pdfrum_render/

device.rs

//! The only seam between the engine and a rasterizer.
//!
//! Two traits and a handful of vocabulary types, all spoken in kurbo/peniko.
//! Nothing PDF-specific crosses this boundary: every decision that both
//! backends must agree on — degeneracy filters, integer rect snapping, the
//! stroke matrix split, dash normalisation, the ±32000 coordinate clamp — is
//! made in the engine *before* a device call, so a backend is free to be dumb
//! and Tier C can demand that the two receive identical geometry.

use kurbo::{Affine, BezPath, Rect, Stroke};
use pdfrum_page::BlendMode;

use crate::pixmap::{AlphaMask, Pixmap};

/// Whether a primitive's edges are antialiased.
///
/// The three modes are AGG's three, and they are genuinely three rather than
/// a flag and its negation:
///
/// - `On` integrates coverage and writes it as alpha.
/// - `Off` is the oracle's `aliased_path`, which thresholds the *same*
///   coverage at `> 127 -> 255` rather than turning the rasterizer off. It is
///   what the axis-aligned rect fast path and a hard-edged clip ask for.
/// - `FullCover` keeps the rasterizer's choice of *which* pixels a span
///   covers and then ignores the coverage value, writing every one of them at
///   the source alpha.
///
/// The distinction between the last two is the whole reason `FullCover`
/// exists. Thresholding drops a pixel two abutting cells each cover halfway,
/// so a subdivided Coons patch shows white pin-holes along every internal
/// seam; `full_cover` paints it from both cells, which is what makes the
/// patch continuous.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum AntiAlias {
    /// Antialiased, the default for every ordinary fill and stroke.
    #[default]
    On,
    /// Hard-edged: coverage thresholded at its midpoint.
    Off,
    /// Every touched pixel at full alpha, whatever its coverage.
    FullCover,
}

/// Which winding rule decides a path's interior.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum FillRule {
    /// Non-zero winding.
    #[default]
    Winding,
    /// Even-odd.
    EvenOdd,
}

/// How an image's samples are reconstructed when it is scaled.
///
/// The *selection* is the engine's — it ports PDFium's `/Interpolate`, its
/// `bNoSmoothing`/`bHalftone` flags, the `kHugeImageSize` forcing rule and
/// the `UseInterpolateBilinear` heuristic — and the kernels are each
/// backend's own. `Nearest` is passed explicitly for an integer-only
/// translation, because `vello_cpu` silently downgrades bilinear to nearest
/// in exactly that case and `tiny-skia` does not.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum ImageQuality {
    /// Nearest neighbour.
    #[default]
    Nearest,
    /// Bilinear.
    Bilinear,
}

/// A premultiplied RGBA8 image ready to be drawn, with its own dimensions.
pub type RasterImage = Pixmap;

/// What a fill or stroke paints with.
#[derive(Debug, Clone, Copy)]
pub enum Brush<'a> {
    /// A single colour.
    Solid(peniko::Color),
    /// An image, mapped onto the primitive by the draw's own transform.
    Image(&'a RasterImage),
}

/// The device a page is drawn into: six primitives plus a hard-edged rect
/// clip, all object-safe.
///
/// The engine holds one of these per render target — the page, each
/// transparency group, each soft mask, each pattern cell — and never learns
/// which rasterizer is behind it.
pub trait RenderDevice {
    /// Fill `path` (in user space) transformed by `t`.
    fn fill_path(
        &mut self,
        path: &BezPath,
        t: Affine,
        brush: &Brush<'_>,
        rule: FillRule,
        aa: AntiAlias,
    );

    /// Stroke `path` (in user space) transformed by `t`.
    ///
    /// The engine has already resolved the stroke's device width — including
    /// the one-device-pixel minimum — and normalised its dash array to an
    /// even-length, all-positive list, so `stroke` is always directly usable.
    fn stroke_path(
        &mut self,
        path: &BezPath,
        t: Affine,
        brush: &Brush<'_>,
        stroke: &Stroke,
        aa: AntiAlias,
    );

    /// Draw `img` with its own **pixel grid** mapped through `t`, at a
    /// constant `alpha`.
    ///
    /// `t` maps image pixel `(0, 0)`'s corner to its device position, so an
    /// identity transform is a texel-for-pixel blit at the origin and a
    /// translation moves it whole pixels — *not* a unit-square mapping.
    /// Reading `t` the other way collapses a whole-page image onto a single
    /// pixel, which is silent and total.
    ///
    /// An image being *reduced* has already been box-filtered down to roughly
    /// its device size by `crate::stretch::prescale`, so `t` scales it by
    /// less than a pixel in each axis and the two-tap kernel below is running
    /// near 1:1. An image being *enlarged* arrives at its own resolution, which
    /// is where the two-tap kernel is the right one and `quality` chooses it.
    fn draw_image(&mut self, img: &RasterImage, t: Affine, quality: ImageQuality, alpha: f32);

