bevy_render/render_resource/

bind_group.rs

use crate::{
    material_bind_groups::FallbackBuffer,
    render_asset::RenderAssets,
    render_resource::{BindGroupLayout, Buffer, PipelineCache, Sampler, TextureView},
    renderer::{wgpu_wrapper, RenderDevice},
    storage::{GpuShaderBuffer, ShaderBuffer},
    texture::GpuImage,
};
use bevy_asset::Handle;
use bevy_derive::{Deref, DerefMut};
use bevy_ecs::system::{SystemParam, SystemParamItem};
use bevy_material::descriptor::BindGroupLayoutDescriptor;
pub use bevy_render_macros::AsBindGroup;
use bevy_utils::define_atomic_id;
use core::ops::Deref;
use encase::ShaderType;
use std::ops::Range;
use thiserror::Error;
use wgpu::{
    BindGroupEntry, BindGroupLayoutEntry, BindingResource, BufferBinding, SamplerBindingType,
    TextureViewDimension,
};

use super::{BindlessDescriptor, BindlessSlabResourceLimit};

define_atomic_id!(BindGroupId);

wgpu_wrapper! {
    #[derive(Clone, Debug)]
    struct WgpuBindGroup(wgpu::BindGroup);
}

/// Bind groups are responsible for binding render resources (e.g. buffers, textures, samplers)
/// to a [`TrackedRenderPass`](crate::render_phase::TrackedRenderPass).
/// This makes them accessible in the pipeline (shaders) as uniforms.
///
/// This is a lightweight thread-safe wrapper around wgpu's own [`BindGroup`](wgpu::BindGroup),
/// which can be cloned as needed to workaround lifetime management issues. It may be converted
/// from and dereferences to wgpu's [`BindGroup`](wgpu::BindGroup).
///
/// Can be created via [`RenderDevice::create_bind_group`](RenderDevice::create_bind_group).
#[derive(Clone, Debug)]
pub struct BindGroup {
    id: BindGroupId,
    value: WgpuBindGroup,
}

impl BindGroup {
    /// Returns the [`BindGroupId`] representing the unique ID of the bind group.
    #[inline]
    pub fn id(&self) -> BindGroupId {
        self.id
    }
}

impl PartialEq for BindGroup {
    fn eq(&self, other: &Self) -> bool {
        self.id == other.id
    }
}

impl Eq for BindGroup {}

impl core::hash::Hash for BindGroup {
    fn hash<H: core::hash::Hasher>(&self, state: &mut H) {
        self.id.0.hash(state);
    }
}

impl From<wgpu::BindGroup> for BindGroup {
    fn from(value: wgpu::BindGroup) -> Self {
        BindGroup {
            id: BindGroupId::new(),
            value: WgpuBindGroup::new(value),
        }
    }
}

impl<'a> From<&'a BindGroup> for Option<&'a wgpu::BindGroup> {
    fn from(value: &'a BindGroup) -> Self {
        Some(value.deref())
    }
}

impl<'a> From<&'a mut BindGroup> for Option<&'a wgpu::BindGroup> {
    fn from(value: &'a mut BindGroup) -> Self {
        Some(&*value)
    }
}

impl Deref for BindGroup {
    type Target = wgpu::BindGroup;

