COUNT),
)
}
}
impl GlobalLightMeta {
pub fn new(buffer_binding_type: BufferBindingType) -> Self {
Self {
gpu_point_lights: GpuPointLights::new(buffer_binding_type),
entity_to_index: HashMap::default(),
}
}
}
Make `Resource` trait opt-in, requiring `#[derive(Resource)]` V2 (#5577) *This PR description is an edited copy of #5007, written by @alice-i-cecile.* # Objective Follow-up to https://github.com/bevyengine/bevy/pull/2254. The `Resource` trait currently has a blanket implementation for all types that meet its bounds. While ergonomic, this results in several drawbacks: * it is possible to make confusing, silent mistakes such as inserting a function pointer (Foo) rather than a value (Foo::Bar) as a resource * it is challenging to discover if a type is intended to be used as a resource * we cannot later add customization options (see the [RFC](https://github.com/bevyengine/rfcs/blob/main/rfcs/27-derive-component.md) for the equivalent choice for Component). * dependencies can use the same Rust type as a resource in invisibly conflicting ways * raw Rust types used as resources cannot preserve privacy appropriately, as anyone able to access that type can read and write to internal values * we cannot capture a definitive list of possible resources to display to users in an editor ## Notes to reviewers * Review this commit-by-commit; there's effectively no back-tracking and there's a lot of churn in some of these commits. *ira: My commits are not as well organized :')* * I've relaxed the bound on Local to Send + Sync + 'static: I don't think these concerns apply there, so this can keep things simple. Storing e.g. a u32 in a Local is fine, because there's a variable name attached explaining what it does. * I think this is a bad place for the Resource trait to live, but I've left it in place to make reviewing easier. IMO that's best tackled with https://github.com/bevyengine/bevy/issues/4981. ## Changelog `Resource` is no longer automatically implemented for all matching types. Instead, use the new `#[derive(Resource)]` macro. ## Migration Guide Add `#[derive(Resource)]` to all types you are using as a resource. If you are using a third party type as a resource, wrap it in a tuple struct to bypass orphan rules. Consider deriving `Deref` and `DerefMut` to improve ergonomics. `ClearColor` no longer implements `Component`. Using `ClearColor` as a component in 0.8 did nothing. Use the `ClearColorConfig` in the `Camera3d` and `Camera2d` components instead. Co-authored-by: Alice <alice.i.cecile@gmail.com> Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com> Co-authored-by: devil-ira <justthecooldude@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-08-08 21:36:35 +00:00
#[derive(Resource, Default)]
2021-06-02 02:59:17 +00:00
pub struct LightMeta {
pub view_gpu_lights: DynamicUniformBuffer<GpuLights>,
2021-06-02 02:59:17 +00:00
}
2021-11-22 23:16:36 +00:00
#[derive(Component)]
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
pub enum LightEntity {
Directional {
light_entity: Entity,
cascade_index: usize,
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
},
Point {
light_entity: Entity,
face_index: usize,
},
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
Spot {
light_entity: Entity,
},
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
}
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
pub fn calculate_cluster_factors(
near: f32,
far: f32,
z_slices: f32,
is_orthographic: bool,
) -> Vec2 {
if is_orthographic {
Vec2::new(-near, z_slices / (-far - -near))
} else {
let z_slices_of_ln_zfar_over_znear = (z_slices - 1.0) / (far / near).ln();
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
Vec2::new(
z_slices_of_ln_zfar_over_znear,
near.ln() * z_slices_of_ln_zfar_over_znear,
)
}
}
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// this method of constructing a basis from a vec3 is used by glam::Vec3::any_orthonormal_pair
// we will also construct it in the fragment shader and need our implementations to match,
// so we reproduce it here to avoid a mismatch if glam changes. we also switch the handedness
// could move this onto transform but it's pretty niche
pub(crate) fn spot_light_view_matrix(transform: &GlobalTransform) -> Mat4 {
// the matrix z_local (opposite of transform.forward())
let fwd_dir = transform.back().extend(0.0);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let sign = 1f32.copysign(fwd_dir.z);
let a = -1.0 / (fwd_dir.z + sign);
let b = fwd_dir.x * fwd_dir.y * a;
let up_dir = Vec4::new(
1.0 + sign * fwd_dir.x * fwd_dir.x * a,
sign * b,
-sign * fwd_dir.x,
0.0,
);
let right_dir = Vec4::new(-b, -sign - fwd_dir.y * fwd_dir.y * a, fwd_dir.y, 0.0);
Mat4::from_cols(
right_dir,
up_dir,
fwd_dir,
transform.translation().extend(1.0),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
)
}
pub(crate) fn spot_light_projection_matrix(angle: f32) -> Mat4 {
// spot light projection FOV is 2x the angle from spot light center to outer edge
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
Mat4::perspective_infinite_reverse_rh(angle * 2.0, 1.0, POINT_LIGHT_NEAR_Z)
}
2021-07-08 03:01:41 +00:00
#[allow(clippy::too_many_arguments)]
2021-06-02 02:59:17 +00:00
pub fn prepare_lights(
mut commands: Commands,
mut texture_cache: ResMut<TextureCache>,
images: Res<RenderAssets<Image>>,
2021-06-21 23:28:52 +00:00
render_device: Res<RenderDevice>,
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
render_queue: Res<RenderQueue>,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
mut global_light_meta: ResMut<GlobalLightMeta>,
2021-06-02 02:59:17 +00:00
mut light_meta: ResMut<LightMeta>,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
views: Query<
(
Entity,
&ExtractedView,
&ExtractedClusterConfig,
Option<&EnvironmentMapLight>,
),
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
With<RenderPhase<Transparent3d>>,
>,
ambient_light: Res<AmbientLight>,
point_light_shadow_map: Res<PointLightShadowMap>,
directional_light_shadow_map: Res<DirectionalLightShadowMap>,
mut max_directional_lights_warning_emitted: Local<bool>,
mut max_cascades_per_light_warning_emitted: Local<bool>,
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
point_lights: Query<(Entity, &ExtractedPointLight)>,
directional_lights: Query<(Entity, &ExtractedDirectionalLight)>,
2021-06-02 02:59:17 +00:00
) {
Sprite Batching (#3060) This implements the following: * **Sprite Batching**: Collects sprites in a vertex buffer to draw many sprites with a single draw call. Sprites are batched by their `Handle<Image>` within a specific z-level. When possible, sprites are opportunistically batched _across_ z-levels (when no sprites with a different texture exist between two sprites with the same texture on different z levels). With these changes, I can now get ~130,000 sprites at 60fps on the `bevymark_pipelined` example. * **Sprite Color Tints**: The `Sprite` type now has a `color` field. Non-white color tints result in a specialized render pipeline that passes the color in as a vertex attribute. I chose to specialize this because passing vertex colors has a measurable price (without colors I get ~130,000 sprites on bevymark, with colors I get ~100,000 sprites). "Colored" sprites cannot be batched with "uncolored" sprites, but I think this is fine because the chance of a "colored" sprite needing to batch with other "colored" sprites is generally probably way higher than an "uncolored" sprite needing to batch with a "colored" sprite. * **Sprite Flipping**: Sprites can be flipped on their x or y axis using `Sprite::flip_x` and `Sprite::flip_y`. This is also true for `TextureAtlasSprite`. * **Simpler BufferVec/UniformVec/DynamicUniformVec Clearing**: improved the clearing interface by removing the need to know the size of the final buffer at the initial clear. ![image](https://user-images.githubusercontent.com/2694663/140001821-99be0d96-025d-489e-9bfa-ba19c1dc9548.png) Note that this moves sprites away from entity-driven rendering and back to extracted lists. We _could_ use entities here, but it necessitates that an intermediate list is allocated / populated to collect and sort extracted sprites. This redundant copy, combined with the normal overhead of spawning extracted sprite entities, brings bevymark down to ~80,000 sprites at 60fps. I think making sprites a bit more fixed (by default) is worth it. I view this as acceptable because batching makes normal entity-driven rendering pretty useless anyway (and we would want to batch most custom materials too). We can still support custom shaders with custom bindings, we'll just need to define a specific interface for it.
2021-11-04 20:28:53 +00:00
light_meta.view_gpu_lights.clear();
2021-06-02 02:59:17 +00:00
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
// Pre-calculate for PointLights
let cube_face_projection =
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
Mat4::perspective_infinite_reverse_rh(std::f32::consts::FRAC_PI_2, 1.0, POINT_LIGHT_NEAR_Z);
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
let cube_face_rotations = CUBE_MAP_FACES
.iter()
.map(|CubeMapFace { target, up }| Transform::IDENTITY.looking_at(*target, *up))
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
.collect::<Vec<_>>();
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
global_light_meta.entity_to_index.clear();
let mut point_lights: Vec<_> = point_lights.iter().collect::<Vec<_>>();
let mut directional_lights: Vec<_> = directional_lights.iter().collect::<Vec<_>>();
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let max_texture_array_layers = render_device.limits().max_texture_array_layers as usize;
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let max_texture_cubes = max_texture_array_layers / 6;
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let max_texture_array_layers = 1;
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let max_texture_cubes = 1;
if !*max_directional_lights_warning_emitted && directional_lights.len() > MAX_DIRECTIONAL_LIGHTS
{
warn!(
"The amount of directional lights of {} is exceeding the supported limit of {}.",
directional_lights.len(),
MAX_DIRECTIONAL_LIGHTS
);
*max_directional_lights_warning_emitted = true;
}
if !*max_cascades_per_light_warning_emitted
&& directional_lights
.iter()
.any(|(_, light)| light.cascade_shadow_config.bounds.len() > MAX_CASCADES_PER_LIGHT)
{
warn!(
"The number of cascades configured for a directional light exceeds the supported limit of {}.",
MAX_CASCADES_PER_LIGHT
);
*max_cascades_per_light_warning_emitted = true;
}
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let point_light_count = point_lights
.iter()
.filter(|light| light.1.spot_light_angles.is_none())
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
.count();
let point_light_shadow_maps_count = point_lights
.iter()
.filter(|light| light.1.shadows_enabled && light.1.spot_light_angles.is_none())
.count()
.min(max_texture_cubes);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let directional_shadow_enabled_count = directional_lights
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
.iter()
.take(MAX_DIRECTIONAL_LIGHTS)
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
.filter(|(_, light)| light.shadows_enabled)
.count()
.min(max_texture_array_layers / MAX_CASCADES_PER_LIGHT);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let spot_light_shadow_maps_count = point_lights
.iter()
.filter(|(_, light)| light.shadows_enabled && light.spot_light_angles.is_some())
.count()
.min(max_texture_array_layers - directional_shadow_enabled_count * MAX_CASCADES_PER_LIGHT);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// Sort lights by
// - point-light vs spot-light, so that we can iterate point lights and spot lights in contiguous blocks in the fragment shader,
// - then those with shadows enabled first, so that the index can be used to render at most `point_light_shadow_maps_count`
// point light shadows and `spot_light_shadow_maps_count` spot light shadow maps,
// - then by entity as a stable key to ensure that a consistent set of lights are chosen if the light count limit is exceeded.
point_lights.sort_by(|(entity_1, light_1), (entity_2, light_2)| {
point_light_order(
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
(
entity_1,
&light_1.shadows_enabled,
&light_1.spot_light_angles.is_some(),
),
(
entity_2,
&light_2.shadows_enabled,
&light_2.spot_light_angles.is_some(),
),
)
});
// Sort lights by
// - those with shadows enabled first, so that the index can be used to render at most `directional_light_shadow_maps_count`
// directional light shadows
// - then by entity as a stable key to ensure that a consistent set of lights are chosen if the light count limit is exceeded.
directional_lights.sort_by(|(entity_1, light_1), (entity_2, light_2)| {
directional_light_order(
(entity_1, &light_1.shadows_enabled),
(entity_2, &light_2.shadows_enabled),
)
});
if global_light_meta.entity_to_index.capacity() < point_lights.len() {
global_light_meta
.entity_to_index
.reserve(point_lights.len());
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
let mut gpu_point_lights = Vec::new();
for (index, &(entity, light)) in point_lights.iter().enumerate() {
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
let mut flags = PointLightFlags::NONE;
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// Lights are sorted, shadow enabled lights are first
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
if light.shadows_enabled
&& (index < point_light_shadow_maps_count
|| (light.spot_light_angles.is_some()
&& index - point_light_count < spot_light_shadow_maps_count))
{
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
flags |= PointLightFlags::SHADOWS_ENABLED;
}
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let (light_custom_data, spot_light_tan_angle) = match light.spot_light_angles {
Some((inner, outer)) => {
let light_direction = light.transform.forward();
if light_direction.y.is_sign_negative() {
flags |= PointLightFlags::SPOT_LIGHT_Y_NEGATIVE;
}
let cos_outer = outer.cos();
let spot_scale = 1.0 / f32::max(inner.cos() - cos_outer, 1e-4);
let spot_offset = -cos_outer * spot_scale;
(
// For spot lights: the direction (x,z), spot_scale and spot_offset
light_direction.xz().extend(spot_scale).extend(spot_offset),
outer.tan(),
)
}
None => {
(
// For point lights: the lower-right 2x2 values of the projection matrix [2][2] [2][3] [3][2] [3][3]
Vec4::new(
cube_face_projection.z_axis.z,
cube_face_projection.z_axis.w,
cube_face_projection.w_axis.z,
cube_face_projection.w_axis.w,
),
// unused
0.0,
)
}
};
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
gpu_point_lights.push(GpuPointLight {
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
light_custom_data,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
// premultiply color by intensity
// we don't use the alpha at all, so no reason to multiply only [0..3]
color_inverse_square_range: (Vec4::from_slice(&light.color.as_linear_rgba_f32())
* light.intensity)
.xyz()
.extend(1.0 / (light.range * light.range)),
position_radius: light.transform.translation().extend(light.radius),
flags: flags.bits(),
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
shadow_depth_bias: light.shadow_depth_bias,
shadow_normal_bias: light.shadow_normal_bias,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
spot_light_tan_angle,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
});
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
global_light_meta.entity_to_index.insert(entity, index);
}
let mut gpu_directional_lights = [GpuDirectionalLight::default(); MAX_DIRECTIONAL_LIGHTS];
let mut num_directional_cascades_enabled = 0usize;
for (index, (_light_entity, light)) in directional_lights
.iter()
.enumerate()
.take(MAX_DIRECTIONAL_LIGHTS)
{
let mut flags = DirectionalLightFlags::NONE;
// Lights are sorted, shadow enabled lights are first
if light.shadows_enabled && (index < directional_shadow_enabled_count) {
flags |= DirectionalLightFlags::SHADOWS_ENABLED;
}
// convert from illuminance (lux) to candelas
//
// exposure is hard coded at the moment but should be replaced
// by values coming from the camera
// see: https://google.github.io/filament/Filament.html#imagingpipeline/physicallybasedcamera/exposuresettings
const APERTURE: f32 = 4.0;
const SHUTTER_SPEED: f32 = 1.0 / 250.0;
const SENSITIVITY: f32 = 100.0;
let ev100 = f32::log2(APERTURE * APERTURE / SHUTTER_SPEED) - f32::log2(SENSITIVITY / 100.0);
let exposure = 1.0 / (f32::powf(2.0, ev100) * 1.2);
let intensity = light.illuminance * exposure;
let num_cascades = light
.cascade_shadow_config
.bounds
.len()
.min(MAX_CASCADES_PER_LIGHT);
gpu_directional_lights[index] = GpuDirectionalLight {
// Filled in later.
cascades: [GpuDirectionalCascade::default(); MAX_CASCADES_PER_LIGHT],
// premultiply color by intensity
// we don't use the alpha at all, so no reason to multiply only [0..3]
color: Vec4::from_slice(&light.color.as_linear_rgba_f32()) * intensity,
// direction is negated to be ready for N.L
dir_to_light: light.transform.back(),
flags: flags.bits(),
shadow_depth_bias: light.shadow_depth_bias,
shadow_normal_bias: light.shadow_normal_bias,
num_cascades: num_cascades as u32,
cascades_overlap_proportion: light.cascade_shadow_config.overlap_proportion,
depth_texture_base_index: num_directional_cascades_enabled as u32,
};
if index < directional_shadow_enabled_count {
num_directional_cascades_enabled += num_cascades;
}
}
global_light_meta.gpu_point_lights.set(gpu_point_lights);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
global_light_meta
.gpu_point_lights
.write_buffer(&render_device, &render_queue);
2021-06-02 02:59:17 +00:00
// set up light data for each view
for (entity, extracted_view, clusters, environment_map) in &views {
let point_light_depth_texture = texture_cache.get(
&render_device,
TextureDescriptor {
size: Extent3d {
width: point_light_shadow_map.size as u32,
height: point_light_shadow_map.size as u32,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
depth_or_array_layers: point_light_shadow_maps_count.max(1) as u32 * 6,
},
mip_level_count: 1,
sample_count: 1,
dimension: TextureDimension::D2,
format: SHADOW_FORMAT,
label: Some("point_light_shadow_map_texture"),
usage: TextureUsages::RENDER_ATTACHMENT | TextureUsages::TEXTURE_BINDING,
Wgpu 0.15 (#7356) # Objective Update Bevy to wgpu 0.15. ## Changelog - Update to wgpu 0.15, wgpu-hal 0.15.1, and naga 0.11 - Users can now use the [DirectX Shader Compiler](https://github.com/microsoft/DirectXShaderCompiler) (DXC) on Windows with DX12 for faster shader compilation and ShaderModel 6.0+ support (requires `dxcompiler.dll` and `dxil.dll`, which are included in DXC downloads from [here](https://github.com/microsoft/DirectXShaderCompiler/releases/latest)) ## Migration Guide ### WGSL Top-Level `let` is now `const` All top level constants are now declared with `const`, catching up with the wgsl spec. `let` is no longer allowed at the global scope, only within functions. ```diff -let SOME_CONSTANT = 12.0; +const SOME_CONSTANT = 12.0; ``` #### `TextureDescriptor` and `SurfaceConfiguration` now requires a `view_formats` field The new `view_formats` field in the `TextureDescriptor` is used to specify a list of formats the texture can be re-interpreted to in a texture view. Currently only changing srgb-ness is allowed (ex. `Rgba8Unorm` <=> `Rgba8UnormSrgb`). You should set `view_formats` to `&[]` (empty) unless you have a specific reason not to. #### The DirectX Shader Compiler (DXC) is now supported on DX12 DXC is now the default shader compiler when using the DX12 backend. DXC is Microsoft's replacement for their legacy FXC compiler, and is faster, less buggy, and allows for modern shader features to be used (ShaderModel 6.0+). DXC requires `dxcompiler.dll` and `dxil.dll` to be available, otherwise it will log a warning and fall back to FXC. You can get `dxcompiler.dll` and `dxil.dll` by downloading the latest release from [Microsoft's DirectXShaderCompiler github repo](https://github.com/microsoft/DirectXShaderCompiler/releases/latest) and copying them into your project's root directory. These must be included when you distribute your Bevy game/app/etc if you plan on supporting the DX12 backend and are using DXC. `WgpuSettings` now has a `dx12_shader_compiler` field which can be used to choose between either FXC or DXC (if you pass None for the paths for DXC, it will check for the .dlls in the working directory).