    /// Blit one glyph whose coverage is **three values per pixel**, one per LCD
    /// stripe, each merged into its own destination channel.
    ///
    /// The oracle's `ClearType` text
    /// (`DrawNormalTextHelper`'s `MergeGammaAdjustRgb` arm): for each pixel and
    /// each channel `c`, `dest_c = (dest_c·(255 − a_c) + colour_c·a_c) / 255`
    /// where `a_c = coverage_c · colour_alpha / 255`, and the destination is
    /// left opaque. Three independent alphas is exactly what
    /// [`RenderDevice::draw_image`] cannot express — one RGBA pixel carries one
    /// — which is why this is its own primitive rather than a flag on that one.
    ///
    /// `origin` is the device position of the bitmap's top-left corner, in
    /// whole pixels: both terms of it are integers by construction, so there is
    /// nothing to resample and a backend blits texel for pixel.
    ///
    /// **Defaulted to grayscale.** The default averages each pixel's three
    /// coverages and draws the result through [`RenderDevice::draw_image`], so
    /// a backend that cannot address channels separately still renders the text
    /// — in grey, without the colour fringes, which is what every *other* run
    /// of text on the page looks like anyway. That is a visible difference from
    /// the oracle on live-edit text and nothing worse: no glyph goes missing and
    /// no geometry moves. A backend that owns its pixels should override it.
    fn draw_glyph_lcd(
        &mut self,
        glyph: &crate::glyph::SubpixelBitmap,
        origin: (f64, f64),
        colour: peniko::Color,
    ) {
        let Some(gray) = crate::glyph::average_to_gray(glyph) else {
            return;
        };
        let Some(pixels) = crate::glyph::recolour(&gray, colour) else {
            return;
        };
        self.draw_image(
            &pixels,
            Affine::translate(origin),
            ImageQuality::Nearest,
            1.0,
        );
    }

    /// Intersect the clip with `path` (already in device space).
    fn push_clip(&mut self, path: &BezPath, rule: FillRule);

    /// Intersect the clip with an axis-aligned rectangle, **hard-edged**.
    ///
    /// PDFium's `SetClip_PathFill` takes a rect fast path that snaps to the
    /// outer integer rect and applies it with no antialiasing at all; `re W n`
    /// is the commonest clip in the corpus, so routing it through
    /// [`RenderDevice::push_clip`] would add a soft pixel along every clipped
    /// edge on a large fraction of the corpus.
    fn push_clip_rect(&mut self, rect: Rect);

    /// Begin a layer that will be composited back with `blend`, scaled by
    /// `alpha` and masked by `mask`.
    ///
    /// `mask`, when present, is device-sized and device-aligned. That is an
    /// invariant, not a convention: a mismatched mask is silently ignored by
    /// `vello_cpu` and merely warned about by `tiny-skia`, so violating it
    /// fails *open*, producing unmasked output. Backends assert it.
    fn push_layer(&mut self, blend: BlendMode, alpha: f32, mask: Option<&AlphaMask>);

    /// End the innermost clip or layer.
    fn pop(&mut self);
}

/// The factory that lets the engine rasterize offscreen: soft masks,
/// transparency groups, pattern cells and the fill+stroke knockout buffer all
/// need a second target.
pub trait RasterBackend {
    /// The device this backend produces.
    type Device: RenderDevice;

    /// A fresh target of the given size, every pixel set to `clear`.
    ///
    /// Both axes must be at most [`MAX_TARGET_DIMENSION`]: `vello_cpu` sizes
    /// its scenes, pixmaps and masks with `u16`. The engine enforces the
    /// bound before calling.
    fn new_target(&self, w: u32, h: u32, clear: peniko::Color) -> Self::Device;

    /// A fresh target seeded with `base`'s pixels.
    ///
    /// This is a non-isolated transparency group's backdrop (ISO 32000
    /// §11.4.6) and the fill+stroke knockout buffer's saved destination.
    fn new_target_with_backdrop(&self, base: &Pixmap) -> Self::Device;

    /// The device's current pixels, without consuming it.
    ///
    /// Only valid with every [`RenderDevice::push_layer`] matched by a
    /// [`RenderDevice::pop`] — `vello_cpu` asserts it. On a retained-scene
    /// backend this costs a full rasterization of the scene so far, not a
    /// rectangle copy, which is why the engine renders each group into its
    /// own target rather than snapshotting sub-rectangles of a shared one.
    fn snapshot(&self, d: &Self::Device) -> Pixmap;