    #[inline]
    fn deref(&self) -> &Self::Target {
        &self.value
    }
}

/// Converts a value to a [`BindGroup`] with a given [`BindGroupLayout`], which can then be used in Bevy shaders.
/// This trait can be derived (and generally should be). Read on for details and examples.
///
/// This is an opinionated trait that is intended to make it easy to generically
/// convert a type into a [`BindGroup`]. It provides access to specific render resources,
/// such as [`RenderAssets<GpuImage>`] and [`crate::texture::FallbackImage`]. If a type has a [`Handle<Image>`](bevy_asset::Handle),
/// these can be used to retrieve the corresponding [`Texture`](crate::render_resource::Texture) resource.
///
/// [`AsBindGroup::as_bind_group`] is intended to be called once, then the result cached somewhere. It is generally
/// ok to do "expensive" work here, such as creating a [`Buffer`] for a uniform.
///
/// If for some reason a [`BindGroup`] cannot be created yet (for example, the [`Texture`](crate::render_resource::Texture)
/// for an [`Image`](bevy_image::Image) hasn't loaded yet), just return [`AsBindGroupError::RetryNextUpdate`], which signals that the caller
/// should retry again later.
///
/// # Deriving
///
/// This trait can be derived. Field attributes like `uniform` and `texture` are used to define which fields should be bindings,
/// what their binding type is, and what index they should be bound at:
///
/// ```
/// # use bevy_render::render_resource::*;
/// # use bevy_image::Image;
/// # use bevy_color::LinearRgba;
/// # use bevy_asset::Handle;
/// # use bevy_render::storage::ShaderBuffer;
///
/// #[derive(AsBindGroup)]
/// struct CoolMaterial {
///     #[uniform(0)]
///     color: LinearRgba,
///     #[texture(1)]
///     #[sampler(2)]
///     color_texture: Handle<Image>,
///     #[storage(3, read_only)]
///     storage_buffer: Handle<ShaderBuffer>,
///     #[storage(4, read_only, buffer)]
///     raw_buffer: Buffer,
///     #[storage_texture(5)]
///     storage_texture: Handle<Image>,
/// }
/// ```
///
/// In WGSL shaders, the binding would look like this:
///
/// ```wgsl
/// @group(#{MATERIAL_BIND_GROUP}) @binding(0) var<uniform> color: vec4<f32>;
/// @group(#{MATERIAL_BIND_GROUP}) @binding(1) var color_texture: texture_2d<f32>;
/// @group(#{MATERIAL_BIND_GROUP}) @binding(2) var color_sampler: sampler;
/// @group(#{MATERIAL_BIND_GROUP}) @binding(3) var<storage> storage_buffer: array<f32>;
/// @group(#{MATERIAL_BIND_GROUP}) @binding(4) var<storage> raw_buffer: array<f32>;
/// @group(#{MATERIAL_BIND_GROUP}) @binding(5) var storage_texture: texture_storage_2d<rgba8unorm, read_write>;
/// ```
/// Note that the "group" index is determined by the usage context. It is not defined in [`AsBindGroup`]. For example, in Bevy material bind groups
/// are generally bound to group 2.
///
/// The following field-level attributes are supported:
///
/// ## `uniform(BINDING_INDEX)`
///
///  * The field will be converted to a shader-compatible type using the [`ShaderType`] trait, written to a [`Buffer`], and bound as a uniform.
///    [`ShaderType`] is implemented for most math types already, such as [`f32`], [`Vec4`](bevy_math::Vec4), and
///    [`LinearRgba`](bevy_color::LinearRgba). It can also be derived for custom structs.
///
/// ## `texture(BINDING_INDEX, arguments)`
///
///  * This field's [`Handle<Image>`](bevy_asset::Handle) will be used to look up the matching [`Texture`](crate::render_resource::Texture)
///    GPU resource, which will be bound as a texture in shaders. The field will be assumed to implement [`Into<Option<Handle<Image>>>`]. In practice,
///    most fields should be a [`Handle<Image>`](bevy_asset::Handle) or [`Option<Handle<Image>>`]. If the value of an [`Option<Handle<Image>>`] is
///    [`None`], the [`crate::texture::FallbackImage`] resource will be used instead. This attribute can be used in conjunction with a `sampler` binding attribute
///    (with a different binding index) if a binding of the sampler for the [`Image`](bevy_image::Image) is also required.
///
/// | Arguments             | Values                                                                  | Default              |
/// |-----------------------|-------------------------------------------------------------------------|----------------------|
/// | `dimension` = "..."   | `"1d"`, `"2d"`, `"2d_array"`, `"3d"`, `"cube"`, `"cube_array"`          | `"2d"`               |
/// | `sample_type` = "..." | `"float"`, `"depth"`, `"s_int"` or `"u_int"`                            | `"float"`            |
/// | `filterable` = ...    | `true`, `false`                                                         | `true`               |
/// | `multisampled` = ...  | `true`, `false`                                                         | `false`              |
/// | `visibility(...)`     | `all`, `none`, or a list-combination of `vertex`, `fragment`, `compute` | `vertex`, `fragment` |
///
/// ## `storage_texture(BINDING_INDEX, arguments)`
///
/// * This field's [`Handle<Image>`](bevy_asset::Handle) will be used to look up the matching [`Texture`](crate::render_resource::Texture)
///   GPU resource, which will be bound as a storage texture in shaders. The field will be assumed to implement [`Into<Option<Handle<Image>>>`]. In practice,
///   most fields should be a [`Handle<Image>`](bevy_asset::Handle) or [`Option<Handle<Image>>`]. If the value of an [`Option<Handle<Image>>`] is
///   [`None`], the [`crate::texture::FallbackImage`] resource will be used instead.
///
/// | Arguments              | Values                                                                                     | Default       |
