agg_gui_wgpu/

lib.rs

//! `WgpuGfxCtx` — a hardware-accelerated [`DrawCtx`](agg_gui::draw_ctx::DrawCtx)
//! implementation via `wgpu`.
//!
//! # Platform coverage
//!
//! | Target | Backend |
//! |---|---|
//! | Windows | Vulkan, DX12 |
//! | macOS / iOS | Metal |
//! | Linux / Android | Vulkan |
//! | WASM (`wasm32-unknown-unknown`) | WebGL2 (via `wgpu` `webgl` feature) |
//!
//! # Scope
//!
//! This crate is the **renderer and nothing else**: the
//! [`DrawCtx`](agg_gui::draw_ctx::DrawCtx)
//! implementation, its pipelines and shaders, offscreen framebuffers,
//! screenshot capture, the [`custom_render`] hook for widgets that want their
//! own GPU passes, and the [`Gpu`] surface bundle that platform shells build
//! their swap-chain on.  It contains no widgets, no event loop, and no demo
//! content, so any app can depend on it directly.
//!
//! Platform shells (winit event loop, browser canvas + rAF loop) live in
//! `demo-wgpu`, which re-exports this crate for backwards compatibility.
//!
//! # Per-frame usage
//!
//! ```text
//! ctx.reset(width, height);
//! ctx.begin_frame(surface_view);   // clear + stash the target view
//! app.paint(&mut ctx);             // 2-D commands accumulate
//! ctx.end_frame();                 // flush everything to the GPU
//! ctx.present(surface_texture);    // drop per-frame handles, then present
//! ```
//!
//! Always present through [`WgpuGfxCtx::present`] rather than
//! `SurfaceTexture::present()` directly: the ctx may hold a clone of the back
//! buffer (for screenshots), and on DX12 a clone surviving present makes the
//! next swap-chain resize fail.
//!
//! # Deferred draw command model
//!
//! Unlike the GL backend which submits draw calls immediately, `WgpuGfxCtx`
//! accumulates `DrawCommand` enums during `fill()` / `stroke()` / etc., then
//! flushes them all in [`WgpuGfxCtx::end_frame`] using a single
//! `wgpu::CommandEncoder`.  This avoids the render-pass borrow lifetime
//! conflict: a `RenderPass` exclusively borrows its encoder, preventing both
//! from living in the same struct simultaneously.
//!
//! # Coordinate system
//!
//! All incoming coordinates are **Y-up pixel space**: origin at the bottom-left
//! of the viewport, positive Y upward.  The vertex shader converts to NDC with
//! `ndc = (pos / resolution) * 2 - 1`.  Scissor rects are stored in Y-up form
//! and converted to wgpu's Y-down framebuffer convention inside `end_frame`.

/// wgpu device + surface bundle and the surface-acquire recovery policy the
/// platform shells share.
#[cfg(not(target_arch = "wasm32"))]
pub mod gpu;
#[cfg(not(target_arch = "wasm32"))]
pub use gpu::{
    clamp_surface_size, surface_acquire_action, CopySrc, FrameAcquire, Gpu, GpuConfig,
    GpuInitError, RetryWake, SurfaceAcquire, SurfaceError,
};
/// When to retry a failed `Surface::configure`, when to give up, and what to
/// log — the pure policy behind [`Gpu`]'s surface recovery.
#[cfg(not(target_arch = "wasm32"))]
mod surface_retry;

pub mod custom_render;
pub use custom_render::{SharedCustomRenderer, WgpuCustomRender, WgpuCustomRenderCtx};

pub mod ssaa;
pub use ssaa::{ssaa_linear_scale, SsaaFramebuffer};

/// Screenshot read-back methods on [`WgpuGfxCtx`] (GPU→CPU frame copy).
mod screenshot_readback;