2023-01-29 20:27:30 +00:00
view_formats: &[],
},
);
let directional_light_depth_texture = texture_cache.get(
2021-06-21 23:28:52 +00:00
&render_device,
2021-06-02 02:59:17 +00:00
TextureDescriptor {
size: Extent3d {
width: (directional_light_shadow_map.size as u32)
.min(render_device.limits().max_texture_dimension_2d),
height: (directional_light_shadow_map.size as u32)
.min(render_device.limits().max_texture_dimension_2d),
depth_or_array_layers: (num_directional_cascades_enabled
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
+ spot_light_shadow_maps_count)
.max(1) as u32,
},
2021-06-02 02:59:17 +00:00
mip_level_count: 1,
sample_count: 1,
dimension: TextureDimension::D2,
format: SHADOW_FORMAT,
label: Some("directional_light_shadow_map_texture"),
usage: TextureUsages::RENDER_ATTACHMENT | TextureUsages::TEXTURE_BINDING,
Wgpu 0.15 (#7356) # Objective Update Bevy to wgpu 0.15. ## Changelog - Update to wgpu 0.15, wgpu-hal 0.15.1, and naga 0.11 - Users can now use the [DirectX Shader Compiler](https://github.com/microsoft/DirectXShaderCompiler) (DXC) on Windows with DX12 for faster shader compilation and ShaderModel 6.0+ support (requires `dxcompiler.dll` and `dxil.dll`, which are included in DXC downloads from [here](https://github.com/microsoft/DirectXShaderCompiler/releases/latest)) ## Migration Guide ### WGSL Top-Level `let` is now `const` All top level constants are now declared with `const`, catching up with the wgsl spec. `let` is no longer allowed at the global scope, only within functions. ```diff -let SOME_CONSTANT = 12.0; +const SOME_CONSTANT = 12.0; ``` #### `TextureDescriptor` and `SurfaceConfiguration` now requires a `view_formats` field The new `view_formats` field in the `TextureDescriptor` is used to specify a list of formats the texture can be re-interpreted to in a texture view. Currently only changing srgb-ness is allowed (ex. `Rgba8Unorm` <=> `Rgba8UnormSrgb`). You should set `view_formats` to `&[]` (empty) unless you have a specific reason not to. #### The DirectX Shader Compiler (DXC) is now supported on DX12 DXC is now the default shader compiler when using the DX12 backend. DXC is Microsoft's replacement for their legacy FXC compiler, and is faster, less buggy, and allows for modern shader features to be used (ShaderModel 6.0+). DXC requires `dxcompiler.dll` and `dxil.dll` to be available, otherwise it will log a warning and fall back to FXC. You can get `dxcompiler.dll` and `dxil.dll` by downloading the latest release from [Microsoft's DirectXShaderCompiler github repo](https://github.com/microsoft/DirectXShaderCompiler/releases/latest) and copying them into your project's root directory. These must be included when you distribute your Bevy game/app/etc if you plan on supporting the DX12 backend and are using DXC. `WgpuSettings` now has a `dx12_shader_compiler` field which can be used to choose between either FXC or DXC (if you pass None for the paths for DXC, it will check for the .dlls in the working directory).
2023-01-29 20:27:30 +00:00
view_formats: &[],
2021-06-02 02:59:17 +00:00
},
);
let mut view_lights = Vec::new();
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
let is_orthographic = extracted_view.projection.w_axis.w == 1.0;
let cluster_factors_zw = calculate_cluster_factors(
clusters.near,
Dynamic light clusters (#3968) # Objective provide some customisation for default cluster setup avoid "cluster index lists is full" in all cases (using a strategy outlined by @superdump) ## Solution Add ClusterConfig enum (which can be inserted into a view at any time) to allow specifying cluster setup with variants: - None (do not do any light assignment - for views which do not require light info, e.g. minimaps etc) - Single (one cluster) - XYZ (explicit cluster counts in each dimension) - FixedZ (most similar to current - specify Z-slices and total, then x and y counts are dynamically determined to give approximately square clusters based on current aspect ratio) Defaults to FixedZ { total: 4096, z: 24 } which is similar to the current setup. Per frame, estimate the number of indices that would be required for the current config and decrease the cluster counts / increase the cluster sizes in the x and y dimensions if the index list would be too small. notes: - I didn't put ClusterConfig in the camera bundles to avoid introducing a dependency from bevy_render to bevy_pbr. the ClusterConfig enum comes with a pbr-centric impl block so i didn't want to move that into bevy_render either. - ~Might want to add None variant to cluster config for views that don't care about lights?~ - Not well tested for orthographic - ~there's a cluster_muck branch on my repo which includes some diagnostics / a modified lighting example which may be useful for tyre-kicking~ (outdated, i will bring it up to date if required) anecdotal timings: FPS on the lighting demo is negligibly better (~5%), maybe due to a small optimisation constraining the light aabb to be in front of the camera FPS on the lighting demo with 100 extra lights added is ~33% faster, and also renders correctly as the cluster index count is no longer exceeded
2022-03-08 04:56:42 +00:00
clusters.far,
clusters.dimensions.z as f32,
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
is_orthographic,
);
let n_clusters = clusters.dimensions.x * clusters.dimensions.y * clusters.dimensions.z;
let mut gpu_lights = GpuLights {
directional_lights: gpu_directional_lights,
ambient_color: Vec4::from_slice(&ambient_light.color.as_linear_rgba_f32())
* ambient_light.brightness,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
cluster_factors: Vec4::new(
clusters.dimensions.x as f32 / extracted_view.viewport.z as f32,
clusters.dimensions.y as f32 / extracted_view.viewport.w as f32,
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
cluster_factors_zw.x,
cluster_factors_zw.y,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
),
cluster_dimensions: clusters.dimensions.extend(n_clusters),
n_directional_lights: directional_lights.iter().len() as u32,
// spotlight shadow maps are stored in the directional light array, starting at num_directional_cascades_enabled.
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// the spot lights themselves start in the light array at point_light_count. so to go from light
// index to shadow map index, we need to subtract point light count and add directional shadowmap count.
spot_light_shadowmap_offset: num_directional_cascades_enabled as i32
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
- point_light_count as i32,
environment_map_smallest_specular_mip_level: environment_map
.and_then(|env_map| images.get(&env_map.specular_map))
.map(|specular_map| specular_map.mip_level_count - 1)
.unwrap_or(0),
2021-06-02 02:59:17 +00:00
};
// TODO: this should select lights based on relevance to the view instead of the first ones that show up in a query
for &(light_entity, light) in point_lights
.iter()
// Lights are sorted, shadow enabled lights are first
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
.take(point_light_shadow_maps_count)
.filter(|(_, light)| light.shadows_enabled)
{
let light_index = *global_light_meta
.entity_to_index
.get(&light_entity)
.unwrap();
// ignore scale because we don't want to effectively scale light radius and range
// by applying those as a view transform to shadow map rendering of objects
// and ignore rotation because we want the shadow map projections to align with the axes
let view_translation = GlobalTransform::from_translation(light.transform.translation());
for (face_index, view_rotation) in cube_face_rotations.iter().enumerate() {
let depth_texture_view =
point_light_depth_texture
.texture
.create_view(&TextureViewDescriptor {
label: Some("point_light_shadow_map_texture_view"),
format: None,
dimension: Some(TextureViewDimension::D2),
aspect: TextureAspect::All,
base_mip_level: 0,
mip_level_count: None,
base_array_layer: (light_index * 6 + face_index) as u32,
array_layer_count: Some(1u32),
});
let view_light_entity = commands
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn((
ShadowView {
depth_texture_view,
pass_name: format!(
"shadow pass point light {} {}",
light_index,
face_index_to_name(face_index)
),
},
ExtractedView {
viewport: UVec4::new(
0,
0,
point_light_shadow_map.size as u32,
point_light_shadow_map.size as u32,
),
transform: view_translation * *view_rotation,
view_projection: None,
projection: cube_face_projection,
separate tonemapping and upscaling passes (#3425) Attempt to make features like bloom https://github.com/bevyengine/bevy/pull/2876 easier to implement. **This PR:** - Moves the tonemapping from `pbr.wgsl` into a separate pass - also add a separate upscaling pass after the tonemapping which writes to the swap chain (enables resolution-independant rendering and post-processing after tonemapping) - adds a `hdr` bool to the camera which controls whether the pbr and sprite shaders render into a `Rgba16Float` texture **Open questions:** - ~should the 2d graph work the same as the 3d one?~ it is the same now - ~The current solution is a bit inflexible because while you can add a post processing pass that writes to e.g. the `hdr_texture`, you can't write to a separate `user_postprocess_texture` while reading the `hdr_texture` and tell the tone mapping pass to read from the `user_postprocess_texture` instead. If the tonemapping and upscaling render graph nodes were to take in a `TextureView` instead of the view entity this would almost work, but the bind groups for their respective input textures are already created in the `Queue` render stage in the hardcoded order.~ solved by creating bind groups in render node **New render graph:** ![render_graph](https://user-images.githubusercontent.com/22177966/147767249-57dd4229-cfab-4ec5-9bf3-dc76dccf8e8b.png) <details> <summary>Before</summary> ![render_graph_old](https://user-images.githubusercontent.com/22177966/147284579-c895fdbd-4028-41cf-914c-e1ffef60e44e.png) </details> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-10-26 20:13:59 +00:00
hdr: false,
color_grading: Default::default(),
},
RenderPhase::<Shadow>::default(),
LightEntity::Point {
light_entity,
face_index,
},
))
.id();
view_lights.push(view_light_entity);
}
2021-06-02 02:59:17 +00:00
}
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// spot lights
for (light_index, &(light_entity, light)) in point_lights
.iter()
.skip(point_light_count)
.take(spot_light_shadow_maps_count)
.enumerate()
{
let spot_view_matrix = spot_light_view_matrix(&light.transform);
let spot_view_transform = spot_view_matrix.into();
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let angle = light.spot_light_angles.expect("lights should be sorted so that \
[point_light_count..point_light_count + spot_light_shadow_maps_count] are spot lights").1;
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let spot_projection = spot_light_projection_matrix(angle);
let depth_texture_view =
directional_light_depth_texture
.texture
.create_view(&TextureViewDescriptor {
label: Some("spot_light_shadow_map_texture_view"),
format: None,
dimension: Some(TextureViewDimension::D2),
aspect: TextureAspect::All,
base_mip_level: 0,
mip_level_count: None,
base_array_layer: (num_directional_cascades_enabled + light_index) as u32,
array_layer_count: Some(1u32),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
});
let view_light_entity = commands
Spawn now takes a Bundle (#6054) # Objective Now that we can consolidate Bundles and Components under a single insert (thanks to #2975 and #6039), almost 100% of world spawns now look like `world.spawn().insert((Some, Tuple, Here))`. Spawning an entity without any components is an extremely uncommon pattern, so it makes sense to give spawn the "first class" ergonomic api. This consolidated api should be made consistent across all spawn apis (such as World and Commands). ## Solution All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input: ```rust // before: commands .spawn() .insert((A, B, C)); world .spawn() .insert((A, B, C); // after commands.spawn((A, B, C)); world.spawn((A, B, C)); ``` All existing instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api. A new `spawn_empty` has been added, replacing the old `spawn` api. By allowing `world.spawn(some_bundle)` to replace `world.spawn().insert(some_bundle)`, this opened the door to removing the initial entity allocation in the "empty" archetype / table done in `spawn()` (and subsequent move to the actual archetype in `.insert(some_bundle)`). This improves spawn performance by over 10%: ![image](https://user-images.githubusercontent.com/2694663/191627587-4ab2f949-4ccd-4231-80eb-80dd4d9ad6b9.png) To take this measurement, I added a new `world_spawn` benchmark. Unfortunately, optimizing `Commands::spawn` is slightly less trivial, as Commands expose the Entity id of spawned entities prior to actually spawning. Doing the optimization would (naively) require assurances that the `spawn(some_bundle)` command is applied before all other commands involving the entity (which would not necessarily be true, if memory serves). Optimizing `Commands::spawn` this way does feel possible, but it will require careful thought (and maybe some additional checks), which deserves its own PR. For now, it has the same performance characteristics of the current `Commands::spawn_bundle` on main. **Note that 99% of this PR is simple renames and refactors. The only code that needs careful scrutiny is the new `World::spawn()` impl, which is relatively straightforward, but it has some new unsafe code (which re-uses battle tested BundlerSpawner code path).** --- ## Changelog - All `spawn` apis (`World::spawn`, `Commands:;spawn`, `ChildBuilder::spawn`, and `WorldChildBuilder::spawn`) now accept a bundle as input - All instances of `spawn_bundle` have been deprecated in favor of the new `spawn` api - World and Commands now have `spawn_empty()`, which is equivalent to the old `spawn()` behavior. ## Migration Guide ```rust // Old (0.8): commands .spawn() .insert_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): commands.spawn_bundle((A, B, C)); // New (0.9) commands.spawn((A, B, C)); // Old (0.8): let entity = commands.spawn().id(); // New (0.9) let entity = commands.spawn_empty().id(); // Old (0.8) let entity = world.spawn().id(); // New (0.9) let entity = world.spawn_empty(); ```
2022-09-23 19:55:54 +00:00
.spawn((
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
ShadowView {
depth_texture_view,
pass_name: format!("shadow pass spot light {light_index}"),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
},
ExtractedView {
viewport: UVec4::new(
0,
0,
directional_light_shadow_map.size as u32,
directional_light_shadow_map.size as u32,
),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
transform: spot_view_transform,
projection: spot_projection,
view_projection: None,
separate tonemapping and upscaling passes (#3425) Attempt to make features like bloom https://github.com/bevyengine/bevy/pull/2876 easier to implement. **This PR:** - Moves the tonemapping from `pbr.wgsl` into a separate pass - also add a separate upscaling pass after the tonemapping which writes to the swap chain (enables resolution-independant rendering and post-processing after tonemapping) - adds a `hdr` bool to the camera which controls whether the pbr and sprite shaders render into a `Rgba16Float` texture **Open questions:** - ~should the 2d graph work the same as the 3d one?~ it is the same now - ~The current solution is a bit inflexible because while you can add a post processing pass that writes to e.g. the `hdr_texture`, you can't write to a separate `user_postprocess_texture` while reading the `hdr_texture` and tell the tone mapping pass to read from the `user_postprocess_texture` instead. If the tonemapping and upscaling render graph nodes were to take in a `TextureView` instead of the view entity this would almost work, but the bind groups for their respective input textures are already created in the `Queue` render stage in the hardcoded order.~ solved by creating bind groups in render node **New render graph:** ![render_graph](https://user-images.githubusercontent.com/22177966/147767249-57dd4229-cfab-4ec5-9bf3-dc76dccf8e8b.png) <details> <summary>Before</summary> ![render_graph_old](https://user-images.githubusercontent.com/22177966/147284579-c895fdbd-4028-41cf-914c-e1ffef60e44e.png) </details> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-10-26 20:13:59 +00:00
hdr: false,
color_grading: Default::default(),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
},
RenderPhase::<Shadow>::default(),
LightEntity::Spot { light_entity },
))
.id();
view_lights.push(view_light_entity);
}
// directional lights
let mut directional_depth_texture_array_index = 0u32;
for (light_index, &(light_entity, light)) in directional_lights
.iter()
.enumerate()
.take(directional_shadow_enabled_count)
{
for (cascade_index, (cascade, bound)) in light
.cascades
.get(&entity)
.unwrap()
.iter()
.take(MAX_CASCADES_PER_LIGHT)
.zip(&light.cascade_shadow_config.bounds)
.enumerate()
{
gpu_lights.directional_lights[light_index].cascades[cascade_index] =
GpuDirectionalCascade {
view_projection: cascade.view_projection,
texel_size: cascade.texel_size,
far_bound: *bound,
};
let depth_texture_view =
directional_light_depth_texture
.texture
.create_view(&TextureViewDescriptor {
label: Some("directional_light_shadow_map_array_texture_view"),
format: None,
dimension: Some(TextureViewDimension::D2),
aspect: TextureAspect::All,
base_mip_level: 0,
mip_level_count: None,
base_array_layer: directional_depth_texture_array_index,
array_layer_count: Some(1u32),
});
directional_depth_texture_array_index += 1;
let view_light_entity = commands
.spawn((
ShadowView {
depth_texture_view,
pass_name: format!(
"shadow pass directional light {light_index} cascade {cascade_index}"),
},
ExtractedView {
viewport: UVec4::new(
0,
0,
directional_light_shadow_map.size as u32,
directional_light_shadow_map.size as u32,
),
transform: GlobalTransform::from(cascade.view_transform),
projection: cascade.projection,
view_projection: Some(cascade.view_projection),
hdr: false,
color_grading: Default::default(),
},
RenderPhase::<Shadow>::default(),
LightEntity::Directional {
light_entity,
cascade_index,
},
))
.id();
view_lights.push(view_light_entity);
}
}
let point_light_depth_texture_view =
point_light_depth_texture
2021-07-01 23:48:55 +00:00
.texture
.create_view(&TextureViewDescriptor {
label: Some("point_light_shadow_map_array_texture_view"),
2021-07-01 23:48:55 +00:00
format: None,
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
2021-07-01 23:48:55 +00:00
dimension: Some(TextureViewDimension::CubeArray),
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
dimension: Some(TextureViewDimension::Cube),
2021-07-01 23:48:55 +00:00
aspect: TextureAspect::All,
base_mip_level: 0,
mip_level_count: None,
2021-07-02 01:05:20 +00:00
base_array_layer: 0,
2021-07-01 23:48:55 +00:00
array_layer_count: None,
});
let directional_light_depth_texture_view = directional_light_depth_texture
.texture
.create_view(&TextureViewDescriptor {
label: Some("directional_light_shadow_map_array_texture_view"),
format: None,
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(any(not(feature = "webgl"), not(target_arch = "wasm32")))]
dimension: Some(TextureViewDimension::D2Array),
Webgpu support (#8336) # Objective - Support WebGPU - alternative to #5027 that doesn't need any async / await - fixes #8315 - Surprise fix #7318 ## Solution ### For async renderer initialisation - Update the plugin lifecycle: - app builds the plugin - calls `plugin.build` - registers the plugin - app starts the event loop - event loop waits for `ready` of all registered plugins in the same order - returns `true` by default - then call all `finish` then all `cleanup` in the same order as registered - then execute the schedule In the case of the renderer, to avoid anything async: - building the renderer plugin creates a detached task that will send back the initialised renderer through a mutex in a resource - `ready` will wait for the renderer to be present in the resource - `finish` will take that renderer and place it in the expected resources by other plugins - other plugins (that expect the renderer to be available) `finish` are called and they are able to set up their pipelines - `cleanup` is called, only custom one is still for pipeline rendering ### For WebGPU support - update the `build-wasm-example` script to support passing `--api webgpu` that will build the example with WebGPU support - feature for webgl2 was always enabled when building for wasm. it's now in the default feature list and enabled on all platforms, so check for this feature must also check that the target_arch is `wasm32` --- ## Migration Guide - `Plugin::setup` has been renamed `Plugin::cleanup` - `Plugin::finish` has been added, and plugins adding pipelines should do it in this function instead of `Plugin::build` ```rust // Before impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>() .init_resource::<OtherRenderResource>(); } } // After impl Plugin for MyPlugin { fn build(&self, app: &mut App) { app.insert_resource::<MyResource> .add_systems(Update, my_system); let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<OtherRenderResource>(); } fn finish(&self, app: &mut App) { let render_app = match app.get_sub_app_mut(RenderApp) { Ok(render_app) => render_app, Err(_) => return, }; render_app .init_resource::<RenderResourceNeedingDevice>(); } } ```
2023-05-04 22:07:57 +00:00
#[cfg(all(feature = "webgl", target_arch = "wasm32"))]
dimension: Some(TextureViewDimension::D2),
aspect: TextureAspect::All,
base_mip_level: 0,
mip_level_count: None,
base_array_layer: 0,
array_layer_count: None,
});
2021-07-01 23:48:55 +00:00
Accept Bundles for insert and remove. Deprecate insert/remove_bundle (#6039) # Objective Take advantage of the "impl Bundle for Component" changes in #2975 / add the follow up changes discussed there. ## Solution - Change `insert` and `remove` to accept a Bundle instead of a Component (for both Commands and World) - Deprecate `insert_bundle`, `remove_bundle`, and `remove_bundle_intersection` - Add `remove_intersection` --- ## Changelog - Change `insert` and `remove` now accept a Bundle instead of a Component (for both Commands and World) - `insert_bundle` and `remove_bundle` are deprecated ## Migration Guide Replace `insert_bundle` with `insert`: ```rust // Old (0.8) commands.spawn().insert_bundle(SomeBundle::default()); // New (0.9) commands.spawn().insert(SomeBundle::default()); ``` Replace `remove_bundle` with `remove`: ```rust // Old (0.8) commands.entity(some_entity).remove_bundle::<SomeBundle>(); // New (0.9) commands.entity(some_entity).remove::<SomeBundle>(); ``` Replace `remove_bundle_intersection` with `remove_intersection`: ```rust // Old (0.8) world.entity_mut(some_entity).remove_bundle_intersection::<SomeBundle>(); // New (0.9) world.entity_mut(some_entity).remove_intersection::<SomeBundle>(); ``` Consider consolidating as many operations as possible to improve ergonomics and cut down on archetype moves: ```rust // Old (0.8) commands.spawn() .insert_bundle(SomeBundle::default()) .insert(SomeComponent); // New (0.9) - Option 1 commands.spawn().insert(( SomeBundle::default(), SomeComponent, )) // New (0.9) - Option 2 commands.spawn_bundle(( SomeBundle::default(), SomeComponent, )) ``` ## Next Steps Consider changing `spawn` to accept a bundle and deprecate `spawn_bundle`.