/// |------------------------|--------------------------------------------------------------------------------------------|---------------|
/// | `dimension` = "..."    | `"1d"`, `"2d"`, `"2d_array"`, `"3d"`, `"cube"`, `"cube_array"`                             | `"2d"`        |
/// | `image_format` = ...   | any member of [`TextureFormat`](crate::render_resource::TextureFormat)                     | `Rgba8Unorm`  |
/// | `access` = ...         | any member of [`StorageTextureAccess`](crate::render_resource::StorageTextureAccess)       | `ReadWrite`   |
/// | `visibility(...)`      | `all`, `none`, or a list-combination of `vertex`, `fragment`, `compute`                    | `compute`     |
///
/// ## `sampler(BINDING_INDEX, arguments)`
///
/// * This field's [`Handle<Image>`](bevy_asset::Handle) will be used to look up the matching [`Sampler`] GPU
///   resource, which will be bound as a sampler in shaders. The field will be assumed to implement [`Into<Option<Handle<Image>>>`]. In practice,
///   most fields should be a [`Handle<Image>`](bevy_asset::Handle) or [`Option<Handle<Image>>`]. If the value of an [`Option<Handle<Image>>`] is
///   [`None`], the [`crate::texture::FallbackImage`] resource will be used instead. This attribute can be used in conjunction with a `texture` binding attribute
///   (with a different binding index) if a binding of the texture for the [`Image`](bevy_image::Image) is also required.
///
/// | Arguments              | Values                                                                  | Default                |
/// |------------------------|-------------------------------------------------------------------------|------------------------|
/// | `sampler_type` = "..." | `"filtering"`, `"non_filtering"`, `"comparison"`.                       |  `"filtering"`         |
/// | `visibility(...)`      | `all`, `none`, or a list-combination of `vertex`, `fragment`, `compute` |   `vertex`, `fragment` |
///
/// ## `storage(BINDING_INDEX, arguments)`
///
/// * The field's [`Handle<Storage>`](bevy_asset::Handle) will be used to look
///   up the matching [`Buffer`] GPU resource, which will be bound as a storage
///   buffer in shaders. If the `storage` attribute is used, the field is expected
///   a raw buffer, and the buffer will be bound as a storage buffer in shaders.
///   In bindless mode, `binding_array()` argument that specifies the binding
///   number of the resulting storage buffer binding array must be present.
///
/// | Arguments              | Values                                                                  | Default                |
/// |------------------------|-------------------------------------------------------------------------|------------------------|
/// | `visibility(...)`      | `all`, `none`, or a list-combination of `vertex`, `fragment`, `compute` | `vertex`, `fragment`   |
/// | `read_only`            | if present then value is true, otherwise false                          | `false`                |
/// | `buffer`               | if present then the field will be assumed to be a raw wgpu buffer       |                        |
/// | `binding_array(...)`   | the binding number of the binding array, for bindless mode              | bindless mode disabled |
///
/// Note that fields without field-level binding attributes will be ignored.
/// ```
/// # use bevy_render::{render_resource::AsBindGroup};
/// # use bevy_color::LinearRgba;
/// # use bevy_asset::Handle;
/// #[derive(AsBindGroup)]
/// struct CoolMaterial {
///     #[uniform(0)]
///     color: LinearRgba,
///     this_field_is_ignored: String,
/// }
/// ```
///
///  As mentioned above, [`Option<Handle<Image>>`] is also supported:
/// ```
/// # use bevy_asset::Handle;
/// # use bevy_color::LinearRgba;
/// # use bevy_image::Image;
/// # use bevy_render::render_resource::AsBindGroup;
/// #[derive(AsBindGroup)]
/// struct CoolMaterial {
///     #[uniform(0)]
///     color: LinearRgba,
///     #[texture(1)]
///     #[sampler(2)]
///     color_texture: Option<Handle<Image>>,
/// }
/// ```
/// This is useful if you want a texture to be optional. When the value is [`None`], the [`crate::texture::FallbackImage`] will be used for the binding instead, which defaults
/// to "pure white".
///
/// Field uniforms with the same index will be combined into a single binding:
/// ```
/// # use bevy_render::{render_resource::AsBindGroup};
/// # use bevy_color::LinearRgba;
/// #[derive(AsBindGroup)]
/// struct CoolMaterial {
///     #[uniform(0)]
///     color: LinearRgba,
///     #[uniform(0)]
///     roughness: f32,
/// }
/// ```
///
/// In WGSL shaders, the binding would look like this:
/// ```wgsl
/// struct CoolMaterial {
///     color: vec4<f32>,
///     roughness: f32,
/// };
///
/// @group(#{MATERIAL_BIND_GROUP}) @binding(0) var<uniform> material: CoolMaterial;
/// ```
///
/// Some less common scenarios will require "struct-level" attributes. These are the currently supported struct-level attributes:
/// ## `uniform(BINDING_INDEX, ConvertedShaderType)`
///
/// * This also creates a [`Buffer`] using [`ShaderType`] and binds it as a
///   uniform, much like the field-level `uniform` attribute. The difference is
///   that the entire [`AsBindGroup`] value is converted to `ConvertedShaderType`,
///   which must implement [`ShaderType`], instead of a specific field
///   implementing [`ShaderType`]. This is useful if more complicated conversion
///   logic is required, or when using bindless mode (see below). The conversion
///   is done using the [`AsBindGroupShaderType<ConvertedShaderType>`] trait,
///   which is automatically implemented if `&Self` implements
///   [`Into<ConvertedShaderType>`]. Outside of bindless mode, only use
///   [`AsBindGroupShaderType`] if access to resources like