/// GPU handle passed to widgets via `DrawCtx::gl_paint` on the wgpu backend.
///
/// All fields are owned (cloned `Arc<...>` for device/queue, `wgpu::TextureView`
/// is internally ref-counted) so the struct is `'static` and works with the
/// `&dyn std::any::Any` plumbing of [`agg_gui::GlPaint`].
///
/// Painters create their own `wgpu::CommandEncoder` and submit it via
/// `queue.submit(...)`.  `WgpuGfxCtx` flushes any pending 2-D commands
/// before invoking the painter, so submissions interleave in the natural
/// paint order without an explicit barrier.
///
/// Constructed by the renderer and handed to painters — `#[non_exhaustive]`
/// so more GPU state can be threaded through without a breaking release.
#[derive(Clone)]
#[non_exhaustive]
pub struct WgpuPaintContext {
    /// Device used to build pipelines, buffers, and textures.
    pub device: Arc<wgpu::Device>,
    /// Queue used to submit the painter's command encoder.
    pub queue: Arc<wgpu::Queue>,
    /// Render-target view — same surface or layer texture the 2-D pipeline
    /// is rendering to this frame.  Painters open render passes against it
    /// with `LoadOp::Load` to overlay on existing content.
    pub target_view: wgpu::TextureView,
    /// Format of `target_view` — needed for pipeline `ColorTargetState`.
    pub surface_format: wgpu::TextureFormat,
    /// Full target dimensions in physical pixels.
    pub target_size: (u32, u32),
}

mod aa_step;
mod buffer_arena;
mod ctx_core;
/// The deferred draw-command enum — see [`DrawCommand`].
mod draw_command;
pub(crate) use draw_command::DrawCommand;
mod draw_ctx_impl;
mod end_frame;
mod end_frame_prepare;
mod gradient;
mod image_blit;
mod layer_mask;
mod layers;
pub mod pipelines;
mod primitives;
mod screenshot_capture;
/// Scaled / region capture readback — small-output companion to
/// `screenshot_capture`'s full-surface readbacks.
mod screenshot_scaled;
pub use screenshot_scaled::RectInPixels;
mod shaders;
mod text_render;

#[cfg(test)]
mod clip_path_readback_tests;
#[cfg(test)]
mod frame_release_tests;
#[cfg(test)]
mod image_blit_readback_tests;
#[cfg(test)]
mod layer_text_readback_tests;
#[cfg(test)]
mod lcd_arc_cache_tests;
#[cfg(test)]
mod screenshot_scaled_tests;

use std::collections::{HashMap, VecDeque};
use std::sync::{Arc, Weak};

use agg_gui::color::Color;
use agg_gui::draw_ctx::{FillRule, LinearGradientPaint, RadialGradientPaint};
use agg_gui::gl_renderer::GlyphCache;
use agg_gui::text::Font;
use agg_gui::TransAffine;
use agg_gui::{LineCap, LineJoin};
use agg_rust::path_storage::PathStorage;

use pipelines::WgpuPipelines;

// ---------------------------------------------------------------------------
// Arc-keyed texture cache entry
// ---------------------------------------------------------------------------

/// One entry in the Arc-keyed wgpu texture cache.  The `Weak` serves as a
/// liveness sentinel: when all strong refs to the source `Vec<u8>` are dropped
/// (typically because the widget's L1 pixel cache evicted the entry),
/// `weak.upgrade()` returns `None` and the entry is swept on the next access.
pub(crate) struct ArcTextureEntry {
    pub(crate) weak: Weak<Vec<u8>>,
    pub(crate) texture: Arc<wgpu::Texture>,
    pub(crate) view: wgpu::TextureView,
    pub(crate) w: u32,
    pub(crate) h: u32,
}