2022-09-21 21:47:53 +00:00
commands.entity(entity).insert((
ViewShadowBindings {
point_light_depth_texture: point_light_depth_texture.texture,
point_light_depth_texture_view,
directional_light_depth_texture: directional_light_depth_texture.texture,
directional_light_depth_texture_view,
},
ViewLightEntities {
lights: view_lights,
},
ViewLightsUniformOffset {
offset: light_meta.view_gpu_lights.push(gpu_lights),
},
));
2021-06-02 02:59:17 +00:00
}
Sprite Batching (#3060) This implements the following: * **Sprite Batching**: Collects sprites in a vertex buffer to draw many sprites with a single draw call. Sprites are batched by their `Handle<Image>` within a specific z-level. When possible, sprites are opportunistically batched _across_ z-levels (when no sprites with a different texture exist between two sprites with the same texture on different z levels). With these changes, I can now get ~130,000 sprites at 60fps on the `bevymark_pipelined` example. * **Sprite Color Tints**: The `Sprite` type now has a `color` field. Non-white color tints result in a specialized render pipeline that passes the color in as a vertex attribute. I chose to specialize this because passing vertex colors has a measurable price (without colors I get ~130,000 sprites on bevymark, with colors I get ~100,000 sprites). "Colored" sprites cannot be batched with "uncolored" sprites, but I think this is fine because the chance of a "colored" sprite needing to batch with other "colored" sprites is generally probably way higher than an "uncolored" sprite needing to batch with a "colored" sprite. * **Sprite Flipping**: Sprites can be flipped on their x or y axis using `Sprite::flip_x` and `Sprite::flip_y`. This is also true for `TextureAtlasSprite`. * **Simpler BufferVec/UniformVec/DynamicUniformVec Clearing**: improved the clearing interface by removing the need to know the size of the final buffer at the initial clear. ![image](https://user-images.githubusercontent.com/2694663/140001821-99be0d96-025d-489e-9bfa-ba19c1dc9548.png) Note that this moves sprites away from entity-driven rendering and back to extracted lists. We _could_ use entities here, but it necessitates that an intermediate list is allocated / populated to collect and sort extracted sprites. This redundant copy, combined with the normal overhead of spawning extracted sprite entities, brings bevymark down to ~80,000 sprites at 60fps. I think making sprites a bit more fixed (by default) is worth it. I view this as acceptable because batching makes normal entity-driven rendering pretty useless anyway (and we would want to batch most custom materials too). We can still support custom shaders with custom bindings, we'll just need to define a specific interface for it.
2021-11-04 20:28:53 +00:00
light_meta
.view_gpu_lights
.write_buffer(&render_device, &render_queue);
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
}
// this must match CLUSTER_COUNT_SIZE in pbr.wgsl
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
// and must be large enough to contain MAX_UNIFORM_BUFFER_POINT_LIGHTS
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
const CLUSTER_COUNT_SIZE: u32 = 9;
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
const CLUSTER_OFFSET_MASK: u32 = (1 << (32 - (CLUSTER_COUNT_SIZE * 2))) - 1;
const CLUSTER_COUNT_MASK: u32 = (1 << CLUSTER_COUNT_SIZE) - 1;
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
// NOTE: With uniform buffer max binding size as 16384 bytes
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// that means we can fit 256 point lights in one uniform
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
// buffer, which means the count can be at most 256 so it
// needs 9 bits.
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
// The array of indices can also use u8 and that means the
// offset in to the array of indices needs to be able to address
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
// 16384 values. log2(16384) = 14 bits.
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
// We use 32 bits to store the offset and counts so
// we pack the offset into the upper 14 bits of a u32,
// the point light count into bits 9-17, and the spot light count into bits 0-8.
// [ 31 .. 18 | 17 .. 9 | 8 .. 0 ]
// [ offset | point light count | spot light count ]
bevy_pbr2: Fix clustering for orthographic projections (#3316) # Objective PBR lighting was broken in the new renderer when using orthographic projections due to the way the depth slicing works for the clusters. Fix it. ## Solution - The default orthographic projection near plane is 0.0. The perspective projection depth slicing does a division by the near plane which gives a floating point NaN and the clustering all breaks down. - Orthographic projections have a linear depth mapping, so it made intuitive sense to me to do depth slicing with a linear mapping too. The alternative I saw was to try to handle the near plane being at 0.0 and using the exponential depth slicing, but that felt like a hack that didn't make sense. - As such, I have added code that detects whether the projection is orthographic based on `projection[3][3] == 1.0` and then implemented the orthographic mapping case throughout (when computing cluster AABBs, and when mapping a view space position (or light) to a cluster id in both the rust and shader code). ## Screenshots Before: ![before](https://user-images.githubusercontent.com/302146/145847278-5b1bca74-fbad-4cc5-8b49-384f6a377fdc.png) After: <img width="1392" alt="Screenshot 2021-12-13 at 16 36 53" src="https://user-images.githubusercontent.com/302146/145847314-6f3a2035-5d87-4896-8032-0c3e35e15b7d.png"> Old renderer (slightly lighter due to slight difference in configured intensity): <img width="1392" alt="Screenshot 2021-12-13 at 16 42 23" src="https://user-images.githubusercontent.com/302146/145847391-6a5e6fe0-22da-4fc1-a6c7-440543689a63.png">
2021-12-14 23:42:35 +00:00
// NOTE: This assumes CPU and GPU endianness are the same which is true
// for all common and tested x86/ARM CPUs and AMD/NVIDIA/Intel/Apple/etc GPUs
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
fn pack_offset_and_counts(offset: usize, point_count: usize, spot_count: usize) -> u32 {
((offset as u32 & CLUSTER_OFFSET_MASK) << (CLUSTER_COUNT_SIZE * 2))
| (point_count as u32 & CLUSTER_COUNT_MASK) << CLUSTER_COUNT_SIZE
| (spot_count as u32 & CLUSTER_COUNT_MASK)
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
#[derive(ShaderType)]
struct GpuClusterLightIndexListsUniform {
data: Box<[UVec4; ViewClusterBindings::MAX_UNIFORM_ITEMS]>,
}
// NOTE: Assert at compile time that GpuClusterLightIndexListsUniform
// fits within the maximum uniform buffer binding size
const _: () = assert!(GpuClusterLightIndexListsUniform::SHADER_SIZE.get() <= 16384);
impl Default for GpuClusterLightIndexListsUniform {
fn default() -> Self {
Self {
data: Box::new([UVec4::ZERO; ViewClusterBindings::MAX_UNIFORM_ITEMS]),
}
}
}
#[derive(ShaderType)]
struct GpuClusterOffsetsAndCountsUniform {
data: Box<[UVec4; ViewClusterBindings::MAX_UNIFORM_ITEMS]>,
}
impl Default for GpuClusterOffsetsAndCountsUniform {
fn default() -> Self {
Self {
data: Box::new([UVec4::ZERO; ViewClusterBindings::MAX_UNIFORM_ITEMS]),
}
}
}
#[derive(ShaderType, Default)]
struct GpuClusterLightIndexListsStorage {
#[size(runtime)]
data: Vec<u32>,
}
#[derive(ShaderType, Default)]
struct GpuClusterOffsetsAndCountsStorage {
#[size(runtime)]
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
data: Vec<UVec4>,
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
enum ViewClusterBuffers {
Uniform {
// NOTE: UVec4 is because all arrays in Std140 layout have 16-byte alignment
cluster_light_index_lists: UniformBuffer<GpuClusterLightIndexListsUniform>,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
// NOTE: UVec4 is because all arrays in Std140 layout have 16-byte alignment
cluster_offsets_and_counts: UniformBuffer<GpuClusterOffsetsAndCountsUniform>,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
},
Storage {
cluster_light_index_lists: StorageBuffer<GpuClusterLightIndexListsStorage>,
cluster_offsets_and_counts: StorageBuffer<GpuClusterOffsetsAndCountsStorage>,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
},
}
impl ViewClusterBuffers {
fn new(buffer_binding_type: BufferBindingType) -> Self {
match buffer_binding_type {
BufferBindingType::Storage { .. } => Self::storage(),
BufferBindingType::Uniform => Self::uniform(),
}
}
fn uniform() -> Self {
ViewClusterBuffers::Uniform {
cluster_light_index_lists: UniformBuffer::default(),
cluster_offsets_and_counts: UniformBuffer::default(),
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
fn storage() -> Self {
ViewClusterBuffers::Storage {
cluster_light_index_lists: StorageBuffer::default(),
cluster_offsets_and_counts: StorageBuffer::default(),
}
}
}
#[derive(Component)]
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
pub struct ViewClusterBindings {
n_indices: usize,
n_offsets: usize,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
buffers: ViewClusterBuffers,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
impl ViewClusterBindings {
pub const MAX_OFFSETS: usize = 16384 / 4;
const MAX_UNIFORM_ITEMS: usize = Self::MAX_OFFSETS / 4;
Dynamic light clusters (#3968) # Objective provide some customisation for default cluster setup avoid "cluster index lists is full" in all cases (using a strategy outlined by @superdump) ## Solution Add ClusterConfig enum (which can be inserted into a view at any time) to allow specifying cluster setup with variants: - None (do not do any light assignment - for views which do not require light info, e.g. minimaps etc) - Single (one cluster) - XYZ (explicit cluster counts in each dimension) - FixedZ (most similar to current - specify Z-slices and total, then x and y counts are dynamically determined to give approximately square clusters based on current aspect ratio) Defaults to FixedZ { total: 4096, z: 24 } which is similar to the current setup. Per frame, estimate the number of indices that would be required for the current config and decrease the cluster counts / increase the cluster sizes in the x and y dimensions if the index list would be too small. notes: - I didn't put ClusterConfig in the camera bundles to avoid introducing a dependency from bevy_render to bevy_pbr. the ClusterConfig enum comes with a pbr-centric impl block so i didn't want to move that into bevy_render either. - ~Might want to add None variant to cluster config for views that don't care about lights?~ - Not well tested for orthographic - ~there's a cluster_muck branch on my repo which includes some diagnostics / a modified lighting example which may be useful for tyre-kicking~ (outdated, i will bring it up to date if required) anecdotal timings: FPS on the lighting demo is negligibly better (~5%), maybe due to a small optimisation constraining the light aabb to be in front of the camera FPS on the lighting demo with 100 extra lights added is ~33% faster, and also renders correctly as the cluster index count is no longer exceeded
2022-03-08 04:56:42 +00:00
pub const MAX_INDICES: usize = 16384;
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
pub fn new(buffer_binding_type: BufferBindingType) -> Self {
Self {
n_indices: 0,
n_offsets: 0,
buffers: ViewClusterBuffers::new(buffer_binding_type),
}
}
pub fn clear(&mut self) {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
match &mut self.buffers {
ViewClusterBuffers::Uniform {
cluster_light_index_lists,
cluster_offsets_and_counts,
} => {
*cluster_light_index_lists.get_mut().data = [UVec4::ZERO; Self::MAX_UNIFORM_ITEMS];
*cluster_offsets_and_counts.get_mut().data = [UVec4::ZERO; Self::MAX_UNIFORM_ITEMS];
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
ViewClusterBuffers::Storage {
cluster_light_index_lists,
cluster_offsets_and_counts,
..
} => {
cluster_light_index_lists.get_mut().data.clear();
cluster_offsets_and_counts.get_mut().data.clear();
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
pub fn push_offset_and_counts(&mut self, offset: usize, point_count: usize, spot_count: usize) {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
match &mut self.buffers {
ViewClusterBuffers::Uniform {
cluster_offsets_and_counts,
..
} => {
let array_index = self.n_offsets >> 2; // >> 2 is equivalent to / 4
if array_index >= Self::MAX_UNIFORM_ITEMS {
warn!("cluster offset and count out of bounds!");
return;
}
let component = self.n_offsets & ((1 << 2) - 1);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let packed = pack_offset_and_counts(offset, point_count, spot_count);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
cluster_offsets_and_counts.get_mut().data[array_index][component] = packed;
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
ViewClusterBuffers::Storage {
cluster_offsets_and_counts,
..
} => {
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
cluster_offsets_and_counts.get_mut().data.push(UVec4::new(
offset as u32,
point_count as u32,
spot_count as u32,
0,
));
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
self.n_offsets += 1;
}
pub fn n_indices(&self) -> usize {
self.n_indices
}
pub fn push_index(&mut self, index: usize) {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
match &mut self.buffers {
ViewClusterBuffers::Uniform {
cluster_light_index_lists,
..
} => {
let array_index = self.n_indices >> 4; // >> 4 is equivalent to / 16
let component = (self.n_indices >> 2) & ((1 << 2) - 1);
let sub_index = self.n_indices & ((1 << 2) - 1);
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
let index = index as u32;
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
cluster_light_index_lists.get_mut().data[array_index][component] |=
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
index << (8 * sub_index);
}
ViewClusterBuffers::Storage {
cluster_light_index_lists,
..
} => {
cluster_light_index_lists.get_mut().data.push(index as u32);
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
self.n_indices += 1;
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
pub fn write_buffers(&mut self, render_device: &RenderDevice, render_queue: &RenderQueue) {
match &mut self.buffers {
ViewClusterBuffers::Uniform {
cluster_light_index_lists,
cluster_offsets_and_counts,
} => {
cluster_light_index_lists.write_buffer(render_device, render_queue);
cluster_offsets_and_counts.write_buffer(render_device, render_queue);
}
ViewClusterBuffers::Storage {
cluster_light_index_lists,
cluster_offsets_and_counts,
} => {
cluster_light_index_lists.write_buffer(render_device, render_queue);
cluster_offsets_and_counts.write_buffer(render_device, render_queue);
}
}
}
pub fn light_index_lists_binding(&self) -> Option<BindingResource> {
match &self.buffers {
ViewClusterBuffers::Uniform {
cluster_light_index_lists,
..
} => cluster_light_index_lists.binding(),
ViewClusterBuffers::Storage {
cluster_light_index_lists,
..
} => cluster_light_index_lists.binding(),
}
}
pub fn offsets_and_counts_binding(&self) -> Option<BindingResource> {
match &self.buffers {
ViewClusterBuffers::Uniform {
cluster_offsets_and_counts,
..
} => cluster_offsets_and_counts.binding(),
ViewClusterBuffers::Storage {
cluster_offsets_and_counts,
..
