bevy/crates/bevy_render/src/render_phase/mod.rs

//! The modular rendering abstraction responsible for queuing, preparing, sorting and drawing
//! entities as part of separate render phases.
//!
//! In Bevy each view (camera, or shadow-casting light, etc.) has one or multiple render phases
//! (e.g. opaque, transparent, shadow, etc).
//! They are used to queue entities for rendering.
//! Multiple phases might be required due to different sorting/batching behaviors
//! (e.g. opaque: front to back, transparent: back to front) or because one phase depends on
//! the rendered texture of the previous phase (e.g. for screen-space reflections).
//!
//! To draw an entity, a corresponding [`PhaseItem`] has to be added to one or multiple of these
//! render phases for each view that it is visible in.
//! This must be done in the [`RenderSystems::Queue`].
//! After that the render phase sorts them in the [`RenderSystems::PhaseSort`].
//! Finally the items are rendered using a single [`TrackedRenderPass`], during
//! the [`RenderSystems::Render`].
//!
//! Therefore each phase item is assigned a [`Draw`] function.
//! These set up the state of the [`TrackedRenderPass`] (i.e. select the
//! [`RenderPipeline`](crate::render_resource::RenderPipeline), configure the
//! [`BindGroup`](crate::render_resource::BindGroup)s, etc.) and then issue a draw call,
//! for the corresponding item.
//!
//! The [`Draw`] function trait can either be implemented directly or such a function can be
//! created by composing multiple [`RenderCommand`]s.

mod draw;
mod draw_state;
mod rangefinder;

use bevy_app::{App, Plugin};
use bevy_derive::{Deref, DerefMut};
use bevy_ecs::component::Tick;
use bevy_ecs::entity::EntityHash;
use bevy_platform::collections::{hash_map::Entry, HashMap};
use bevy_utils::default;
pub use draw::*;
pub use draw_state::*;
use encase::{internal::WriteInto, ShaderSize};
use fixedbitset::{Block, FixedBitSet};
use indexmap::IndexMap;
use nonmax::NonMaxU32;
pub use rangefinder::*;
use wgpu::Features;

use crate::batching::gpu_preprocessing::{
    GpuPreprocessingMode, GpuPreprocessingSupport, PhaseBatchedInstanceBuffers,
    PhaseIndirectParametersBuffers,
};
use crate::renderer::RenderDevice;
use crate::sync_world::{MainEntity, MainEntityHashMap};
use crate::view::RetainedViewEntity;
use crate::RenderDebugFlags;
use crate::{
    batching::{
        self,
        gpu_preprocessing::{self, BatchedInstanceBuffers},
        no_gpu_preprocessing::{self, BatchedInstanceBuffer},
        GetFullBatchData,
    },
    render_resource::{CachedRenderPipelineId, GpuArrayBufferIndex, PipelineCache},
    Render, RenderApp, RenderSystems,
};
use bevy_ecs::{
    prelude::*,
    system::{lifetimeless::SRes, SystemParamItem},
};
use core::{fmt::Debug, hash::Hash, iter, marker::PhantomData, ops::Range, slice::SliceIndex};
use smallvec::SmallVec;
use tracing::warn;

/// Stores the rendering instructions for a single phase that uses bins in all
/// views.
///
/// They're cleared out every frame, but storing them in a resource like this
/// allows us to reuse allocations.
#[derive(Resource, Deref, DerefMut)]
pub struct ViewBinnedRenderPhases<BPI>(pub HashMap<RetainedViewEntity, BinnedRenderPhase<BPI>>)
where
    BPI: BinnedPhaseItem;

/// A collection of all rendering instructions, that will be executed by the GPU, for a
/// single render phase for a single view.
///
/// Each view (camera, or shadow-casting light, etc.) can have one or multiple render phases.
/// They are used to queue entities for rendering.
/// Multiple phases might be required due to different sorting/batching behaviors
/// (e.g. opaque: front to back, transparent: back to front) or because one phase depends on
/// the rendered texture of the previous phase (e.g. for screen-space reflections).
/// All [`PhaseItem`]s are then rendered using a single [`TrackedRenderPass`].
/// The render pass might be reused for multiple phases to reduce GPU overhead.
///
/// This flavor of render phase is used for phases in which the ordering is less
/// critical: for example, `Opaque3d`. It's generally faster than the
/// alternative [`SortedRenderPhase`].
pub struct BinnedRenderPhase<BPI>
where
    BPI: BinnedPhaseItem,
{
    /// The multidrawable bins.
    ///
    /// Each batch set key maps to a *batch set*, which in this case is a set of
    /// meshes that can be drawn together in one multidraw call. Each batch set
    /// is subdivided into *bins*, each of which represents a particular mesh.
    /// Each bin contains the entity IDs of instances of that mesh.
    ///
    /// So, for example, if there are two cubes and a sphere present in the
    /// scene, we would generally have one batch set containing two bins,
    /// assuming that the cubes and sphere meshes are allocated together and use
    /// the same pipeline. The first bin, corresponding to the cubes, will have
    /// two entities in it. The second bin, corresponding to the sphere, will
    /// have one entity in it.
    pub multidrawable_meshes: IndexMap<BPI::BatchSetKey, IndexMap<BPI::BinKey, RenderBin>>,

    /// The bins corresponding to batchable items that aren't multidrawable.
    ///
    /// For multidrawable entities, use `multidrawable_meshes`; for
    /// unbatchable entities, use `unbatchable_values`.
    pub batchable_meshes: IndexMap<(BPI::BatchSetKey, BPI::BinKey), RenderBin>,

    /// The unbatchable bins.
    ///
    /// Each entity here is rendered in a separate drawcall.
    pub unbatchable_meshes: IndexMap<(BPI::BatchSetKey, BPI::BinKey), UnbatchableBinnedEntities>,

    /// Items in the bin that aren't meshes at all.
    ///
    /// Bevy itself doesn't place anything in this list, but plugins or your app
    /// can in order to execute custom drawing commands. Draw functions for each
    /// entity are simply called in order at rendering time.
    ///
    /// See the `custom_phase_item` example for an example of how to use this.
    pub non_mesh_items: IndexMap<(BPI::BatchSetKey, BPI::BinKey), NonMeshEntities>,

    /// Information on each batch set.
    ///
    /// A *batch set* is a set of entities that will be batched together unless
    /// we're on a platform that doesn't support storage buffers (e.g. WebGL 2)
    /// and differing dynamic uniform indices force us to break batches. On
    /// platforms that support storage buffers, a batch set always consists of
    /// at most one batch.
    ///
    /// Multidrawable entities come first, then batchable entities, then
    /// unbatchable entities.
    pub(crate) batch_sets: BinnedRenderPhaseBatchSets<BPI::BinKey>,

    /// The batch and bin key for each entity.
    ///
    /// We retain these so that, when the entity changes,
    /// [`Self::sweep_old_entities`] can quickly find the bin it was located in
    /// and remove it.
    cached_entity_bin_keys: IndexMap<MainEntity, CachedBinnedEntity<BPI>, EntityHash>,

    /// The set of indices in [`Self::cached_entity_bin_keys`] that are
    /// confirmed to be up to date.
    ///
    /// Note that each bit in this bit set refers to an *index* in the
    /// [`IndexMap`] (i.e. a bucket in the hash table). They aren't entity IDs.
    valid_cached_entity_bin_keys: FixedBitSet,

    /// The set of entities that changed bins this frame.
    ///
    /// An entity will only be present in this list if it was in one bin on the
    /// previous frame and is in a new bin on this frame. Each list entry
    /// specifies the bin the entity used to be in. We use this in order to
    /// remove the entity from the old bin during
    /// [`BinnedRenderPhase::sweep_old_entities`].
    entities_that_changed_bins: Vec<EntityThatChangedBins<BPI>>,
    /// The gpu preprocessing mode configured for the view this phase is associated
    /// with.
    gpu_preprocessing_mode: GpuPreprocessingMode,
}

/// All entities that share a mesh and a material and can be batched as part of
/// a [`BinnedRenderPhase`].
#[derive(Default)]
pub struct RenderBin {
    /// A list of the entities in each bin, along with their cached
    /// [`InputUniformIndex`].
    entities: IndexMap<MainEntity, InputUniformIndex, EntityHash>,
}

/// Information that we track about an entity that was in one bin on the
/// previous frame and is in a different bin this frame.
struct EntityThatChangedBins<BPI>
where
    BPI: BinnedPhaseItem,
{
    /// The entity.
    main_entity: MainEntity,
    /// The key that identifies the bin that this entity used to be in.
    old_cached_binned_entity: CachedBinnedEntity<BPI>,
}