/// One entry in the **LCD** Arc-keyed wgpu texture cache.
///
/// Distinct from [`ArcTextureEntry`] because the LCD cache is keyed on the
/// source `Vec`'s heap **buffer address** (`data.as_ref().as_ptr()`), not on
/// `Arc::as_ptr`.  The TextArea dirty-strip path mutates a retained LCD plane
/// *in place* via `Arc::make_mut`; when a `Weak` is outstanding that relocates
/// the Arc control block (so `Arc::as_ptr` changes every keystroke) while only
/// *moving* the `Vec` header — the byte buffer stays put.  Keying on the buffer
/// address therefore keeps the entry stable across in-place edits so the texture
/// allocation is reused instead of destroyed and rebuilt every keystroke.
///
/// `version` is the backbuffer's globally-unique `content_version` (from
/// `next_content_version()` in agg-gui, one atomic counter across all caches).
/// A freed buffer address can be recycled by an unrelated allocation, but the
/// version it carries will always differ from any this entry has seen, so a
/// stale entry can never falsely match — the mismatch forces a fresh write.
/// Dimensions are checked as an additional guard.  Content-addressed *immutable*
/// masks pass `version == 0` and are instead verified by Arc pointer identity
/// via the stored `weak` (those buffers never relocate).
pub(crate) struct LcdArcTextureEntry {
    pub(crate) weak: Weak<Vec<u8>>,
    pub(crate) texture: Arc<wgpu::Texture>,
    pub(crate) view: wgpu::TextureView,
    pub(crate) w: u32,
    pub(crate) h: u32,
    pub(crate) version: u64,
}

// ---------------------------------------------------------------------------
// Saved draw state (for push_layer)
// ---------------------------------------------------------------------------

#[derive(Clone)]
pub(crate) struct SavedWgpuDrawState {
    pub(crate) viewport: (f32, f32),
    pub(crate) fill_color: Color,
    pub(crate) fill_linear_gradient: Option<LinearGradientPaint>,
    pub(crate) fill_radial_gradient: Option<RadialGradientPaint>,
    pub(crate) stroke_color: Color,
    pub(crate) stroke_linear_gradient: Option<LinearGradientPaint>,
    pub(crate) stroke_radial_gradient: Option<RadialGradientPaint>,
    pub(crate) line_width: f64,
    pub(crate) line_join: LineJoin,
    pub(crate) line_cap: LineCap,
    pub(crate) fill_rule: FillRule,
    pub(crate) miter_limit: f64,
    pub(crate) line_dash: Vec<f64>,
    pub(crate) dash_offset: f64,
    pub(crate) global_alpha: f64,
    pub(crate) state_stack: Vec<(TransAffine, Option<[i32; 4]>)>,
    pub(crate) font: Option<Arc<Font>>,
    pub(crate) font_size: f64,
    pub(crate) lcd_mode: bool,
}

// ---------------------------------------------------------------------------
// Layer types
// ---------------------------------------------------------------------------

#[derive(Clone, Copy)]
pub(crate) struct LayerRoundedClip {
    pub(crate) x: f32,
    pub(crate) y: f32,
    pub(crate) w: f32,
    pub(crate) h: f32,
    pub(crate) r: f32,
}

/// One transient wgpu compositing layer.
pub(crate) struct WgpuLayerEntry {
    /// Render-attachment + sampler texture for this layer.
    pub(crate) texture: Arc<wgpu::Texture>,
    pub(crate) view: wgpu::TextureView,
    pub(crate) width: u32,
    pub(crate) height: u32,
    pub(crate) origin_x: f64,
    pub(crate) origin_y: f64,
    pub(crate) alpha: f64,
    pub(crate) saved: SavedWgpuDrawState,
    /// Non-None when this layer will be stored in `retained_layers` on pop.
    pub(crate) retained_key: Option<u64>,
    pub(crate) rounded_clip: Option<LayerRoundedClip>,
    /// Scissor active in the parent at push time — applied to the composite
    /// blit on pop so the layer can't paint outside the parent's clip.
    pub(crate) parent_clip: Option<[i32; 4]>,
    /// Set via `set_layer_opaque_backdrop` once the caller has covered this
    /// layer with opaque content; re-enables LCD subpixel text inside it.
    ///
    /// The default (`false`) is the safe one. A fresh layer texture is
    /// cleared to alpha 0 and the 3-pass subpixel composite writes only
    /// colour, so glyphs would keep alpha 0 and the pop composite would
    /// blend them additively toward white — the wash-out this flag's
    /// absence is designed to avoid. Once opaque content covers the
    /// region, destination alpha is already 1 where glyphs land and
    /// subpixel geometry is exactly as valid as against the backbuffer.
    pub(crate) opaque_backdrop: bool,
    /// `Some` when this is a **clip layer** created by `DrawCtx::clip_path`:
    /// the tessellated clip path (parent-space positions + per-vertex
    /// coverage) that composites the layer back through the path instead of
    /// through a plain quad.  See `layer_mask.rs`.
    pub(crate) clip_mask: Option<crate::layer_mask::ClipMaskMesh>,
    /// For a clip layer, the current path as it was when `clip_path()` ran.
    /// Canvas `clip()` leaves the path alone, so it is restored on pop (after
    /// `restore_draw_state`, which resets `path`).
    pub(crate) clip_saved_path: Option<PathStorage>,
}

