// Copyright © SixtyFPS GmbH <info@slint.dev>
// SPDX-License-Identifier: GPL-3.0-only OR LicenseRef-Slint-Royalty-free-2.0 OR LicenseRef-Slint-Software-3.0

#[cfg(feature = "unstable-wgpu-30")]
use i_slint_core::api::GraphicsAPI;
use i_slint_core::api::{PhysicalSize as PhysicalWindowSize, Window};
use i_slint_core::graphics::RequestedGraphicsAPI;
use i_slint_core::partial_renderer::DirtyRegion;
use i_slint_core::platform::PlatformError;
use i_slint_core::renderer::DrawOutcome;

use std::cell::RefCell;
use std::rc::Rc;
use std::sync::Arc;

use wgpu_30 as wgpu;

use crate::SkiaSharedContext;

// Jank-hunt: per-stage timing of the Vulkan/wgpu calls inside render().
// Stalls reproduce ONLY on the Mali G715 Vulkan driver (never on the
// SwiftShader emulator), so we time each Vulkan call site and log the slow
// ones to the perf file (nblog mirrors `VK ` lines). See
// docs/BUGS/scroll-jank-120hz.md.
pub mod vkprobe {
    use std::sync::atomic::{AtomicU64, Ordering};
    /// Log stages slower than this (microseconds). 3 ms by default; the
    /// jank-hunt apps lower it (e.g. 300) to see the whole per-frame
    /// breakdown of a ~4 ms render.
    pub static THRESHOLD_US: AtomicU64 = AtomicU64::new(3_000);
    pub static FRAME: AtomicU64 = AtomicU64::new(0);
    pub static ACQ: AtomicU64 = AtomicU64::new(0);
    pub static CLR: AtomicU64 = AtomicU64::new(0);
    pub static SKD: AtomicU64 = AtomicU64::new(0);
    pub static SKF: AtomicU64 = AtomicU64::new(0);
    pub static ORD: AtomicU64 = AtomicU64::new(0);
    pub static PRE: AtomicU64 = AtomicU64::new(0);
    /// 120Hz experiment: when set, skip the separate wgpu CLEAR pass (the
    /// scene's opaque background covers the whole window) and move the wgpu
    /// PRESENT-ordering submit to AFTER Skia's draw, so it only does the
    /// texture transition. Result: 3 submits/frame instead of 4, and no
    /// full-screen GPU clear. Safe while the root background is opaque.
    /// NOTE: DISPROVEN — without the wgpu clear/write pass, `Queue::present`
    /// clears the frame (black screen). Kept for reference; do not enable.
    pub static REORDER_NOCLR: std::sync::atomic::AtomicBool =
        std::sync::atomic::AtomicBool::new(false);
    /// 120Hz experiment: render-ahead pipeline. When set, each `render()`
    /// call first PRESENTS the frame prepared by the previous call (phase =
    /// vsync wake + ~1.1 ms present, i.e. locked to the vsync edge), and
    /// then prepares the NEXT frame (ACQ + CLR + Skia draw + ORD, ~3.5 ms)
    /// without presenting it. The display shows one slot (8.3 ms) older
    /// content at a constant, deterministic latency. Requires the loop to
    /// call render() every vsync while animating (VSYNC_LOCKED_RENDER).
    pub static RENDER_AHEAD: std::sync::atomic::AtomicBool =
        std::sync::atomic::AtomicBool::new(false);
    /// RENDER_AHEAD per-frame activation: the event loop sets this to the
    /// current `locked` state before each `do_render()`, so the render-ahead
    /// pipeline (hold-without-present) is only used while animating. An
    /// unlocked render (idle redraw, model update) presents immediately —
    /// otherwise its content would sit in the held slot forever (no further
    /// renders coming) and never reach the screen.
    pub static RENDER_AHEAD_ACTIVE: std::sync::atomic::AtomicBool =
        std::sync::atomic::AtomicBool::new(false);
    // ------------------------------------------------------------------
    // PRESENT GATE (vsync-grid quantization, the real 120Hz fix).
    //
    // Measured on the phone: the Choreographer vsync callback reaches the
    // UI thread 2.4–6 ms (median 3.3 ms) after the vsync edge, and the
    // render takes 3–9 ms — so each frame's present lands at a RANDOM
    // phase 8–14 ms after the edge against a 13.8 ms latch deadline.
    // 51% of frames land one slot late (judder) and ~4% two slots late
    // (visible hitch), independent of where in the render() call the
    // present happens.
    //
    // Fix: quantize the PRESENT to the vsync grid. The vsync driver hands
    // us the exact edge (Choreographer `frame_time_nanos`, CLOCK_MONOTONIC)
    // via `note_vsync_edge`; before `Queue::present` we sleep (≤ ~2 ms) so
    // the present lands at (next edge + 1.5 ms) — constant phase, ~4 ms of
    // deadline margin, frame-to-frame gaps of exactly one period. See
    // docs/BUGS/scroll-jank-120hz.md §9.
    /// Latest Choreographer `frame_time_nanos` (CLOCK_MONOTONIC ns).
    pub static VSYNC_EDGE_NS: AtomicU64 = AtomicU64::new(0);
    /// Detected vsync period in ns (from consecutive callback deltas).
    pub static VSYNC_PERIOD_NS: AtomicU64 = AtomicU64::new(0);
    /// Gate enabled for the current render (set per-frame by the event
    /// loop; true while animating/locked, false for one-off redraws so
    /// taps and model updates keep minimum latency).
    pub static PRESENT_GATE_ON: std::sync::atomic::AtomicBool =
        std::sync::atomic::AtomicBool::new(false);
    /// Microseconds the most recent `present_gate()` slept (0 = no sleep).
    pub static GATE_US: AtomicU64 = AtomicU64::new(0);
    /// CLOCK_MONOTONIC ns when the last vsync-wake (locked) render STARTED.
    /// The event loop uses it to skip a second wake (e.g. the Choreographer's
    /// late fallback wake arriving ~5 ms after the early-waker's edge wake
    /// already triggered a render) instead of double-rendering the slot.
    pub static LAST_RENDER_START_NS: AtomicU64 = AtomicU64::new(0);