/// Information that we keep about an entity currently within a bin.
pub struct CachedBinnedEntity<BPI>
where
    BPI: BinnedPhaseItem,
{
    /// Information that we use to identify a cached entity in a bin.
    pub cached_bin_key: Option<CachedBinKey<BPI>>,
    /// The last modified tick of the entity.
    ///
    /// We use this to detect when the entity needs to be invalidated.
    pub change_tick: Tick,
}

/// Information that we use to identify a cached entity in a bin.
pub struct CachedBinKey<BPI>
where
    BPI: BinnedPhaseItem,
{
    /// The key of the batch set containing the entity.
    pub batch_set_key: BPI::BatchSetKey,
    /// The key of the bin containing the entity.
    pub bin_key: BPI::BinKey,
    /// The type of render phase that we use to render the entity: multidraw,
    /// plain batch, etc.
    pub phase_type: BinnedRenderPhaseType,
}

impl<BPI> Clone for CachedBinnedEntity<BPI>
where
    BPI: BinnedPhaseItem,
{
    fn clone(&self) -> Self {
        CachedBinnedEntity {
            cached_bin_key: self.cached_bin_key.clone(),
            change_tick: self.change_tick,
        }
    }
}

impl<BPI> Clone for CachedBinKey<BPI>
where
    BPI: BinnedPhaseItem,
{
    fn clone(&self) -> Self {
        CachedBinKey {
            batch_set_key: self.batch_set_key.clone(),
            bin_key: self.bin_key.clone(),
            phase_type: self.phase_type,
        }
    }
}

impl<BPI> PartialEq for CachedBinKey<BPI>
where
    BPI: BinnedPhaseItem,
{
    fn eq(&self, other: &Self) -> bool {
        self.batch_set_key == other.batch_set_key
            && self.bin_key == other.bin_key
            && self.phase_type == other.phase_type
    }
}

/// How we store and render the batch sets.
///
/// Each one of these corresponds to a [`GpuPreprocessingMode`].
pub enum BinnedRenderPhaseBatchSets<BK> {
    /// Batches are grouped into batch sets based on dynamic uniforms.
    ///
    /// This corresponds to [`GpuPreprocessingMode::None`].
    DynamicUniforms(Vec<SmallVec<[BinnedRenderPhaseBatch; 1]>>),

    /// Batches are never grouped into batch sets.
    ///
    /// This corresponds to [`GpuPreprocessingMode::PreprocessingOnly`].
    Direct(Vec<BinnedRenderPhaseBatch>),

    /// Batches are grouped together into batch sets based on their ability to
    /// be multi-drawn together.
    ///
    /// This corresponds to [`GpuPreprocessingMode::Culling`].
    MultidrawIndirect(Vec<BinnedRenderPhaseBatchSet<BK>>),
}

/// A group of entities that will be batched together into a single multi-draw
/// call.
pub struct BinnedRenderPhaseBatchSet<BK> {
    /// The first batch in this batch set.
    pub(crate) first_batch: BinnedRenderPhaseBatch,
    /// The key of the bin that the first batch corresponds to.
    pub(crate) bin_key: BK,
    /// The number of batches.
    pub(crate) batch_count: u32,
    /// The index of the batch set in the GPU buffer.
    pub(crate) index: u32,
}

impl<BK> BinnedRenderPhaseBatchSets<BK> {
    fn clear(&mut self) {
        match *self {
            BinnedRenderPhaseBatchSets::DynamicUniforms(ref mut vec) => vec.clear(),
            BinnedRenderPhaseBatchSets::Direct(ref mut vec) => vec.clear(),
            BinnedRenderPhaseBatchSets::MultidrawIndirect(ref mut vec) => vec.clear(),
        }
    }
}

/// Information about a single batch of entities rendered using binned phase
/// items.
#[derive(Debug)]
pub struct BinnedRenderPhaseBatch {
    /// An entity that's *representative* of this batch.
    ///
    /// Bevy uses this to fetch the mesh. It can be any entity in the batch.
    pub representative_entity: (Entity, MainEntity),
    /// The range of instance indices in this batch.
    pub instance_range: Range<u32>,

    /// The dynamic offset of the batch.
    ///
    /// Note that dynamic offsets are only used on platforms that don't support
    /// storage buffers.
    pub extra_index: PhaseItemExtraIndex,
}

/// Information about the unbatchable entities in a bin.
pub struct UnbatchableBinnedEntities {
    /// The entities.
    pub entities: MainEntityHashMap<Entity>,

    /// The GPU array buffer indices of each unbatchable binned entity.
    pub(crate) buffer_indices: UnbatchableBinnedEntityIndexSet,
}

/// Information about [`BinnedRenderPhaseType::NonMesh`] entities.
pub struct NonMeshEntities {
    /// The entities.
    pub entities: MainEntityHashMap<Entity>,
}

/// Stores instance indices and dynamic offsets for unbatchable entities in a
/// binned render phase.
///
/// This is conceptually `Vec<UnbatchableBinnedEntityDynamicOffset>`, but it
/// avoids the overhead of storing dynamic offsets on platforms that support
/// them. In other words, this allows a fast path that avoids allocation on
/// platforms that aren't WebGL 2.
#[derive(Default)]

pub(crate) enum UnbatchableBinnedEntityIndexSet {
    /// There are no unbatchable entities in this bin (yet).
    #[default]
    NoEntities,

    /// The instances for all unbatchable entities in this bin are contiguous,
    /// and there are no dynamic uniforms.
    ///
    /// This is the typical case on platforms other than WebGL 2. We special
    /// case this to avoid allocation on those platforms.
    Sparse {
        /// The range of indices.
        instance_range: Range<u32>,
        /// The index of the first indirect instance parameters.
        ///
        /// The other indices immediately follow these.
        first_indirect_parameters_index: Option<NonMaxU32>,
    },

    /// Dynamic uniforms are present for unbatchable entities in this bin.
    ///
    /// We fall back to this on WebGL 2.
    Dense(Vec<UnbatchableBinnedEntityIndices>),
}

/// The instance index and dynamic offset (if present) for an unbatchable entity.
///
/// This is only useful on platforms that don't support storage buffers.
#[derive(Clone)]
pub(crate) struct UnbatchableBinnedEntityIndices {
    /// The instance index.
    pub(crate) instance_index: u32,
    /// The [`PhaseItemExtraIndex`], if present.
    pub(crate) extra_index: PhaseItemExtraIndex,
}

/// Identifies the list within [`BinnedRenderPhase`] that a phase item is to be
/// placed in.
#[derive(Clone, Copy, PartialEq, Debug)]
pub enum BinnedRenderPhaseType {
    /// The item is a mesh that's eligible for multi-draw indirect rendering and
    /// can be batched with other meshes of the same type.
    MultidrawableMesh,

    /// The item is a mesh that can be batched with other meshes of the same type and
    /// drawn in a single draw call.
    BatchableMesh,

    /// The item is a mesh that's eligible for indirect rendering, but can't be
    /// batched with other meshes of the same type.
    UnbatchableMesh,

    /// The item isn't a mesh at all.
    ///
    /// Bevy will simply invoke the drawing commands for such items one after
    /// another, with no further processing.
    ///
    /// The engine itself doesn't enqueue any items of this type, but it's
    /// available for use in your application and/or plugins.
    NonMesh,
}

impl<T> From<GpuArrayBufferIndex<T>> for UnbatchableBinnedEntityIndices
where
    T: Clone + ShaderSize + WriteInto,
{
    fn from(value: GpuArrayBufferIndex<T>) -> Self {
        UnbatchableBinnedEntityIndices {
            instance_index: value.index,
            extra_index: PhaseItemExtraIndex::maybe_dynamic_offset(value.dynamic_offset),
        }
    }
}

impl<BPI> Default for ViewBinnedRenderPhases<BPI>
where
    BPI: BinnedPhaseItem,
{
    fn default() -> Self {
        Self(default())
    }
}

impl<BPI> ViewBinnedRenderPhases<BPI>
where
    BPI: BinnedPhaseItem,
{
    pub fn prepare_for_new_frame(
        &mut self,
        retained_view_entity: RetainedViewEntity,
        gpu_preprocessing: GpuPreprocessingMode,
    ) {
        match self.entry(retained_view_entity) {
            Entry::Occupied(mut entry) => entry.get_mut().prepare_for_new_frame(),
            Entry::Vacant(entry) => {
                entry.insert(BinnedRenderPhase::<BPI>::new(gpu_preprocessing));
            }
        }
    }
}