/// A retained layer that persists across frames (keyed by `u64` handle).
pub(crate) struct RetainedWgpuLayer {
    pub(crate) texture: Arc<wgpu::Texture>,
    pub(crate) view: wgpu::TextureView,
    pub(crate) width: u32,
    pub(crate) height: u32,
    pub(crate) rounded_clip: Option<LayerRoundedClip>,
}

// ---------------------------------------------------------------------------
// WgpuGfxCtx
// ---------------------------------------------------------------------------

/// An in-flight, non-blocking readback of the capture texture for the **web**
/// screen-share sender.  A blocking `map_async` + `poll(Wait)` (as
/// `DrawCtx::read_captured_screenshot` does for native Save/Copy) deadlocks the
/// single-threaded browser: the map can only complete once control returns to the
/// JS event loop, but the blocking wait never yields.  So the sender starts a
/// readback one frame and harvests it a later frame via
/// [`WgpuGfxCtx::poll_capture_readback`].
pub(crate) struct PendingReadback {
    staging: wgpu::Buffer,
    w: u32,
    h: u32,
    padded_bpr: u32,
    done: std::sync::mpsc::Receiver<Result<(), wgpu::BufferAsyncError>>,
}

/// A [`DrawCtx`](agg_gui::draw_ctx::DrawCtx) that renders via `wgpu`
/// (Vulkan / DX12 / Metal / WebGL2).
///
/// Create with [`WgpuGfxCtx::new`], passing a `wgpu::Device` and `wgpu::Queue`
/// that were obtained by the platform shell.  Each frame: call
/// `render_app_frame` (which calls [`reset`][WgpuGfxCtx::reset] and
/// `app.paint(ctx)`), then call [`end_frame`][WgpuGfxCtx::end_frame] with the
/// current surface texture view to flush all deferred draw commands.
pub struct WgpuGfxCtx {
    // ── wgpu core ────────────────────────────────────────────────────────────
    pub(crate) device: Arc<wgpu::Device>,
    pub(crate) queue: Arc<wgpu::Queue>,
    pub(crate) surface_format: wgpu::TextureFormat,
    pub(crate) viewport: (f32, f32),

    // ── render pipelines (created in Phase 2) ────────────────────────────────
    pub(crate) pipelines: WgpuPipelines,

    // ── deferred draw commands (flushed in end_frame — Phase 4) ──────────────
    pub(crate) commands: Vec<DrawCommand>,

    // ── texture caches (populated in Phase 6) ────────────────────────────────
    /// Generic slice-keyed cache: key is a FNV-like hash of (ptr, len, w, h, head/tail bytes).
    pub(crate) texture_cache: HashMap<u64, (Arc<wgpu::Texture>, wgpu::TextureView, u32, u32)>,
    pub(crate) texture_cache_order: VecDeque<u64>,
    /// Arc-pointer-keyed cache for `draw_image_rgba_arc` (Label backbuffers).
    pub(crate) arc_texture_cache: HashMap<usize, ArcTextureEntry>,
    /// Buffer-address-keyed cache for LCD coverage masks and backbuffer planes.
    pub(crate) lcd_arc_texture_cache: HashMap<usize, LcdArcTextureEntry>,

    // ── layer stack (wired in Phase 8) ────────────────────────────────────────
    pub(crate) layer_stack: Vec<WgpuLayerEntry>,
    pub(crate) retained_layers: HashMap<u64, RetainedWgpuLayer>,