///   [`RenderAssets<GpuImage>`] is required.
///
/// * In bindless mode (see `bindless(COUNT)`), this attribute becomes
///   `uniform(BINDLESS_INDEX, ConvertedShaderType,
///   binding_array(BINDING_INDEX))`. The resulting uniform buffers will be
///   available in the shader as a binding array at the given `BINDING_INDEX`. The
///   `BINDLESS_INDEX` specifies the offset of the buffer in the bindless index
///   table.
///
///   For example, suppose that the material slot is stored in a variable named
///   `slot`, the bindless index table is named `material_indices`, and that the
///   first field (index 0) of the bindless index table type is named
///   `material`. Then specifying `#[uniform(0, StandardMaterialUniform,
///   binding_array(10)]` will create a binding array buffer declared in the
///   shader as `var<storage> material_array:
///   binding_array<StandardMaterialUniform>` and accessible as
///   `material_array[material_indices[slot].material]`.
///
/// ## `data(BINDING_INDEX, ConvertedShaderType, binding_array(BINDING_INDEX))`
///
/// * This is very similar to `uniform(BINDING_INDEX, ConvertedShaderType,
///   binding_array(BINDING_INDEX)` and in fact is identical if bindless mode
///   isn't being used. The difference is that, in bindless mode, the `data`
///   attribute produces a single buffer containing an array, not an array of
///   buffers. For example, suppose you had the following declaration:
///
/// ```ignore
/// #[uniform(0, StandardMaterialUniform, binding_array(10))]
/// struct StandardMaterial { ... }
/// ```
///
/// In bindless mode, this will produce a binding matching the following WGSL
/// declaration:
///
/// ```wgsl
/// @group(#{MATERIAL_BIND_GROUP}) @binding(10) var<storage> material_array: binding_array<StandardMaterial>;
/// ```
///
/// On the other hand, if you write this declaration:
///
/// ```ignore
/// #[data(0, StandardMaterialUniform, binding_array(10))]
/// struct StandardMaterial { ... }
/// ```
///
/// Then Bevy produces a binding that matches this WGSL declaration instead:
///
/// ```wgsl
/// @group(#{MATERIAL_BIND_GROUP}) @binding(10) var<storage> material_array: array<StandardMaterial>;
/// ```
///
/// * Just as with the structure-level `uniform` attribute, Bevy converts the
///   entire [`AsBindGroup`] to `ConvertedShaderType`, using the
///   [`AsBindGroupShaderType<ConvertedShaderType>`] trait.
///
/// * In non-bindless mode, the structure-level `data` attribute is the same as
///   the structure-level `uniform` attribute and produces a single uniform buffer
///   in the shader. The above example would result in a binding that looks like
///   this in WGSL in non-bindless mode:
///
/// ```wgsl
/// @group(#{MATERIAL_BIND_GROUP}) @binding(0) var<uniform> material: StandardMaterial;
/// ```
///
/// * For efficiency reasons, `data` is generally preferred over `uniform`
///   unless you need to place your data in individual buffers.
///
/// ## `bind_group_data(DataType)`
///
/// * The [`AsBindGroup`] type will be converted to some `DataType` using [`Into<DataType>`] and stored
///   as [`AsBindGroup::Data`] as part of the [`AsBindGroup::as_bind_group`] call. This is useful if data needs to be stored alongside
///   the generated bind group, such as a unique identifier for a material's bind group. The most common use case for this attribute
///   is "shader pipeline specialization". See [`SpecializedRenderPipeline`](crate::render_resource::SpecializedRenderPipeline).
///
/// ## `bindless`
///
/// * This switch enables *bindless resources*, which changes the way Bevy
///   supplies resources (textures, and samplers) to the shader.  When bindless
///   resources are enabled, and the current platform supports them, Bevy will
///   allocate textures, and samplers into *binding arrays*, separated based on
///   type and will supply your shader with indices into those arrays.
/// * Bindless textures and samplers are placed into the appropriate global
///   array defined in `bevy_render::bindless` (`bindless.wesl`).
/// * Bevy doesn't currently support bindless buffers, except for those created
///   with the `uniform(BINDLESS_INDEX, ConvertedShaderType,
///   binding_array(BINDING_INDEX))` attribute. If you need to include a buffer in
///   your object, and you can't create the data in that buffer with the `uniform`
///   attribute, consider a non-bindless object instead.
/// * If bindless mode is enabled, the `BINDLESS` definition will be
///   available. Because not all platforms support bindless resources, you
///   should check for the presence of this definition via `@if` and fall
///   back to standard bindings if it isn't present.
/// * By default, in bindless mode, binding 0 becomes the *bindless index
///   table*, which is an array of structures, each of which contains as many
///   fields of type `u32` as the highest binding number in the structure
///   annotated with `#[derive(AsBindGroup)]`. Again by default, the *i*th field
///   of the bindless index table contains the index of the resource with binding
///   *i* within the appropriate binding array.
/// * In the case of materials, the index of the applicable table within the
///   bindless index table list corresponding to the mesh currently being drawn
///   can be retrieved with
///   `mesh[in.instance_index].material_and_lightmap_bind_group_slot & 0xffffu`.
/// * You can limit the size of the bindless slabs to N resources with the
///   `limit(N)` declaration. For example, `#[bindless(limit(16))]` ensures that
///   each slab will have no more than 16 total resources in it. If you don't
///   specify a limit, Bevy automatically picks a reasonable one for the current
///   platform.
/// * The `index_table(range(M..N), binding(B))` declaration allows you to
