Skip to content

Commit

Permalink
Improve push constant documentation, including internal docs.
Browse files Browse the repository at this point in the history
  • Loading branch information
jimblandy committed Jun 11, 2022
1 parent 5e07078 commit 76a3afc
Show file tree
Hide file tree
Showing 5 changed files with 108 additions and 37 deletions.
42 changes: 33 additions & 9 deletions wgpu-core/src/command/bundle.rs
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,17 @@ times, on different encoders. Constructing a render bundle lets `wgpu` validate
and analyze its commands up front, so that replaying a bundle can be more
efficient than simply re-recording its commands each time.
Not all commands are available in bundles; for example, a render bundle may not
contain a [`RenderCommand::SetViewport`] command.
Most of `wgpu`'s backend graphics APIs have something like bundles. For example,
Vulkan calls them "secondary command buffers", and Metal calls them "indirect
command buffers". However, `wgpu`'s implementation of render bundles does not
take advantage of those underlying platform features. At the hal level, `wgpu`
render bundles just replay the commands.
## Render Bundle Isolation
One important property of render bundles is that the draw calls in a render
bundle depend solely on the pipeline and state established within the render
bundle itself. A draw call in a bundle will never use a vertex buffer, say, that
Expand All @@ -17,14 +28,11 @@ Render passes are also isolated from the effects of bundles. After executing a
render bundle, a render pass's pipeline, bind groups, and vertex and index
buffers are are unset, so the bundle cannot affect later draw calls in the pass.
Not all commands are available in bundles; for example, a render bundle may not
contain a [`RenderCommand::SetViewport`] command.
Most of `wgpu`'s backend graphics APIs have something like bundles. For example,
Vulkan calls them "secondary command buffers", and Metal calls them "indirect
command buffers". However, `wgpu`'s implementation of render bundles does not
take advantage of those underlying platform features. At the hal level, `wgpu`
render bundles just replay the commands.
A render pass is not fully isolated from a bundle's effects on push constant
values. Draw calls following a bundle's execution will see whatever values the
bundle writes to push constant storage. Setting a pipeline initializes any push
constant storage it could access to zero, and this initialization may also be
visible after bundle execution.
## Render Bundle Lifecycle
Expand Down Expand Up @@ -372,6 +380,11 @@ impl RenderBundleEncoder {
&layout.push_constant_ranges,
);
commands.push(command);

// If this pipeline's push constant ranges aren't the same
// as the ones we were using previously (or if this is the
// first pipeline to use push constants at all), then emit
// commands to zero out the push constant values it will use.
if let Some(iter) = state.flush_push_constants() {
commands.extend(iter)
}
Expand Down Expand Up @@ -1044,7 +1057,15 @@ struct BindState {

#[derive(Debug)]
struct PushConstantState {
/// Push constant ranges used by the most recently set pipeline.
///
/// Before any pipeline has been set, this is empty.
ranges: ArrayVec<wgt::PushConstantRange, { SHADER_STAGE_COUNT }>,

/// True if this bundle has ever set a pipeline that uses push constants.
///
/// If this is true, then every time we set a pipeline, we will emit
/// `SetPushConstant` commands to clear the push constants it uses.
is_dirty: bool,
}
impl PushConstantState {
Expand Down Expand Up @@ -1246,6 +1267,9 @@ impl<A: HalApi> State<A> {
self.index.as_mut().and_then(|index| index.flush())
}

/// Return a sequence of commands to zero the push constant ranges that will
/// be used by the current pipeline. If no initialization is necessary,
/// return `None`.
fn flush_push_constants(&mut self) -> Option<impl Iterator<Item = RenderCommand>> {
let is_dirty = self.push_constant_ranges.is_dirty;

Expand All @@ -1260,7 +1284,7 @@ impl<A: HalApi> State<A> {
stages: range.stages,
offset: range.range.start,
size_bytes: range.range.end - range.range.start,
values_offset: None,
values_offset: None, // write zeros
}),
)
} else {
Expand Down
13 changes: 13 additions & 0 deletions wgpu-core/src/command/compute.rs
Original file line number Diff line number Diff line change
Expand Up @@ -43,11 +43,24 @@ pub enum ComputeCommand {
bind_group_id: id::BindGroupId,
},
SetPipeline(id::ComputePipelineId),

/// Set a range of push constants to values stored in [`BasePass::push_constant_data`].
SetPushConstant {
/// The byte offset within the push constant storage to write to. This
/// must be a multiple of four.
offset: u32,

/// The number of bytes to write. This must be a multiple of four.
size_bytes: u32,

/// Index in [`BasePass::push_constant_data`] of the start of the data
/// to be written.
///
/// Note: this is not a byte offset like `offset`. Rather, it is the
/// index of the first `u32` element in `push_constant_data` to read.
values_offset: u32,
},

Dispatch([u32; 3]),
DispatchIndirect {
buffer_id: id::BufferId,
Expand Down
23 changes: 21 additions & 2 deletions wgpu-core/src/command/draw.rs
Original file line number Diff line number Diff line change
Expand Up @@ -172,13 +172,32 @@ pub enum RenderCommand {
depth_max: f32,
},
SetScissor(Rect<u32>),

/// Set a range of push constants to values stored in [`BasePass::push_constant_data`].
///
/// See [`wgpu::RenderPass::set_push_constants`] for a detailed explanation
/// of the restrictions these commands must satisfy.
SetPushConstant {
/// Which stages we are setting push constant values for.
stages: wgt::ShaderStages,

/// The byte offset within the push constant storage to write to. This
/// must be a multiple of four.
offset: u32,

/// The number of bytes to write. This must be a multiple of four.
size_bytes: u32,
/// None means there is no data and the data should be an array of zeros.