    // ── drawing state ────────────────────────────────────────────────────────
    pub(crate) fill_color: Color,
    pub(crate) fill_linear_gradient: Option<LinearGradientPaint>,
    pub(crate) fill_radial_gradient: Option<RadialGradientPaint>,
    pub(crate) stroke_color: Color,
    pub(crate) stroke_linear_gradient: Option<LinearGradientPaint>,
    pub(crate) stroke_radial_gradient: Option<RadialGradientPaint>,
    pub(crate) line_width: f64,
    pub(crate) line_join: LineJoin,
    pub(crate) line_cap: LineCap,
    pub(crate) fill_rule: FillRule,
    pub(crate) miter_limit: f64,
    pub(crate) line_dash: Vec<f64>,
    pub(crate) dash_offset: f64,
    pub(crate) global_alpha: f64,
    /// Each entry is `(transform, scissor_yup)` — scissor stored in Y-up screen
    /// coordinates; converted to Y-down at `end_frame` time.
    pub(crate) state_stack: Vec<(TransAffine, Option<[i32; 4]>)>,
    /// Path builder — stored in local Y-up coordinates.
    pub(crate) path: PathStorage,
    pub(crate) font: Option<Arc<Font>>,
    pub(crate) font_size: f64,
    pub(crate) lcd_mode: bool,

    /// Tessellated-glyph cache shared with the GL backend — produces XY
    /// triangles per `(font, glyph_id, size)` key.  Lives on the context so
    /// glyph tessellations persist across frames.
    pub(crate) glyph_cache: GlyphCache,

    /// Surface texture view for the current frame — set by [`begin_frame`],
    /// cleared by [`Self::end_frame`].  Required so widgets that issue raw GPU
    /// draws via `DrawCtx::gl_paint` can target the same attachment as the
    /// deferred 2-D pipeline without the platform shell having to plumb the
    /// view through every call.
    pub(crate) surface_view: Option<wgpu::TextureView>,
    /// Cloned handle to the current frame's surface texture (the underlying
    /// resource is internally ref-counted, so cloning the handle is cheap).
    /// Used by [`Self::read_screenshot`] and `capture_screenshot` to copy
    /// from the frame after `end_frame` has flushed the render — the
    /// platform shell wires this up by calling `set_surface_texture` before
    /// paint.  It must NOT outlive the frame: holding a back-buffer clone
    /// past present breaks DX12 swap-chain resize, so it is dropped by
    /// [`Self::present`] / [`Self::release_frame_texture`].
    pub(crate) surface_texture: Option<wgpu::Texture>,
    /// Set once `set_surface_texture` has found the previous frame's texture
    /// still stashed (the shell never presented via [`Self::present`]) and
    /// logged the warning — keeps that warning to one per ctx.
    pub(crate) warned_unreleased_frame: bool,
    /// Pixels captured during the active frame for the screenshot UI.  The
    /// platform shell must call [`Self::read_screenshot`] BEFORE
    /// [`Self::present`] (which releases the stashed texture),
    /// stash the result here, then the screenshot orchestration picks it
    /// up via [`Self::take_pending_screenshot`] in its read-back closure.
    pub(crate) pending_screenshot: Option<(Vec<u8>, u32, u32)>,

    /// GPU-resident copy of the most recent surface contents — populated by
    /// [`DrawCtx::capture_screenshot`], sampled directly by
    /// [`DrawCtx::draw_captured_screenshot`].  Lives on the GPU so the
    /// screenshot preview pane can render it every frame with no CPU
    /// readback (the previous `Vec<u8>` + re-upload + mipmap gen path was
    /// blowing the frame budget under continuous capture).
    ///
    /// Pixels are pulled back to system memory only when the user clicks
    /// Save or Copy — see [`DrawCtx::read_captured_screenshot`].
    pub(crate) capture_texture: Option<(Arc<wgpu::Texture>, wgpu::TextureView, u32, u32)>,

    /// In-flight async capture readback for the web screen-share sender.
    /// See [`PendingReadback`].
    pub(crate) pending_readback: Option<PendingReadback>,