///   customize the layout of the bindless index table. This is useful for
///   materials that are composed of multiple bind groups, such as
///   `ExtendedMaterial`. In such cases, there will be multiple bindless index
///   tables, so they can't both be assigned to binding 0 or their bindings will
///   conflict.
///   - The `binding(B)` attribute of the `index_table` attribute allows you to
///     customize the binding (`@binding(B)`, in the shader) at which the index
///     table will be bound.
///   - The `range(M, N)` attribute of the `index_table` attribute allows you to
///     change the mapping from the field index in the bindless index table to the
///     bindless index. Instead of the field at index $i$ being mapped to the
///     bindless index $i$, with the `range(M, N)` attribute the field at index
///     $i$ in the bindless index table is mapped to the bindless index $i$ + M.
///     The size of the index table will be set to N - M. Note that this may
///     result in the table being too small to contain all the bindless bindings.
/// * The purpose of bindless mode is to improve performance by reducing
///   state changes. By grouping resources together into binding arrays, Bevy
///   doesn't have to modify GPU state as often, decreasing API and driver
///   overhead.
/// * See the `shaders/shader_material_bindless` example for an example of how
///   to use bindless mode. See the `shaders/extended_material_bindless` example
///   for a more exotic example of bindless mode that demonstrates the
///   `index_table` attribute.
/// * The following diagram illustrates how bindless mode works using a subset
///   of `StandardMaterial`:
///
/// ```text
///      Shader Bindings                          Sampler Binding Array
///     +----+-----------------------------+     +-----------+-----------+-----+
/// +---|  0 | material_indices            |  +->| sampler 0 | sampler 1 | ... |
/// |   +----+-----------------------------+  |  +-----------+-----------+-----+
/// |   |  1 | bindless_samplers_filtering +--+        ^
/// |   +----+-----------------------------+           +-------------------------------+
/// |   | .. |            ...              |                                           |
/// |   +----+-----------------------------+      Texture Binding Array                |
/// |   |  5 | bindless_textures_2d        +--+  +-----------+-----------+-----+       |
/// |   +----+-----------------------------+  +->| texture 0 | texture 1 | ... |       |
/// |   | .. |            ...              |     +-----------+-----------+-----+       |
/// |   +----+-----------------------------+           ^                               |
/// |   + 10 | material_array              +--+        +---------------------------+   |
/// |   +----+-----------------------------+  |                                    |   |
/// |                                         |   Buffer Binding Array             |   |
/// |                                         |  +----------+----------+-----+     |   |
/// |                                         +->| buffer 0 | buffer 1 | ... |     |   |
/// |    Material Bindless Indices               +----------+----------+-----+     |   |
/// |   +----+-----------------------------+          ^                            |   |
/// +-->|  0 | material                    +----------+                            |   |
///     +----+-----------------------------+                                       |   |
///     |  1 | base_color_texture          +---------------------------------------+   |
///     +----+-----------------------------+                                           |
///     |  2 | base_color_sampler          +-------------------------------------------+
///     +----+-----------------------------+
///     | .. |            ...              |
///     +----+-----------------------------+
/// ```
///
/// The previous `CoolMaterial` example illustrating "combining multiple field-level uniform attributes with the same binding index" can
/// also be equivalently represented with a single struct-level uniform attribute:
/// ```
/// # use bevy_render::{render_resource::{AsBindGroup, ShaderType}};
/// # use bevy_color::LinearRgba;
/// #[derive(AsBindGroup)]
/// #[uniform(0, CoolMaterialUniform)]
/// struct CoolMaterial {
///     color: LinearRgba,
///     roughness: f32,
/// }
///
/// #[derive(ShaderType)]
/// struct CoolMaterialUniform {
///     color: LinearRgba,
///     roughness: f32,
/// }
///
/// impl From<&CoolMaterial> for CoolMaterialUniform {
///     fn from(material: &CoolMaterial) -> CoolMaterialUniform {
///         CoolMaterialUniform {
///             color: material.color,
///             roughness: material.roughness,
///         }
///     }
/// }
/// ```
///
/// Setting `bind_group_data` looks like this:
/// ```
/// # use bevy_render::{render_resource::AsBindGroup};
/// # use bevy_color::LinearRgba;
/// #[derive(AsBindGroup)]
/// #[bind_group_data(CoolMaterialKey)]
/// struct CoolMaterial {
///     #[uniform(0)]
///     color: LinearRgba,
///     is_shaded: bool,
/// }
///
/// // Materials keys are intended to be small, cheap to hash, and
/// // uniquely identify a specific material permutation.
/// #[repr(C)]
/// #[derive(Copy, Clone, Hash, Eq, PartialEq)]
/// struct CoolMaterialKey {
///     is_shaded: bool,
/// }
///
/// impl From<&CoolMaterial> for CoolMaterialKey {
///     fn from(material: &CoolMaterial) -> CoolMaterialKey {
///         CoolMaterialKey {
///             is_shaded: material.is_shaded,
///         }
///     }
/// }
/// ```
pub trait AsBindGroup {
    /// Data that will be stored alongside the "prepared" bind group.
    type Data: Send + Sync;

