bevy/examples/shader/custom_render_phase.rs
Patrick Walton 0ede857103
Build batches across phases in parallel. (#17764)
Currently, invocations of `batch_and_prepare_binned_render_phase` and
`batch_and_prepare_sorted_render_phase` can't run in parallel because
they write to scene-global GPU buffers. After PR #17698,
`batch_and_prepare_binned_render_phase` started accounting for the
lion's share of the CPU time, causing us to be strongly CPU bound on
scenes like Caldera when occlusion culling was on (because of the
overhead of batching for the Z-prepass). Although I eventually plan to
optimize `batch_and_prepare_binned_render_phase`, we can obtain
significant wins now by parallelizing that system across phases.

This commit splits all GPU buffers that
`batch_and_prepare_binned_render_phase` and
`batch_and_prepare_sorted_render_phase` touches into separate buffers
for each phase so that the scheduler will run those phases in parallel.
At the end of batch preparation, we gather the render phases up into a
single resource with a new *collection* phase. Because we already run
mesh preprocessing separately for each phase in order to make occlusion
culling work, this is actually a cleaner separation. For example, mesh
output indices (the unique ID that identifies each mesh instance on GPU)
are now guaranteed to be sequential starting from 0, which will simplify
the forthcoming work to remove them in favor of the compute dispatch ID.

On Caldera, this brings the frame time down to approximately 9.1 ms with
occlusion culling on.

![Screenshot 2025-02-08
210720](https://github.com/user-attachments/assets/44bed500-e323-4786-b40c-828b75bc7d3f)
2025-02-13 00:02:20 +00:00

633 lines
23 KiB
Rust

//! This example demonstrates how to write a custom phase
//!
//! Render phases in bevy are used whenever you need to draw a group of meshes in a specific way.
//! For example, bevy's main pass has an opaque phase, a transparent phase for both 2d and 3d.
//! Sometimes, you may want to only draw a subset of meshes before or after the builtin phase. In
//! those situations you need to write your own phase.
//!
//! This example showcases how writing a custom phase to draw a stencil of a bevy mesh could look
//! like. Some shortcuts have been used for simplicity.
//!
//! This example was made for 3d, but a 2d equivalent would be almost identical.
use std::ops::Range;
use bevy::{
core_pipeline::core_3d::graph::{Core3d, Node3d},
ecs::{
query::QueryItem,
system::{lifetimeless::SRes, SystemParamItem},
},
math::FloatOrd,
pbr::{
DrawMesh, MeshInputUniform, MeshPipeline, MeshPipelineKey, MeshPipelineViewLayoutKey,
MeshUniform, RenderMeshInstances, SetMeshBindGroup, SetMeshViewBindGroup,
},
platform_support::collections::HashSet,
prelude::*,
render::{
batching::{
gpu_preprocessing::{
batch_and_prepare_sorted_render_phase, IndirectParametersMetadata,
UntypedPhaseIndirectParametersBuffers,
},
GetBatchData, GetFullBatchData,
},
camera::ExtractedCamera,
extract_component::{ExtractComponent, ExtractComponentPlugin},
mesh::{allocator::MeshAllocator, MeshVertexBufferLayoutRef, RenderMesh},
render_asset::RenderAssets,
render_graph::{
NodeRunError, RenderGraphApp, RenderGraphContext, RenderLabel, ViewNode, ViewNodeRunner,
},
render_phase::{
sort_phase_system, AddRenderCommand, CachedRenderPipelinePhaseItem, DrawFunctionId,
DrawFunctions, InputUniformIndex, PhaseItem, PhaseItemExtraIndex, SetItemPipeline,
SortedPhaseItem, ViewSortedRenderPhases,
},
render_resource::{