    /// In-flight async *scaled / region* capture readback — a separate slot so
    /// periodic small thumbnails and a full-surface screenshot can be in
    /// flight at the same time.  See `screenshot_scaled`.
    pub(crate) pending_scaled_readback: Option<PendingReadback>,

    /// Pipeline + sampler + destination texture for the scaled capture blit.
    /// Built on first scaled capture; `None` until then.
    pub(crate) scaled_blit: Option<screenshot_scaled::ScaledBlit>,

    /// Per-frame chunked buffer pool — see [`buffer_arena`] module docs.
    /// All `DrawCommand`s in a single flush share these three buffers
    /// instead of allocating their own, which is the single biggest lever
    /// against the per-command `create_buffer_init` cost.
    pub(crate) frame_arenas: buffer_arena::FrameArenas,

    /// 1024×1 RGBA8 alpha-step texture for the AA-texture pipeline.
    /// Column 0 = `(255, 255, 255, 0)`, columns 1..1023 =
    /// `(255, 255, 255, 255)`.  Sampled LINEAR — produces the
    /// sub-texel-wide AA transition right on the polygon edge, exactly
    /// like agg-sharp's `aATextureImages[255]`.  See
    /// `agg_gui::gl_renderer::aa_texture_mesh` for the texcoord scheme
    /// that drives it.
    #[allow(dead_code)]
    pub(crate) aa_step_texture: Arc<wgpu::Texture>,
    #[allow(dead_code)]
    pub(crate) aa_step_view: wgpu::TextureView,
    pub(crate) aa_step_bg1: Arc<wgpu::BindGroup>,

    /// Per-phase wall-clock timings from the most recent `end_frame`. Populated
    /// inside `flush_to_surface` so platform shells (atomartist, marbles) can
    /// surface a true breakdown of where wgpu-side time goes without needing
    /// to fork the renderer for instrumentation. All numbers are wall-clock
    /// microseconds; `command_count` is the number of `DrawCommand`s walked.
    pub(crate) last_end_frame_stats: LastEndFrameStats,
}

/// Wall-clock breakdown of the most recent `WgpuGfxCtx::end_frame` call.
///
/// `prepare_us` is the per-command CPU walk that allocates wgpu buffers and
/// bind groups (often the dominant cost for command-heavy scenes).
/// `execute_us` is the render-pass walk that issues draw calls into the
/// command encoder. `submit_us` is the `queue.submit` cost — usually tiny on
/// native, occasionally large on WebGPU or when the driver is back-pressured.
#[derive(Debug, Default, Clone, Copy)]
pub struct LastEndFrameStats {
    pub prepare_us: u32,
    pub execute_us: u32,
    pub submit_us: u32,
    pub command_count: u32,
}

impl WgpuGfxCtx {
    /// Create a new `WgpuGfxCtx`.
    ///
    /// `device` and `queue` must come from a `wgpu::Adapter` whose surface
    /// is already configured with `surface_format`.  The caller retains
    /// ownership of the surface; this struct only receives `Arc` refs to
    /// the device and queue so both it and the platform shell can drive
    /// buffer writes and texture uploads on the same queue.
    pub fn new(
        device: Arc<wgpu::Device>,
        queue: Arc<wgpu::Queue>,
        surface_format: wgpu::TextureFormat,
        width: f32,
        height: f32,
    ) -> Self {
        // 2-D pipelines stay sample_count=1: text and shapes already have
        // analytic edge AA via the tess2 halo strip + per-vertex alpha, so
        // hardware MSAA wouldn't add visible quality and would cost a full-
        // surface MSAA buffer (and per-layer ones) every frame.  MSAA belongs
        // scoped to the bar-grid renderer, which manages its own multi-sample
        // attachments and resolves into the active 1-sample target view.
        let pipelines = WgpuPipelines::new(&device, surface_format, 1);
        let frame_arenas = buffer_arena::FrameArenas::new(&device);