    type Param: SystemParam + 'static;

    /// The number of slots per bind group, if bindless mode is enabled.
    ///
    /// If this bind group doesn't use bindless, then this will be `None`.
    ///
    /// Note that the *actual* slot count may be different from this value, due
    /// to platform limitations. For example, if bindless resources aren't
    /// supported on this platform, the actual slot count will be 1.
    fn bindless_slot_count() -> Option<BindlessSlabResourceLimit> {
        None
    }

    /// True if the hardware *actually* supports bindless textures for this
    /// type, taking the device and driver capabilities into account.
    ///
    /// If this type doesn't use bindless textures, then the return value from
    /// this function is meaningless.
    fn bindless_supported(_: &RenderDevice) -> bool {
        true
    }

    /// label
    fn label() -> &'static str;

    /// Creates a bind group for `self` matching the layout defined in [`AsBindGroup::bind_group_layout`].
    fn as_bind_group(
        &self,
        layout_descriptor: &BindGroupLayoutDescriptor,
        render_device: &RenderDevice,
        pipeline_cache: &PipelineCache,
        fallback_buffer: &FallbackBuffer,
        shader_buffer_assets: &RenderAssets<GpuShaderBuffer>,
        param: &mut SystemParamItem<'_, '_, Self::Param>,
    ) -> Result<PreparedBindGroup, AsBindGroupError> {
        let layout = &pipeline_cache.get_bind_group_layout(layout_descriptor);

        let mut bind_group_builder = BindGroupBuilder::default();
        Self::build_bind_group(
            self,
            layout,
            render_device,
            param,
            false,
            &mut bind_group_builder,
        )?;

        let entries = bind_group_builder
            .binding_resources
            .iter()
            .map(|(index, binding)| BindGroupEntry {
                binding: *index,
                resource: binding.get_binding(fallback_buffer, shader_buffer_assets),
            })
            .collect::<Vec<_>>();

        let bind_group = render_device.create_bind_group(Self::label(), layout, &entries);

        Ok(PreparedBindGroup {
            bindings: bind_group_builder.into_binding_resources(),
            bind_group,
        })
    }

    fn bind_group_data(&self) -> Self::Data;

    /// Populates a [`BindGroupBuilder`] with all bindings that this type
    /// exposes.
    ///
    /// If the bind group builder API isn't sufficient to produce bind groups
    /// for this type (e.g. if the type uses its own bindless texture arrays),
    /// an implementor may return `AsBindGroupError::CreateBindGroupDirectly`
    /// from this function and instead define `as_bind_group` directly. This may
    /// prevent certain features, such as bindless mode, from working correctly.
    ///
    /// Set `force_no_bindless` to true to require that bindless textures *not*
    /// be used. `ExtendedMaterial` uses this in order to ensure that the base
    /// material doesn't use bindless mode if the extension doesn't.
    fn build_bind_group(
        &self,
        layout: &BindGroupLayout,
        render_device: &RenderDevice,
        param: &mut SystemParamItem<'_, '_, Self::Param>,
        force_no_bindless: bool,
        output: &mut BindGroupBuilder,
    ) -> Result<(), AsBindGroupError>;

    /// Creates the bind group layout matching all bind groups returned by
    /// [`AsBindGroup::as_bind_group`]
    fn bind_group_layout(render_device: &RenderDevice) -> BindGroupLayout
    where
        Self: Sized,
    {
        render_device.create_bind_group_layout(
            Self::label(),
            &Self::bind_group_layout_entries(render_device, false),
        )
    }

    /// Creates the bind group layout descriptor matching all bind groups returned by
    /// [`AsBindGroup::as_bind_group`]
    /// TODO: we only need `RenderDevice` to determine if bindless is supported
    fn bind_group_layout_descriptor(render_device: &RenderDevice) -> BindGroupLayoutDescriptor
    where
        Self: Sized,
    {
        BindGroupLayoutDescriptor {
            label: Self::label().into(),
            entries: Self::bind_group_layout_entries(render_device, false),
        }
    }

    /// Returns a vec of bind group layout entries.
    ///
    /// Set `force_no_bindless` to true to require that bindless textures *not*
    /// be used. `ExtendedMaterial` uses this in order to ensure that the base
    /// material doesn't use bindless mode if the extension doesn't.
    fn bind_group_layout_entries(
        render_device: &RenderDevice,
        force_no_bindless: bool,
    ) -> Vec<BindGroupLayoutEntry>
    where
        Self: Sized;

    fn bindless_descriptor() -> Option<BindlessDescriptor> {
        None
    }
}

/// An error that occurs during [`AsBindGroup::as_bind_group`] calls.
#[derive(Debug, Error)]
pub enum AsBindGroupError {
    /// The bind group could not be generated. Try again next frame.
    #[error("The bind group could not be generated")]
    RetryNextUpdate,
    #[error("Create the bind group via `as_bind_group()` instead")]
    CreateBindGroupDirectly,
    #[error("At binding index {0}, the provided image sampler `{1}` does not match the required sampler type(s) `{2}`.")]
    InvalidSamplerType(u32, String, String),
}