/// The index of the uniform describing this object in the GPU buffer, when GPU
/// preprocessing is enabled.
///
/// For example, for 3D meshes, this is the index of the `MeshInputUniform` in
/// the buffer.
///
/// This field is ignored if GPU preprocessing isn't in use, such as (currently)
/// in the case of 2D meshes. In that case, it can be safely set to
/// [`core::default::Default::default`].
#[derive(Clone, Copy, PartialEq, Default, Deref, DerefMut)]
#[repr(transparent)]
pub struct InputUniformIndex(pub u32);

impl<BPI> BinnedRenderPhase<BPI>
where
    BPI: BinnedPhaseItem,
{
    /// Bins a new entity.
    ///
    /// The `phase_type` parameter specifies whether the entity is a
    /// preprocessable mesh and whether it can be binned with meshes of the same
    /// type.
    pub fn add(
        &mut self,
        batch_set_key: BPI::BatchSetKey,
        bin_key: BPI::BinKey,
        (entity, main_entity): (Entity, MainEntity),
        input_uniform_index: InputUniformIndex,
        mut phase_type: BinnedRenderPhaseType,
        change_tick: Tick,
    ) {
        // If the user has overridden indirect drawing for this view, we need to
        // force the phase type to be batchable instead.
        if self.gpu_preprocessing_mode == GpuPreprocessingMode::PreprocessingOnly
            && phase_type == BinnedRenderPhaseType::MultidrawableMesh
        {
            phase_type = BinnedRenderPhaseType::BatchableMesh;
        }

        match phase_type {
            BinnedRenderPhaseType::MultidrawableMesh => {
                match self.multidrawable_meshes.entry(batch_set_key.clone()) {
                    indexmap::map::Entry::Occupied(mut entry) => {
                        entry
                            .get_mut()
                            .entry(bin_key.clone())
                            .or_default()
                            .insert(main_entity, input_uniform_index);
                    }
                    indexmap::map::Entry::Vacant(entry) => {
                        let mut new_batch_set = IndexMap::default();
                        new_batch_set.insert(
                            bin_key.clone(),
                            RenderBin::from_entity(main_entity, input_uniform_index),
                        );
                        entry.insert(new_batch_set);
                    }
                }
            }

            BinnedRenderPhaseType::BatchableMesh => {
                match self
                    .batchable_meshes
                    .entry((batch_set_key.clone(), bin_key.clone()).clone())
                {
                    indexmap::map::Entry::Occupied(mut entry) => {
                        entry.get_mut().insert(main_entity, input_uniform_index);
                    }
                    indexmap::map::Entry::Vacant(entry) => {
                        entry.insert(RenderBin::from_entity(main_entity, input_uniform_index));
                    }
                }
            }

            BinnedRenderPhaseType::UnbatchableMesh => {
                match self
                    .unbatchable_meshes
                    .entry((batch_set_key.clone(), bin_key.clone()))
                {
                    indexmap::map::Entry::Occupied(mut entry) => {
                        entry.get_mut().entities.insert(main_entity, entity);
                    }
                    indexmap::map::Entry::Vacant(entry) => {
                        let mut entities = MainEntityHashMap::default();
                        entities.insert(main_entity, entity);
                        entry.insert(UnbatchableBinnedEntities {
                            entities,
                            buffer_indices: default(),
                        });
                    }
                }
            }

            BinnedRenderPhaseType::NonMesh => {
                // We don't process these items further.
                match self
                    .non_mesh_items
                    .entry((batch_set_key.clone(), bin_key.clone()).clone())
                {
                    indexmap::map::Entry::Occupied(mut entry) => {
                        entry.get_mut().entities.insert(main_entity, entity);
                    }
                    indexmap::map::Entry::Vacant(entry) => {
                        let mut entities = MainEntityHashMap::default();
                        entities.insert(main_entity, entity);
                        entry.insert(NonMeshEntities { entities });
                    }
                }
            }
        }

        // Update the cache.
        self.update_cache(
            main_entity,
            Some(CachedBinKey {
                batch_set_key,
                bin_key,
                phase_type,
            }),
            change_tick,
        );
    }

    /// Inserts an entity into the cache with the given change tick.
    pub fn update_cache(
        &mut self,
        main_entity: MainEntity,
        cached_bin_key: Option<CachedBinKey<BPI>>,
        change_tick: Tick,
    ) {
        let new_cached_binned_entity = CachedBinnedEntity {
            cached_bin_key,
            change_tick,
        };

        let (index, old_cached_binned_entity) = self
            .cached_entity_bin_keys
            .insert_full(main_entity, new_cached_binned_entity.clone());

        // If the entity changed bins, record its old bin so that we can remove
        // the entity from it.
        if let Some(old_cached_binned_entity) = old_cached_binned_entity {
            if old_cached_binned_entity.cached_bin_key != new_cached_binned_entity.cached_bin_key {
                self.entities_that_changed_bins.push(EntityThatChangedBins {
                    main_entity,
                    old_cached_binned_entity,
                });
            }
        }

        // Mark the entity as valid.
        self.valid_cached_entity_bin_keys.grow_and_insert(index);
    }

    /// Encodes the GPU commands needed to render all entities in this phase.
    pub fn render<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
    ) -> Result<(), DrawError> {
        {
            let draw_functions = world.resource::<DrawFunctions<BPI>>();
            let mut draw_functions = draw_functions.write();
            draw_functions.prepare(world);
            // Make sure to drop the reader-writer lock here to avoid recursive
            // locks.
        }

        self.render_batchable_meshes(render_pass, world, view)?;
        self.render_unbatchable_meshes(render_pass, world, view)?;
        self.render_non_meshes(render_pass, world, view)?;

        Ok(())
    }

    /// Renders all batchable meshes queued in this phase.
    fn render_batchable_meshes<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
    ) -> Result<(), DrawError> {
        let draw_functions = world.resource::<DrawFunctions<BPI>>();
        let mut draw_functions = draw_functions.write();

        let render_device = world.resource::<RenderDevice>();
        let multi_draw_indirect_count_supported = render_device
            .features()
            .contains(Features::MULTI_DRAW_INDIRECT_COUNT);

        match self.batch_sets {
            BinnedRenderPhaseBatchSets::DynamicUniforms(ref batch_sets) => {
                debug_assert_eq!(self.batchable_meshes.len(), batch_sets.len());

                for ((batch_set_key, bin_key), batch_set) in
                    self.batchable_meshes.keys().zip(batch_sets.iter())
                {
                    for batch in batch_set {
                        let binned_phase_item = BPI::new(
                            batch_set_key.clone(),
                            bin_key.clone(),
                            batch.representative_entity,
                            batch.instance_range.clone(),
                            batch.extra_index.clone(),
                        );

                        // Fetch the draw function.
                        let Some(draw_function) =
                            draw_functions.get_mut(binned_phase_item.draw_function())
                        else {
                            continue;
                        };

                        draw_function.draw(world, render_pass, view, &binned_phase_item)?;
                    }
                }
            }

            BinnedRenderPhaseBatchSets::Direct(ref batch_set) => {
                for (batch, (batch_set_key, bin_key)) in
                    batch_set.iter().zip(self.batchable_meshes.keys())
                {
                    let binned_phase_item = BPI::new(
                        batch_set_key.clone(),
                        bin_key.clone(),
                        batch.representative_entity,
                        batch.instance_range.clone(),
                        batch.extra_index.clone(),
                    );

                    // Fetch the draw function.
                    let Some(draw_function) =
                        draw_functions.get_mut(binned_phase_item.draw_function())
                    else {
                        continue;
                    };

                    draw_function.draw(world, render_pass, view, &binned_phase_item)?;
                }
            }

            BinnedRenderPhaseBatchSets::MultidrawIndirect(ref batch_sets) => {
                for (batch_set_key, batch_set) in self
                    .multidrawable_meshes
                    .keys()
                    .chain(
                        self.batchable_meshes
                            .keys()
                            .map(|(batch_set_key, _)| batch_set_key),
                    )
                    .zip(batch_sets.iter())
                {
                    let batch = &batch_set.first_batch;

                    let batch_set_index = if multi_draw_indirect_count_supported {
                        NonMaxU32::new(batch_set.index)
                    } else {
                        None
                    };

                    let binned_phase_item = BPI::new(
                        batch_set_key.clone(),
                        batch_set.bin_key.clone(),
                        batch.representative_entity,
                        batch.instance_range.clone(),
                        match batch.extra_index {
                            PhaseItemExtraIndex::None => PhaseItemExtraIndex::None,
                            PhaseItemExtraIndex::DynamicOffset(ref dynamic_offset) => {
                                PhaseItemExtraIndex::DynamicOffset(*dynamic_offset)
                            }
                            PhaseItemExtraIndex::IndirectParametersIndex { ref range, .. } => {
                                PhaseItemExtraIndex::IndirectParametersIndex {
                                    range: range.start..(range.start + batch_set.batch_count),
                                    batch_set_index,
                                }
                            }
                        },
                    );