        // Build the 1024×1 RGBA8 alpha-step texture once and stash a
        // ready-to-bind bind group on the context — every AA-texture
        // draw can reuse this exact `bg1` without ever rebuilding it,
        // since the texture itself is immutable.
        let aa_step_texture = aa_step::build_aa_step_texture(&device, &queue);
        let aa_step_view = aa_step_texture.create_view(&wgpu::TextureViewDescriptor::default());
        let aa_step_bg1 = Arc::new(device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("aa_step_bg1"),
            layout: &pipelines.aa_texture_bgl1,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: wgpu::BindingResource::TextureView(&aa_step_view),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::Sampler(&pipelines.linear_sampler),
                },
            ],
        }));
        let aa_step_texture = Arc::new(aa_step_texture);

        Self {
            device,
            queue,
            surface_format,
            viewport: (width, height),
            pipelines,
            commands: Vec::new(),
            texture_cache: HashMap::new(),
            texture_cache_order: VecDeque::new(),
            arc_texture_cache: HashMap::new(),
            lcd_arc_texture_cache: HashMap::new(),
            layer_stack: Vec::new(),
            retained_layers: HashMap::new(),
            fill_color: Color::rgba(0.0, 0.0, 0.0, 1.0),
            fill_linear_gradient: None,
            fill_radial_gradient: None,
            stroke_color: Color::rgba(0.0, 0.0, 0.0, 1.0),
            stroke_linear_gradient: None,
            stroke_radial_gradient: None,
            line_width: 1.0,
            line_join: LineJoin::Miter,
            line_cap: LineCap::Butt,
            fill_rule: FillRule::NonZero,
            miter_limit: 4.0,
            line_dash: Vec::new(),
            dash_offset: 0.0,
            global_alpha: 1.0,
            state_stack: vec![(TransAffine::new(), None)],
            path: PathStorage::new(),
            font: None,
            font_size: 16.0,
            lcd_mode: false,
            glyph_cache: GlyphCache::new(),
            surface_view: None,
            surface_texture: None,
            warned_unreleased_frame: false,
            pending_screenshot: None,
            capture_texture: None,
            pending_readback: None,
            pending_scaled_readback: None,
            scaled_blit: None,
            frame_arenas,
            aa_step_texture,
            aa_step_view,
            aa_step_bg1,
            last_end_frame_stats: LastEndFrameStats::default(),
        }
    }

    /// Wall-clock breakdown of the most recent `end_frame` flush. Returns
    /// zeroes until the first frame has been flushed. Designed for live perf
    /// HUDs in platform shells — see atomartist's `record_frame_timings` for
    /// an example consumer.
    pub fn last_end_frame_stats(&self) -> LastEndFrameStats {
        self.last_end_frame_stats
    }

    /// Reset drawing state for a new frame.  Preserves GPU resources.
    pub fn reset(&mut self, width: f32, height: f32) {
        self.viewport = (width, height);
        self.fill_color = Color::rgba(0.0, 0.0, 0.0, 1.0);
        self.fill_linear_gradient = None;
        self.fill_radial_gradient = None;
        self.stroke_color = Color::rgba(0.0, 0.0, 0.0, 1.0);
        self.stroke_linear_gradient = None;
        self.stroke_radial_gradient = None;
        self.line_width = 1.0;
        self.fill_rule = FillRule::NonZero;
        self.miter_limit = 4.0;
        self.line_dash.clear();
        self.dash_offset = 0.0;
        self.global_alpha = 1.0;
        self.state_stack = vec![(TransAffine::new(), None)];
        self.path = PathStorage::new();
        self.font = None;
        self.font_size = 16.0;
        self.commands.clear();
        self.layer_stack.clear();
    }

    /// Enable / disable LCD subpixel text for this context.  Called each frame
    /// from `render_app_frame` with `font_settings::lcd_enabled()`.
    pub fn set_lcd_mode(&mut self, on: bool) {
        self.lcd_mode = on;
    }