} => cluster_offsets_and_counts.binding(),
}
}
pub fn min_size_cluster_light_index_lists(
buffer_binding_type: BufferBindingType,
) -> NonZeroU64 {
match buffer_binding_type {
BufferBindingType::Storage { .. } => GpuClusterLightIndexListsStorage::min_size(),
BufferBindingType::Uniform => GpuClusterLightIndexListsUniform::min_size(),
}
}
pub fn min_size_cluster_offsets_and_counts(
buffer_binding_type: BufferBindingType,
) -> NonZeroU64 {
match buffer_binding_type {
BufferBindingType::Storage { .. } => GpuClusterOffsetsAndCountsStorage::min_size(),
BufferBindingType::Uniform => GpuClusterOffsetsAndCountsUniform::min_size(),
}
}
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
}
pub fn prepare_clusters(
mut commands: Commands,
render_device: Res<RenderDevice>,
render_queue: Res<RenderQueue>,
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
mesh_pipeline: Res<MeshPipeline>,
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
global_light_meta: Res<GlobalLightMeta>,
views: Query<
(
Entity,
&ExtractedClusterConfig,
&ExtractedClustersPointLights,
),
With<RenderPhase<Transparent3d>>,
>,
) {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
let render_device = render_device.into_inner();
let supports_storage_buffers = matches!(
mesh_pipeline.clustered_forward_buffer_binding_type,
BufferBindingType::Storage { .. }
);
for (entity, cluster_config, extracted_clusters) in &views {
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
let mut view_clusters_bindings =
ViewClusterBindings::new(mesh_pipeline.clustered_forward_buffer_binding_type);
view_clusters_bindings.clear();
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
let mut indices_full = false;
let mut cluster_index = 0;
for _y in 0..cluster_config.dimensions.y {
for _x in 0..cluster_config.dimensions.x {
for _z in 0..cluster_config.dimensions.z {
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
let offset = view_clusters_bindings.n_indices();
let cluster_lights = &extracted_clusters.data[cluster_index];
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
view_clusters_bindings.push_offset_and_counts(
offset,
cluster_lights.point_light_count,
cluster_lights.spot_light_count,
);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
if !indices_full {
for entity in cluster_lights.iter() {
if let Some(light_index) = global_light_meta.entity_to_index.get(entity)
{
if view_clusters_bindings.n_indices()
>= ViewClusterBindings::MAX_INDICES
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
&& !supports_storage_buffers
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
{
warn!("Cluster light index lists is full! The PointLights in the view are affecting too many clusters.");
indices_full = true;
break;
}
view_clusters_bindings.push_index(*light_index);
}
}
}
cluster_index += 1;
}
}
}
Use storage buffers for clustered forward point lights (#3989) # Objective - Make use of storage buffers, where they are available, for clustered forward bindings to support far more point lights in a scene - Fixes #3605 - Based on top of #4079 This branch on an M1 Max can keep 60fps with about 2150 point lights of radius 1m in the Sponza scene where I've been testing. The bottleneck is mostly assigning lights to clusters which grows faster than linearly (I think 1000 lights was about 1.5ms and 5000 was 7.5ms). I have seen papers and presentations leveraging compute shaders that can get this up to over 1 million. That said, I think any further optimisations should probably be done in a separate PR. ## Solution - Add `RenderDevice` to the `Material` and `SpecializedMaterial` trait `::key()` functions to allow setting flags on the keys depending on feature/limit availability - Make `GpuPointLights` and `ViewClusterBuffers` into enums containing `UniformVec` and `StorageBuffer` variants. Implement the necessary API on them to make usage the same for both cases, and the only difference is at initialisation time. - Appropriate shader defs in the shader code to handle the two cases ## Context on some decisions / open questions - I'm using `max_storage_buffers_per_shader_stage >= 3` as a check to see if storage buffers are supported. I was thinking about diving into 'binding resource management' but it feels like we don't have enough use cases to understand the problem yet, and it is mostly a separate concern to this PR, so I think it should be handled separately. - Should `ViewClusterBuffers` and `ViewClusterBindings` be merged, duplicating the count variables into the enum variants? Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-04-07 16:16:35 +00:00
view_clusters_bindings.write_buffers(render_device, &render_queue);
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
commands.get_or_spawn(entity).insert(view_clusters_bindings);
}
}
#[allow(clippy::too_many_arguments)]
pub fn queue_shadows<M: Material>(
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
shadow_draw_functions: Res<DrawFunctions<Shadow>>,
prepass_pipeline: Res<PrepassPipeline<M>>,
Reorder render sets, refactor bevy_sprite to take advantage (#9236) This is a continuation of this PR: #8062 # Objective - Reorder render schedule sets to allow data preparation when phase item order is known to support improved batching - Part of the batching/instancing etc plan from here: https://github.com/bevyengine/bevy/issues/89#issuecomment-1379249074 - The original idea came from @inodentry and proved to be a good one. Thanks! - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the new ordering ## Solution - Move `Prepare` and `PrepareFlush` after `PhaseSortFlush` - Add a `PrepareAssets` set that runs in parallel with other systems and sets in the render schedule. - Put prepare_assets systems in the `PrepareAssets` set - If explicit dependencies are needed on Mesh or Material RenderAssets then depend on the appropriate system. - Add `ManageViews` and `ManageViewsFlush` sets between `ExtractCommands` and Queue - Move `queue_mesh*_bind_group` to the Prepare stage - Rename them to `prepare_` - Put systems that prepare resources (buffers, textures, etc.) into a `PrepareResources` set inside `Prepare` - Put the `prepare_..._bind_group` systems into a `PrepareBindGroup` set after `PrepareResources` - Move `prepare_lights` to the `ManageViews` set - `prepare_lights` creates views and this must happen before `Queue` - This system needs refactoring to stop handling all responsibilities - Gather lights, sort, and create shadow map views. Store sorted light entities in a resource - Remove `BatchedPhaseItem` - Replace `batch_range` with `batch_size` representing how many items to skip after rendering the item or to skip the item entirely if `batch_size` is 0. - `queue_sprites` has been split into `queue_sprites` for queueing phase items and `prepare_sprites` for batching after the `PhaseSort` - `PhaseItem`s are still inserted in `queue_sprites` - After sorting adjacent compatible sprite phase items are accumulated into `SpriteBatch` components on the first entity of each batch, containing a range of vertex indices. The associated `PhaseItem`'s `batch_size` is updated appropriately. - `SpriteBatch` items are then drawn skipping over the other items in the batch based on the value in `batch_size` - A very similar refactor was performed on `bevy_ui` --- ## Changelog Changed: - Reordered and reworked render app schedule sets. The main change is that data is extracted, queued, sorted, and then prepared when the order of data is known. - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the reordering. ## Migration Guide - Assets such as materials and meshes should now be created in `PrepareAssets` e.g. `prepare_assets<Mesh>` - Queueing entities to `RenderPhase`s continues to be done in `Queue` e.g. `queue_sprites` - Preparing resources (textures, buffers, etc.) should now be done in `PrepareResources`, e.g. `prepare_prepass_textures`, `prepare_mesh_uniforms` - Prepare bind groups should now be done in `PrepareBindGroups` e.g. `prepare_mesh_bind_group` - Any batching or instancing can now be done in `Prepare` where the order of the phase items is known e.g. `prepare_sprites` ## Next Steps - Introduce some generic mechanism to ensure items that can be batched are grouped in the phase item order, currently you could easily have `[sprite at z 0, mesh at z 0, sprite at z 0]` preventing batching. - Investigate improved orderings for building the MeshUniform buffer - Implementing batching across the rest of bevy --------- Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
2023-08-27 14:33:49 +00:00
casting_meshes: Query<(&Handle<Mesh>, &Handle<M>), Without<NotShadowCaster>>,
render_meshes: Res<RenderAssets<Mesh>>,
render_materials: Res<RenderMaterials<M>>,
mut pipelines: ResMut<SpecializedMeshPipelines<PrepassPipeline<M>>>,
pipeline_cache: Res<PipelineCache>,
view_lights: Query<(Entity, &ViewLightEntities)>,
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
mut view_light_shadow_phases: Query<(&LightEntity, &mut RenderPhase<Shadow>)>,
point_light_entities: Query<&CubemapVisibleEntities, With<ExtractedPointLight>>,
directional_light_entities: Query<&CascadesVisibleEntities, With<ExtractedDirectionalLight>>,
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
spot_light_entities: Query<&VisibleEntities, With<ExtractedPointLight>>,
) where
M::Data: PartialEq + Eq + Hash + Clone,
{
for (entity, view_lights) in &view_lights {
let draw_shadow_mesh = shadow_draw_functions.read().id::<DrawPrepass<M>>();
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
for view_light_entity in view_lights.lights.iter().copied() {
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
let (light_entity, mut shadow_phase) =
view_light_shadow_phases.get_mut(view_light_entity).unwrap();
let is_directional_light = matches!(light_entity, LightEntity::Directional { .. });
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
let visible_entities = match light_entity {
LightEntity::Directional {
light_entity,
cascade_index,
} => directional_light_entities
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
.get(*light_entity)
.expect("Failed to get directional light visible entities")
.entities
.get(&entity)
.expect("Failed to get directional light visible entities for view")
.get(*cascade_index)
.expect("Failed to get directional light visible entities for cascade"),
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
LightEntity::Point {
light_entity,
face_index,
} => point_light_entities
.get(*light_entity)
.expect("Failed to get point light visible entities")
.get(*face_index),
Spotlights (#4715) # Objective add spotlight support ## Solution / Changelog - add spotlight angles (inner, outer) to ``PointLight`` struct. emitted light is linearly attenuated from 100% to 0% as angle tends from inner to outer. Direction is taken from the existing transform rotation. - add spotlight direction (vec3) and angles (f32,f32) to ``GpuPointLight`` struct (60 bytes -> 80 bytes) in ``pbr/render/lights.rs`` and ``mesh_view_bind_group.wgsl`` - reduce no-buffer-support max point light count to 204 due to above - use spotlight data to attenuate light in ``pbr.wgsl`` - do additional cluster culling on spotlights to minimise cost in ``assign_lights_to_clusters`` - changed one of the lights in the lighting demo to a spotlight - also added a ``spotlight`` demo - probably not justified but so reviewers can see it more easily ## notes increasing the size of the GpuPointLight struct on my machine reduces the FPS of ``many_lights -- sphere`` from ~150fps to 140fps. i thought this was a reasonable tradeoff, and felt better than handling spotlights separately which is possible but would mean introducing a new bind group, refactoring light-assignment code and adding new spotlight-specific code in pbr.wgsl. the FPS impact for smaller numbers of lights should be very small. the cluster culling strategy reintroduces the cluster aabb code which was recently removed... sorry. the aabb is used to get a cluster bounding sphere, which can then be tested fairly efficiently using the strategy described at the end of https://bartwronski.com/2017/04/13/cull-that-cone/. this works well with roughly cubic clusters (where the cluster z size is close to the same as x/y size), less well for other cases like single Z slice / tiled forward rendering. In the worst case we will end up just keeping the culling of the equivalent point light. Co-authored-by: François <mockersf@gmail.com>
2022-07-08 19:57:43 +00:00
LightEntity::Spot { light_entity } => spot_light_entities
.get(*light_entity)
.expect("Failed to get spot light visible entities"),
Frustum culling (#2861) # Objective Implement frustum culling for much better performance on more complex scenes. With the Amazon Lumberyard Bistro scene, I was getting roughly 15fps without frustum culling and 60+fps with frustum culling on a MacBook Pro 16 with i9 9980HK 8c/16t CPU and Radeon Pro 5500M. macOS does weird things with vsync so even though vsync was off, it really looked like sometimes other applications or the desktop window compositor were interfering, but the difference could be even more as I even saw up to 90+fps sometimes. ## Solution - Until the https://github.com/bevyengine/rfcs/pull/12 RFC is completed, I wanted to implement at least some of the bounding volume functionality we needed to be able to unblock a bunch of rendering features and optimisations such as frustum culling, fitting the directional light orthographic projection to the relevant meshes in the view, clustered forward rendering, etc. - I have added `Aabb`, `Frustum`, and `Sphere` types with only the necessary intersection tests for the algorithms used. I also added `CubemapFrusta` which contains a `[Frustum; 6]` and can be used by cube maps such as environment maps, and point light shadow maps. - I did do a bit of benchmarking and optimisation on the intersection tests. I compared the [rafx parallel-comparison bitmask approach](https://github.com/aclysma/rafx/blob/c91bd5fcfdfa3f4d1b43507c32d84b94ffdf1b2e/rafx-visibility/src/geometry/frustum.rs#L64-L92) with a naïve loop that has an early-out in case of a bounding volume being outside of any one of the `Frustum` planes and found them to be very similar, so I chose the simpler and more readable option. I also compared using Vec3 and Vec3A and it turned out that promoting Vec3s to Vec3A improved performance of the culling significantly due to Vec3A operations using SIMD optimisations where Vec3 uses plain scalar operations. - When loading glTF models, the vertex attribute accessors generally store the minimum and maximum values, which allows for adding AABBs to meshes loaded from glTF for free. - For meshes without an AABB (`PbrBundle` deliberately does not have an AABB by default), a system is executed that scans over the vertex positions to find the minimum and maximum values along each axis. This is used to construct the AABB. - The `Frustum::intersects_obb` and `Sphere::insersects_obb` algorithm is from Foundations of Game Engine Development 2: Rendering by Eric Lengyel. There is no OBB type, yet, rather an AABB and the model matrix are passed in as arguments. This calculates a 'relative radius' of the AABB with respect to the plane normal (the plane normal in the Sphere case being something I came up with as the direction pointing from the centre of the sphere to the centre of the AABB) such that it can then do a sphere-sphere intersection test in practice. - `RenderLayers` were copied over from the current renderer. - `VisibleEntities` was copied over from the current renderer and a `CubemapVisibleEntities` was added to support `PointLight`s for now. `VisibleEntities` are added to views (cameras and lights) and contain a `Vec<Entity>` that is populated by culling/visibility systems that run in PostUpdate of the app world, and are iterated over in the render world for, for example, queuing up meshes to be drawn by lights for shadow maps and the main pass for cameras. - `Visibility` and `ComputedVisibility` components were added. The `Visibility` component is user-facing so that, for example, the entity can be marked as not visible in an editor. `ComputedVisibility` on the other hand is the result of the culling/visibility systems and takes `Visibility` into account. So if an entity is marked as not being visible in its `Visibility` component, that will skip culling/visibility intersection tests and just mark the `ComputedVisibility` as false. - The `ComputedVisibility` is used to decide which meshes to extract. - I had to add a way to get the far plane from the `CameraProjection` in order to define an explicit far frustum plane for culling. This should perhaps be optional as it is not always desired and in that case, testing 5 planes instead of 6 is a performance win. I think that's about all. I discussed some of the design with @cart on Discord already so hopefully it's not too far from being mergeable. It works well at least. 😄
2021-11-07 21:45:52 +00:00
};
// NOTE: Lights with shadow mapping disabled will have no visible entities
Clustered forward rendering (#3153) # Objective Implement clustered-forward rendering. ## Solution ~~FIXME - in the interest of keeping the merge train moving, I'm submitting this PR now before the description is ready. I want to add in some comments into the code with references for the various bits and pieces and I want to describe some of the key decisions I made here. I'll do that as soon as I can.~~ Anyone reviewing is welcome to add review comments where you want to know more about how something or other works. * The summary of the technique is that the view frustum is divided into a grid of sub-volumes called clusters, point lights are tested against each of the clusters to see if they would affect that volume within the scene and if so, added to a list of lights affecting that cluster. Then when shading a fragment which is a point on the surface of a mesh within the scene, the point is mapped to a cluster and only the lights affecting that clusters are used in lighting calculations. This brings huge performance and scalability benefits as most of the time lights are placed so that there are not that many that overlap each other in terms of their sphere of influence, but there may be many distinct point lights visible in the scene. Doing all the lighting calculations for all visible lights in the scene for every pixel on the screen quickly becomes a performance limitation. Clustered forward rendering allows us to make an approximate list of lights that affect each pixel, indeed each surface in the scene (as it works along the view z axis too, unlike tiled/forward+). * WebGL2 is a platform we want to support and it does not support storage buffers. Uniform buffer bindings are limited to a maximum of 16384 bytes per binding. I used bit shifting and masking to pack the cluster light lists and various indices into a uniform buffer and the 16kB limit is very likely the first bottleneck in scaling the number of lights in a scene at the moment if the lights can affect many clusters due to their range or proximity to the camera (there are a lot of clusters close to the camera, which is an area for improvement). We could store the information in textures instead of uniform buffers to remove this bottleneck though I don’t know if there are performance implications to reading from textures instead if uniform buffers. * Because of the uniform buffer binding size limitations we can support a maximum of 256 lights with the current size of the PointLight struct * The z-slicing method (i.e. the mapping from view space z to a depth slice which defines the near and far planes of a cluster) is using the Doom 2016 method. I need to add comments with references to this. It’s an exponential function that simplifies well for the purposes of optimising the fragment shader. xy grid divisions are regular in screen space. * Some optimisation work was done on the allocation of lights to clusters, which involves intersection tests, and for this number of clusters and lights the system has insignificant cost using a fairly naïve algorithm. I think for more lights / finer-grained clusters we could use a BVH, but at some point it would be just much better to use compute shaders and storage buffers. * Something else to note is that it is absolutely infeasible to use plain cube map point light shadow mapping for many lights. It does not scale in terms of performance nor memory usage. There are some interesting methods I saw discussed in reference material that I will add a link to which render and update shadow maps piece-wise, but they also need compute shaders to work well. Basically for now you need to sacrifice point light shadows for all but a handful of point lights if you don’t want to kill performance. I set the limit to 10 but that’s just what we had from before where 10 was the maximum number of point lights before this PR. * I added a couple of debug visualisations behind a shader def that were useful for seeing performance impact of light distribution - I should make the debug mode configurable without modifying the shader code. One mode shows the number of lights affecting each cluster by tinting toward red for few lights or green for many lights (maxes out at 16, but not sure that’s a reasonable max). The other shows which cluster the surface at a fragment belongs to by tinting it with a randomish colour. This can help to understand deeper performance issues due to screen space tiles spanning multiple clusters in depth with divergent shader execution times. Also, there are more things that could be done as improvements, and I will document those somewhere (I'm not sure where will be the best place... in a todo alongside the code, a GitHub issue, somewhere else?) but I think it works well enough and brings significant performance and scalability benefits that it's worth integrating already now and then iterating on. * Calculate the light’s effective range based on its intensity and physical falloff and either just use this, or take the minimum of the user-supplied range and this. This would avoid unnecessary lighting calculations for clusters that cannot be affected. This would need to take into account HDR tone mapping as in my not-fully-understanding-the-details understanding, the threshold is relative to how bright the scene is. * Improve the z-slicing to use a larger first slice. * More gracefully handle the cluster light list uniform buffer binding size limitations by prioritising which lights are included (some heuristic for most significant like closest to the camera, brightest, affecting the most pixels, …) * Switch to using a texture instead of uniform buffer * Figure out the / a better story for shadows I will also probably add an example that demonstrates some of the issues: * What situations exhaust the space available in the uniform buffers * Light range too large making lights affect many clusters and so exhausting the space for the lists of lights that affect clusters * Light range set to be too small producing visible artifacts where clusters the light would physically affect are not affected by the light * Perhaps some performance issues * How many lights can be closely packed or affect large portions of the view before performance drops?
2021-12-09 03:08:54 +00:00
// so no meshes will be queued
for entity in visible_entities.iter().copied() {
Reorder render sets, refactor bevy_sprite to take advantage (#9236) This is a continuation of this PR: #8062 # Objective - Reorder render schedule sets to allow data preparation when phase item order is known to support improved batching - Part of the batching/instancing etc plan from here: https://github.com/bevyengine/bevy/issues/89#issuecomment-1379249074 - The original idea came from @inodentry and proved to be a good one. Thanks! - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the new ordering ## Solution - Move `Prepare` and `PrepareFlush` after `PhaseSortFlush` - Add a `PrepareAssets` set that runs in parallel with other systems and sets in the render schedule. - Put prepare_assets systems in the `PrepareAssets` set - If explicit dependencies are needed on Mesh or Material RenderAssets then depend on the appropriate system. - Add `ManageViews` and `ManageViewsFlush` sets between `ExtractCommands` and Queue - Move `queue_mesh*_bind_group` to the Prepare stage - Rename them to `prepare_` - Put systems that prepare resources (buffers, textures, etc.) into a `PrepareResources` set inside `Prepare` - Put the `prepare_..._bind_group` systems into a `PrepareBindGroup` set after `PrepareResources` - Move `prepare_lights` to the `ManageViews` set - `prepare_lights` creates views and this must happen before `Queue` - This system needs refactoring to stop handling all responsibilities - Gather lights, sort, and create shadow map views. Store sorted light entities in a resource - Remove `BatchedPhaseItem` - Replace `batch_range` with `batch_size` representing how many items to skip after rendering the item or to skip the item entirely if `batch_size` is 0. - `queue_sprites` has been split into `queue_sprites` for queueing phase items and `prepare_sprites` for batching after the `PhaseSort` - `PhaseItem`s are still inserted in `queue_sprites` - After sorting adjacent compatible sprite phase items are accumulated into `SpriteBatch` components on the first entity of each batch, containing a range of vertex indices. The associated `PhaseItem`'s `batch_size` is updated appropriately. - `SpriteBatch` items are then drawn skipping over the other items in the batch based on the value in `batch_size` - A very similar refactor was performed on `bevy_ui` --- ## Changelog Changed: - Reordered and reworked render app schedule sets. The main change is that data is extracted, queued, sorted, and then prepared when the order of data is known. - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the reordering. ## Migration Guide - Assets such as materials and meshes should now be created in `PrepareAssets` e.g. `prepare_assets<Mesh>` - Queueing entities to `RenderPhase`s continues to be done in `Queue` e.g. `queue_sprites` - Preparing resources (textures, buffers, etc.) should now be done in `PrepareResources`, e.g. `prepare_prepass_textures`, `prepare_mesh_uniforms` - Prepare bind groups should now be done in `PrepareBindGroups` e.g. `prepare_mesh_bind_group` - Any batching or instancing can now be done in `Prepare` where the order of the phase items is known e.g. `prepare_sprites` ## Next Steps - Introduce some generic mechanism to ensure items that can be batched are grouped in the phase item order, currently you could easily have `[sprite at z 0, mesh at z 0, sprite at z 0]` preventing batching. - Investigate improved orderings for building the MeshUniform buffer - Implementing batching across the rest of bevy --------- Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
2023-08-27 14:33:49 +00:00
if let Ok((mesh_handle, material_handle)) = casting_meshes.get(entity) {
if let (Some(mesh), Some(material)) = (
render_meshes.get(mesh_handle),