/// Index in [`BasePass::push_constant_data`] of the start of the data
/// to be written.
///
/// Note: this is not a byte offset like `offset`. Rather, it is the
/// index of the first `u32` element in `push_constant_data` to read.
///
/// Facilitates clears in renderbundles which explicitly do their clears.
/// `None` means zeros should be written to the destination range, and
/// there is no corresponding data in `push_constant_data`. This is used
/// by render bundles, which explicitly clear out any state that
/// post-bundle code might see.
values_offset: Option<u32>,
},
Draw {
Expand Down
4 changes: 4 additions & 0 deletions wgpu-core/src/command/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -287,6 +287,10 @@ pub struct BasePass<C> {
/// instruction consumes the next `len` bytes from this vector.
pub string_data: Vec<u8>,

/// Data used by `SetPushConstant` instructions.
///
/// See the documentation for [`RenderCommand::SetPushConstant`]
/// and [`ComputeCommand::SetPushConstant`] for details.
pub push_constant_data: Vec<u32>,
}

Expand Down
63 changes: 37 additions & 26 deletions wgpu/src/lib.rs
Original file line number Diff line number Diff line change
Expand Up @@ -3016,34 +3016,44 @@ impl<'a> RenderPass<'a> {

/// [`Features::PUSH_CONSTANTS`] must be enabled on the device in order to call these functions.
impl<'a> RenderPass<'a> {
/// Set push constant data.
/// Set push constant data for subsequent draw calls.
///
/// Offset is measured in bytes, but must be a multiple of [`PUSH_CONSTANT_ALIGNMENT`].
/// Write the bytes in `data` at offset `offset` within push constant
/// storage, all of which are accessible by all the pipeline stages in
/// `stages`, and no others. Both `offset` and the length of `data` must be
/// multiples of [`PUSH_CONSTANT_ALIGNMENT`], which is always 4.
///
/// Data size must be a multiple of 4 and must have an alignment of 4.
/// For example, with an offset of 4 and an array of `[u8; 8]`, that will write to the range
/// of 4..12.
/// For example, if `offset` is `4` and `data` is eight bytes long, this
/// call will write `data` to bytes `4..12` of push constant storage.
///
/// For each byte in the range of push constant data written, the union of the stages of all push constant
/// ranges that covers that byte must be exactly `stages`. There's no good way of explaining this simply,
/// so here are some examples:
/// # Stage matching
///
/// ```text
/// For the given ranges:
/// - 0..4 Vertex
/// - 4..8 Fragment
/// ```
/// Every byte in the affected range of push constant storage must be
/// accessible to exactly the same set of pipeline stages, which must match
/// `stages`. If there are two bytes of storage that are accessible by
/// different sets of pipeline stages - say, one is accessible by fragment
/// shaders, and the other is accessible by both fragment shaders and vertex
/// shaders - then no single `set_push_constants` call may affect both of
/// them; to write both, you must make multiple calls, each with the
/// appropriate `stages` value.
///
/// You would need to upload this in two set_push_constants calls. First for the `Vertex` range, second for the `Fragment` range.
/// Which pipeline stages may access a given byte is determined by the
/// pipeline's [`PushConstant`] global variable and (if it is a struct) its
/// members' offsets.
///
/// ```text
/// For the given ranges:
/// - 0..8 Vertex
/// - 4..12 Fragment
/// ```
/// For example, suppose you have twelve bytes of push constant storage,
/// where bytes `0..8` are accessed by the vertex shader, and bytes `4..12`
/// are accessed by the fragment shader. This means there are three byte
/// ranges each accessed by a different set of stages:
///
/// You would need to upload this in three set_push_constants calls. First for the `Vertex` only range 0..4, second
/// for the `Vertex | Fragment` range 4..8, third for the `Fragment` range 8..12.
/// - Bytes `0..4` are accessed only by the fragment shader.
///
/// - Bytes `4..8` are accessed by both the fragment shader and the vertex shader.
///
/// - Bytes `8..12 are accessed only by the vertex shader.
///
/// To write all twelve bytes requires three `set_push_constants` calls, one
/// for each range, each passing the matching `stages` mask.
pub fn set_push_constants(&mut self, stages: ShaderStages, offset: u32, data: &[u8]) {
self.id.set_push_constants(stages, offset, data);
}
Expand Down Expand Up @@ -3151,13 +3161,14 @@ impl<'a> ComputePass<'a> {

/// [`Features::PUSH_CONSTANTS`] must be enabled on the device in order to call these functions.
impl<'a> ComputePass<'a> {
/// Set push constant data.
/// Set push constant data for subsequent dispatch calls.
///
/// Offset is measured in bytes, but must be a multiple of [`PUSH_CONSTANT_ALIGNMENT`].
/// Write the bytes in `data` at offset `offset` within push constant
/// storage. Both `offset` and the length of `data` must be
/// multiples of [`PUSH_CONSTANT_ALIGNMENT`], which is always 4.
///
/// Data size must be a multiple of 4 and must have an alignment of 4.
/// For example, with an offset of 4 and an array of `[u8; 8]`, that will write to the range
/// of 4..12.
/// For example, if `offset` is `4` and `data` is eight bytes long, this
/// call will write `data` to bytes `4..12` of push constant storage.
pub fn set_push_constants(&mut self, offset: u32, data: &[u8]) {
self.id.set_push_constants(offset, data);
}
Expand Down

0 comments on commit 76a3afc

Please sign in to comment.