                    // Fetch the draw function.
                    let Some(draw_function) =
                        draw_functions.get_mut(binned_phase_item.draw_function())
                    else {
                        continue;
                    };

                    draw_function.draw(world, render_pass, view, &binned_phase_item)?;
                }
            }
        }

        Ok(())
    }

    /// Renders all unbatchable meshes queued in this phase.
    fn render_unbatchable_meshes<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
    ) -> Result<(), DrawError> {
        let draw_functions = world.resource::<DrawFunctions<BPI>>();
        let mut draw_functions = draw_functions.write();

        for (batch_set_key, bin_key) in self.unbatchable_meshes.keys() {
            let unbatchable_entities =
                &self.unbatchable_meshes[&(batch_set_key.clone(), bin_key.clone())];
            for (entity_index, entity) in unbatchable_entities.entities.iter().enumerate() {
                let unbatchable_dynamic_offset = match &unbatchable_entities.buffer_indices {
                    UnbatchableBinnedEntityIndexSet::NoEntities => {
                        // Shouldn't happen…
                        continue;
                    }
                    UnbatchableBinnedEntityIndexSet::Sparse {
                        instance_range,
                        first_indirect_parameters_index,
                    } => UnbatchableBinnedEntityIndices {
                        instance_index: instance_range.start + entity_index as u32,
                        extra_index: match first_indirect_parameters_index {
                            None => PhaseItemExtraIndex::None,
                            Some(first_indirect_parameters_index) => {
                                let first_indirect_parameters_index_for_entity =
                                    u32::from(*first_indirect_parameters_index)
                                        + entity_index as u32;
                                PhaseItemExtraIndex::IndirectParametersIndex {
                                    range: first_indirect_parameters_index_for_entity
                                        ..(first_indirect_parameters_index_for_entity + 1),
                                    batch_set_index: None,
                                }
                            }
                        },
                    },
                    UnbatchableBinnedEntityIndexSet::Dense(dynamic_offsets) => {
                        dynamic_offsets[entity_index].clone()
                    }
                };

                let binned_phase_item = BPI::new(
                    batch_set_key.clone(),
                    bin_key.clone(),
                    (*entity.1, *entity.0),
                    unbatchable_dynamic_offset.instance_index
                        ..(unbatchable_dynamic_offset.instance_index + 1),
                    unbatchable_dynamic_offset.extra_index,
                );

                // Fetch the draw function.
                let Some(draw_function) = draw_functions.get_mut(binned_phase_item.draw_function())
                else {
                    continue;
                };

                draw_function.draw(world, render_pass, view, &binned_phase_item)?;
            }
        }
        Ok(())
    }

    /// Renders all objects of type [`BinnedRenderPhaseType::NonMesh`].
    ///
    /// These will have been added by plugins or the application.
    fn render_non_meshes<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
    ) -> Result<(), DrawError> {
        let draw_functions = world.resource::<DrawFunctions<BPI>>();
        let mut draw_functions = draw_functions.write();

        for ((batch_set_key, bin_key), non_mesh_entities) in &self.non_mesh_items {
            for (main_entity, entity) in non_mesh_entities.entities.iter() {
                // Come up with a fake batch range and extra index. The draw
                // function is expected to manage any sort of batching logic itself.
                let binned_phase_item = BPI::new(
                    batch_set_key.clone(),
                    bin_key.clone(),
                    (*entity, *main_entity),
                    0..1,
                    PhaseItemExtraIndex::None,
                );

                let Some(draw_function) = draw_functions.get_mut(binned_phase_item.draw_function())
                else {
                    continue;
                };

                draw_function.draw(world, render_pass, view, &binned_phase_item)?;
            }
        }

        Ok(())
    }

    pub fn is_empty(&self) -> bool {
        self.multidrawable_meshes.is_empty()
            && self.batchable_meshes.is_empty()
            && self.unbatchable_meshes.is_empty()
            && self.non_mesh_items.is_empty()
    }

    pub fn prepare_for_new_frame(&mut self) {
        self.batch_sets.clear();

        self.valid_cached_entity_bin_keys.clear();
        self.valid_cached_entity_bin_keys
            .grow(self.cached_entity_bin_keys.len());
        self.valid_cached_entity_bin_keys
            .set_range(self.cached_entity_bin_keys.len().., true);

        self.entities_that_changed_bins.clear();

        for unbatchable_bin in self.unbatchable_meshes.values_mut() {
            unbatchable_bin.buffer_indices.clear();
        }
    }

    /// Checks to see whether the entity is in a bin and returns true if it's
    /// both in a bin and up to date.
    ///
    /// If this function returns true, we also add the entry to the
    /// `valid_cached_entity_bin_keys` list.
    pub fn validate_cached_entity(
        &mut self,
        visible_entity: MainEntity,
        current_change_tick: Tick,
    ) -> bool {
        if let indexmap::map::Entry::Occupied(entry) =
            self.cached_entity_bin_keys.entry(visible_entity)
        {
            if entry.get().change_tick == current_change_tick {
                self.valid_cached_entity_bin_keys.insert(entry.index());
                return true;
            }
        }

        false
    }

    /// Removes all entities not marked as clean from the bins.
    ///
    /// During `queue_material_meshes`, we process all visible entities and mark
    /// each as clean as we come to it. Then, in [`sweep_old_entities`], we call
    /// this method, which removes entities that aren't marked as clean from the
    /// bins.
    pub fn sweep_old_entities(&mut self) {
        // Search for entities not marked as valid. We have to do this in
        // reverse order because `swap_remove_index` will potentially invalidate
        // all indices after the one we remove.
        for index in ReverseFixedBitSetZeroesIterator::new(&self.valid_cached_entity_bin_keys) {
            let Some((entity, cached_binned_entity)) =
                self.cached_entity_bin_keys.swap_remove_index(index)
            else {
                continue;
            };

            if let Some(ref cached_bin_key) = cached_binned_entity.cached_bin_key {
                remove_entity_from_bin(
                    entity,
                    cached_bin_key,
                    &mut self.multidrawable_meshes,
                    &mut self.batchable_meshes,
                    &mut self.unbatchable_meshes,
                    &mut self.non_mesh_items,
                );
            }
        }

        // If an entity changed bins, we need to remove it from its old bin.
        for entity_that_changed_bins in self.entities_that_changed_bins.drain(..) {
            let Some(ref old_cached_bin_key) = entity_that_changed_bins
                .old_cached_binned_entity
                .cached_bin_key
            else {
                continue;
            };
            remove_entity_from_bin(
                entity_that_changed_bins.main_entity,
                old_cached_bin_key,
                &mut self.multidrawable_meshes,
                &mut self.batchable_meshes,
                &mut self.unbatchable_meshes,
                &mut self.non_mesh_items,
            );
        }
    }
}

/// Removes an entity from a bin.
///
/// If this makes the bin empty, this function removes the bin as well.
///
/// This is a standalone function instead of a method on [`BinnedRenderPhase`]
/// for borrow check reasons.
fn remove_entity_from_bin<BPI>(
    entity: MainEntity,
    entity_bin_key: &CachedBinKey<BPI>,
    multidrawable_meshes: &mut IndexMap<BPI::BatchSetKey, IndexMap<BPI::BinKey, RenderBin>>,
    batchable_meshes: &mut IndexMap<(BPI::BatchSetKey, BPI::BinKey), RenderBin>,
    unbatchable_meshes: &mut IndexMap<(BPI::BatchSetKey, BPI::BinKey), UnbatchableBinnedEntities>,
    non_mesh_items: &mut IndexMap<(BPI::BatchSetKey, BPI::BinKey), NonMeshEntities>,
) where
    BPI: BinnedPhaseItem,
{
    match entity_bin_key.phase_type {
        BinnedRenderPhaseType::MultidrawableMesh => {
            if let indexmap::map::Entry::Occupied(mut batch_set_entry) =
                multidrawable_meshes.entry(entity_bin_key.batch_set_key.clone())
            {
                if let indexmap::map::Entry::Occupied(mut bin_entry) = batch_set_entry
                    .get_mut()
                    .entry(entity_bin_key.bin_key.clone())
                {
                    bin_entry.get_mut().remove(entity);