Bevy Asset V2 (#8624) # Bevy Asset V2 Proposal ## Why Does Bevy Need A New Asset System? Asset pipelines are a central part of the gamedev process. Bevy's current asset system is missing a number of features that make it non-viable for many classes of gamedev. After plenty of discussions and [a long community feedback period](https://github.com/bevyengine/bevy/discussions/3972), we've identified a number missing features: * **Asset Preprocessing**: it should be possible to "preprocess" / "compile" / "crunch" assets at "development time" rather than when the game starts up. This enables offloading expensive work from deployed apps, faster asset loading, less runtime memory usage, etc. * **Per-Asset Loader Settings**: Individual assets cannot define their own loaders that override the defaults. Additionally, they cannot provide per-asset settings to their loaders. This is a huge limitation, as many asset types don't provide all information necessary for Bevy _inside_ the asset. For example, a raw PNG image says nothing about how it should be sampled (ex: linear vs nearest). * **Asset `.meta` files**: assets should have configuration files stored adjacent to the asset in question, which allows the user to configure asset-type-specific settings. These settings should be accessible during the pre-processing phase. Modifying a `.meta` file should trigger a re-processing / re-load of the asset. It should be possible to configure asset loaders from the meta file. * **Processed Asset Hot Reloading**: Changes to processed assets (or their dependencies) should result in re-processing them and re-loading the results in live Bevy Apps. * **Asset Dependency Tracking**: The current bevy_asset has no good way to wait for asset dependencies to load. It punts this as an exercise for consumers of the loader apis, which is unreasonable and error prone. There should be easy, ergonomic ways to wait for assets to load and block some logic on an asset's entire dependency tree loading. * **Runtime Asset Loading**: it should be (optionally) possible to load arbitrary assets dynamically at runtime. This necessitates being able to deploy and run the asset server alongside Bevy Apps on _all platforms_. For example, we should be able to invoke the shader compiler at runtime, stream scenes from sources like the internet, etc. To keep deployed binaries (and startup times) small, the runtime asset server configuration should be configurable with different settings compared to the "pre processor asset server". * **Multiple Backends**: It should be possible to load assets from arbitrary sources (filesystems, the internet, remote asset serves, etc). * **Asset Packing**: It should be possible to deploy assets in compressed "packs", which makes it easier and more efficient to distribute assets with Bevy Apps. * **Asset Handoff**: It should be possible to hold a "live" asset handle, which correlates to runtime data, without actually holding the asset in memory. Ex: it must be possible to hold a reference to a GPU mesh generated from a "mesh asset" without keeping the mesh data in CPU memory * **Per-Platform Processed Assets**: Different platforms and app distributions have different capabilities and requirements. Some platforms need lower asset resolutions or different asset formats to operate within the hardware constraints of the platform. It should be possible to define per-platform asset processing profiles. And it should be possible to deploy only the assets required for a given platform. These features have architectural implications that are significant enough to require a full rewrite. The current Bevy Asset implementation got us this far, but it can take us no farther. This PR defines a brand new asset system that implements most of these features, while laying the foundations for the remaining features to be built. ## Bevy Asset V2 Here is a quick overview of the features introduced in this PR. * **Asset Preprocessing**: Preprocess assets at development time into more efficient (and configurable) representations * **Dependency Aware**: Dependencies required to process an asset are tracked. If an asset's processed dependency changes, it will be reprocessed * **Hot Reprocessing/Reloading**: detect changes to asset source files, reprocess them if they have changed, and then hot-reload them in Bevy Apps. * **Only Process Changes**: Assets are only re-processed when their source file (or meta file) has changed. This uses hashing and timestamps to avoid processing assets that haven't changed. * **Transactional and Reliable**: Uses write-ahead logging (a technique commonly used by databases) to recover from crashes / forced-exits. Whenever possible it avoids full-reprocessing / only uncompleted transactions will be reprocessed. When the processor is running in parallel with a Bevy App, processor asset writes block Bevy App asset reads. Reading metadata + asset bytes is guaranteed to be transactional / correctly paired. * **Portable / Run anywhere / Database-free**: The processor does not rely on an in-memory database (although it uses some database techniques for reliability). This is important because pretty much all in-memory databases have unsupported platforms or build complications. * **Configure Processor Defaults Per File Type**: You can say "use this processor for all files of this type". * **Custom Processors**: The `Processor` trait is flexible and unopinionated. It can be implemented by downstream plugins. * **LoadAndSave Processors**: Most asset processing scenarios can be expressed as "run AssetLoader A, save the results using AssetSaver X, and then load the result using AssetLoader B". For example, load this png image using `PngImageLoader`, which produces an `Image` asset and then save it using `CompressedImageSaver` (which also produces an `Image` asset, but in a compressed format), which takes an `Image` asset as input. This means if you have an `AssetLoader` for an asset, you are already half way there! It also means that you can share AssetSavers across multiple loaders. Because `CompressedImageSaver` accepts Bevy's generic Image asset as input, it means you can also use it with some future `JpegImageLoader`. * **Loader and Saver Settings**: Asset Loaders and Savers can now define their own settings types, which are passed in as input when an asset is loaded / saved. Each asset can define its own settings. * **Asset `.meta` files**: configure asset loaders, their settings, enable/disable processing, and configure processor settings * **Runtime Asset Dependency Tracking** Runtime asset dependencies (ex: if an asset contains a `Handle<Image>`) are tracked by the asset server. An event is emitted when an asset and all of its dependencies have been loaded * **Unprocessed Asset Loading**: Assets do not require preprocessing. They can be loaded directly. A processed asset is just a "normal" asset with some extra metadata. Asset Loaders don't need to know or care about whether or not an asset was processed. * **Async Asset IO**: Asset readers/writers use async non-blocking interfaces. Note that because Rust doesn't yet support async traits, there is a bit of manual Boxing / Future boilerplate. This will hopefully be removed in the near future when Rust gets async traits. * **Pluggable Asset Readers and Writers**: Arbitrary asset source readers/writers are supported, both by the processor and the asset server. * **Better Asset Handles** * **Single Arc Tree**: Asset Handles now use a single arc tree that represents the lifetime of the asset. This makes their implementation simpler, more efficient, and allows us to cheaply attach metadata to handles. Ex: the AssetPath of a handle is now directly accessible on the handle itself! * **Const Typed Handles**: typed handles can be constructed in a const context. No more weird "const untyped converted to typed at runtime" patterns! * **Handles and Ids are Smaller / Faster To Hash / Compare**: Typed `Handle<T>` is now much smaller in memory and `AssetId<T>` is even smaller. * **Weak Handle Usage Reduction**: In general Handles are now considered to be "strong". Bevy features that previously used "weak `Handle<T>`" have been ported to `AssetId<T>`, which makes it statically clear that the features do not hold strong handles (while retaining strong type information). Currently Handle::Weak still exists, but it is very possible that we can remove that entirely. * **Efficient / Dense Asset Ids**: Assets now have efficient dense runtime asset ids, which means we can avoid expensive hash lookups. Assets are stored in Vecs instead of HashMaps. There are now typed and untyped ids, which means we no longer need to store dynamic type information in the ID for typed handles. "AssetPathId" (which was a nightmare from a performance and correctness standpoint) has been entirely removed in favor of dense ids (which are retrieved for a path on load) * **Direct Asset Loading, with Dependency Tracking**: Assets that are defined at runtime can still have their dependencies tracked by the Asset Server (ex: if you create a material at runtime, you can still wait for its textures to load). This is accomplished via the (currently optional) "asset dependency visitor" trait. This system can also be used to define a set of assets to load, then wait for those assets to load. * **Async folder loading**: Folder loading also uses this system and immediately returns a handle to the LoadedFolder asset, which means folder loading no longer blocks on directory traversals. * **Improved Loader Interface**: Loaders now have a specific "top level asset type", which makes returning the top-level asset simpler and statically typed. * **Basic Image Settings and Processing**: Image assets can now be processed into the gpu-friendly Basic Universal format. The ImageLoader now has a setting to define what format the image should be loaded as. Note that this is just a minimal MVP ... plenty of additional work to do here. To demo this, enable the `basis-universal` feature and turn on asset processing. * **Simpler Audio Play / AudioSink API**: Asset handle providers are cloneable, which means the Audio resource can mint its own handles. This means you can now do `let sink_handle = audio.play(music)` instead of `let sink_handle = audio_sinks.get_handle(audio.play(music))`. Note that this might still be replaced by https://github.com/bevyengine/bevy/pull/8424. **Removed Handle Casting From Engine Features**: Ex: FontAtlases no longer use casting between handle types ## Using The New Asset System ### Normal Unprocessed Asset Loading By default the `AssetPlugin` does not use processing. It behaves pretty much the same way as the old system. If you are defining a custom asset, first derive `Asset`: ```rust #[derive(Asset)] struct Thing { value: String, } ``` Initialize the asset: ```rust app.init_asset:<Thing>() ``` Implement a new `AssetLoader` for it: ```rust #[derive(Default)] struct ThingLoader; #[derive(Serialize, Deserialize, Default)] pub struct ThingSettings { some_setting: bool, } impl AssetLoader for ThingLoader { type Asset = Thing; type Settings = ThingSettings; fn load<'a>( &'a self, reader: &'a mut Reader, settings: &'a ThingSettings, load_context: &'a mut LoadContext, ) -> BoxedFuture<'a, Result<Thing, anyhow::Error>> { Box::pin(async move { let mut bytes = Vec::new(); reader.read_to_end(&mut bytes).await?; // convert bytes to value somehow Ok(Thing { value }) }) } fn extensions(&self) -> &[&str] { &["thing"] } } ``` Note that this interface will get much cleaner once Rust gets support for async traits. `Reader` is an async futures_io::AsyncRead. You can stream bytes as they come in or read them all into a `Vec<u8>`, depending on the context. You can use `let handle = load_context.load(path)` to kick off a dependency load, retrieve a handle, and register the dependency for the asset. Then just register the loader in your Bevy app: ```rust app.init_asset_loader::<ThingLoader>() ``` Now just add your `Thing` asset files into the `assets` folder and load them like this: ```rust fn system(asset_server: Res<AssetServer>) { let handle = Handle<Thing> = asset_server.load("cool.thing"); } ``` You can check load states directly via the asset server: ```rust if asset_server.load_state(&handle) == LoadState::Loaded { } ``` You can also listen for events: ```rust fn system(mut events: EventReader<AssetEvent<Thing>>, handle: Res<SomeThingHandle>) { for event in events.iter() { if event.is_loaded_with_dependencies(&handle) { } } } ``` Note the new `AssetEvent::LoadedWithDependencies`, which only fires when the asset is loaded _and_ all dependencies (and their dependencies) have loaded. Unlike the old asset system, for a given asset path all `Handle<T>` values point to the same underlying Arc. This means Handles can cheaply hold more asset information, such as the AssetPath: ```rust // prints the AssetPath of the handle info!("{:?}", handle.path()) ``` ### Processed Assets Asset processing can be enabled via the `AssetPlugin`. When developing Bevy Apps with processed assets, do this: ```rust app.add_plugins(DefaultPlugins.set(AssetPlugin::processed_dev())) ``` This runs the `AssetProcessor` in the background with hot-reloading. It reads assets from the `assets` folder, processes them, and writes them to the `.imported_assets` folder. Asset loads in the Bevy App will wait for a processed version of the asset to become available. If an asset in the `assets` folder changes, it will be reprocessed and hot-reloaded in the Bevy App. When deploying processed Bevy apps, do this: ```rust app.add_plugins(DefaultPlugins.set(AssetPlugin::processed())) ``` This does not run the `AssetProcessor` in the background. It behaves like `AssetPlugin::unprocessed()`, but reads assets from `.imported_assets`. When the `AssetProcessor` is running, it will populate sibling `.meta` files for assets in the `assets` folder. Meta files for assets that do not have a processor configured look like this: ```rust ( meta_format_version: "1.0", asset: Load( loader: "bevy_render::texture::image_loader::ImageLoader", settings: ( format: FromExtension, ), ), ) ``` This is metadata for an image asset. For example, if you have `assets/my_sprite.png`, this could be the metadata stored at `assets/my_sprite.png.meta`. Meta files are totally optional. If no metadata exists, the default settings will be used. In short, this file says "load this asset with the ImageLoader and use the file extension to determine the image type". This type of meta file is supported in all AssetPlugin modes. If in `Unprocessed` mode, the asset (with the meta settings) will be loaded directly. If in `ProcessedDev` mode, the asset file will be copied directly to the `.imported_assets` folder. The meta will also be copied directly to the `.imported_assets` folder, but with one addition: ```rust ( meta_format_version: "1.0", processed_info: Some(( hash: 12415480888597742505, full_hash: 14344495437905856884, process_dependencies: [], )), asset: Load( loader: "bevy_render::texture::image_loader::ImageLoader", settings: ( format: FromExtension, ), ), ) ``` `processed_info` contains `hash` (a direct hash of the asset and meta bytes), `full_hash` (a hash of `hash` and the hashes of all `process_dependencies`), and `process_dependencies` (the `path` and `full_hash` of every process_dependency). A "process dependency" is an asset dependency that is _directly_ used when processing the asset. Images do not have process dependencies, so this is empty. When the processor is enabled, you can use the `Process` metadata config: ```rust ( meta_format_version: "1.0", asset: Process( processor: "bevy_asset::processor::process::LoadAndSave<bevy_render::texture::image_loader::ImageLoader, bevy_render::texture::compressed_image_saver::CompressedImageSaver>", settings: ( loader_settings: ( format: FromExtension, ), saver_settings: ( generate_mipmaps: true, ), ), ), ) ``` This configures the asset to use the `LoadAndSave` processor, which runs an AssetLoader and feeds the result into an AssetSaver (which saves the given Asset and defines a loader to load it with). (for terseness LoadAndSave will likely get a shorter/friendlier type name when [Stable Type Paths](#7184) lands). `LoadAndSave` is likely to be the most common processor type, but arbitrary processors are supported. `CompressedImageSaver` saves an `Image` in the Basis Universal format and configures the ImageLoader to load it as basis universal. The `AssetProcessor` will read this meta, run it through the LoadAndSave processor, and write the basis-universal version of the image to `.imported_assets`. The final metadata will look like this: ```rust ( meta_format_version: "1.0", processed_info: Some(( hash: 905599590923828066, full_hash: 9948823010183819117, process_dependencies: [], )), asset: Load( loader: "bevy_render::texture::image_loader::ImageLoader", settings: ( format: Format(Basis), ), ), ) ``` To try basis-universal processing out in Bevy examples, (for example `sprite.rs`), change `add_plugins(DefaultPlugins)` to `add_plugins(DefaultPlugins.set(AssetPlugin::processed_dev()))` and run with the `basis-universal` feature enabled: `cargo run --features=basis-universal --example sprite`. To create a custom processor, there are two main paths: 1. Use the `LoadAndSave` processor with an existing `AssetLoader`. Implement the `AssetSaver` trait, register the processor using `asset_processor.register_processor::<LoadAndSave<ImageLoader, CompressedImageSaver>>(image_saver.into())`. 2. Implement the `Process` trait directly and register it using: `asset_processor.register_processor(thing_processor)`. You can configure default processors for file extensions like this: ```rust asset_processor.set_default_processor::<ThingProcessor>("thing") ``` There is one more metadata type to be aware of: ```rust ( meta_format_version: "1.0", asset: Ignore, ) ``` This will ignore the asset during processing / prevent it from being written to `.imported_assets`. The AssetProcessor stores a transaction log at `.imported_assets/log` and uses it to gracefully recover from unexpected stops. This means you can force-quit the processor (and Bevy Apps running the processor in parallel) at arbitrary times! `.imported_assets` is "local state". It should _not_ be checked into source control. It should also be considered "read only". In practice, you _can_ modify processed assets and processed metadata if you really need to test something. But those modifications will not be represented in the hashes of the assets, so the processed state will be "out of sync" with the source assets. The processor _will not_ fix this for you. Either revert the change after you have tested it, or delete the processed files so they can be re-populated. ## Open Questions There are a number of open questions to be discussed. We should decide if they need to be addressed in this PR and if so, how we will address them: ### Implied Dependencies vs Dependency Enumeration There are currently two ways to populate asset dependencies: * **Implied via AssetLoaders**: if an AssetLoader loads an asset (and retrieves a handle), a dependency is added to the list. * **Explicit via the optional Asset::visit_dependencies**: if `server.load_asset(my_asset)` is called, it will call `my_asset.visit_dependencies`, which will grab dependencies that have been manually defined for the asset via the Asset trait impl (which can be derived). This means that defining explicit dependencies is optional for "loaded assets". And the list of dependencies is always accurate because loaders can only produce Handles if they register dependencies. If an asset was loaded with an AssetLoader, it only uses the implied dependencies. If an asset was created at runtime and added with `asset_server.load_asset(MyAsset)`, it will use `Asset::visit_dependencies`. However this can create a behavior mismatch between loaded assets and equivalent "created at runtime" assets if `Assets::visit_dependencies` doesn't exactly match the dependencies produced by the AssetLoader. This behavior mismatch can be resolved by completely removing "implied loader dependencies" and requiring `Asset::visit_dependencies` to supply dependency data. But this creates two problems: * It makes defining loaded assets harder and more error prone: Devs must remember to manually annotate asset dependencies with `#[dependency]` when deriving `Asset`. For more complicated assets (such as scenes), the derive likely wouldn't be sufficient and a manual `visit_dependencies` impl would be required. * Removes the ability to immediately kick off dependency loads: When AssetLoaders retrieve a Handle, they also immediately kick off an asset load for the handle, which means it can start loading in parallel _before_ the asset finishes loading. For large assets, this could be significant. (although this could be mitigated for processed assets if we store dependencies in the processed meta file and load them ahead of time) ### Eager ProcessorDev Asset Loading I made a controversial call in the interest of fast startup times ("time to first pixel") for the "processor dev mode configuration". When initializing the AssetProcessor, current processed versions of unchanged assets are yielded immediately, even if their dependencies haven't been checked yet for reprocessing. This means that non-current-state-of-filesystem-but-previously-valid assets might be returned to the App first, then hot-reloaded if/when their dependencies change and the asset is reprocessed. Is this behavior desirable? There is largely one alternative: do not yield an asset from the processor to the app until all of its dependencies have been checked for changes. In some common cases (load dependency has not changed since last run) this will increase startup time. The main question is "by how much" and is that slower startup time worth it in the interest of only yielding assets that are true to the current state of the filesystem. Should this be configurable? I'm starting to think we should only yield an asset after its (historical) dependencies have been checked for changes + processed as necessary, but I'm curious what you all think. ### Paths Are Currently The Only Canonical ID / Do We Want Asset UUIDs? In this implementation AssetPaths are the only canonical asset identifier (just like the previous Bevy Asset system and Godot). Moving assets will result in re-scans (and currently reprocessing, although reprocessing can easily be avoided with some changes). Asset renames/moves will break code and assets that rely on specific paths, unless those paths are fixed up. Do we want / need "stable asset uuids"? Introducing them is very possible: 1. Generate a UUID and include it in .meta files 2. Support UUID in AssetPath 3. Generate "asset indices" which are loaded on startup and map UUIDs to paths. 4 (maybe). Consider only supporting UUIDs for processed assets so we can generate quick-to-load indices instead of scanning meta files. The main "pro" is that assets referencing UUIDs don't need to be migrated when a path changes. The main "con" is that UUIDs cannot be "lazily resolved" like paths. They need a full view of all assets to answer the question "does this UUID exist". Which means UUIDs require the AssetProcessor to fully finish startup scans before saying an asset doesnt exist. And they essentially require asset pre-processing to use in apps, because scanning all asset metadata files at runtime to resolve a UUID is not viable for medium-to-large apps. It really requires a pre-generated UUID index, which must be loaded before querying for assets. I personally think this should be investigated in a separate PR. Paths aren't going anywhere ... _everyone_ uses filesystems (and filesystem-like apis) to manage their asset source files. I consider them permanent canonical asset information. Additionally, they behave well for both processed and unprocessed asset modes. Given that Bevy is supporting both, this feels like the right canonical ID to start with. UUIDS (and maybe even other indexed-identifier types) can be added later as necessary. ### Folder / File Naming Conventions All asset processing config currently lives in the `.imported_assets` folder. The processor transaction log is in `.imported_assets/log`. Processed assets are added to `.imported_assets/Default`, which will make migrating to processed asset profiles (ex: a `.imported_assets/Mobile` profile) a non-breaking change. It also allows us to create top-level files like `.imported_assets/log` without it being interpreted as an asset. Meta files currently have a `.meta` suffix. Do we like these names and conventions? ### Should the `AssetPlugin::processed_dev` configuration enable `watch_for_changes` automatically? Currently it does (which I think makes sense), but it does make it the only configuration that enables watch_for_changes by default. ### Discuss on_loaded High Level Interface: This PR includes a very rough "proof of concept" `on_loaded` system adapter that uses the `LoadedWithDependencies` event in combination with `asset_server.load_asset` dependency tracking to support this pattern ```rust fn main() { App::new() .init_asset::<MyAssets>() .add_systems(Update, on_loaded(create_array_texture)) .run(); } #[derive(Asset, Clone)] struct MyAssets { #[dependency] picture_of_my_cat: Handle<Image>, #[dependency] picture_of_my_other_cat: Handle<Image>, } impl FromWorld for ArrayTexture { fn from_world(world: &mut World) -> Self { picture_of_my_cat: server.load("meow.png"), picture_of_my_other_cat: server.load("meeeeeeeow.png"), } } fn spawn_cat(In(my_assets): In<MyAssets>, mut commands: Commands) { commands.spawn(SpriteBundle { texture: my_assets.picture_of_my_cat.clone(), ..default() }); commands.spawn(SpriteBundle { texture: my_assets.picture_of_my_other_cat.clone(), ..default() }); } ``` The implementation is _very_ rough. And it is currently unsafe because `bevy_ecs` doesn't expose some internals to do this safely from inside `bevy_asset`. There are plenty of unanswered questions like: * "do we add a Loadable" derive? (effectively automate the FromWorld implementation above) * Should `MyAssets` even be an Asset? (largely implemented this way because it elegantly builds on `server.load_asset(MyAsset { .. })` dependency tracking). We should think hard about what our ideal API looks like (and if this is a pattern we want to support). Not necessarily something we need to solve in this PR. The current `on_loaded` impl should probably be removed from this PR before merging. ## Clarifying Questions ### What about Assets as Entities? This Bevy Asset V2 proposal implementation initially stored Assets as ECS Entities. Instead of `AssetId<T>` + the `Assets<T>` resource it used `Entity` as the asset id and Asset values were just ECS components. There are plenty of compelling reasons to do this: 1. Easier to inline assets in Bevy Scenes (as they are "just" normal entities + components) 2. More flexible queries: use the power of the ECS to filter assets (ex: `Query<Mesh, With<Tree>>`). 3. Extensible. Users can add arbitrary component data to assets. 