frust_engine/filters/

mod.rs

//! Layer filters: a layer rendered in isolation, run through a sequence of
//! filter passes, and composited from the page the last of them wrote.
//!
//! An opacity layer needs one page and one extra pass ([`schedule`]); a
//! filtered layer needs a *second* page as well, because a filter pass reads a
//! whole image and writes a whole image and a render pass cannot do both to one
//! texture. The two pages the scheduler already ping-pongs between are exactly
//! that pair: the layer's contents land in one, each pass writes the one it did
//! not read, and the sequence is arranged so the result ends back in the page
//! the parent's composite samples.
//!
//! [`schedule`]: crate::schedule
//!
//! ## What is served
//!
//! Two filters: [`LayerFilter::Blur`], the Gaussian blur that the theme
//! layer's `GlassMaterial` backdrop needs, and [`LayerFilter::DropShadow`], a
//! shadow-only drop shadow (`vello_common`'s `DropShadowOnly`, never
//! compositing the layer's own unfiltered content back over the shadow — see
//! [`drop_shadow`]'s own doc for why). Every other filter a recording can
//! carry — a flood, a standalone offset, a drop shadow that composites the
//! original back over itself, an edge mode other than [`SERVED_EDGE_MODE`], or
//! a graph of more than one primitive — is refused by [`served_filter`] with a
//! reason naming what was found, and the frame is skipped rather than rendered
//! without its filter. That refusal is not a placeholder for a fallback: the
//! engine tier carries no second renderer (see the [`schedule`] module header).
//!
//! The scene seam is deliberately absent. `frust_scene` has no filter command
//! and gains none here — [`push_filter_layer`] is how the engine's own compiler
//! will open a filtered layer once a later plan gives the display list
//! something to lower from. This module proves the engine can render one.
//!
//! ## What is pure, and what is not
//!
//! Everything here is a decision over plain values, in the crate's usual split:
//! which passes a blur costs and at what extents ([`blur::blur_passes`]), what
//! the fragment stage reads out of the filter-data texture
//! ([`blur::GpuGaussianBlur`]), and how one pass's instance is packed
//! ([`blur::FilterInstanceData`]). No device is touched and no texture is
//! allocated here.
//!
//! The three device-side halves live where every other engine pipeline's do:
//! the WGSL program is assembled by [`crate::gpu::shader_src::FILTER`], the
//! pipeline it is drawn through is
//! [`crate::gpu::pipelines::EnginePipeline::Filter`], and the texture holding
//! the frame's parameter blocks plus the bilinear sampler the kernels read
//! through are [`crate::gpu::targets`]'. The renderer binds the pages the
//! scheduler named and issues one instanced quad per pass
//! ([`crate::renderer::FilterResources`]).

pub mod blur;
pub mod drop_shadow;

use kurbo::Affine;
use peniko::color::{AlphaColor, Srgb};
use vello_common::filter::drop_shadow::DropShadow;
use vello_common::filter::gaussian_blur::GaussianBlur;
use vello_common::filter::{FilterData, PreparedFilter};
use vello_common::filter_effects::{EdgeMode, Filter, FilterPrimitive};
use vello_common::geometry::SizeU16;
use vello_common::record::{CommandRecorder, LayerProps, RecordedLayerKind};
use vello_common::util::extract_scales;

use crate::error::EngineError;

/// What [`served_filter`] says this engine renders when a recording's
/// primitive is neither of its two, so a refusal names the whole served set
/// rather than one filter's half of it.
const SERVED_FILTERS: &str = "the only filters it renders are a single Gaussian blur or a single, \
     shadow-only drop shadow (no original content composited back over it)";