                    // If the bin is now empty, remove the bin.
                    if bin_entry.get_mut().is_empty() {
                        bin_entry.swap_remove();
                    }
                }

                // If the batch set is now empty, remove it. This will perturb
                // the order, but that's OK because we're going to sort the bin
                // afterwards.
                if batch_set_entry.get_mut().is_empty() {
                    batch_set_entry.swap_remove();
                }
            }
        }

        BinnedRenderPhaseType::BatchableMesh => {
            if let indexmap::map::Entry::Occupied(mut bin_entry) = batchable_meshes.entry((
                entity_bin_key.batch_set_key.clone(),
                entity_bin_key.bin_key.clone(),
            )) {
                bin_entry.get_mut().remove(entity);

                // If the bin is now empty, remove the bin.
                if bin_entry.get_mut().is_empty() {
                    bin_entry.swap_remove();
                }
            }
        }

        BinnedRenderPhaseType::UnbatchableMesh => {
            if let indexmap::map::Entry::Occupied(mut bin_entry) = unbatchable_meshes.entry((
                entity_bin_key.batch_set_key.clone(),
                entity_bin_key.bin_key.clone(),
            )) {
                bin_entry.get_mut().entities.remove(&entity);

                // If the bin is now empty, remove the bin.
                if bin_entry.get_mut().entities.is_empty() {
                    bin_entry.swap_remove();
                }
            }
        }

        BinnedRenderPhaseType::NonMesh => {
            if let indexmap::map::Entry::Occupied(mut bin_entry) = non_mesh_items.entry((
                entity_bin_key.batch_set_key.clone(),
                entity_bin_key.bin_key.clone(),
            )) {
                bin_entry.get_mut().entities.remove(&entity);

                // If the bin is now empty, remove the bin.
                if bin_entry.get_mut().entities.is_empty() {
                    bin_entry.swap_remove();
                }
            }
        }
    }
}

impl<BPI> BinnedRenderPhase<BPI>
where
    BPI: BinnedPhaseItem,
{
    fn new(gpu_preprocessing: GpuPreprocessingMode) -> Self {
        Self {
            multidrawable_meshes: IndexMap::default(),
            batchable_meshes: IndexMap::default(),
            unbatchable_meshes: IndexMap::default(),
            non_mesh_items: IndexMap::default(),
            batch_sets: match gpu_preprocessing {
                GpuPreprocessingMode::Culling => {
                    BinnedRenderPhaseBatchSets::MultidrawIndirect(vec![])
                }
                GpuPreprocessingMode::PreprocessingOnly => {
                    BinnedRenderPhaseBatchSets::Direct(vec![])
                }
                GpuPreprocessingMode::None => BinnedRenderPhaseBatchSets::DynamicUniforms(vec![]),
            },
            cached_entity_bin_keys: IndexMap::default(),
            valid_cached_entity_bin_keys: FixedBitSet::new(),
            entities_that_changed_bins: vec![],
            gpu_preprocessing_mode: gpu_preprocessing,
        }
    }
}

impl UnbatchableBinnedEntityIndexSet {
    /// Returns the [`UnbatchableBinnedEntityIndices`] for the given entity.
    fn indices_for_entity_index(
        &self,
        entity_index: u32,
    ) -> Option<UnbatchableBinnedEntityIndices> {
        match self {
            UnbatchableBinnedEntityIndexSet::NoEntities => None,
            UnbatchableBinnedEntityIndexSet::Sparse { instance_range, .. }
                if entity_index >= instance_range.len() as u32 =>
            {
                None
            }
            UnbatchableBinnedEntityIndexSet::Sparse {
                instance_range,
                first_indirect_parameters_index: None,
            } => Some(UnbatchableBinnedEntityIndices {
                instance_index: instance_range.start + entity_index,
                extra_index: PhaseItemExtraIndex::None,
            }),
            UnbatchableBinnedEntityIndexSet::Sparse {
                instance_range,
                first_indirect_parameters_index: Some(first_indirect_parameters_index),
            } => {
                let first_indirect_parameters_index_for_this_batch =
                    u32::from(*first_indirect_parameters_index) + entity_index;
                Some(UnbatchableBinnedEntityIndices {
                    instance_index: instance_range.start + entity_index,
                    extra_index: PhaseItemExtraIndex::IndirectParametersIndex {
                        range: first_indirect_parameters_index_for_this_batch
                            ..(first_indirect_parameters_index_for_this_batch + 1),
                        batch_set_index: None,
                    },
                })
            }
            UnbatchableBinnedEntityIndexSet::Dense(indices) => {
                indices.get(entity_index as usize).cloned()
            }
        }
    }
}

/// A convenient abstraction for adding all the systems necessary for a binned
/// render phase to the render app.
///
/// This is the version used when the pipeline supports GPU preprocessing: e.g.
/// 3D PBR meshes.
pub struct BinnedRenderPhasePlugin<BPI, GFBD>
where
    BPI: BinnedPhaseItem,
    GFBD: GetFullBatchData,
{
    /// Debugging flags that can optionally be set when constructing the renderer.
    pub debug_flags: RenderDebugFlags,
    phantom: PhantomData<(BPI, GFBD)>,
}

impl<BPI, GFBD> BinnedRenderPhasePlugin<BPI, GFBD>
where
    BPI: BinnedPhaseItem,
    GFBD: GetFullBatchData,
{
    pub fn new(debug_flags: RenderDebugFlags) -> Self {
        Self {
            debug_flags,
            phantom: PhantomData,
        }
    }
}

impl<BPI, GFBD> Plugin for BinnedRenderPhasePlugin<BPI, GFBD>
where
    BPI: BinnedPhaseItem,
    GFBD: GetFullBatchData + Sync + Send + 'static,
{
    fn build(&self, app: &mut App) {
        let Some(render_app) = app.get_sub_app_mut(RenderApp) else {
            return;
        };

        render_app
            .init_resource::<ViewBinnedRenderPhases<BPI>>()
            .init_resource::<PhaseBatchedInstanceBuffers<BPI, GFBD::BufferData>>()
            .insert_resource(PhaseIndirectParametersBuffers::<BPI>::new(
                self.debug_flags
                    .contains(RenderDebugFlags::ALLOW_COPIES_FROM_INDIRECT_PARAMETERS),
            ))
            .add_systems(
                Render,
                (
                    batching::sort_binned_render_phase::<BPI>.in_set(RenderSystems::PhaseSort),
                    (
                        no_gpu_preprocessing::batch_and_prepare_binned_render_phase::<BPI, GFBD>
                            .run_if(resource_exists::<BatchedInstanceBuffer<GFBD::BufferData>>),
                        gpu_preprocessing::batch_and_prepare_binned_render_phase::<BPI, GFBD>
                            .run_if(
                                resource_exists::<
                                    BatchedInstanceBuffers<GFBD::BufferData, GFBD::BufferInputData>,
                                >,
                            ),
                    )
                        .in_set(RenderSystems::PrepareResources),
                    sweep_old_entities::<BPI>.in_set(RenderSystems::QueueSweep),
                    gpu_preprocessing::collect_buffers_for_phase::<BPI, GFBD>
                        .run_if(
                            resource_exists::<
                                BatchedInstanceBuffers<GFBD::BufferData, GFBD::BufferInputData>,
                            >,
                        )
                        .in_set(RenderSystems::PrepareResourcesCollectPhaseBuffers),
                ),
            );
    }
}

/// Stores the rendering instructions for a single phase that sorts items in all
/// views.
///
/// They're cleared out every frame, but storing them in a resource like this
/// allows us to reuse allocations.
#[derive(Resource, Deref, DerefMut)]
pub struct ViewSortedRenderPhases<SPI>(pub HashMap<RetainedViewEntity, SortedRenderPhase<SPI>>)
where
    SPI: SortedPhaseItem;

impl<SPI> Default for ViewSortedRenderPhases<SPI>
where
    SPI: SortedPhaseItem,
{
    fn default() -> Self {
        Self(default())
    }
}

impl<SPI> ViewSortedRenderPhases<SPI>
where
    SPI: SortedPhaseItem,
{
    pub fn insert_or_clear(&mut self, retained_view_entity: RetainedViewEntity) {
        match self.entry(retained_view_entity) {
            Entry::Occupied(mut entry) => entry.get_mut().clear(),
            Entry::Vacant(entry) => {
                entry.insert(default());
            }
        }
    }
}