CachedRenderPipelineId, ColorTargetState, ColorWrites, Face, FragmentState, FrontFace,
MultisampleState, PipelineCache, PolygonMode, PrimitiveState, RenderPassDescriptor,
RenderPipelineDescriptor, SpecializedMeshPipeline, SpecializedMeshPipelineError,
SpecializedMeshPipelines, TextureFormat, VertexState,
},
renderer::RenderContext,
sync_world::MainEntity,
view::{ExtractedView, RenderVisibleEntities, RetainedViewEntity, ViewTarget},
Extract, Render, RenderApp, RenderSet,
},
};
use nonmax::NonMaxU32;
const SHADER_ASSET_PATH: &str = "shaders/custom_stencil.wgsl";
fn main() {
App::new()
.add_plugins((DefaultPlugins, MeshStencilPhasePlugin))
.add_systems(Startup, setup)
.run();
}
fn setup(
mut commands: Commands,
mut meshes: ResMut<Assets<Mesh>>,
mut materials: ResMut<Assets<StandardMaterial>>,
) {
// circular base
commands.spawn((
Mesh3d(meshes.add(Circle::new(4.0))),
MeshMaterial3d(materials.add(Color::WHITE)),
Transform::from_rotation(Quat::from_rotation_x(-std::f32::consts::FRAC_PI_2)),
));
// cube
// This cube will be rendered by the main pass, but it will also be rendered by our custom
// pass. This should result in an unlit red cube
commands.spawn((
Mesh3d(meshes.add(Cuboid::new(1.0, 1.0, 1.0))),
MeshMaterial3d(materials.add(Color::srgb_u8(124, 144, 255))),
Transform::from_xyz(0.0, 0.5, 0.0),
// This marker component is used to identify which mesh will be used in our custom pass
// The circle doesn't have it so it won't be rendered in our pass
DrawStencil,
));
// light
commands.spawn((
PointLight {
shadows_enabled: true,
..default()
},
Transform::from_xyz(4.0, 8.0, 4.0),
));
// camera
commands.spawn((
Camera3d::default(),
Transform::from_xyz(-2.0, 4.5, 9.0).looking_at(Vec3::ZERO, Vec3::Y),
// disable msaa for simplicity
Msaa::Off,
));
}
#[derive(Component, ExtractComponent, Clone, Copy, Default)]
struct DrawStencil;
struct MeshStencilPhasePlugin;
impl Plugin for MeshStencilPhasePlugin {
fn build(&self, app: &mut App) {
app.add_plugins((ExtractComponentPlugin::<DrawStencil>::default(),));
// We need to get the render app from the main app
let Some(render_app) = app.get_sub_app_mut(RenderApp) else {
return;
};
render_app
.init_resource::<SpecializedMeshPipelines<StencilPipeline>>()
.init_resource::<DrawFunctions<Stencil3d>>()
.add_render_command::<Stencil3d, DrawMesh3dStencil>()
.init_resource::<ViewSortedRenderPhases<Stencil3d>>()
.add_systems(ExtractSchedule, extract_camera_phases)
.add_systems(
Render,
(
queue_custom_meshes.in_set(RenderSet::QueueMeshes),
sort_phase_system::<Stencil3d>.in_set(RenderSet::PhaseSort),
batch_and_prepare_sorted_render_phase::<Stencil3d, StencilPipeline>
.in_set(RenderSet::PrepareResources),
),
);
render_app
.add_render_graph_node::<ViewNodeRunner<CustomDrawNode>>(Core3d, CustomDrawPassLabel)
// Tell the node to run after the main pass
.add_render_graph_edges(Core3d, (Node3d::MainOpaquePass, CustomDrawPassLabel));
}
fn finish(&self, app: &mut App) {
// We need to get the render app from the main app
let Some(render_app) = app.get_sub_app_mut(RenderApp) else {
return;
};
// The pipeline needs the RenderDevice to be created and it's only available once plugins
// are initialized
render_app.init_resource::<StencilPipeline>();
}
}
#[derive(Resource)]
struct StencilPipeline {
/// The base mesh pipeline defined by bevy
///
/// Since we want to draw a stencil of an existing bevy mesh we want to reuse the default
/// pipeline as much as possible
mesh_pipeline: MeshPipeline,
/// Stores the shader used for this pipeline directly on the pipeline.