/// The largest standard deviation a blur layer is served at, in the layer's
/// own space and in device space alike.
///
/// A blur expands its layer by 3σ on every side, snapped up to the tile grid
/// and carried as a `u16`; a σ past this bound would wrap that padding instead
/// of expanding the layer. It sits an order of magnitude above the largest page
/// the scheduler will ever size (the default ceiling is 4096 texels per axis),
/// so nothing a frame could have rendered is refused by it.
///
/// Both the recorded σ and the device-space σ the layer's transform scales it
/// into are held to it (see [`device_blur_sigma`]); the second is the one that
/// reaches `vello_common`'s planner, and the planner does not terminate on an
/// infinite one.
pub const MAX_BLUR_SIGMA: f32 = 4096.0;

/// The one edge mode this engine's filter kernels implement.
///
/// Every kernel bounds its taps against the region it is filtering and reads
/// transparent black outside it (`sample_region_bilinear` in
/// `shaders/filters_blur.wgsl`, `drop_shadow_load_checked` in
/// `shaders/filters_drop_shadow.wgsl`), which is exactly
/// [`EdgeMode::None`] — the mode
/// [`LayerFilter::filter_data`] records and the only one
/// [`served_filter`] serves.
pub const SERVED_EDGE_MODE: EdgeMode = EdgeMode::None;

/// A filter the engine can apply to an isolated layer.
///
/// Engine-internal on purpose: this is the vocabulary
/// [`push_filter_layer`] records with, not a `frust_scene` command.
#[derive(Debug, Clone, Copy, PartialEq)]
pub enum LayerFilter {
    /// A Gaussian blur of standard deviation `sigma`, in the layer's own user
    /// space — the transform in force when the layer is opened scales it into
    /// device space.
    Blur {
        /// The blur's standard deviation. Zero blurs nothing.
        sigma: f32,
    },
    /// A shadow-only drop shadow: the layer's alpha, blurred by `sigma`,
    /// shifted by `offset`, and recoloured to `color` — never the layer's own
    /// unfiltered content composited back over it (see [`drop_shadow`]'s own
    /// doc for why).
    DropShadow {
        /// Horizontal, vertical offset of the shadow, in the layer's own user
        /// space — the transform in force when the layer is opened scales it
        /// into device space, exactly as `sigma` is.
        offset: (f32, f32),
        /// The shadow's blur standard deviation. Zero blurs nothing, leaving a
        /// hard-edged, offset silhouette.
        sigma: f32,
        /// The shadow's own colour. Its alpha scales the shadow's own opacity;
        /// the layer's colour never reaches the shadow.
        color: AlphaColor<Srgb>,
    },
}

impl LayerFilter {
    /// The recorder-side description of this filter under `transform`.
    ///
    /// This is what carries the filter's *expansion* into the recording: a
    /// blurred layer's tile-aligned bounds grow by 3σ on every side, because
    /// the blur paints outside the contents that produced it, and the page the
    /// layer is rendered into is sized from those grown bounds.
    ///
    /// # Errors
    ///
    /// [`EngineError::InvalidGeometry`] for a σ that is not finite, is
    /// negative, or is past [`MAX_BLUR_SIGMA`] — in the layer's own user space
    /// *or* in the device space `transform` scales it into (a
    /// [`Self::DropShadow`]'s own offset held to the same finiteness) — and
    /// [`EngineError::InvalidTransform`] for a transform that is not finite —
    /// the same terms the compiler already refuses a draw's own geometry and
    /// transform on, checked here because every quantity below is derived from
    /// both.
    pub fn filter_data(self, transform: Affine) -> Result<FilterData, EngineError> {
        if !transform.as_coeffs().iter().all(|coeff| coeff.is_finite()) {
            return Err(EngineError::InvalidTransform);
        }

        match self {
            Self::Blur { sigma } => {
                // Both spaces, through the same predicate the served gate
                // takes: what the planner is handed is σ scaled by the
                // transform, and only the recorded half of that is visible
                // here without asking.
                if !served_sigmas(sigma, transform) {
                    return Err(EngineError::InvalidGeometry);
                }

                let filter = Filter::from_primitive(FilterPrimitive::GaussianBlur {
                    std_deviation: sigma,
                    edge_mode: SERVED_EDGE_MODE,
                });