/// A prepared bind group returned as a result of [`AsBindGroup::as_bind_group`].
pub struct PreparedBindGroup {
    pub bindings: BindingResources,
    pub bind_group: BindGroup,
}

impl PreparedBindGroup {
    pub(crate) fn unprepare(&self) -> BindGroupBuilder {
        let mut data_buffer = vec![];
        BindGroupBuilder {
            binding_resources: UnpreparedBindingResources(
                self.bindings
                    .iter()
                    .map(|(binding, owned_binding_resource)| {
                        let unprepared_binding_resource = match owned_binding_resource {
                            OwnedBindingResource::Buffer(buffer) => {
                                UnpreparedBindingResource::Buffer(buffer.clone())
                            }
                            OwnedBindingResource::ShaderBuffer(handle) => {
                                UnpreparedBindingResource::ShaderBuffer(handle.clone())
                            }
                            OwnedBindingResource::TextureView(
                                texture_view_dimension,
                                texture_view,
                            ) => UnpreparedBindingResource::TextureView(
                                *texture_view_dimension,
                                texture_view.clone(),
                            ),
                            OwnedBindingResource::Sampler(sampler_binding_type, sampler) => {
                                UnpreparedBindingResource::Sampler(
                                    *sampler_binding_type,
                                    sampler.clone(),
                                )
                            }
                            OwnedBindingResource::Data(owned_data) => {
                                let start_offset = data_buffer.len() as u32;
                                data_buffer.extend_from_slice(&owned_data.0);
                                let end_offset = data_buffer.len() as u32;
                                UnpreparedBindingResource::Data(start_offset..end_offset)
                            }
                        };
                        (*binding, unprepared_binding_resource)
                    })
                    .collect(),
            ),
            data_buffer,
        }
    }
}

/// A raw list of binding resources, suitable for either preparing into a bind
/// group or for combining with other bind groups (perhaps via bindless).
///
/// This can be cleared and reused for different bind groups, in order to reduce
/// allocations.
#[derive(Default, Deref, DerefMut)]
pub struct BindGroupBuilder {
    /// The resources that this bind group builder holds.
    #[deref]
    pub binding_resources: UnpreparedBindingResources,

    /// A shared buffer that holds all the POD that this bind group uses.
    ///
    /// Data binding resources contain ranges into this buffer.
    pub data_buffer: Vec<u8>,
}

/// A pair of binding index and binding resource, used as part of
/// [`PreparedBindGroup`].
#[derive(Default, Deref, DerefMut)]
pub struct BindingResources(pub Vec<(u32, OwnedBindingResource)>);

/// A pair of binding index and binding resource, used as part of
/// [`BindGroupBuilder`].
#[derive(Default, Deref, DerefMut)]
pub struct UnpreparedBindingResources(pub Vec<(u32, UnpreparedBindingResource)>);

impl BindGroupBuilder {
    /// Turns this [`BindGroupBuilder`] into a set of [`BindingResources`]
    /// suitable for embedding into a [`PreparedBindGroup`].
    pub(crate) fn into_binding_resources(self) -> BindingResources {
        BindingResources(
            self.binding_resources
                .0
                .into_iter()
                .map(|(binding_index, unprepared_binding_resource)| {
                    let owned_binding_resource = match unprepared_binding_resource {
                        UnpreparedBindingResource::Buffer(buffer) => {
                            OwnedBindingResource::Buffer(buffer)
                        }
                        UnpreparedBindingResource::ShaderBuffer(handle) => {
                            OwnedBindingResource::ShaderBuffer(handle)
                        }
                        UnpreparedBindingResource::TextureView(
                            texture_view_dimension,
                            texture_view,
                        ) => {
                            OwnedBindingResource::TextureView(texture_view_dimension, texture_view)
                        }
                        UnpreparedBindingResource::Sampler(sampler_binding_type, sampler) => {
                            OwnedBindingResource::Sampler(sampler_binding_type, sampler)
                        }
                        UnpreparedBindingResource::Data(range) => {
                            OwnedBindingResource::Data(OwnedData(
                                self.data_buffer[(range.start as usize)..(range.end as usize)]
                                    .to_vec(),
                            ))
                        }
                    };
                    (binding_index, owned_binding_resource)
                })
                .collect(),
        )
    }

    /// Clears out this [`BindGroupBuilder`] so that it can be used to construct
    /// a new bind group, while preserving the allocations.
    pub fn clear(&mut self) {
        self.binding_resources.clear();
        self.data_buffer.clear();
    }
}

/// An owned binding resource of any type (ex: a [`Buffer`], [`TextureView`], etc).
/// This is used by types like [`PreparedBindGroup`] to hold a single list of all
/// render resources used by bindings.
#[derive(Debug)]
pub enum OwnedBindingResource {
    Buffer(Buffer),
    ShaderBuffer(Handle<ShaderBuffer>),
    TextureView(TextureViewDimension, TextureView),
    Sampler(SamplerBindingType, Sampler),
    Data(OwnedData),
}

/// A raw binding resource inside a [`BindGroupBuilder`].
///
/// This is the same as [`OwnedBindingResource`], except it references the
/// [`BindGroupBuilder::data_buffer`] instead of requiring separate allocations
/// for POD.
///
/// The name "unprepared" refers to the fact that this represents an individual
/// binding resource that hasn't been added to a bind group (i.e. "prepared")
/// yet.
#[derive(Debug)]
pub enum UnpreparedBindingResource {
    Buffer(Buffer),
    ShaderBuffer(Handle<ShaderBuffer>),
    TextureView(TextureViewDimension, TextureView),
    Sampler(SamplerBindingType, Sampler),
    /// Plain old data (POD).
    ///
    /// The given range represents the byte range within the
    /// [`BindGroupBuilder::data_buffer`].
    Data(Range<u32>),
}