/// This isn't required, it's only done like this for simplicity.
shader_handle: Handle<Shader>,
}
impl FromWorld for StencilPipeline {
fn from_world(world: &mut World) -> Self {
Self {
mesh_pipeline: MeshPipeline::from_world(world),
shader_handle: world.resource::<AssetServer>().load(SHADER_ASSET_PATH),
}
}
}
// For more information on how SpecializedMeshPipeline work, please look at the
// specialized_mesh_pipeline example
impl SpecializedMeshPipeline for StencilPipeline {
type Key = MeshPipelineKey;
fn specialize(
&self,
key: Self::Key,
layout: &MeshVertexBufferLayoutRef,
) -> Result<RenderPipelineDescriptor, SpecializedMeshPipelineError> {
// We will only use the position of the mesh in our shader so we only need to specify that
let mut vertex_attributes = Vec::new();
if layout.0.contains(Mesh::ATTRIBUTE_POSITION) {
// Make sure this matches the shader location
vertex_attributes.push(Mesh::ATTRIBUTE_POSITION.at_shader_location(0));
}
// This will automatically generate the correct `VertexBufferLayout` based on the vertex attributes
let vertex_buffer_layout = layout.0.get_layout(&vertex_attributes)?;
Ok(RenderPipelineDescriptor {
label: Some("Specialized Mesh Pipeline".into()),
// We want to reuse the data from bevy so we use the same bind groups as the default
// mesh pipeline
layout: vec![
// Bind group 0 is the view uniform
self.mesh_pipeline
.get_view_layout(MeshPipelineViewLayoutKey::from(key))
.clone(),
// Bind group 1 is the mesh uniform
self.mesh_pipeline.mesh_layouts.model_only.clone(),
],
push_constant_ranges: vec![],
vertex: VertexState {
shader: self.shader_handle.clone(),
shader_defs: vec![],
entry_point: "vertex".into(),
buffers: vec![vertex_buffer_layout],
},
fragment: Some(FragmentState {
shader: self.shader_handle.clone(),
shader_defs: vec![],
entry_point: "fragment".into(),
targets: vec![Some(ColorTargetState {
format: TextureFormat::bevy_default(),
blend: None,
write_mask: ColorWrites::ALL,
})],
}),
primitive: PrimitiveState {
topology: key.primitive_topology(),
front_face: FrontFace::Ccw,
cull_mode: Some(Face::Back),
polygon_mode: PolygonMode::Fill,
..default()
},
depth_stencil: None,
// It's generally recommended to specialize your pipeline for MSAA,
// but it's not always possible
multisample: MultisampleState::default(),
zero_initialize_workgroup_memory: false,
})
}
}
// We will reuse render commands already defined by bevy to draw a 3d mesh
type DrawMesh3dStencil = (
SetItemPipeline,
// This will set the view bindings in group 0
SetMeshViewBindGroup<0>,
// This will set the mesh bindings in group 1
SetMeshBindGroup<1>,
// This will draw the mesh
DrawMesh,
);
// This is the data required per entity drawn in a custom phase in bevy. More specifically this is the
// data required when using a ViewSortedRenderPhase. This would look differently if we wanted a
// batched render phase. Sorted phases are a bit easier to implement, but a batched phase would
// look similar.
//
// If you want to see how a batched phase implementation looks, you should look at the Opaque2d
// phase.
struct Stencil3d {
pub sort_key: FloatOrd,
pub entity: (Entity, MainEntity),
pub pipeline: CachedRenderPipelineId,
pub draw_function: DrawFunctionId,
pub batch_range: Range<u32>,
pub extra_index: PhaseItemExtraIndex,
/// Whether the mesh in question is indexed (uses an index buffer in
/// addition to its vertex buffer).