    /// Consume the device, yielding its premultiplied RGBA8 pixels.
    fn finish(&self, d: Self::Device) -> Pixmap;

    /// [`snapshot`][Self::snapshot] cropped to `(origin_x, origin_y, width, height)`.
    ///
    /// The default is a full snapshot and a host crop. A retained-scene GPU
    /// backend still has to rasterize the scene so far — that is the
    /// structural cost — but can copy only this rectangle off the device,
    /// which is the whole of a non-isolated group's backdrop.
    ///
    /// ```
    /// use pdfrum_raster_vello_cpu::VelloCpuBackend;
    /// use pdfrum_render::RasterBackend;
    ///
    /// let backend = VelloCpuBackend;
    /// let device = backend.new_target(4, 4, peniko::Color::WHITE);
    /// let crop = backend.snapshot_rect(&device, 1, 1, 2, 2);
    /// assert_eq!((crop.width(), crop.height()), (2, 2));
    /// ```
    fn snapshot_rect(
        &self,
        d: &Self::Device,
        origin_x: u32,
        origin_y: u32,
        width: u32,
        height: u32,
    ) -> Pixmap {
        self.snapshot(d).cropped(origin_x, origin_y, width, height)
    }

    /// Whether isolated groups should composite as native layers instead of
    /// an offscreen pixmap round trip.
    ///
    /// The pixmap path is the CPU goldens: `finish`, `multiply_alpha_mask`,
    /// `remove_backdrop`. Native layers skip that host stall, which is the
    /// GPU win, but they are not bit-identical to the pixmap arithmetic —
    /// so the default is `false` and only a backend that has opted in (the
    /// GPU one) takes it. CPU backends keep the board still.
    ///
    /// ```
    /// use pdfrum_raster_vello_cpu::VelloCpuBackend;
    /// use pdfrum_render::RasterBackend;
    ///
    /// assert!(!VelloCpuBackend.composite_isolated_groups_as_layers());
    /// assert!(!(&VelloCpuBackend).composite_isolated_groups_as_layers());
    /// ```
    fn composite_isolated_groups_as_layers(&self) -> bool {
        false
    }
}

/// A reference to a backend is a backend.
///
/// Stateless CPU backends are unit structs, so a caller writes
/// `page.render(VelloCpuBackend)`. A backend that holds a device is passed
/// by reference, `page.render(&gpu)`, through this impl — the rasterizer is
/// still named, and nothing is moved.
impl<T: RasterBackend + ?Sized> RasterBackend for &T {
    type Device = T::Device;

    fn new_target(&self, w: u32, h: u32, clear: peniko::Color) -> Self::Device {
        (**self).new_target(w, h, clear)
    }

    fn new_target_with_backdrop(&self, base: &Pixmap) -> Self::Device {
        (**self).new_target_with_backdrop(base)
    }

    fn snapshot(&self, d: &Self::Device) -> Pixmap {
        (**self).snapshot(d)
    }

    fn finish(&self, d: Self::Device) -> Pixmap {
        (**self).finish(d)
    }

    fn snapshot_rect(
        &self,
        d: &Self::Device,
        origin_x: u32,
        origin_y: u32,
        width: u32,
        height: u32,
    ) -> Pixmap {
        (**self).snapshot_rect(d, origin_x, origin_y, width, height)
    }

    fn composite_isolated_groups_as_layers(&self) -> bool {
        (**self).composite_isolated_groups_as_layers()
    }
}

/// The largest render target either backend accepts in one axis.
///
/// `vello_cpu`'s `RenderContext::new`, `Pixmap::new` and `Mask` are all
/// `u16`-dimensioned; `tiny-skia` is `u32` but must agree for Tier C.
pub const MAX_TARGET_DIMENSION: u32 = u16::MAX as u32;

impl From<pdfrum_page::FillRule> for FillRule {
    /// `kWinding` maps to non-zero and **everything else, `kNoFill` included,
    /// maps to even-odd**.
    fn from(value: pdfrum_page::FillRule) -> Self {
        match value {
            pdfrum_page::FillRule::Winding => Self::Winding,
            pdfrum_page::FillRule::EvenOdd | pdfrum_page::FillRule::None => Self::EvenOdd,
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn no_fill_rule_reads_as_even_odd() {
        // cfx_agg_devicedriver.cpp:395-400: only kWinding is non-zero.
        assert_eq!(
            FillRule::from(pdfrum_page::FillRule::None),
            FillRule::EvenOdd
        );
        assert_eq!(
            FillRule::from(pdfrum_page::FillRule::EvenOdd),
            FillRule::EvenOdd
        );
        assert_eq!(
            FillRule::from(pdfrum_page::FillRule::Winding),
            FillRule::Winding
        );
    }
}