/// Data that will be copied into a GPU buffer.
///
/// This corresponds to the `#[data]` attribute in `AsBindGroup`.
#[derive(Debug, Deref, DerefMut)]
pub struct OwnedData(pub Vec<u8>);

impl OwnedBindingResource {
    /// Creates a [`BindingResource`] reference to this
    /// [`OwnedBindingResource`].
    ///
    /// Note that this operation panics if passed a
    /// [`OwnedBindingResource::Data`], because [`OwnedData`] doesn't itself
    /// correspond to any binding and instead requires the
    /// `MaterialBindGroupAllocator` to pack it into a buffer.
    pub fn get_binding(&self) -> BindingResource<'_> {
        match self {
            OwnedBindingResource::Buffer(buffer) => buffer.as_entire_binding(),
            OwnedBindingResource::ShaderBuffer(_) => {
                panic!(
                    "You can't use `get_binding` with a `ShaderBuffer`; fetch the buffer from \
                     the `RenderAssets<GpuShaderBuffer>` instead"
                )
            }
            OwnedBindingResource::TextureView(_, view) => BindingResource::TextureView(view),
            OwnedBindingResource::Sampler(_, sampler) => BindingResource::Sampler(sampler),
            OwnedBindingResource::Data(_) => panic!("`OwnedData` has no binding resource"),
        }
    }
}

impl UnpreparedBindingResource {
    /// Creates a [`BindingResource`] reference to this
    /// [`UnpreparedBindingResource`].
    ///
    /// Note that this operation panics if passed a
    /// [`UnpreparedBindingResource::Data`], because the range doesn't itself
    /// correspond to any binding and instead requires the
    /// `MaterialBindGroupAllocator` to pack it into a buffer.
    pub fn get_binding<'a>(
        &'a self,
        fallback_buffer: &'a FallbackBuffer,
        shader_buffer_assets: &'a RenderAssets<GpuShaderBuffer>,
    ) -> BindingResource<'a> {
        match self {
            UnpreparedBindingResource::Buffer(buffer) => buffer.as_entire_binding(),
            UnpreparedBindingResource::TextureView(_, view) => BindingResource::TextureView(view),
            UnpreparedBindingResource::Sampler(_, sampler) => BindingResource::Sampler(sampler),
            UnpreparedBindingResource::ShaderBuffer(shader_buffer) => {
                // Fetch the raw buffer from the
                // `shader_buffer_assets`. If it's not there,
                // use the fallback buffer.
                match shader_buffer_assets.get(shader_buffer.id()) {
                    Some(shader_buffer) => BindingResource::Buffer(BufferBinding {
                        buffer: &shader_buffer.buffer,
                        offset: 0,
                        size: None,
                    }),
                    None => BindingResource::Buffer(BufferBinding {
                        buffer: fallback_buffer,
                        offset: 0,
                        size: None,
                    }),
                }
            }
            UnpreparedBindingResource::Data(_) => panic!("Data ranges have no binding resource"),
        }
    }
}

/// Converts a value to a [`ShaderType`] for use in a bind group.
///
/// This is automatically implemented for references that implement [`Into`].
/// Generally normal [`Into`] / [`From`] impls should be preferred, but
/// sometimes additional runtime metadata is required.
/// This exists largely to make some [`AsBindGroup`] use cases easier.
pub trait AsBindGroupShaderType<T: ShaderType> {
    /// Return the `T` [`ShaderType`] for `self`. When used in [`AsBindGroup`]
    /// derives, it is safe to assume that all images in `self` exist.
    fn as_bind_group_shader_type(&self, images: &RenderAssets<GpuImage>) -> T;
}

impl<T, U: ShaderType> AsBindGroupShaderType<U> for T
where
    for<'a> &'a T: Into<U>,
{
    #[inline]
    fn as_bind_group_shader_type(&self, _images: &RenderAssets<GpuImage>) -> U {
        self.into()
    }
}

#[cfg(test)]
mod test {
    use super::*;
    use bevy_asset::Handle;
    use bevy_image::Image;

    #[test]
    fn texture_visibility() {
        #[expect(
            dead_code,
            reason = "This is a derive macro compilation test. It will not be constructed."
        )]
        #[derive(AsBindGroup)]
        pub struct TextureVisibilityTest {
            #[texture(0, visibility(all))]
            pub all: Handle<Image>,
            #[texture(1, visibility(none))]
            pub none: Handle<Image>,
            #[texture(2, visibility(fragment))]
            pub fragment: Handle<Image>,
            #[texture(3, visibility(vertex))]
            pub vertex: Handle<Image>,
            #[texture(4, visibility(compute))]
            pub compute: Handle<Image>,
            #[texture(5, visibility(vertex, fragment))]
            pub vertex_fragment: Handle<Image>,
            #[texture(6, visibility(vertex, compute))]
            pub vertex_compute: Handle<Image>,
            #[texture(7, visibility(fragment, compute))]
            pub fragment_compute: Handle<Image>,
            #[texture(8, visibility(vertex, fragment, compute))]
            pub vertex_fragment_compute: Handle<Image>,
        }
    }
}