                Ok(FilterData::new(filter, transform))
            }
            Self::DropShadow {
                offset: (dx, dy),
                sigma,
                color,
            } => {
                if !served_sigmas(sigma, transform) || !dx.is_finite() || !dy.is_finite() {
                    return Err(EngineError::InvalidGeometry);
                }

                // `DropShadowOnly`, never `DropShadow`: the shadow-only shape
                // is the whole of what this engine serves (see this module's
                // own doc and `served_filter` below).
                let filter = Filter::from_primitive(FilterPrimitive::DropShadowOnly {
                    dx,
                    dy,
                    std_deviation: sigma,
                    color,
                    edge_mode: SERVED_EDGE_MODE,
                });

                Ok(FilterData::new(filter, transform))
            }
        }
    }
}

/// Opens a filtered layer on `recorder` — the engine-internal counterpart of
/// [`CommandRecorder::push_layer`] for a layer that carries a filter.
///
/// Close it with [`CommandRecorder::pop_layer`], exactly like a regular layer.
/// Popping is what computes the layer's bounds, and for a filtered layer those
/// bounds are the *expanded* ones the filter paints into, which is what the
/// scheduler sizes the layer's pages from.
///
/// # Errors
///
/// Whatever [`LayerFilter::filter_data`] refuses; nothing is recorded when it
/// does, so a refused filter leaves the recording exactly as it was rather than
/// opening a layer no `pop_layer` will balance.
pub fn push_filter_layer<D>(
    recorder: &mut CommandRecorder<D>,
    props: LayerProps,
    filter: LayerFilter,
    transform: Affine,
) -> Result<(), EngineError> {
    let filter_data = filter.filter_data(transform)?;
    recorder.push_layer(props, Some(filter_data));
    Ok(())
}

/// The device-space standard deviation a blur of `sigma` runs at under
/// `transform`.
///
/// The quantity `vello_common`'s own `transform_blur_params` computes and hands
/// its planner — recomputed here, from the same `extract_scales` singular-value
/// pair averaged the same way, because upstream keeps that function
/// `pub(crate)` and because the whole point of the bound it feeds is to be
/// taken *before* the value reaches the planner. Keep it in step with
/// `vello_common::filter::gaussian_blur::transform_blur_params`.
#[must_use]
pub fn device_blur_sigma(sigma: f32, transform: Affine) -> f32 {
    let (scale_x, scale_y) = extract_scales(&transform);
    let uniform_scale = (scale_x + scale_y) / 2.0;
    sigma * uniform_scale
}

/// Whether `sigma` is a standard deviation this engine blurs at both as
/// recorded and as `transform` scales it into device space.
///
/// A range check rather than a chain of comparisons, which also settles the
/// non-finite cases: neither a NaN nor an infinity is contained by it. Both
/// ends matter — the recorded σ is what the layer's bounds are expanded by, and
/// the device-space σ is what `vello_common`'s `plan_decimated_blur` halves
/// down (`while remaining_variance > 4.0 { (v - 1.5) * 0.25 }`), a loop an
/// infinite variance never leaves.
#[must_use]
fn served_sigmas(sigma: f32, transform: Affine) -> bool {
    let served = 0.0..=MAX_BLUR_SIGMA;

    served.contains(&sigma) && served.contains(&device_blur_sigma(sigma, transform))
}

/// One of the two filters this engine renders, read back off a recording.
///
/// The shared answer of [`served_filter`], so the three sites that have to know
/// which filter a layer carries — [`crate::schedule`]'s `layer_role` and
/// `filter_rounds`, and [`crate::renderer`]'s `filter_block` — dispatch on one
/// decision rather than each re-deriving it from a try-blur-then-drop-shadow
/// pair of their own.
#[derive(Debug)]
pub enum ServedFilter {
    /// A Gaussian blur, prepared into device space.
    Blur(GaussianBlur),
    /// A shadow-only drop shadow, prepared into device space.
    DropShadow(DropShadow),
}