pub indexed: bool,
}
// For more information about writing a phase item, please look at the custom_phase_item example
impl PhaseItem for Stencil3d {
#[inline]
fn entity(&self) -> Entity {
self.entity.0
}
#[inline]
fn main_entity(&self) -> MainEntity {
self.entity.1
}
#[inline]
fn draw_function(&self) -> DrawFunctionId {
self.draw_function
}
#[inline]
fn batch_range(&self) -> &Range<u32> {
&self.batch_range
}
#[inline]
fn batch_range_mut(&mut self) -> &mut Range<u32> {
&mut self.batch_range
}
#[inline]
fn extra_index(&self) -> PhaseItemExtraIndex {
self.extra_index.clone()
}
#[inline]
fn batch_range_and_extra_index_mut(&mut self) -> (&mut Range<u32>, &mut PhaseItemExtraIndex) {
(&mut self.batch_range, &mut self.extra_index)
}
}
impl SortedPhaseItem for Stencil3d {
type SortKey = FloatOrd;
#[inline]
fn sort_key(&self) -> Self::SortKey {
self.sort_key
}
#[inline]
fn sort(items: &mut [Self]) {
// bevy normally uses radsort instead of the std slice::sort_by_key
// radsort is a stable radix sort that performed better than `slice::sort_by_key` or `slice::sort_unstable_by_key`.
// Since it is not re-exported by bevy, we just use the std sort for the purpose of the example
items.sort_by_key(SortedPhaseItem::sort_key);
}
#[inline]
fn indexed(&self) -> bool {
self.indexed
}
}
impl CachedRenderPipelinePhaseItem for Stencil3d {
#[inline]
fn cached_pipeline(&self) -> CachedRenderPipelineId {
self.pipeline
}
}
impl GetBatchData for StencilPipeline {
type Param = (
SRes<RenderMeshInstances>,
SRes<RenderAssets<RenderMesh>>,
SRes<MeshAllocator>,
);
type CompareData = AssetId<Mesh>;
type BufferData = MeshUniform;
fn get_batch_data(
(mesh_instances, _render_assets, mesh_allocator): &SystemParamItem<Self::Param>,
(_entity, main_entity): (Entity, MainEntity),
) -> Option<(Self::BufferData, Option<Self::CompareData>)> {
let RenderMeshInstances::CpuBuilding(ref mesh_instances) = **mesh_instances else {
error!(
"`get_batch_data` should never be called in GPU mesh uniform \
building mode"
);
return None;
};
let mesh_instance = mesh_instances.get(&main_entity)?;
let first_vertex_index =
match mesh_allocator.mesh_vertex_slice(&mesh_instance.mesh_asset_id) {
Some(mesh_vertex_slice) => mesh_vertex_slice.range.start,
None => 0,
};
let mesh_uniform = {
let mesh_transforms = &mesh_instance.transforms;
let (local_from_world_transpose_a, local_from_world_transpose_b) =
mesh_transforms.world_from_local.inverse_transpose_3x3();
MeshUniform {
world_from_local: mesh_transforms.world_from_local.to_transpose(),
previous_world_from_local: mesh_transforms.previous_world_from_local.to_transpose(),
lightmap_uv_rect: UVec2::ZERO,
local_from_world_transpose_a,
local_from_world_transpose_b,
flags: mesh_transforms.flags,
first_vertex_index,
current_skin_index: u32::MAX,
previous_skin_index: u32::MAX,
material_and_lightmap_bind_group_slot: 0,
tag: 0,
}
};
Some((mesh_uniform, None))
}
}
impl GetFullBatchData for StencilPipeline {
type BufferInputData = MeshInputUniform;
fn get_index_and_compare_data(
(mesh_instances, _, _): &SystemParamItem<Self::Param>,
main_entity: MainEntity,
) -> Option<(NonMaxU32, Option<Self::CompareData>)> {
// This should only be called during GPU building.