/// A convenient abstraction for adding all the systems necessary for a sorted
/// render phase to the render app.
///
/// This is the version used when the pipeline supports GPU preprocessing: e.g.
/// 3D PBR meshes.
pub struct SortedRenderPhasePlugin<SPI, GFBD>
where
    SPI: SortedPhaseItem,
    GFBD: GetFullBatchData,
{
    /// Debugging flags that can optionally be set when constructing the renderer.
    pub debug_flags: RenderDebugFlags,
    phantom: PhantomData<(SPI, GFBD)>,
}

impl<SPI, GFBD> SortedRenderPhasePlugin<SPI, GFBD>
where
    SPI: SortedPhaseItem,
    GFBD: GetFullBatchData,
{
    pub fn new(debug_flags: RenderDebugFlags) -> Self {
        Self {
            debug_flags,
            phantom: PhantomData,
        }
    }
}

impl<SPI, GFBD> Plugin for SortedRenderPhasePlugin<SPI, GFBD>
where
    SPI: SortedPhaseItem + CachedRenderPipelinePhaseItem,
    GFBD: GetFullBatchData + Sync + Send + 'static,
{
    fn build(&self, app: &mut App) {
        let Some(render_app) = app.get_sub_app_mut(RenderApp) else {
            return;
        };

        render_app
            .init_resource::<ViewSortedRenderPhases<SPI>>()
            .init_resource::<PhaseBatchedInstanceBuffers<SPI, GFBD::BufferData>>()
            .insert_resource(PhaseIndirectParametersBuffers::<SPI>::new(
                self.debug_flags
                    .contains(RenderDebugFlags::ALLOW_COPIES_FROM_INDIRECT_PARAMETERS),
            ))
            .add_systems(
                Render,
                (
                    (
                        no_gpu_preprocessing::batch_and_prepare_sorted_render_phase::<SPI, GFBD>
                            .run_if(resource_exists::<BatchedInstanceBuffer<GFBD::BufferData>>),
                        gpu_preprocessing::batch_and_prepare_sorted_render_phase::<SPI, GFBD>
                            .run_if(
                                resource_exists::<
                                    BatchedInstanceBuffers<GFBD::BufferData, GFBD::BufferInputData>,
                                >,
                            ),
                    )
                        .in_set(RenderSystems::PrepareResources),
                    gpu_preprocessing::collect_buffers_for_phase::<SPI, GFBD>
                        .run_if(
                            resource_exists::<
                                BatchedInstanceBuffers<GFBD::BufferData, GFBD::BufferInputData>,
                            >,
                        )
                        .in_set(RenderSystems::PrepareResourcesCollectPhaseBuffers),
                ),
            );
    }
}

impl UnbatchableBinnedEntityIndexSet {
    /// Adds a new entity to the list of unbatchable binned entities.
    pub fn add(&mut self, indices: UnbatchableBinnedEntityIndices) {
        match self {
            UnbatchableBinnedEntityIndexSet::NoEntities => {
                match indices.extra_index {
                    PhaseItemExtraIndex::DynamicOffset(_) => {
                        // This is the first entity we've seen, and we don't have
                        // compute shaders. Initialize an array.
                        *self = UnbatchableBinnedEntityIndexSet::Dense(vec![indices]);
                    }
                    PhaseItemExtraIndex::None => {
                        // This is the first entity we've seen, and we have compute
                        // shaders. Initialize the fast path.
                        *self = UnbatchableBinnedEntityIndexSet::Sparse {
                            instance_range: indices.instance_index..indices.instance_index + 1,
                            first_indirect_parameters_index: None,
                        }
                    }
                    PhaseItemExtraIndex::IndirectParametersIndex {
                        range: ref indirect_parameters_index,
                        ..
                    } => {
                        // This is the first entity we've seen, and we have compute
                        // shaders. Initialize the fast path.
                        *self = UnbatchableBinnedEntityIndexSet::Sparse {
                            instance_range: indices.instance_index..indices.instance_index + 1,
                            first_indirect_parameters_index: NonMaxU32::new(
                                indirect_parameters_index.start,
                            ),
                        }
                    }
                }
            }

            UnbatchableBinnedEntityIndexSet::Sparse {
                instance_range,
                first_indirect_parameters_index,
            } if instance_range.end == indices.instance_index
                && ((first_indirect_parameters_index.is_none()
                    && indices.extra_index == PhaseItemExtraIndex::None)
                    || first_indirect_parameters_index.is_some_and(
                        |first_indirect_parameters_index| match indices.extra_index {
                            PhaseItemExtraIndex::IndirectParametersIndex {
                                range: ref this_range,
                                ..
                            } => {
                                u32::from(first_indirect_parameters_index) + instance_range.end
                                    - instance_range.start
                                    == this_range.start
                            }
                            PhaseItemExtraIndex::DynamicOffset(_) | PhaseItemExtraIndex::None => {
                                false
                            }
                        },
                    )) =>
            {
                // This is the normal case on non-WebGL 2.
                instance_range.end += 1;
            }

            UnbatchableBinnedEntityIndexSet::Sparse { instance_range, .. } => {
                // We thought we were in non-WebGL 2 mode, but we got a dynamic
                // offset or non-contiguous index anyway. This shouldn't happen,
                // but let's go ahead and do the sensible thing anyhow: demote
                // the compressed `NoDynamicOffsets` field to the full
                // `DynamicOffsets` array.
                warn!(
                    "Unbatchable binned entity index set was demoted from sparse to dense. \
                    This is a bug in the renderer. Please report it.",
                );
                let new_dynamic_offsets = (0..instance_range.len() as u32)
                    .flat_map(|entity_index| self.indices_for_entity_index(entity_index))
                    .chain(iter::once(indices))
                    .collect();
                *self = UnbatchableBinnedEntityIndexSet::Dense(new_dynamic_offsets);
            }

            UnbatchableBinnedEntityIndexSet::Dense(dense_indices) => {
                dense_indices.push(indices);
            }
        }
    }

    /// Clears the unbatchable binned entity index set.
    fn clear(&mut self) {
        match self {
            UnbatchableBinnedEntityIndexSet::Dense(dense_indices) => dense_indices.clear(),
            UnbatchableBinnedEntityIndexSet::Sparse { .. } => {
                *self = UnbatchableBinnedEntityIndexSet::NoEntities;
            }
            _ => {}
        }
    }
}

/// A collection of all items to be rendered that will be encoded to GPU
/// commands for a single render phase for a single view.
///
/// Each view (camera, or shadow-casting light, etc.) can have one or multiple render phases.
/// They are used to queue entities for rendering.
/// Multiple phases might be required due to different sorting/batching behaviors
/// (e.g. opaque: front to back, transparent: back to front) or because one phase depends on
/// the rendered texture of the previous phase (e.g. for screen-space reflections).
/// All [`PhaseItem`]s are then rendered using a single [`TrackedRenderPass`].
/// The render pass might be reused for multiple phases to reduce GPU overhead.
///
/// This flavor of render phase is used only for meshes that need to be sorted
/// back-to-front, such as transparent meshes. For items that don't need strict
/// sorting, [`BinnedRenderPhase`] is preferred, for performance.
pub struct SortedRenderPhase<I>
where
    I: SortedPhaseItem,
{
    /// The items within this [`SortedRenderPhase`].
    pub items: Vec<I>,
}

impl<I> Default for SortedRenderPhase<I>
where
    I: SortedPhaseItem,
{
    fn default() -> Self {
        Self { items: Vec::new() }
    }
}

impl<I> SortedRenderPhase<I>
where
    I: SortedPhaseItem,
{
    /// Adds a [`PhaseItem`] to this render phase.
    #[inline]
    pub fn add(&mut self, item: I) {
        self.items.push(item);
    }

    /// Removes all [`PhaseItem`]s from this render phase.
    #[inline]
    pub fn clear(&mut self) {
        self.items.clear();
    }

    /// Sorts all of its [`PhaseItem`]s.
    pub fn sort(&mut self) {
        I::sort(&mut self.items);
    }

    /// An [`Iterator`] through the associated [`Entity`] for each [`PhaseItem`] in order.
    #[inline]
    pub fn iter_entities(&'_ self) -> impl Iterator<Item = Entity> + '_ {
        self.items.iter().map(PhaseItem::entity)
    }

    /// Renders all of its [`PhaseItem`]s using their corresponding draw functions.
    pub fn render<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
    ) -> Result<(), DrawError> {
        self.render_range(render_pass, world, view, ..)
    }