4. Things like "component visualization tools" work out of the box to visualize asset data. However Assets as Entities has a ton of caveats right now: * We need to be able to allocate entity ids without a direct World reference (aka rework id allocator in Entities ... i worked around this in my prototypes by just pre allocating big chunks of entities) * We want asset change events in addition to ECS change tracking ... how do we populate them when mutations can come from anywhere? Do we use Changed queries? This would require iterating over the change data for all assets every frame. Is this acceptable or should we implement a new "event based" component change detection option? * Reconciling manually created assets with asset-system managed assets has some nuance (ex: are they "loaded" / do they also have that component metadata?) * "how do we handle "static" / default entity handles" (ties in to the Entity Indices discussion: https://github.com/bevyengine/bevy/discussions/8319). This is necessary for things like "built in" assets and default handles in things like SpriteBundle. * Storing asset information as a component makes it easy to "invalidate" asset state by removing the component (or forcing modifications). Ideally we have ways to lock this down (some combination of Rust type privacy and ECS validation) In practice, how we store and identify assets is a reasonably superficial change (porting off of Assets as Entities and implementing dedicated storage + ids took less than a day). So once we sort out the remaining challenges the flip should be straightforward. Additionally, I do still have "Assets as Entities" in my commit history, so we can reuse that work. I personally think "assets as entities" is a good endgame, but it also doesn't provide _significant_ value at the moment and it certainly isn't ready yet with the current state of things. ### Why not Distill? [Distill](https://github.com/amethyst/distill) is a high quality fully featured asset system built in Rust. It is very natural to ask "why not just use Distill?". It is also worth calling out that for awhile, [we planned on adopting Distill / I signed off on it](https://github.com/bevyengine/bevy/issues/708). However I think Bevy has a number of constraints that make Distill adoption suboptimal: * **Architectural Simplicity:** * Distill's processor requires an in-memory database (lmdb) and RPC networked API (using Cap'n Proto). Each of these introduces API complexity that increases maintenance burden and "code grokability". Ignoring tests, documentation, and examples, Distill has 24,237 lines of Rust code (including generated code for RPC + database interactions). If you ignore generated code, it has 11,499 lines. * Bevy builds the AssetProcessor and AssetServer using pluggable AssetReader/AssetWriter Rust traits with simple io interfaces. They do not necessitate databases or RPC interfaces (although Readers/Writers could use them if that is desired). Bevy Asset V2 (at the time of writing this PR) is 5,384 lines of Rust code (ignoring tests, documentation, and examples). Grain of salt: Distill does have more features currently (ex: Asset Packing, GUIDS, remote-out-of-process asset processor). I do plan to implement these features in Bevy Asset V2 and I personally highly doubt they will meaningfully close the 6115 lines-of-code gap. * This complexity gap (which while illustrated by lines of code, is much bigger than just that) is noteworthy to me. Bevy should be hackable and there are pillars of Distill that are very hard to understand and extend. This is a matter of opinion (and Bevy Asset V2 also has complicated areas), but I think Bevy Asset V2 is much more approachable for the average developer. * Necessary disclaimer: counting lines of code is an extremely rough complexity metric. Read the code and form your own opinions. * **Optional Asset Processing:** Not all Bevy Apps (or Bevy App developers) need / want asset preprocessing. Processing increases the complexity of the development environment by introducing things like meta files, imported asset storage, running processors in the background, waiting for processing to finish, etc. Distill _requires_ preprocessing to work. With Bevy Asset V2 processing is fully opt-in. The AssetServer isn't directly aware of asset processors at all. AssetLoaders only care about converting bytes to runtime Assets ... they don't know or care if the bytes were pre-processed or not. Processing is "elegantly" (forgive my self-congratulatory phrasing) layered on top and builds on the existing Asset system primitives. * **Direct Filesystem Access to Processed Asset State:** Distill stores processed assets in a database. This makes debugging / inspecting the processed outputs harder (either requires special tooling to query the database or they need to be "deployed" to be inspected). Bevy Asset V2, on the other hand, stores processed assets in the filesystem (by default ... this is configurable). This makes interacting with the processed state more natural. Note that both Godot and Unity's new asset system store processed assets in the filesystem. * **Portability**: Because Distill's processor uses lmdb and RPC networking, it cannot be run on certain platforms (ex: lmdb is a non-rust dependency that cannot run on the web, some platforms don't support running network servers). Bevy should be able to process assets everywhere (ex: run the Bevy Editor on the web, compile + process shaders on mobile, etc). Distill does partially mitigate this problem by supporting "streaming" assets via the RPC protocol, but this is not a full solve from my perspective. And Bevy Asset V2 can (in theory) also stream assets (without requiring RPC, although this isn't implemented yet) Note that I _do_ still think Distill would be a solid asset system for Bevy. But I think the approach in this PR is a better solve for Bevy's specific "asset system requirements". ### Doesn't async-fs just shim requests to "sync" `std::fs`? What is the point? "True async file io" has limited / spotty platform support. async-fs (and the rust async ecosystem generally ... ex Tokio) currently use async wrappers over std::fs that offload blocking requests to separate threads. This may feel unsatisfying, but it _does_ still provide value because it prevents our task pools from blocking on file system operations (which would prevent progress when there are many tasks to do, but all threads in a pool are currently blocking on file system ops). Additionally, using async APIs for our AssetReaders and AssetWriters also provides value because we can later add support for "true async file io" for platforms that support it. _And_ we can implement other "true async io" asset backends (such as networked asset io). ## Draft TODO - [x] Fill in missing filesystem event APIs: file removed event (which is expressed as dangling RenameFrom events in some cases), file/folder renamed event - [x] Assets without loaders are not moved to the processed folder. This breaks things like referenced `.bin` files for GLTFs. This should be configurable per-non-asset-type. - [x] Initial implementation of Reflect and FromReflect for Handle. The "deserialization" parity bar is low here as this only worked with static UUIDs in the old impl ... this is a non-trivial problem. Either we add a Handle::AssetPath variant that gets "upgraded" to a strong handle on scene load or we use a separate AssetRef type for Bevy scenes (which is converted to a runtime Handle on load). This deserves its own discussion in a different pr. - [x] Populate read_asset_bytes hash when run by the processor (a bit of a special case .. when run by the processor the processed meta will contain the hash so we don't need to compute it on the spot, but we don't want/need to read the meta when run by the main AssetServer) - [x] Delay hot reloading: currently filesystem events are handled immediately, which creates timing issues in some cases. For example hot reloading images can sometimes break because the image isn't finished writing. We should add a delay, likely similar to the [implementation in this PR](https://github.com/bevyengine/bevy/pull/8503). - [x] Port old platform-specific AssetIo implementations to the new AssetReader interface (currently missing Android and web) - [x] Resolve on_loaded unsafety (either by removing the API entirely or removing the unsafe) - [x] Runtime loader setting overrides - [x] Remove remaining unwraps that should be error-handled. There are number of TODOs here - [x] Pretty AssetPath Display impl - [x] Document more APIs - [x] Resolve spurious "reloading because it has changed" events (to repro run load_gltf with `processed_dev()`) - [x] load_dependency hot reloading currently only works for processed assets. If processing is disabled, load_dependency changes are not hot reloaded. - [x] Replace AssetInfo dependency load/fail counters with `loading_dependencies: HashSet<UntypedAssetId>` to prevent reloads from (potentially) breaking counters. Storing this will also enable "dependency reloaded" events (see [Next Steps](#next-steps)) - [x] Re-add filesystem watcher cargo feature gate (currently it is not optional) - [ ] Migration Guide - [ ] Changelog ## Followup TODO - [ ] Replace "eager unchanged processed asset loading" behavior with "don't returned unchanged processed asset until dependencies have been checked". - [ ] Add true `Ignore` AssetAction that does not copy the asset to the imported_assets folder. - [ ] Finish "live asset unloading" (ex: free up CPU asset memory after uploading an image to the GPU), rethink RenderAssets, and port renderer features. The `Assets` collection uses `Option<T>` for asset storage to support its removal. (1) the Option might not actually be necessary ... might be able to just remove from the collection entirely (2) need to finalize removal apis - [ ] Try replacing the "channel based" asset id recycling with something a bit more efficient (ex: we might be able to use raw atomic ints with some cleverness) - [ ] Consider adding UUIDs to processed assets (scoped just to helping identify moved assets ... not exposed to load queries ... see [Next Steps](#next-steps)) - [ ] Store "last modified" source asset and meta timestamps in processed meta files to enable skipping expensive hashing when the file wasn't changed - [ ] Fix "slow loop" handle drop fix - [ ] Migrate to TypeName - [x] Handle "loader preregistration". See #9429 ## Next Steps * **Configurable per-type defaults for AssetMeta**: It should be possible to add configuration like "all png image meta should default to using nearest sampling" (currently this hard-coded per-loader/processor Settings::default() impls). Also see the "Folder Meta" bullet point. * **Avoid Reprocessing on Asset Renames / Moves**: See the "canonical asset ids" discussion in [Open Questions](#open-questions) and the relevant bullet point in [Draft TODO](#draft-todo). Even without canonical ids, folder renames could avoid reprocessing in some cases. * **Multiple Asset Sources**: Expand AssetPath to support "asset source names" and support multiple AssetReaders in the asset server (ex: `webserver://some_path/image.png` backed by an Http webserver AssetReader). The "default" asset reader would use normal `some_path/image.png` paths. Ideally this works in combination with multiple AssetWatchers for hot-reloading * **Stable Type Names**: this pr removes the TypeUuid requirement from assets in favor of `std::any::type_name`. This makes defining assets easier (no need to generate a new uuid / use weird proc macro syntax). It also makes reading meta files easier (because things have "friendly names"). We also use type names for components in scene files. If they are good enough for components, they are good enough for assets. And consistency across Bevy pillars is desirable. However, `std::any::type_name` is not guaranteed to be stable (although in practice it is). We've developed a [stable type path](https://github.com/bevyengine/bevy/pull/7184) to resolve this, which should be adopted when it is ready. * **Command Line Interface**: It should be possible to run the asset processor in a separate process from the command line. This will also require building a network-server-backed AssetReader to communicate between the app and the processor. We've been planning to build a "bevy cli" for awhile. This seems like a good excuse to build it. * **Asset Packing**: This is largely an additive feature, so it made sense to me to punt this until we've laid the foundations in this PR. * **Per-Platform Processed Assets**: It should be possible to generate assets for multiple platforms by supporting multiple "processor profiles" per asset (ex: compress with format X on PC and Y on iOS). I think there should probably be arbitrary "profiles" (which can be separate from actual platforms), which are then assigned to a given platform when generating the final asset distribution for that platform. Ex: maybe devs want a "Mobile" profile that is shared between iOS and Android. Or a "LowEnd" profile shared between web and mobile. * **Versioning and Migrations**: Assets, Loaders, Savers, and Processors need to have versions to determine if their schema is valid. If an asset / loader version is incompatible with the current version expected at runtime, the processor should be able to migrate them. I think we should try using Bevy Reflect for this, as it would allow us to load the old version as a dynamic Reflect type without actually having the old Rust type. It would also allow us to define "patches" to migrate between versions (Bevy Reflect devs are currently working on patching). The `.meta` file already has its own format version. Migrating that to new versions should also be possible. * **Real Copy-on-write AssetPaths**: Rust's actual Cow (clone-on-write type) currently used by AssetPath can still result in String clones that aren't actually necessary (cloning an Owned Cow clones the contents). Bevy's asset system requires cloning AssetPaths in a number of places, which result in actual clones of the internal Strings. This is not efficient. AssetPath internals should be reworked to exhibit truer cow-like-behavior that reduces String clones to the absolute minimum. * **Consider processor-less processing**: In theory the AssetServer could run processors "inline" even if the background AssetProcessor is disabled. If we decide this is actually desirable, we could add this. But I don't think its a priority in the short or medium term. * **Pre-emptive dependency loading**: We could encode dependencies in processed meta files, which could then be used by the Asset Server to kick of dependency loads as early as possible (prior to starting the actual asset load). Is this desirable? How much time would this save in practice? * **Optimize Processor With UntypedAssetIds**: The processor exclusively uses AssetPath to identify assets currently. It might be possible to swap these out for UntypedAssetIds in some places, which are smaller / cheaper to hash and compare. * **One to Many Asset Processing**: An asset source file that produces many assets currently must be processed into a single "processed" asset source. If labeled assets can be written separately they can each have their own configured savers _and_ they could be loaded more granularly. Definitely worth exploring! * **Automatically Track "Runtime-only" Asset Dependencies**: Right now, tracking "created at runtime" asset dependencies requires adding them via `asset_server.load_asset(StandardMaterial::default())`. I think with some cleverness we could also do this for `materials.add(StandardMaterial::default())`, making tracking work "everywhere". There are challenges here relating to change detection / ensuring the server is made aware of dependency changes. This could be expensive in some cases. * **"Dependency Changed" events**: Some assets have runtime artifacts that need to be re-generated when one of their dependencies change (ex: regenerate a material's bind group when a Texture needs to change). We are generating the dependency graph so we can definitely produce these events. Buuuuut generating these events will have a cost / they could be high frequency for some assets, so we might want this to be opt-in for specific cases. * **Investigate Storing More Information In Handles**: Handles can now store arbitrary information, which makes it cheaper and easier to access. How much should we move into them? Canonical asset load states (via atomics)? (`handle.is_loaded()` would be very cool). Should we store the entire asset and remove the `Assets<T>` collection? (`Arc<RwLock<Option<Image>>>`?) * **Support processing and loading files without extensions**: This is a pretty arbitrary restriction and could be supported with very minimal changes. * **Folder Meta**: It would be nice if we could define per folder processor configuration defaults (likely in a `.meta` or `.folder_meta` file). Things like "default to linear filtering for all Images in this folder". * **Replace async_broadcast with event-listener?** This might be approximately drop-in for some uses and it feels more light weight * **Support Running the AssetProcessor on the Web**: Most of the hard work is done here, but there are some easy straggling TODOs (make the transaction log an interface instead of a direct file writer so we can write a web storage backend, implement an AssetReader/AssetWriter that reads/writes to something like LocalStorage). * **Consider identifying and preventing circular dependencies**: This is especially important for "processor dependencies", as processing will silently never finish in these cases. * **Built-in/Inlined Asset Hot Reloading**: This PR regresses "built-in/inlined" asset hot reloading (previously provided by the DebugAssetServer). I'm intentionally punting this because I think it can be cleanly implemented with "multiple asset sources" by registering a "debug asset source" (ex: `debug://bevy_pbr/src/render/pbr.wgsl` asset paths) in combination with an AssetWatcher for that asset source and support for "manually loading pats with asset bytes instead of AssetReaders". The old DebugAssetServer was quite nasty and I'd love to avoid that hackery going forward. * **Investigate ways to remove double-parsing meta files**: Parsing meta files currently involves parsing once with "minimal" versions of the meta file to extract the type name of the loader/processor config, then parsing again to parse the "full" meta. This is suboptimal. We should be able to define custom deserializers that (1) assume the loader/processor type name comes first (2) dynamically looks up the loader/processor registrations to deserialize settings in-line (similar to components in the bevy scene format). Another alternative: deserialize as dynamic Reflect objects and then convert. * **More runtime loading configuration**: Support using the Handle type as a hint to select an asset loader (instead of relying on AssetPath extensions) * **More high level Processor trait implementations**: For example, it might be worth adding support for arbitrary chains of "asset transforms" that modify an in-memory asset representation between loading and saving. (ex: load a Mesh, run a `subdivide_mesh` transform, followed by a `flip_normals` transform, then save the mesh to an efficient compressed format). * **Bevy Scene Handle Deserialization**: (see the relevant [Draft TODO item](#draft-todo) for context) * **Explore High Level Load Interfaces**: See [this discussion](#discuss-on_loaded-high-level-interface) for one prototype. * **Asset Streaming**: It would be great if we could stream Assets (ex: stream a long video file piece by piece) * **ID Exchanging**: In this PR Asset Handles/AssetIds are bigger than they need to be because they have a Uuid enum variant. If we implement an "id exchanging" system that trades Uuids for "efficient runtime ids", we can cut down on the size of AssetIds, making them more efficient. This has some open design questions, such as how to spawn entities with "default" handle values (as these wouldn't have access to the exchange api in the current system). * **Asset Path Fixup Tooling**: Assets that inline asset paths inside them will break when an asset moves. The asset system provides the functionality to detect when paths break. We should build a framework that enables formats to define "path migrations". This is especially important for scene files. For editor-generated files, we should also consider using UUIDs (see other bullet point) to avoid the need to migrate in these cases. --------- Co-authored-by: BeastLe9enD <beastle9end@outlook.de> Co-authored-by: Mike <mike.hsu@gmail.com> Co-authored-by: Nicola Papale <nicopap@users.noreply.github.com>
2023-09-07 02:07:27 +00:00
render_materials.get(&material_handle.id()),
) {
let mut mesh_key =
MeshPipelineKey::from_primitive_topology(mesh.primitive_topology)
| MeshPipelineKey::DEPTH_PREPASS;
Add morph targets (#8158) # Objective - Add morph targets to `bevy_pbr` (closes #5756) & load them from glTF - Supersedes #3722 - Fixes #6814 [Morph targets][1] (also known as shape interpolation, shape keys, or blend shapes) allow animating individual vertices with fine grained controls. This is typically used for facial expressions. By specifying multiple poses as vertex offset, and providing a set of weight of each pose, it is possible to define surprisingly realistic transitions between poses. Blending between multiple poses also allow composition. Morph targets are part of the [gltf standard][2] and are a feature of Unity and Unreal, and babylone.js, it is only natural to implement them in bevy. ## Solution This implementation of morph targets uses a 3d texture where each pixel is a component of an animated attribute. Each layer is a different target. We use a 2d texture for each target, because the number of attribute×components×animated vertices is expected to always exceed the maximum pixel row size limit of webGL2. It copies fairly closely the way skinning is implemented on the CPU side, while on the GPU side, the shader morph target implementation is a relatively trivial detail. We add an optional `morph_texture` to the `Mesh` struct. The `morph_texture` is built through a method that accepts an iterator over attribute buffers. The `MorphWeights` component, user-accessible, controls the blend of poses used by mesh instances (so that multiple copy of the same mesh may have different weights), all the weights are uploaded to a uniform buffer of 256 `f32`. We limit to 16 poses per mesh, and a total of 256 poses. More literature: * Old babylone.js implementation (vertex attribute-based): https://www.eternalcoding.com/dev-log-1-morph-targets/ * Babylone.js implementation (similar to ours): https://www.youtube.com/watch?v=LBPRmGgU0PE * GPU gems 3: https://developer.nvidia.com/gpugems/gpugems3/part-i-geometry/chapter-3-directx-10-blend-shapes-breaking-limits * Development discord thread https://discord.com/channels/691052431525675048/1083325980615114772 https://user-images.githubusercontent.com/26321040/231181046-3bca2ab2-d4d9-472e-8098-639f1871ce2e.mp4 https://github.com/bevyengine/bevy/assets/26321040/d2a0c544-0ef8-45cf-9f99-8c3792f5a258 ## Acknowledgements * Thanks to `storytold` for sponsoring the feature * Thanks to `superdump` and `james7132` for guidance and help figuring out stuff ## Future work - Handling of less and more attributes (eg: animated uv, animated arbitrary attributes) - Dynamic pose allocation (so that zero-weighted poses aren't uploaded to GPU for example, enables much more total poses) - Better animation API, see #8357 ---- ## Changelog - Add morph targets to bevy meshes - Support up to 64 poses per mesh of individually up to 116508 vertices, animation currently strictly limited to the position, normal and tangent attributes. - Load a morph target using `Mesh::set_morph_targets` - Add `VisitMorphTargets` and `VisitMorphAttributes` traits to `bevy_render`, this allows defining morph targets (a fairly complex and nested data structure) through iterators (ie: single copy instead of passing around buffers), see documentation of those traits for details - Add `MorphWeights` component exported by `bevy_render` - `MorphWeights` control mesh's morph target weights, blending between various poses defined as morph targets. - `MorphWeights` are directly inherited by direct children (single level of hierarchy) of an entity. This allows controlling several mesh primitives through a unique entity _as per GLTF spec_. - Add `MorphTargetNames` component, naming each indices of loaded morph targets. - Load morph targets weights and buffers in `bevy_gltf` - handle morph targets animations in `bevy_animation` (previously, it was a `warn!` log) - Add the `MorphStressTest.gltf` asset for morph targets testing, taken from the glTF samples repo, CC0. - Add morph target manipulation to `scene_viewer` - Separate the animation code in `scene_viewer` from the rest of the code, reducing `#[cfg(feature)]` noise - Add the `morph_targets.rs` example to show off how to manipulate morph targets, loading `MorpStressTest.gltf` ## Migration Guide - (very specialized, unlikely to be touched by 3rd parties) - `MeshPipeline` now has a single `mesh_layouts` field rather than separate `mesh_layout` and `skinned_mesh_layout` fields. You should handle all possible mesh bind group layouts in your implementation - You should also handle properly the new `MORPH_TARGETS` shader def and mesh pipeline key. A new function is exposed to make this easier: `setup_moprh_and_skinning_defs` - The `MeshBindGroup` is now `MeshBindGroups`, cached bind groups are now accessed through the `get` method. [1]: https://en.wikipedia.org/wiki/Morph_target_animation [2]: https://registry.khronos.org/glTF/specs/2.0/glTF-2.0.html#morph-targets --------- Co-authored-by: François <mockersf@gmail.com> Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2023-06-22 20:00:01 +00:00
if mesh.morph_targets.is_some() {
mesh_key |= MeshPipelineKey::MORPH_TARGETS;
}
if is_directional_light {
mesh_key |= MeshPipelineKey::DEPTH_CLAMP_ORTHO;
}
let alpha_mode = material.properties.alpha_mode;
match alpha_mode {
Add `MAY_DISCARD` shader def, enabling early depth tests for most cases (#6697) # Objective - Right now we can't really benefit from [early depth testing](https://www.khronos.org/opengl/wiki/Early_Fragment_Test) in our PBR shader because it includes codepaths with `discard`, even for situations where they are not necessary. ## Solution - This PR introduces a new `MeshPipelineKey` and shader def, `MAY_DISCARD`; - All possible material/mesh options that that may result in `discard`s being needed must set `MAY_DISCARD` ahead of time: - Right now, this is only `AlphaMode::Mask(f32)`, but in the future might include other options/effects; (e.g. one effect I'm personally interested in is bayer dither pseudo-transparency for LOD transitions of opaque meshes) - Shader codepaths that can `discard` are guarded by an `#ifdef MAY_DISCARD` preprocessor directive: - Right now, this is just one branch in `alpha_discard()`; - If `MAY_DISCARD` is _not_ set, the `@early_depth_test` attribute is added to the PBR fragment shader. This is a not yet documented, possibly non-standard WGSL extension I found browsing Naga's source code. [I opened a PR to document it there](https://github.com/gfx-rs/naga/pull/2132). My understanding is that for backends where this attribute is supported, it will force an explicit opt-in to early depth test. (e.g. via `layout(early_fragment_tests) in;` in GLSL) ## Caveats - I included `@early_depth_test` for the sake of us being explicit, and avoiding the need for the driver to be “smart” about enabling this feature. That way, if we make a mistake and include a `discard` unguarded by `MAY_DISCARD`, it will either produce errors or noticeable visual artifacts so that we'll catch early, instead of causing a performance regression. - I'm not sure explicit early depth test is supported on the naga Metal backend, which is what I'm currently using, so I can't really test the explicit early depth test enable, I would like others with Vulkan/GL hardware to test it if possible; - I would like some guidance on how to measure/verify the performance benefits of this; - If I understand it correctly, this, or _something like this_ is needed to fully reap the performance gains enabled by #6284; - This will _most definitely_ conflict with #6284 and #6644. I can fix the conflicts as needed, depending on whether/the order they end up being merging in. --- ## Changelog ### Changed - Early depth tests are now enabled whenever possible for meshes using `StandardMaterial`, reducing the number of fragments evaluated for scenes with lots of occlusions.