let RenderMeshInstances::GpuBuilding(ref mesh_instances) = **mesh_instances else {
error!(
"`get_index_and_compare_data` should never be called in CPU mesh uniform building \
mode"
);
return None;
};
let mesh_instance = mesh_instances.get(&main_entity)?;
Some((
mesh_instance.current_uniform_index,
mesh_instance
.should_batch()
.then_some(mesh_instance.mesh_asset_id),
))
}
fn get_binned_batch_data(
(mesh_instances, _render_assets, mesh_allocator): &SystemParamItem<Self::Param>,
main_entity: MainEntity,
) -> Option<Self::BufferData> {
let RenderMeshInstances::CpuBuilding(ref mesh_instances) = **mesh_instances else {
error!(
"`get_binned_batch_data` should never be called in GPU mesh uniform building mode"
);
return None;
};
let mesh_instance = mesh_instances.get(&main_entity)?;
let first_vertex_index =
match mesh_allocator.mesh_vertex_slice(&mesh_instance.mesh_asset_id) {
Some(mesh_vertex_slice) => mesh_vertex_slice.range.start,
None => 0,
};
Some(MeshUniform::new(
&mesh_instance.transforms,
first_vertex_index,
mesh_instance.material_bindings_index.slot,
None,
None,
None,
None,
))
}
fn write_batch_indirect_parameters_metadata(
mesh_index: InputUniformIndex,
indexed: bool,
base_output_index: u32,
batch_set_index: Option<NonMaxU32>,
indirect_parameters_buffers: &mut UntypedPhaseIndirectParametersBuffers,
indirect_parameters_offset: u32,
) {
// Note that `IndirectParameters` covers both of these structures, even
// though they actually have distinct layouts. See the comment above that
// type for more information.
let indirect_parameters = IndirectParametersMetadata {
mesh_index: *mesh_index,
base_output_index,
batch_set_index: match batch_set_index {
None => !0,
Some(batch_set_index) => u32::from(batch_set_index),
},
early_instance_count: 0,
late_instance_count: 0,
};
if indexed {
indirect_parameters_buffers
.set_indexed(indirect_parameters_offset, indirect_parameters);
} else {
indirect_parameters_buffers
.set_non_indexed(indirect_parameters_offset, indirect_parameters);
}
}
fn get_binned_index(
_param: &SystemParamItem<Self::Param>,
_query_item: MainEntity,
) -> Option<NonMaxU32> {
None
}
}
// When defining a phase, we need to extract it from the main world and add it to a resource
// that will be used by the render world. We need to give that resource all views that will use
// that phase
fn extract_camera_phases(
mut stencil_phases: ResMut<ViewSortedRenderPhases<Stencil3d>>,
cameras: Extract<Query<(Entity, &Camera), With<Camera3d>>>,
mut live_entities: Local<HashSet<RetainedViewEntity>>,
) {
live_entities.clear();
for (main_entity, camera) in &cameras {
if !camera.is_active {
continue;
}
// This is the main camera, so we use the first subview index (0)
let retained_view_entity = RetainedViewEntity::new(main_entity.into(), None, 0);
stencil_phases.insert_or_clear(retained_view_entity);
live_entities.insert(retained_view_entity);
}
// Clear out all dead views.
stencil_phases.retain(|camera_entity, _| live_entities.contains(camera_entity));
}
// This is a very important step when writing a custom phase.
//
// This system determines which meshes will be added to the phase.
fn queue_custom_meshes(
custom_draw_functions: Res<DrawFunctions<Stencil3d>>,
mut pipelines: ResMut<SpecializedMeshPipelines<StencilPipeline>>,
pipeline_cache: Res<PipelineCache>,
custom_draw_pipeline: Res<StencilPipeline>,
render_meshes: Res<RenderAssets<RenderMesh>>,
render_mesh_instances: Res<RenderMeshInstances>,
mut custom_render_phases: ResMut<ViewSortedRenderPhases<Stencil3d>>,
mut views: Query<(&ExtractedView, &RenderVisibleEntities, &Msaa)>,
has_marker: Query<(), With<DrawStencil>>,
) {
for (view, visible_entities, msaa) in &mut views {
let Some(custom_phase) = custom_render_phases.get_mut(&view.retained_view_entity) else {
continue;
};
let draw_custom = custom_draw_functions.read().id::<DrawMesh3dStencil>();
// Create the key based on the view.