/// The filter layer `id`'s recorded `kind` describes, or a reason naming what
/// was found instead.
///
/// The reason is prose rather than an error variant because it reaches a log
/// through [`EngineError::SchedulerEscalation`], where a caller reading it is
/// trying to find out which recorded shape froze a surface. It names the
/// primitive's *own* fault — an unserved edge mode, an unserved σ — rather than
/// the fault of whichever filter was tried second, which is what a pair of
/// independent gates tried in sequence can only report.
///
/// # Errors
///
/// A `kind` that is not a filter at all; a filter under a transform that is not
/// finite; a filter graph of anything but one Gaussian blur or one shadow-only
/// drop shadow (a `DropShadow` that composites its original content back over
/// the shadow included — this engine serves only `DropShadowOnly`, see this
/// module's own doc); an edge mode other than [`SERVED_EDGE_MODE`]; a σ outside
/// [`MAX_BLUR_SIGMA`] as recorded or as the transform scales it; or a primitive
/// whose prepared kernel is not the one its own tag named.
pub fn served_filter(id: u32, kind: &RecordedLayerKind) -> Result<ServedFilter, String> {
    let RecordedLayerKind::Filter { filter_data, .. } = kind else {
        return Err(format!("layer {id} carries no filter"));
    };

    // Checked before preparing rather than after: every quantity the reference
    // derives below is derived from the transform as well as from the
    // primitive's own parameters, and a recording is not trusted to hold only
    // the shapes the engine's own compiler records (E17).
    let transform = filter_data.transform;
    if !transform.as_coeffs().iter().all(|coeff| coeff.is_finite()) {
        return Err(format!(
            "layer {id} carries a filter under a non-finite transform ({:?}), which is what \
             scales its σ into the device space the blur is planned in",
            transform.as_coeffs()
        ));
    }

    // Checked before preparing rather than after for the second reason too:
    // `PreparedFilter::new` panics on a graph the reference has not
    // implemented.
    let primitives = filter_data.filter.graph.primitives.as_slice();
    match primitives {
        [
            FilterPrimitive::GaussianBlur {
                std_deviation,
                edge_mode,
            },
        ] => {
            served_edge_mode(id, "a Gaussian blur", *edge_mode)?;
            served_sigma(id, "a Gaussian blur", *std_deviation, transform)?;

            match PreparedFilter::new(&filter_data.filter, &transform) {
                PreparedFilter::GaussianBlur(blur) => Ok(ServedFilter::Blur(blur)),
                _ => Err(format!(
                    "layer {id} carries a Gaussian blur that prepared as another filter"
                )),
            }
        }
        [
            FilterPrimitive::DropShadowOnly {
                std_deviation,
                edge_mode,
                ..
            },
        ] => {
            served_edge_mode(id, "a drop shadow", *edge_mode)?;
            served_sigma(id, "a drop shadow", *std_deviation, transform)?;

            match PreparedFilter::new(&filter_data.filter, &transform) {
                PreparedFilter::DropShadow(shadow) => Ok(ServedFilter::DropShadow(shadow)),
                _ => Err(format!(
                    "layer {id} carries a drop shadow that prepared as another filter"
                )),
            }
        }
        _ => Err(format!(
            "layer {id} carries a filter graph of {} primitive(s) the engine serves as neither of \
             its two filters; {SERVED_FILTERS}",
            primitives.len()
        )),
    }
}

/// Refuses an edge mode this engine's kernels do not implement, by name.
///
/// Only [`SERVED_EDGE_MODE`] is implemented: `Wrap` and `Mirror` would have to
/// address the region's opposite or reflected texels, and `Duplicate` would
/// have to replicate its edge ones, while every kernel here reads transparent
/// black outside the region unconditionally. Serving a recording that asks for
/// one of the other three would render it as `None` and call it that filter,
/// which is the silent-wrong-pixels answer this tier does not give.
fn served_edge_mode(id: u32, filter: &str, edge_mode: EdgeMode) -> Result<(), String> {
    if edge_mode == SERVED_EDGE_MODE {
        return Ok(());
    }