    /// Renders all [`PhaseItem`]s in the provided `range` (based on their index in `self.items`) using their corresponding draw functions.
    pub fn render_range<'w>(
        &self,
        render_pass: &mut TrackedRenderPass<'w>,
        world: &'w World,
        view: Entity,
        range: impl SliceIndex<[I], Output = [I]>,
    ) -> Result<(), DrawError> {
        let items = self
            .items
            .get(range)
            .expect("`Range` provided to `render_range()` is out of bounds");

        let draw_functions = world.resource::<DrawFunctions<I>>();
        let mut draw_functions = draw_functions.write();
        draw_functions.prepare(world);

        let mut index = 0;
        while index < items.len() {
            let item = &items[index];
            let batch_range = item.batch_range();
            if batch_range.is_empty() {
                index += 1;
            } else {
                let draw_function = draw_functions.get_mut(item.draw_function()).unwrap();
                draw_function.draw(world, render_pass, view, item)?;
                index += batch_range.len();
            }
        }
        Ok(())
    }
}

/// An item (entity of the render world) which will be drawn to a texture or the screen,
/// as part of a render phase.
///
/// The data required for rendering an entity is extracted from the main world in the
/// [`ExtractSchedule`](crate::ExtractSchedule).
/// Then it has to be queued up for rendering during the [`RenderSystems::Queue`],
/// by adding a corresponding phase item to a render phase.
/// Afterwards it will be possibly sorted and rendered automatically in the
/// [`RenderSystems::PhaseSort`] and [`RenderSystems::Render`], respectively.
///
/// `PhaseItem`s come in two flavors: [`BinnedPhaseItem`]s and
/// [`SortedPhaseItem`]s.
///
/// * Binned phase items have a `BinKey` which specifies what bin they're to be
///   placed in. All items in the same bin are eligible to be batched together.
///   The `BinKey`s are sorted, but the individual bin items aren't. Binned phase
///   items are good for opaque meshes, in which the order of rendering isn't
///   important. Generally, binned phase items are faster than sorted phase items.
///
/// * Sorted phase items, on the other hand, are placed into one large buffer
///   and then sorted all at once. This is needed for transparent meshes, which
///   have to be sorted back-to-front to render with the painter's algorithm.
///   These types of phase items are generally slower than binned phase items.
pub trait PhaseItem: Sized + Send + Sync + 'static {
    /// Whether or not this `PhaseItem` should be subjected to automatic batching. (Default: `true`)
    const AUTOMATIC_BATCHING: bool = true;

    /// The corresponding entity that will be drawn.
    ///
    /// This is used to fetch the render data of the entity, required by the draw function,
    /// from the render world .
    fn entity(&self) -> Entity;

    /// The main world entity represented by this `PhaseItem`.
    fn main_entity(&self) -> MainEntity;

    /// Specifies the [`Draw`] function used to render the item.
    fn draw_function(&self) -> DrawFunctionId;

    /// The range of instances that the batch covers. After doing a batched draw, batch range
    /// length phase items will be skipped. This design is to avoid having to restructure the
    /// render phase unnecessarily.
    fn batch_range(&self) -> &Range<u32>;
    fn batch_range_mut(&mut self) -> &mut Range<u32>;

    /// Returns the [`PhaseItemExtraIndex`].
    ///
    /// If present, this is either a dynamic offset or an indirect parameters
    /// index.
    fn extra_index(&self) -> PhaseItemExtraIndex;

    /// Returns a pair of mutable references to both the batch range and extra
    /// index.
    fn batch_range_and_extra_index_mut(&mut self) -> (&mut Range<u32>, &mut PhaseItemExtraIndex);
}

/// The "extra index" associated with some [`PhaseItem`]s, alongside the
/// indirect instance index.
///
/// Sometimes phase items require another index in addition to the range of
/// instances they already have. These can be:
///
/// * The *dynamic offset*: a `wgpu` dynamic offset into the uniform buffer of
///   instance data. This is used on platforms that don't support storage
///   buffers, to work around uniform buffer size limitations.
///
/// * The *indirect parameters index*: an index into the buffer that specifies
///   the indirect parameters for this [`PhaseItem`]'s drawcall. This is used when
///   indirect mode is on (as used for GPU culling).
///
/// Note that our indirect draw functionality requires storage buffers, so it's
/// impossible to have both a dynamic offset and an indirect parameters index.
/// This convenient fact allows us to pack both indices into a single `u32`.
#[derive(Clone, PartialEq, Eq, Hash, Debug)]
pub enum PhaseItemExtraIndex {
    /// No extra index is present.
    None,
    /// A `wgpu` dynamic offset into the uniform buffer of instance data. This
    /// is used on platforms that don't support storage buffers, to work around
    /// uniform buffer size limitations.
    DynamicOffset(u32),
    /// An index into the buffer that specifies the indirect parameters for this
    /// [`PhaseItem`]'s drawcall. This is used when indirect mode is on (as used
    /// for GPU culling).
    IndirectParametersIndex {
        /// The range of indirect parameters within the indirect parameters array.
        ///
        /// If we're using `multi_draw_indirect_count`, this specifies the
        /// maximum range of indirect parameters within that array. If batches
        /// are ultimately culled out on the GPU, the actual number of draw
        /// commands might be lower than the length of this range.
        range: Range<u32>,
        /// If `multi_draw_indirect_count` is in use, and this phase item is
        /// part of a batch set, specifies the index of the batch set that this
        /// phase item is a part of.
        ///
        /// If `multi_draw_indirect_count` isn't in use, or this phase item
        /// isn't part of a batch set, this is `None`.
        batch_set_index: Option<NonMaxU32>,
    },
}

impl PhaseItemExtraIndex {
    /// Returns either an indirect parameters index or
    /// [`PhaseItemExtraIndex::None`], as appropriate.
    pub fn maybe_indirect_parameters_index(
        indirect_parameters_index: Option<NonMaxU32>,
    ) -> PhaseItemExtraIndex {
        match indirect_parameters_index {
            Some(indirect_parameters_index) => PhaseItemExtraIndex::IndirectParametersIndex {
                range: u32::from(indirect_parameters_index)
                    ..(u32::from(indirect_parameters_index) + 1),
                batch_set_index: None,
            },
            None => PhaseItemExtraIndex::None,
        }
    }

    /// Returns either a dynamic offset index or [`PhaseItemExtraIndex::None`],
    /// as appropriate.
    pub fn maybe_dynamic_offset(dynamic_offset: Option<NonMaxU32>) -> PhaseItemExtraIndex {
        match dynamic_offset {
            Some(dynamic_offset) => PhaseItemExtraIndex::DynamicOffset(dynamic_offset.into()),
            None => PhaseItemExtraIndex::None,
        }
    }
}

/// Represents phase items that are placed into bins. The `BinKey` specifies
/// which bin they're to be placed in. Bin keys are sorted, and items within the
/// same bin are eligible to be batched together. The elements within the bins
/// aren't themselves sorted.
///
/// An example of a binned phase item is `Opaque3d`, for which the rendering
/// order isn't critical.
pub trait BinnedPhaseItem: PhaseItem {
    /// The key used for binning [`PhaseItem`]s into bins. Order the members of
    /// [`BinnedPhaseItem::BinKey`] by the order of binding for best
    /// performance. For example, pipeline id, draw function id, mesh asset id,
    /// lowest variable bind group id such as the material bind group id, and
    /// its dynamic offsets if any, next bind group and offsets, etc. This
    /// reduces the need for rebinding between bins and improves performance.
    type BinKey: Clone + Send + Sync + PartialEq + Eq + Ord + Hash;

    /// The key used to combine batches into batch sets.
    ///
    /// A *batch set* is a set of meshes that can potentially be multi-drawn
    /// together.
    type BatchSetKey: PhaseItemBatchSetKey;

    /// Creates a new binned phase item from the key and per-entity data.
    ///
    /// Unlike [`SortedPhaseItem`]s, this is generally called "just in time"
    /// before rendering. The resulting phase item isn't stored in any data
    /// structures, resulting in significant memory savings.
    fn new(
        batch_set_key: Self::BatchSetKey,
        bin_key: Self::BinKey,
        representative_entity: (Entity, MainEntity),
        batch_range: Range<u32>,
        extra_index: PhaseItemExtraIndex,
    ) -> Self;
}

/// A key used to combine batches into batch sets.
///
/// A *batch set* is a set of meshes that can potentially be multi-drawn
/// together.
pub trait PhaseItemBatchSetKey: Clone + Send + Sync + PartialEq + Eq + Ord + Hash {
    /// Returns true if this batch set key describes indexed meshes or false if
    /// it describes non-indexed meshes.
    ///
    /// Bevy uses this in order to determine which kind of indirect draw
    /// parameters to use, if indirect drawing is enabled.
    fn indexed(&self) -> bool;
}