2023-05-29 15:15:01 +00:00
AlphaMode::Mask(_)
| AlphaMode::Blend
| AlphaMode::Premultiplied
| AlphaMode::Add => {
mesh_key |= MeshPipelineKey::MAY_DISCARD;
}
_ => {}
}
Mesh vertex buffer layouts (#3959) This PR makes a number of changes to how meshes and vertex attributes are handled, which the goal of enabling easy and flexible custom vertex attributes: * Reworks the `Mesh` type to use the newly added `VertexAttribute` internally * `VertexAttribute` defines the name, a unique `VertexAttributeId`, and a `VertexFormat` * `VertexAttributeId` is used to produce consistent sort orders for vertex buffer generation, replacing the more expensive and often surprising "name based sorting" * Meshes can be used to generate a `MeshVertexBufferLayout`, which defines the layout of the gpu buffer produced by the mesh. `MeshVertexBufferLayouts` can then be used to generate actual `VertexBufferLayouts` according to the requirements of a specific pipeline. This decoupling of "mesh layout" vs "pipeline vertex buffer layout" is what enables custom attributes. We don't need to standardize _mesh layouts_ or contort meshes to meet the needs of a specific pipeline. As long as the mesh has what the pipeline needs, it will work transparently. * Mesh-based pipelines now specialize on `&MeshVertexBufferLayout` via the new `SpecializedMeshPipeline` trait (which behaves like `SpecializedPipeline`, but adds `&MeshVertexBufferLayout`). The integrity of the pipeline cache is maintained because the `MeshVertexBufferLayout` is treated as part of the key (which is fully abstracted from implementers of the trait ... no need to add any additional info to the specialization key). * Hashing `MeshVertexBufferLayout` is too expensive to do for every entity, every frame. To make this scalable, I added a generalized "pre-hashing" solution to `bevy_utils`: `Hashed<T>` keys and `PreHashMap<K, V>` (which uses `Hashed<T>` internally) . Why didn't I just do the quick and dirty in-place "pre-compute hash and use that u64 as a key in a hashmap" that we've done in the past? Because its wrong! Hashes by themselves aren't enough because two different values can produce the same hash. Re-hashing a hash is even worse! I decided to build a generalized solution because this pattern has come up in the past and we've chosen to do the wrong thing. Now we can do the right thing! This did unfortunately require pulling in `hashbrown` and using that in `bevy_utils`, because avoiding re-hashes requires the `raw_entry_mut` api, which isn't stabilized yet (and may never be ... `entry_ref` has favor now, but also isn't available yet). If std's HashMap ever provides the tools we need, we can move back to that. Note that adding `hashbrown` doesn't increase our dependency count because it was already in our tree. I will probably break these changes out into their own PR. * Specializing on `MeshVertexBufferLayout` has one non-obvious behavior: it can produce identical pipelines for two different MeshVertexBufferLayouts. To optimize the number of active pipelines / reduce re-binds while drawing, I de-duplicate pipelines post-specialization using the final `VertexBufferLayout` as the key. For example, consider a pipeline that needs the layout `(position, normal)` and is specialized using two meshes: `(position, normal, uv)` and `(position, normal, other_vec2)`. If both of these meshes result in `(position, normal)` specializations, we can use the same pipeline! Now we do. Cool! To briefly illustrate, this is what the relevant section of `MeshPipeline`'s specialization code looks like now: ```rust impl SpecializedMeshPipeline for MeshPipeline { type Key = MeshPipelineKey; fn specialize( &self, key: Self::Key, layout: &MeshVertexBufferLayout, ) -> RenderPipelineDescriptor { let mut vertex_attributes = vec![ Mesh::ATTRIBUTE_POSITION.at_shader_location(0), Mesh::ATTRIBUTE_NORMAL.at_shader_location(1), Mesh::ATTRIBUTE_UV_0.at_shader_location(2), ]; let mut shader_defs = Vec::new(); if layout.contains(Mesh::ATTRIBUTE_TANGENT) { shader_defs.push(String::from("VERTEX_TANGENTS")); vertex_attributes.push(Mesh::ATTRIBUTE_TANGENT.at_shader_location(3)); } let vertex_buffer_layout = layout .get_layout(&vertex_attributes) .expect("Mesh is missing a vertex attribute"); ``` Notice that this is _much_ simpler than it was before. And now any mesh with any layout can be used with this pipeline, provided it has vertex postions, normals, and uvs. We even got to remove `HAS_TANGENTS` from MeshPipelineKey and `has_tangents` from `GpuMesh`, because that information is redundant with `MeshVertexBufferLayout`. This is still a draft because I still need to: * Add more docs * Experiment with adding error handling to mesh pipeline specialization (which would print errors at runtime when a mesh is missing a vertex attribute required by a pipeline). If it doesn't tank perf, we'll keep it. * Consider breaking out the PreHash / hashbrown changes into a separate PR. * Add an example illustrating this change * Verify that the "mesh-specialized pipeline de-duplication code" works properly Please dont yell at me for not doing these things yet :) Just trying to get this in peoples' hands asap. Alternative to #3120 Fixes #3030 Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-02-23 23:21:13 +00:00
let pipeline_id = pipelines.specialize(
&pipeline_cache,
&prepass_pipeline,
MaterialPipelineKey {
mesh_key,
bind_group_data: material.key.clone(),
},
Mesh vertex buffer layouts (#3959) This PR makes a number of changes to how meshes and vertex attributes are handled, which the goal of enabling easy and flexible custom vertex attributes: * Reworks the `Mesh` type to use the newly added `VertexAttribute` internally * `VertexAttribute` defines the name, a unique `VertexAttributeId`, and a `VertexFormat` * `VertexAttributeId` is used to produce consistent sort orders for vertex buffer generation, replacing the more expensive and often surprising "name based sorting" * Meshes can be used to generate a `MeshVertexBufferLayout`, which defines the layout of the gpu buffer produced by the mesh. `MeshVertexBufferLayouts` can then be used to generate actual `VertexBufferLayouts` according to the requirements of a specific pipeline. This decoupling of "mesh layout" vs "pipeline vertex buffer layout" is what enables custom attributes. We don't need to standardize _mesh layouts_ or contort meshes to meet the needs of a specific pipeline. As long as the mesh has what the pipeline needs, it will work transparently. * Mesh-based pipelines now specialize on `&MeshVertexBufferLayout` via the new `SpecializedMeshPipeline` trait (which behaves like `SpecializedPipeline`, but adds `&MeshVertexBufferLayout`). The integrity of the pipeline cache is maintained because the `MeshVertexBufferLayout` is treated as part of the key (which is fully abstracted from implementers of the trait ... no need to add any additional info to the specialization key). * Hashing `MeshVertexBufferLayout` is too expensive to do for every entity, every frame. To make this scalable, I added a generalized "pre-hashing" solution to `bevy_utils`: `Hashed<T>` keys and `PreHashMap<K, V>` (which uses `Hashed<T>` internally) . Why didn't I just do the quick and dirty in-place "pre-compute hash and use that u64 as a key in a hashmap" that we've done in the past? Because its wrong! Hashes by themselves aren't enough because two different values can produce the same hash. Re-hashing a hash is even worse! I decided to build a generalized solution because this pattern has come up in the past and we've chosen to do the wrong thing. Now we can do the right thing! This did unfortunately require pulling in `hashbrown` and using that in `bevy_utils`, because avoiding re-hashes requires the `raw_entry_mut` api, which isn't stabilized yet (and may never be ... `entry_ref` has favor now, but also isn't available yet). If std's HashMap ever provides the tools we need, we can move back to that. Note that adding `hashbrown` doesn't increase our dependency count because it was already in our tree. I will probably break these changes out into their own PR. * Specializing on `MeshVertexBufferLayout` has one non-obvious behavior: it can produce identical pipelines for two different MeshVertexBufferLayouts. To optimize the number of active pipelines / reduce re-binds while drawing, I de-duplicate pipelines post-specialization using the final `VertexBufferLayout` as the key. For example, consider a pipeline that needs the layout `(position, normal)` and is specialized using two meshes: `(position, normal, uv)` and `(position, normal, other_vec2)`. If both of these meshes result in `(position, normal)` specializations, we can use the same pipeline! Now we do. Cool! To briefly illustrate, this is what the relevant section of `MeshPipeline`'s specialization code looks like now: ```rust impl SpecializedMeshPipeline for MeshPipeline { type Key = MeshPipelineKey; fn specialize( &self, key: Self::Key, layout: &MeshVertexBufferLayout, ) -> RenderPipelineDescriptor { let mut vertex_attributes = vec![ Mesh::ATTRIBUTE_POSITION.at_shader_location(0), Mesh::ATTRIBUTE_NORMAL.at_shader_location(1), Mesh::ATTRIBUTE_UV_0.at_shader_location(2), ]; let mut shader_defs = Vec::new(); if layout.contains(Mesh::ATTRIBUTE_TANGENT) { shader_defs.push(String::from("VERTEX_TANGENTS")); vertex_attributes.push(Mesh::ATTRIBUTE_TANGENT.at_shader_location(3)); } let vertex_buffer_layout = layout .get_layout(&vertex_attributes) .expect("Mesh is missing a vertex attribute"); ``` Notice that this is _much_ simpler than it was before. And now any mesh with any layout can be used with this pipeline, provided it has vertex postions, normals, and uvs. We even got to remove `HAS_TANGENTS` from MeshPipelineKey and `has_tangents` from `GpuMesh`, because that information is redundant with `MeshVertexBufferLayout`. This is still a draft because I still need to: * Add more docs * Experiment with adding error handling to mesh pipeline specialization (which would print errors at runtime when a mesh is missing a vertex attribute required by a pipeline). If it doesn't tank perf, we'll keep it. * Consider breaking out the PreHash / hashbrown changes into a separate PR. * Add an example illustrating this change * Verify that the "mesh-specialized pipeline de-duplication code" works properly Please dont yell at me for not doing these things yet :) Just trying to get this in peoples' hands asap. Alternative to #3120 Fixes #3030 Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-02-23 23:21:13 +00:00
&mesh.layout,
);
let pipeline_id = match pipeline_id {
Ok(id) => id,
Err(err) => {
error!("{}", err);
continue;
}
};
shadow_phase.add(Shadow {
draw_function: draw_shadow_mesh,
pipeline: pipeline_id,
entity,
Use GpuArrayBuffer for MeshUniform (#9254) # Objective - Reduce the number of rebindings to enable batching of draw commands ## Solution - Use the new `GpuArrayBuffer` for `MeshUniform` data to store all `MeshUniform` data in arrays within fewer bindings - Sort opaque/alpha mask prepass, opaque/alpha mask main, and shadow phases also by the batch per-object data binding dynamic offset to improve performance on WebGL2. --- ## Changelog - Changed: Per-object `MeshUniform` data is now managed by `GpuArrayBuffer` as arrays in buffers that need to be indexed into. ## Migration Guide Accessing the `model` member of an individual mesh object's shader `Mesh` struct the old way where each `MeshUniform` was stored at its own dynamic offset: ```rust struct Vertex { @location(0) position: vec3<f32>, }; fn vertex(vertex: Vertex) -> VertexOutput { var out: VertexOutput; out.clip_position = mesh_position_local_to_clip( mesh.model, vec4<f32>(vertex.position, 1.0) ); return out; } ``` The new way where one needs to index into the array of `Mesh`es for the batch: ```rust struct Vertex { @builtin(instance_index) instance_index: u32, @location(0) position: vec3<f32>, }; fn vertex(vertex: Vertex) -> VertexOutput { var out: VertexOutput; out.clip_position = mesh_position_local_to_clip( mesh[vertex.instance_index].model, vec4<f32>(vertex.position, 1.0) ); return out; } ``` Note that using the instance_index is the default way to pass the per-object index into the shader, but if you wish to do custom rendering approaches you can pass it in however you like. --------- Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com> Co-authored-by: Elabajaba <Elabajaba@users.noreply.github.com>
2023-07-30 13:17:08 +00:00
distance: 0.0, // TODO: sort front-to-back
Mesh vertex buffer layouts (#3959) This PR makes a number of changes to how meshes and vertex attributes are handled, which the goal of enabling easy and flexible custom vertex attributes: * Reworks the `Mesh` type to use the newly added `VertexAttribute` internally * `VertexAttribute` defines the name, a unique `VertexAttributeId`, and a `VertexFormat` * `VertexAttributeId` is used to produce consistent sort orders for vertex buffer generation, replacing the more expensive and often surprising "name based sorting" * Meshes can be used to generate a `MeshVertexBufferLayout`, which defines the layout of the gpu buffer produced by the mesh. `MeshVertexBufferLayouts` can then be used to generate actual `VertexBufferLayouts` according to the requirements of a specific pipeline. This decoupling of "mesh layout" vs "pipeline vertex buffer layout" is what enables custom attributes. We don't need to standardize _mesh layouts_ or contort meshes to meet the needs of a specific pipeline. As long as the mesh has what the pipeline needs, it will work transparently. * Mesh-based pipelines now specialize on `&MeshVertexBufferLayout` via the new `SpecializedMeshPipeline` trait (which behaves like `SpecializedPipeline`, but adds `&MeshVertexBufferLayout`). The integrity of the pipeline cache is maintained because the `MeshVertexBufferLayout` is treated as part of the key (which is fully abstracted from implementers of the trait ... no need to add any additional info to the specialization key). * Hashing `MeshVertexBufferLayout` is too expensive to do for every entity, every frame. To make this scalable, I added a generalized "pre-hashing" solution to `bevy_utils`: `Hashed<T>` keys and `PreHashMap<K, V>` (which uses `Hashed<T>` internally) . Why didn't I just do the quick and dirty in-place "pre-compute hash and use that u64 as a key in a hashmap" that we've done in the past? Because its wrong! Hashes by themselves aren't enough because two different values can produce the same hash. Re-hashing a hash is even worse! I decided to build a generalized solution because this pattern has come up in the past and we've chosen to do the wrong thing. Now we can do the right thing! This did unfortunately require pulling in `hashbrown` and using that in `bevy_utils`, because avoiding re-hashes requires the `raw_entry_mut` api, which isn't stabilized yet (and may never be ... `entry_ref` has favor now, but also isn't available yet). If std's HashMap ever provides the tools we need, we can move back to that. Note that adding `hashbrown` doesn't increase our dependency count because it was already in our tree. I will probably break these changes out into their own PR. * Specializing on `MeshVertexBufferLayout` has one non-obvious behavior: it can produce identical pipelines for two different MeshVertexBufferLayouts. To optimize the number of active pipelines / reduce re-binds while drawing, I de-duplicate pipelines post-specialization using the final `VertexBufferLayout` as the key. For example, consider a pipeline that needs the layout `(position, normal)` and is specialized using two meshes: `(position, normal, uv)` and `(position, normal, other_vec2)`. If both of these meshes result in `(position, normal)` specializations, we can use the same pipeline! Now we do. Cool! To briefly illustrate, this is what the relevant section of `MeshPipeline`'s specialization code looks like now: ```rust impl SpecializedMeshPipeline for MeshPipeline { type Key = MeshPipelineKey; fn specialize( &self, key: Self::Key, layout: &MeshVertexBufferLayout, ) -> RenderPipelineDescriptor { let mut vertex_attributes = vec![ Mesh::ATTRIBUTE_POSITION.at_shader_location(0), Mesh::ATTRIBUTE_NORMAL.at_shader_location(1), Mesh::ATTRIBUTE_UV_0.at_shader_location(2), ]; let mut shader_defs = Vec::new(); if layout.contains(Mesh::ATTRIBUTE_TANGENT) { shader_defs.push(String::from("VERTEX_TANGENTS")); vertex_attributes.push(Mesh::ATTRIBUTE_TANGENT.at_shader_location(3)); } let vertex_buffer_layout = layout .get_layout(&vertex_attributes) .expect("Mesh is missing a vertex attribute"); ``` Notice that this is _much_ simpler than it was before. And now any mesh with any layout can be used with this pipeline, provided it has vertex postions, normals, and uvs. We even got to remove `HAS_TANGENTS` from MeshPipelineKey and `has_tangents` from `GpuMesh`, because that information is redundant with `MeshVertexBufferLayout`. This is still a draft because I still need to: * Add more docs * Experiment with adding error handling to mesh pipeline specialization (which would print errors at runtime when a mesh is missing a vertex attribute required by a pipeline). If it doesn't tank perf, we'll keep it. * Consider breaking out the PreHash / hashbrown changes into a separate PR. * Add an example illustrating this change * Verify that the "mesh-specialized pipeline de-duplication code" works properly Please dont yell at me for not doing these things yet :) Just trying to get this in peoples' hands asap. Alternative to #3120 Fixes #3030 Co-authored-by: Carter Anderson <mcanders1@gmail.com>
2022-02-23 23:21:13 +00:00
});
}
}
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
}
}
}
}
pub struct Shadow {
pub distance: f32,
pub entity: Entity,
pub pipeline: CachedRenderPipelineId,
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
pub draw_function: DrawFunctionId,
2021-06-02 02:59:17 +00:00
}
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
impl PhaseItem for Shadow {
Reorder render sets, refactor bevy_sprite to take advantage (#9236) This is a continuation of this PR: #8062 # Objective - Reorder render schedule sets to allow data preparation when phase item order is known to support improved batching - Part of the batching/instancing etc plan from here: https://github.com/bevyengine/bevy/issues/89#issuecomment-1379249074 - The original idea came from @inodentry and proved to be a good one. Thanks! - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the new ordering ## Solution - Move `Prepare` and `PrepareFlush` after `PhaseSortFlush` - Add a `PrepareAssets` set that runs in parallel with other systems and sets in the render schedule. - Put prepare_assets systems in the `PrepareAssets` set - If explicit dependencies are needed on Mesh or Material RenderAssets then depend on the appropriate system. - Add `ManageViews` and `ManageViewsFlush` sets between `ExtractCommands` and Queue - Move `queue_mesh*_bind_group` to the Prepare stage - Rename them to `prepare_` - Put systems that prepare resources (buffers, textures, etc.) into a `PrepareResources` set inside `Prepare` - Put the `prepare_..._bind_group` systems into a `PrepareBindGroup` set after `PrepareResources` - Move `prepare_lights` to the `ManageViews` set - `prepare_lights` creates views and this must happen before `Queue` - This system needs refactoring to stop handling all responsibilities - Gather lights, sort, and create shadow map views. Store sorted light entities in a resource - Remove `BatchedPhaseItem` - Replace `batch_range` with `batch_size` representing how many items to skip after rendering the item or to skip the item entirely if `batch_size` is 0. - `queue_sprites` has been split into `queue_sprites` for queueing phase items and `prepare_sprites` for batching after the `PhaseSort` - `PhaseItem`s are still inserted in `queue_sprites` - After sorting adjacent compatible sprite phase items are accumulated into `SpriteBatch` components on the first entity of each batch, containing a range of vertex indices. The associated `PhaseItem`'s `batch_size` is updated appropriately. - `SpriteBatch` items are then drawn skipping over the other items in the batch based on the value in `batch_size` - A very similar refactor was performed on `bevy_ui` --- ## Changelog Changed: - Reordered and reworked render app schedule sets. The main change is that data is extracted, queued, sorted, and then prepared when the order of data is known. - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the reordering. ## Migration Guide - Assets such as materials and meshes should now be created in `PrepareAssets` e.g. `prepare_assets<Mesh>` - Queueing entities to `RenderPhase`s continues to be done in `Queue` e.g. `queue_sprites` - Preparing resources (textures, buffers, etc.) should now be done in `PrepareResources`, e.g. `prepare_prepass_textures`, `prepare_mesh_uniforms` - Prepare bind groups should now be done in `PrepareBindGroups` e.g. `prepare_mesh_bind_group` - Any batching or instancing can now be done in `Prepare` where the order of the phase items is known e.g. `prepare_sprites` ## Next Steps - Introduce some generic mechanism to ensure items that can be batched are grouped in the phase item order, currently you could easily have `[sprite at z 0, mesh at z 0, sprite at z 0]` preventing batching. - Investigate improved orderings for building the MeshUniform buffer - Implementing batching across the rest of bevy --------- Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