    Err(format!(
        "layer {id} carries {filter} with edge mode {edge_mode:?}, which this engine does not \
         render; its filter kernels read transparent black outside the region they filter, which \
         is {SERVED_EDGE_MODE:?}, and no other mode"
    ))
}

/// Refuses a σ this engine does not blur at, by name, in both spaces.
fn served_sigma(id: u32, filter: &str, sigma: f32, transform: Affine) -> Result<(), String> {
    if served_sigmas(sigma, transform) {
        return Ok(());
    }

    let device = device_blur_sigma(sigma, transform);
    Err(format!(
        "layer {id} carries {filter} of σ {sigma}, which its own transform scales to a \
         device-space σ of {device}; this engine blurs at a standard deviation of 0 to \
         {MAX_BLUR_SIGMA} in both spaces"
    ))
}

/// The blur that layer `id`'s recorded filter describes, or a reason naming
/// what was found instead.
///
/// [`served_filter`] narrowed to its blur arm, for a caller that has no answer
/// for a drop shadow.
///
/// # Errors
///
/// Whatever [`served_filter`] refuses, and a layer that carries the
/// shadow-only drop shadow instead.
pub fn served_blur(id: u32, kind: &RecordedLayerKind) -> Result<GaussianBlur, String> {
    match served_filter(id, kind)? {
        ServedFilter::Blur(blur) => Ok(blur),
        ServedFilter::DropShadow(_) => Err(format!(
            "layer {id} carries a shadow-only drop shadow rather than a Gaussian blur"
        )),
    }
}

/// The drop shadow that layer `id`'s recorded filter describes, or a reason
/// naming what was found instead.
///
/// [`served_filter`] narrowed to its drop-shadow arm, the counterpart of
/// [`served_blur`].
///
/// # Errors
///
/// Whatever [`served_filter`] refuses, and a layer that carries the Gaussian
/// blur instead.
pub fn served_drop_shadow(id: u32, kind: &RecordedLayerKind) -> Result<DropShadow, String> {
    match served_filter(id, kind)? {
        ServedFilter::DropShadow(shadow) => Ok(shadow),
        ServedFilter::Blur(_) => Err(format!(
            "layer {id} carries a Gaussian blur rather than a shadow-only drop shadow"
        )),
    }
}

/// One pass of a filter's sequence.
///
/// The numbering is the wire format the fragment stage switches on; it matches
/// the reference renderer's own — flood (1) and the drop-shadow composite that
/// reads the layer's own unfiltered content back (7) are still reserved and
/// still unserved (see [`drop_shadow`]'s own doc), and can be added later
/// without renumbering.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum FilterPassKind {
    /// Copy the source region through unchanged.
    Copy,
    /// Shift the source region by a drop shadow's own device-space offset.
    Offset,
    /// Halve both axes.
    Downscale,
    /// Convolve horizontally with the blur kernel.
    BlurH,
    /// Convolve vertically with the blur kernel.
    BlurV,
    /// Double both axes.
    Upscale,
    /// Recolour a blurred, offset alpha mask into a drop shadow's own
    /// premultiplied colour.
    Colorize,
}

impl FilterPassKind {
    /// The value `filter_pass_kind` carries into the fragment stage.
    #[must_use]
    pub const fn code(self) -> u32 {
        match self {
            Self::Copy => 0,
            Self::Offset => 2,
            Self::Downscale => 3,
            Self::BlurH => 4,
            Self::BlurV => 5,
            Self::Upscale => 6,
            Self::Colorize => 8,
        }
    }
}

/// One pass of a filter's sequence, with the extents it reads and writes.
///
/// A pass reads the page the pass before it wrote and writes the other one, so
/// `source` is the previous step's `dest`; the two differ only for the
/// rescaling passes.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct FilterStep {
    /// Which pass this is.
    pub kind: FilterPassKind,
    /// Extent of the region read, at the source page's origin.
    pub source: SizeU16,
    /// Extent of the region written, at the destination page's origin.
    pub dest: SizeU16,
}