/// Represents phase items that must be sorted. The `SortKey` specifies the
/// order that these items are drawn in. These are placed into a single array,
/// and the array as a whole is then sorted.
///
/// An example of a sorted phase item is `Transparent3d`, which must be sorted
/// back to front in order to correctly render with the painter's algorithm.
pub trait SortedPhaseItem: PhaseItem {
    /// The type used for ordering the items. The smallest values are drawn first.
    /// This order can be calculated using the [`ViewRangefinder3d`],
    /// based on the view-space `Z` value of the corresponding view matrix.
    type SortKey: Ord;

    /// Determines the order in which the items are drawn.
    fn sort_key(&self) -> Self::SortKey;

    /// Sorts a slice of phase items into render order. Generally if the same type
    /// is batched this should use a stable sort like [`slice::sort_by_key`].
    /// In almost all other cases, this should not be altered from the default,
    /// which uses an unstable sort, as this provides the best balance of CPU and GPU
    /// performance.
    ///
    /// Implementers can optionally not sort the list at all. This is generally advisable if and
    /// only if the renderer supports a depth prepass, which is by default not supported by
    /// the rest of Bevy's first party rendering crates. Even then, this may have a negative
    /// impact on GPU-side performance due to overdraw.
    ///
    /// It's advised to always profile for performance changes when changing this implementation.
    #[inline]
    fn sort(items: &mut [Self]) {
        items.sort_unstable_by_key(Self::sort_key);
    }

    /// Whether this phase item targets indexed meshes (those with both vertex
    /// and index buffers as opposed to just vertex buffers).
    ///
    /// Bevy needs this information in order to properly group phase items
    /// together for multi-draw indirect, because the GPU layout of indirect
    /// commands differs between indexed and non-indexed meshes.
    ///
    /// If you're implementing a custom phase item that doesn't describe a mesh,
    /// you can safely return false here.
    fn indexed(&self) -> bool;
}

/// A [`PhaseItem`] item, that automatically sets the appropriate render pipeline,
/// cached in the [`PipelineCache`].
///
/// You can use the [`SetItemPipeline`] render command to set the pipeline for this item.
pub trait CachedRenderPipelinePhaseItem: PhaseItem {
    /// The id of the render pipeline, cached in the [`PipelineCache`], that will be used to draw
    /// this phase item.
    fn cached_pipeline(&self) -> CachedRenderPipelineId;
}

/// A [`RenderCommand`] that sets the pipeline for the [`CachedRenderPipelinePhaseItem`].
pub struct SetItemPipeline;

impl<P: CachedRenderPipelinePhaseItem> RenderCommand<P> for SetItemPipeline {
    type Param = SRes<PipelineCache>;
    type ViewQuery = ();
    type ItemQuery = ();
    #[inline]
    fn render<'w>(
        item: &P,
        _view: (),
        _entity: Option<()>,
        pipeline_cache: SystemParamItem<'w, '_, Self::Param>,
        pass: &mut TrackedRenderPass<'w>,
    ) -> RenderCommandResult {
        if let Some(pipeline) = pipeline_cache
            .into_inner()
            .get_render_pipeline(item.cached_pipeline())
        {
            pass.set_render_pipeline(pipeline);
            RenderCommandResult::Success
        } else {
            RenderCommandResult::Skip
        }
    }
}

/// This system sorts the [`PhaseItem`]s of all [`SortedRenderPhase`]s of this
/// type.
pub fn sort_phase_system<I>(mut render_phases: ResMut<ViewSortedRenderPhases<I>>)
where
    I: SortedPhaseItem,
{
    for phase in render_phases.values_mut() {
        phase.sort();
    }
}

/// Removes entities that became invisible or changed phases from the bins.
///
/// This must run after queuing.
pub fn sweep_old_entities<BPI>(mut render_phases: ResMut<ViewBinnedRenderPhases<BPI>>)
where
    BPI: BinnedPhaseItem,
{
    for phase in render_phases.0.values_mut() {
        phase.sweep_old_entities();
    }
}

impl BinnedRenderPhaseType {
    pub fn mesh(
        batchable: bool,
        gpu_preprocessing_support: &GpuPreprocessingSupport,
    ) -> BinnedRenderPhaseType {
        match (batchable, gpu_preprocessing_support.max_supported_mode) {
            (true, GpuPreprocessingMode::Culling) => BinnedRenderPhaseType::MultidrawableMesh,
            (true, _) => BinnedRenderPhaseType::BatchableMesh,
            (false, _) => BinnedRenderPhaseType::UnbatchableMesh,
        }
    }
}

impl RenderBin {
    /// Creates a [`RenderBin`] containing a single entity.
    fn from_entity(entity: MainEntity, uniform_index: InputUniformIndex) -> RenderBin {
        let mut entities = IndexMap::default();
        entities.insert(entity, uniform_index);
        RenderBin { entities }
    }

    /// Inserts an entity into the bin.
    fn insert(&mut self, entity: MainEntity, uniform_index: InputUniformIndex) {
        self.entities.insert(entity, uniform_index);
    }

    /// Removes an entity from the bin.
    fn remove(&mut self, entity_to_remove: MainEntity) {
        self.entities.swap_remove(&entity_to_remove);
    }

    /// Returns true if the bin contains no entities.
    fn is_empty(&self) -> bool {
        self.entities.is_empty()
    }

    /// Returns the [`IndexMap`] containing all the entities in the bin, along
    /// with the cached [`InputUniformIndex`] of each.
    #[inline]
    pub fn entities(&self) -> &IndexMap<MainEntity, InputUniformIndex, EntityHash> {
        &self.entities
    }
}

/// An iterator that efficiently finds the indices of all zero bits in a
/// [`FixedBitSet`] and returns them in reverse order.
///
/// [`FixedBitSet`] doesn't natively offer this functionality, so we have to
/// implement it ourselves.
#[derive(Debug)]
struct ReverseFixedBitSetZeroesIterator<'a> {
    /// The bit set.
    bitset: &'a FixedBitSet,
    /// The next bit index we're going to scan when [`Iterator::next`] is
    /// called.
    bit_index: isize,
}

impl<'a> ReverseFixedBitSetZeroesIterator<'a> {
    fn new(bitset: &'a FixedBitSet) -> ReverseFixedBitSetZeroesIterator<'a> {
        ReverseFixedBitSetZeroesIterator {
            bitset,
            bit_index: (bitset.len() as isize) - 1,
        }
    }
}

impl<'a> Iterator for ReverseFixedBitSetZeroesIterator<'a> {
    type Item = usize;

    fn next(&mut self) -> Option<usize> {
        while self.bit_index >= 0 {
            // Unpack the bit index into block and bit.
            let block_index = self.bit_index / (Block::BITS as isize);
            let bit_pos = self.bit_index % (Block::BITS as isize);

            // Grab the block. Mask off all bits above the one we're scanning
            // from by setting them all to 1.
            let mut block = self.bitset.as_slice()[block_index as usize];
            if bit_pos + 1 < (Block::BITS as isize) {
                block |= (!0) << (bit_pos + 1);
            }

            // Search for the next unset bit. Note that the `leading_ones`
            // function counts from the MSB to the LSB, so we need to flip it to
            // get the bit number.
            let pos = (Block::BITS as isize) - (block.leading_ones() as isize) - 1;

            // If we found an unset bit, return it.
            if pos != -1 {
                let result = block_index * (Block::BITS as isize) + pos;
                self.bit_index = result - 1;
                return Some(result as usize);
            }

            // Otherwise, go to the previous block.
            self.bit_index = block_index * (Block::BITS as isize) - 1;
        }

        None
    }
}

#[cfg(test)]
mod test {
    use super::ReverseFixedBitSetZeroesIterator;
    use fixedbitset::FixedBitSet;
    use proptest::{collection::vec, prop_assert_eq, proptest};

    proptest! {
        #[test]
        fn reverse_fixed_bit_set_zeroes_iterator(
            bits in vec(0usize..1024usize, 0usize..1024usize),
            size in 0usize..1024usize,
        ) {
            // Build a random bit set.
            let mut bitset = FixedBitSet::new();
            bitset.grow(size);
            for bit in bits {
                if bit < size {
                    bitset.set(bit, true);
                }
            }

            // Iterate over the bit set backwards in a naive way, and check that
            // that iteration sequence corresponds to the optimized one.
            let mut iter = ReverseFixedBitSetZeroesIterator::new(&bitset);
            for bit_index in (0..size).rev() {
                if !bitset.contains(bit_index) {
                    prop_assert_eq!(iter.next(), Some(bit_index));
                }
            }

            prop_assert_eq!(iter.next(), None);
        }
    }
}