// In this case we only care about MSAA and HDR
let view_key = MeshPipelineKey::from_msaa_samples(msaa.samples())
| MeshPipelineKey::from_hdr(view.hdr);
let rangefinder = view.rangefinder3d();
// Since our phase can work on any 3d mesh we can reuse the default mesh 3d filter
for (render_entity, visible_entity) in visible_entities.iter::<Mesh3d>() {
// We only want meshes with the marker component to be queued to our phase.
if has_marker.get(*render_entity).is_err() {
continue;
}
let Some(mesh_instance) = render_mesh_instances.render_mesh_queue_data(*visible_entity)
else {
continue;
};
let Some(mesh) = render_meshes.get(mesh_instance.mesh_asset_id) else {
continue;
};
// Specialize the key for the current mesh entity
// For this example we only specialize based on the mesh topology
// but you could have more complex keys and that's where you'd need to create those keys
let mut mesh_key = view_key;
mesh_key |= MeshPipelineKey::from_primitive_topology(mesh.primitive_topology());
let pipeline_id = pipelines.specialize(
&pipeline_cache,
&custom_draw_pipeline,
mesh_key,
&mesh.layout,
);
let pipeline_id = match pipeline_id {
Ok(id) => id,
Err(err) => {
error!("{}", err);
continue;
}
};
let distance = rangefinder.distance_translation(&mesh_instance.translation);
// At this point we have all the data we need to create a phase item and add it to our
// phase
custom_phase.add(Stencil3d {
// Sort the data based on the distance to the view
sort_key: FloatOrd(distance),
entity: (*render_entity, *visible_entity),
pipeline: pipeline_id,
draw_function: draw_custom,
// Sorted phase items aren't batched
batch_range: 0..1,
extra_index: PhaseItemExtraIndex::None,
indexed: mesh.indexed(),
});
}
}
}
// Render label used to order our render graph node that will render our phase
#[derive(RenderLabel, Debug, Clone, Hash, PartialEq, Eq)]
struct CustomDrawPassLabel;
#[derive(Default)]
struct CustomDrawNode;
impl ViewNode for CustomDrawNode {
type ViewQuery = (
&'static ExtractedCamera,
&'static ExtractedView,
&'static ViewTarget,
);
fn run<'w>(
&self,
graph: &mut RenderGraphContext,
render_context: &mut RenderContext<'w>,
(camera, view, target): QueryItem<'w, Self::ViewQuery>,
world: &'w World,
) -> Result<(), NodeRunError> {
// First, we need to get our phases resource
let Some(stencil_phases) = world.get_resource::<ViewSortedRenderPhases<Stencil3d>>() else {
return Ok(());
};
// Get the view entity from the graph
let view_entity = graph.view_entity();
// Get the phase for the current view running our node
let Some(stencil_phase) = stencil_phases.get(&view.retained_view_entity) else {
return Ok(());
};
// Render pass setup
let mut render_pass = render_context.begin_tracked_render_pass(RenderPassDescriptor {
label: Some("stencil pass"),
// For the purpose of the example, we will write directly to the view target. A real
// stencil pass would write to a custom texture and that texture would be used in later
// passes to render custom effects using it.
color_attachments: &[Some(target.get_color_attachment())],
// We don't bind any depth buffer for this pass
depth_stencil_attachment: None,
timestamp_writes: None,
occlusion_query_set: None,
});
if let Some(viewport) = camera.viewport.as_ref() {
render_pass.set_camera_viewport(viewport);
}
// Render the phase
// This will execute each draw functions of each phase items queued in this phase
if let Err(err) = stencil_phase.render(&mut render_pass, world, view_entity) {
error!("Error encountered while rendering the stencil phase {err:?}");
}
Ok(())
}
}