2023-08-27 14:33:49 +00:00
type SortKey = usize;
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
#[inline]
fn entity(&self) -> Entity {
self.entity
}
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
#[inline]
fn sort_key(&self) -> Self::SortKey {
Reorder render sets, refactor bevy_sprite to take advantage (#9236) This is a continuation of this PR: #8062 # Objective - Reorder render schedule sets to allow data preparation when phase item order is known to support improved batching - Part of the batching/instancing etc plan from here: https://github.com/bevyengine/bevy/issues/89#issuecomment-1379249074 - The original idea came from @inodentry and proved to be a good one. Thanks! - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the new ordering ## Solution - Move `Prepare` and `PrepareFlush` after `PhaseSortFlush` - Add a `PrepareAssets` set that runs in parallel with other systems and sets in the render schedule. - Put prepare_assets systems in the `PrepareAssets` set - If explicit dependencies are needed on Mesh or Material RenderAssets then depend on the appropriate system. - Add `ManageViews` and `ManageViewsFlush` sets between `ExtractCommands` and Queue - Move `queue_mesh*_bind_group` to the Prepare stage - Rename them to `prepare_` - Put systems that prepare resources (buffers, textures, etc.) into a `PrepareResources` set inside `Prepare` - Put the `prepare_..._bind_group` systems into a `PrepareBindGroup` set after `PrepareResources` - Move `prepare_lights` to the `ManageViews` set - `prepare_lights` creates views and this must happen before `Queue` - This system needs refactoring to stop handling all responsibilities - Gather lights, sort, and create shadow map views. Store sorted light entities in a resource - Remove `BatchedPhaseItem` - Replace `batch_range` with `batch_size` representing how many items to skip after rendering the item or to skip the item entirely if `batch_size` is 0. - `queue_sprites` has been split into `queue_sprites` for queueing phase items and `prepare_sprites` for batching after the `PhaseSort` - `PhaseItem`s are still inserted in `queue_sprites` - After sorting adjacent compatible sprite phase items are accumulated into `SpriteBatch` components on the first entity of each batch, containing a range of vertex indices. The associated `PhaseItem`'s `batch_size` is updated appropriately. - `SpriteBatch` items are then drawn skipping over the other items in the batch based on the value in `batch_size` - A very similar refactor was performed on `bevy_ui` --- ## Changelog Changed: - Reordered and reworked render app schedule sets. The main change is that data is extracted, queued, sorted, and then prepared when the order of data is known. - Refactor `bevy_sprite` and `bevy_ui` to take advantage of the reordering. ## Migration Guide - Assets such as materials and meshes should now be created in `PrepareAssets` e.g. `prepare_assets<Mesh>` - Queueing entities to `RenderPhase`s continues to be done in `Queue` e.g. `queue_sprites` - Preparing resources (textures, buffers, etc.) should now be done in `PrepareResources`, e.g. `prepare_prepass_textures`, `prepare_mesh_uniforms` - Prepare bind groups should now be done in `PrepareBindGroups` e.g. `prepare_mesh_bind_group` - Any batching or instancing can now be done in `Prepare` where the order of the phase items is known e.g. `prepare_sprites` ## Next Steps - Introduce some generic mechanism to ensure items that can be batched are grouped in the phase item order, currently you could easily have `[sprite at z 0, mesh at z 0, sprite at z 0]` preventing batching. - Investigate improved orderings for building the MeshUniform buffer - Implementing batching across the rest of bevy --------- Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com>
2023-08-27 14:33:49 +00:00
self.pipeline.id()
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
}
#[inline]
fn draw_function(&self) -> DrawFunctionId {
self.draw_function
}
Allow unbatched render phases to use unstable sorts (#5049) # Objective Partially addresses #4291. Speed up the sort phase for unbatched render phases. ## Solution Split out one of the optimizations in #4899 and allow implementors of `PhaseItem` to change what kind of sort is used when sorting the items in the phase. This currently includes Stable, Unstable, and Unsorted. Each of these corresponds to `Vec::sort_by_key`, `Vec::sort_unstable_by_key`, and no sorting at all. The default is `Unstable`. The last one can be used as a default if users introduce a preliminary depth prepass. ## Performance This will not impact the performance of any batched phases, as it is still using a stable sort. 2D's only phase is unchanged. All 3D phases are unbatched currently, and will benefit from this change. On `many_cubes`, where the primary phase is opaque, this change sees a speed up from 907.02us -> 477.62us, a 47.35% reduction. ![image](https://user-images.githubusercontent.com/3137680/174471253-22424874-30d5-4db5-b5b4-65fb2c612a9c.png) ## Future Work There were prior discussions to add support for faster radix sorts in #4291, which in theory should be a `O(n)` instead of a `O(nlog(n))` time. [`voracious`](https://crates.io/crates/voracious_radix_sort) has been proposed, but it seems to be optimize for use cases with more than 30,000 items, which may be atypical for most systems. Another optimization included in #4899 is to reduce the size of a few of the IDs commonly used in `PhaseItem` implementations to shrink the types to make swapping/sorting faster. Both `CachedPipelineId` and `DrawFunctionId` could be reduced to `u32` instead of `usize`. Ideally, this should automatically change to use stable sorts when `BatchedPhaseItem` is implemented on the same phase item type, but this requires specialization, which may not land in stable Rust for a short while. --- ## Changelog Added: `PhaseItem::sort` ## Migration Guide RenderPhases now default to a unstable sort (via `slice::sort_unstable_by_key`). This can typically improve sort phase performance, but may produce incorrect batching results when implementing `BatchedPhaseItem`. To revert to the older stable sort, manually implement `PhaseItem::sort` to implement a stable sort (i.e. via `slice::sort_by_key`). Co-authored-by: Federico Rinaldi <gisquerin@gmail.com> Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: colepoirier <colepoirier@gmail.com>
2022-06-23 10:52:49 +00:00
#[inline]
fn sort(items: &mut [Self]) {
// The shadow phase is sorted by pipeline id for performance reasons.
// Grouping all draw commands using the same pipeline together performs
// better than rebinding everything at a high rate.
Use GpuArrayBuffer for MeshUniform (#9254) # Objective - Reduce the number of rebindings to enable batching of draw commands ## Solution - Use the new `GpuArrayBuffer` for `MeshUniform` data to store all `MeshUniform` data in arrays within fewer bindings - Sort opaque/alpha mask prepass, opaque/alpha mask main, and shadow phases also by the batch per-object data binding dynamic offset to improve performance on WebGL2. --- ## Changelog - Changed: Per-object `MeshUniform` data is now managed by `GpuArrayBuffer` as arrays in buffers that need to be indexed into. ## Migration Guide Accessing the `model` member of an individual mesh object's shader `Mesh` struct the old way where each `MeshUniform` was stored at its own dynamic offset: ```rust struct Vertex { @location(0) position: vec3<f32>, }; fn vertex(vertex: Vertex) -> VertexOutput { var out: VertexOutput; out.clip_position = mesh_position_local_to_clip( mesh.model, vec4<f32>(vertex.position, 1.0) ); return out; } ``` The new way where one needs to index into the array of `Mesh`es for the batch: ```rust struct Vertex { @builtin(instance_index) instance_index: u32, @location(0) position: vec3<f32>, }; fn vertex(vertex: Vertex) -> VertexOutput { var out: VertexOutput; out.clip_position = mesh_position_local_to_clip( mesh[vertex.instance_index].model, vec4<f32>(vertex.position, 1.0) ); return out; } ``` Note that using the instance_index is the default way to pass the per-object index into the shader, but if you wish to do custom rendering approaches you can pass it in however you like. --------- Co-authored-by: robtfm <50659922+robtfm@users.noreply.github.com> Co-authored-by: Elabajaba <Elabajaba@users.noreply.github.com>
2023-07-30 13:17:08 +00:00
radsort::sort_by_key(items, |item| item.sort_key());
Allow unbatched render phases to use unstable sorts (#5049) # Objective Partially addresses #4291. Speed up the sort phase for unbatched render phases. ## Solution Split out one of the optimizations in #4899 and allow implementors of `PhaseItem` to change what kind of sort is used when sorting the items in the phase. This currently includes Stable, Unstable, and Unsorted. Each of these corresponds to `Vec::sort_by_key`, `Vec::sort_unstable_by_key`, and no sorting at all. The default is `Unstable`. The last one can be used as a default if users introduce a preliminary depth prepass. ## Performance This will not impact the performance of any batched phases, as it is still using a stable sort. 2D's only phase is unchanged. All 3D phases are unbatched currently, and will benefit from this change. On `many_cubes`, where the primary phase is opaque, this change sees a speed up from 907.02us -> 477.62us, a 47.35% reduction. ![image](https://user-images.githubusercontent.com/3137680/174471253-22424874-30d5-4db5-b5b4-65fb2c612a9c.png) ## Future Work There were prior discussions to add support for faster radix sorts in #4291, which in theory should be a `O(n)` instead of a `O(nlog(n))` time. [`voracious`](https://crates.io/crates/voracious_radix_sort) has been proposed, but it seems to be optimize for use cases with more than 30,000 items, which may be atypical for most systems. Another optimization included in #4899 is to reduce the size of a few of the IDs commonly used in `PhaseItem` implementations to shrink the types to make swapping/sorting faster. Both `CachedPipelineId` and `DrawFunctionId` could be reduced to `u32` instead of `usize`. Ideally, this should automatically change to use stable sorts when `BatchedPhaseItem` is implemented on the same phase item type, but this requires specialization, which may not land in stable Rust for a short while. --- ## Changelog Added: `PhaseItem::sort` ## Migration Guide RenderPhases now default to a unstable sort (via `slice::sort_unstable_by_key`). This can typically improve sort phase performance, but may produce incorrect batching results when implementing `BatchedPhaseItem`. To revert to the older stable sort, manually implement `PhaseItem::sort` to implement a stable sort (i.e. via `slice::sort_by_key`). Co-authored-by: Federico Rinaldi <gisquerin@gmail.com> Co-authored-by: Robert Swain <robert.swain@gmail.com> Co-authored-by: colepoirier <colepoirier@gmail.com>
2022-06-23 10:52:49 +00:00
}
Modular Rendering (#2831) This changes how render logic is composed to make it much more modular. Previously, all extraction logic was centralized for a given "type" of rendered thing. For example, we extracted meshes into a vector of ExtractedMesh, which contained the mesh and material asset handles, the transform, etc. We looked up bindings for "drawn things" using their index in the `Vec<ExtractedMesh>`. This worked fine for built in rendering, but made it hard to reuse logic for "custom" rendering. It also prevented us from reusing things like "extracted transforms" across contexts. To make rendering more modular, I made a number of changes: * Entities now drive rendering: * We extract "render components" from "app components" and store them _on_ entities. No more centralized uber lists! We now have true "ECS-driven rendering" * To make this perform well, I implemented #2673 in upstream Bevy for fast batch insertions into specific entities. This was merged into the `pipelined-rendering` branch here: #2815 * Reworked the `Draw` abstraction: * Generic `PhaseItems`: each draw phase can define its own type of "rendered thing", which can define its own "sort key" * Ported the 2d, 3d, and shadow phases to the new PhaseItem impl (currently Transparent2d, Transparent3d, and Shadow PhaseItems) * `Draw` trait and and `DrawFunctions` are now generic on PhaseItem * Modular / Ergonomic `DrawFunctions` via `RenderCommands` * RenderCommand is a trait that runs an ECS query and produces one or more RenderPass calls. Types implementing this trait can be composed to create a final DrawFunction. For example the DrawPbr DrawFunction is created from the following DrawCommand tuple. Const generics are used to set specific bind group locations: ```rust pub type DrawPbr = ( SetPbrPipeline, SetMeshViewBindGroup<0>, SetStandardMaterialBindGroup<1>, SetTransformBindGroup<2>, DrawMesh, ); ``` * The new `custom_shader_pipelined` example illustrates how the commands above can be reused to create a custom draw function: ```rust type DrawCustom = ( SetCustomMaterialPipeline, SetMeshViewBindGroup<0>, SetTransformBindGroup<2>, DrawMesh, ); ``` * ExtractComponentPlugin and UniformComponentPlugin: * Simple, standardized ways to easily extract individual components and write them to GPU buffers * Ported PBR and Sprite rendering to the new primitives above. * Removed staging buffer from UniformVec in favor of direct Queue usage * Makes UniformVec much easier to use and more ergonomic. Completely removes the need for custom render graph nodes in these contexts (see the PbrNode and view Node removals and the much simpler call patterns in the relevant Prepare systems). * Added a many_cubes_pipelined example to benchmark baseline 3d rendering performance and ensure there were no major regressions during this port. Avoiding regressions was challenging given that the old approach of extracting into centralized vectors is basically the "optimal" approach. However thanks to a various ECS optimizations and render logic rephrasing, we pretty much break even on this benchmark! * Lifetimeless SystemParams: this will be a bit divisive, but as we continue to embrace "trait driven systems" (ex: ExtractComponentPlugin, UniformComponentPlugin, DrawCommand), the ergonomics of `(Query<'static, 'static, (&'static A, &'static B, &'static)>, Res<'static, C>)` were getting very hard to bear. As a compromise, I added "static type aliases" for the relevant SystemParams. The previous example can now be expressed like this: `(SQuery<(Read<A>, Read<B>)>, SRes<C>)`. If anyone has better ideas / conflicting opinions, please let me know! * RunSystem trait: a way to define Systems via a trait with a SystemParam associated type. This is used to implement the various plugins mentioned above. I also added SystemParamItem and QueryItem type aliases to make "trait stye" ecs interactions nicer on the eyes (and fingers). * RenderAsset retrying: ensures that render assets are only created when they are "ready" and allows us to create bind groups directly inside render assets (which significantly simplified the StandardMaterial code). I think ultimately we should swap this out on "asset dependency" events to wait for dependencies to load, but this will require significant asset system changes. * Updated some built in shaders to account for missing MeshUniform fields
2021-09-23 06:16:11 +00:00
}
2021-06-02 02:59:17 +00:00
impl CachedRenderPipelinePhaseItem for Shadow {
#[inline]
fn cached_pipeline(&self) -> CachedRenderPipelineId {
self.pipeline
}
}
2021-06-02 02:59:17 +00:00
pub struct ShadowPassNode {
main_view_query: QueryState<&'static ViewLightEntities>,
view_light_query: QueryState<(&'static ShadowView, &'static RenderPhase<Shadow>)>,
2021-06-02 02:59:17 +00:00
}
impl ShadowPassNode {
pub fn new(world: &mut World) -> Self {
Self {
main_view_query: QueryState::new(world),
view_light_query: QueryState::new(world),
}
}
}
impl Node for ShadowPassNode {
fn update(&mut self, world: &mut World) {
self.main_view_query.update_archetypes(world);
self.view_light_query.update_archetypes(world);
}
fn run(
&self,
graph: &mut RenderGraphContext,
2021-06-21 23:28:52 +00:00
render_context: &mut RenderContext,
2021-06-02 02:59:17 +00:00
world: &World,
) -> Result<(), NodeRunError> {
Make render graph slots optional for most cases (#8109) # Objective - Currently, the render graph slots are only used to pass the view_entity around. This introduces significant boilerplate for very little value. Instead of using slots for this, make the view_entity part of the `RenderGraphContext`. This also means we won't need to have `IN_VIEW` on every node and and we'll be able to use the default impl of `Node::input()`. ## Solution - Add `view_entity: Option<Entity>` to the `RenderGraphContext` - Update all nodes to use this instead of entity slot input --- ## Changelog - Add optional `view_entity` to `RenderGraphContext` ## Migration Guide You can now get the view_entity directly from the `RenderGraphContext`. When implementing the Node: ```rust // 0.10 struct FooNode; impl FooNode { const IN_VIEW: &'static str = "view"; } impl Node for FooNode { fn input(&self) -> Vec<SlotInfo> { vec![SlotInfo::new(Self::IN_VIEW, SlotType::Entity)] } fn run( &self, graph: &mut RenderGraphContext, // ... ) -> Result<(), NodeRunError> { let view_entity = graph.get_input_entity(Self::IN_VIEW)?; // ... Ok(()) } } // 0.11 struct FooNode; impl Node for FooNode { fn run( &self, graph: &mut RenderGraphContext, // ... ) -> Result<(), NodeRunError> { let view_entity = graph.view_entity(); // ... Ok(()) } } ``` When adding the node to the graph, you don't need to specify a slot_edge for the view_entity. ```rust // 0.10 let mut graph = RenderGraph::default(); graph.add_node(FooNode::NAME, node); let input_node_id = draw_2d_graph.set_input(vec![SlotInfo::new( graph::input::VIEW_ENTITY, SlotType::Entity, )]); graph.add_slot_edge( input_node_id, graph::input::VIEW_ENTITY, FooNode::NAME, FooNode::IN_VIEW, ); // add_node_edge ... // 0.11 let mut graph = RenderGraph::default(); graph.add_node(FooNode::NAME, node); // add_node_edge ... ``` ## Notes This PR paired with #8007 will help reduce a lot of annoying boilerplate with the render nodes. Depending on which one gets merged first. It will require a bit of clean up work to make both compatible. I tagged this as a breaking change, because using the old system to get the view_entity will break things because it's not a node input slot anymore. ## Notes for reviewers A lot of the diffs are just removing the slots in every nodes and graph creation. The important part is mostly in the graph_runner/CameraDriverNode.
2023-03-21 20:11:13 +00:00
let view_entity = graph.view_entity();
2021-07-02 01:05:20 +00:00
if let Ok(view_lights) = self.main_view_query.get_manual(world, view_entity) {
for view_light_entity in view_lights.lights.iter().copied() {
let (view_light, shadow_phase) = self
.view_light_query
.get_manual(world, view_light_entity)
.unwrap();
if shadow_phase.items.is_empty() {
continue;
}
let mut render_pass =
render_context.begin_tracked_render_pass(RenderPassDescriptor {
label: Some(&view_light.pass_name),
color_attachments: &[],
depth_stencil_attachment: Some(RenderPassDepthStencilAttachment {
view: &view_light.depth_texture_view,
depth_ops: Some(Operations {
load: LoadOp::Clear(0.0),
store: true,
}),
stencil_ops: None,
}),
});
shadow_phase.render(&mut render_pass, world, view_light_entity);
}
2021-06-02 02:59:17 +00:00
}
Ok(())
}
}