    /// Start a frame: take ownership of this frame's render target and queue
    /// the clear.
    ///
    /// `view` is the `wgpu::TextureView` for the surface texture (or any other
    /// target) being painted this frame.  It is stashed on the context so that
    /// mid-frame GPU work — custom render passes, layer compositing — targets
    /// the same attachment as the deferred 2-D pipeline without the caller
    /// plumbing it through every method; [`Self::end_frame`] consumes it.
    ///
    /// The clear itself happens inside `end_frame`, when the leading
    /// `DrawCommand::Clear` is flushed into the first render pass; this
    /// method only pushes the current theme's background colour onto the
    /// command list so the frame starts from a clean framebuffer.
    pub fn begin_frame(&mut self, view: wgpu::TextureView) {
        self.surface_view = Some(view);
        let bg = agg_gui::current_visuals().bg_color;
        self.commands.push(DrawCommand::Clear(bg));
        // Sync the per-ctx LCD flag from the global typography setting so
        // any subsequent `fill_text` routes through the cached subpixel mask
        // path when LCD is enabled (default at scale ≤ 1.25).  Without this
        // shells that drive the loop via `begin_frame` instead of a helper
        // that sets it silently fall through to the aliased tessellated-
        // outline path even when the global says LCD is on.
        self.set_lcd_mode(agg_gui::font_settings::lcd_enabled());
    }

    /// Flush all deferred draw commands into a single wgpu command submission.
    ///
    /// Must be called after the frame's painting and before
    /// [`Self::present`].  The render target was stashed by
    /// [`Self::begin_frame`] — the platform shell does not need to pass it
    /// again here.
    pub fn end_frame(&mut self) {
        let Some(view) = self.surface_view.take() else {
            return;
        };
        self.flush_to_surface(&view);
    }

    /// Borrow the shared 2-D pipeline collection.  Exposed so platform
    /// shells (currently `demo-wasm`) can drive a [`SsaaFramebuffer::blit_to`]
    /// when they need to composite an intermediate scene texture onto the
    /// real swap-chain surface — see the comment on
    /// [`SsaaFramebuffer::resolve_texture`] for the WebGL2 scene-buffer
    /// pattern.  The returned `&WgpuPipelines` only exposes the fields
    /// `pub(crate)` library code already uses; external callers can pass
    /// it back into other library APIs but cannot reach in directly.
    pub fn pipelines(&self) -> &pipelines::WgpuPipelines {
        &self.pipelines
    }

    /// Borrow a clone-able handle to the wgpu device used for resource
    /// allocation.  Exposed alongside [`Self::pipelines`] so platform
    /// shells driving an external blit pass (currently `demo-wasm`'s
    /// scene-buffer → surface composite) don't need to hold a duplicate
    /// `Arc<wgpu::Device>` themselves.
    pub fn device(&self) -> &Arc<wgpu::Device> {
        &self.device
    }

    /// Borrow a clone-able handle to the wgpu queue.  Same rationale as
    /// [`Self::device`] — the shell submits the scene-blit encoder
    /// through this queue.
    pub fn queue(&self) -> &Arc<wgpu::Queue> {
        &self.queue
    }

    /// Format of the render target this context was configured for.  Custom
    /// renderers need it to build pipelines whose `ColorTargetState` matches
    /// the surface they will be composited onto.
    pub fn surface_format(&self) -> wgpu::TextureFormat {
        self.surface_format
    }

    /// Queue a custom wgpu render pass to run at the current point in the
    /// frame's draw order. The user-supplied [`WgpuCustomRender`] is invoked
    /// from `end_frame` after the active 2-D pass closes; subsequent 2-D
    /// content reopens with `LoadOp::Load` so it composites on top.
    ///
    /// `screen_rect` is the widget's logical-pixel rect in agg-gui Y-up
    /// coords. The implementor receives it via
    /// [`WgpuCustomRenderCtx::screen_rect`] and is responsible for any
    /// scissor / viewport conversions to wgpu's Y-down convention.
    pub fn push_custom_render(
        &mut self,
        renderer: custom_render::SharedCustomRenderer,
        screen_rect: agg_gui::Rect,
    ) {
        let parent_clip = self.current_clip();
        self.commands.push(DrawCommand::Custom {
            renderer,
            screen_rect,
            parent_clip,
        });
    }
}