    /// Called from the vsync driver on every Choreographer callback.
    pub fn note_vsync_edge(ftn_ns: u64) {
        let prev = VSYNC_EDGE_NS.swap(ftn_ns, Ordering::Relaxed);
        if prev > 0 {
            let d = ftn_ns.saturating_sub(prev);
            if (4_000_000..20_000_000).contains(&d) {
                VSYNC_PERIOD_NS.store(d, Ordering::Relaxed);
            }
        }
    }

    /// Sleep so the `Queue::present` that follows lands at (latest vsync
    /// edge + one period + 1.5 ms). No-op when disabled, when the grid is
    /// unknown/stale (no recent callbacks), or when already past target.
    pub fn present_gate() {
        if !PRESENT_GATE_ON.load(Ordering::Acquire) {
            return;
        }
        let edge = VSYNC_EDGE_NS.load(Ordering::Relaxed);
        let period = VSYNC_PERIOD_NS.load(Ordering::Relaxed);
        if edge == 0 || period == 0 {
            return;
        }
        let now = monotonic_ns();
        if now == 0 || now > edge + 4 * period {
            return; // stale grid (e.g. after resume) — don't gate
        }
        // Earliest grid phase (a vsync edge + 1.5 ms) strictly after now:
        //   T = edge + m*period + 1.5 ms  with  m = floor((now-edge-1.5ms)
        //   / period) + 1.
        // NOTE: a Choreographer callback can fire MID-RENDER (the next
        // edge's callback arrives at edge+~11.6 ms while a slow render is
        // still running), which advances `edge` by one period — the
        // smallest-phase rule above then just yields the same absolute T, so
        // the mid-render callback can never add an extra slot of delay.
        const PHASE_NS: u64 = 1_500_000;
        let rel = now.saturating_sub(edge).saturating_sub(PHASE_NS) / period;
        let target = edge + (rel + 1) * period + PHASE_NS;
        let wait = target.saturating_sub(now);
        // Only sleep for close phases; if the next grid phase is >4 ms out
        // this frame is already running late for it — present immediately
        // (the next frame re-syncs to the grid).
        let slept = if wait > 200_000 && wait <= 4_000_000 { wait } else { 0 };
        GATE_US.store(slept / 1000, Ordering::Relaxed);
        if slept > 0 {
            std::thread::sleep(std::time::Duration::from_nanos(slept));
        }
    }

    pub fn scn_reset() {
        crate::scn::reset();
    }

    pub fn scn_report(frame: u64) {
        crate::scn::report(frame);
    }

    pub fn now_us() -> u64 {
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .map(|d| d.as_micros() as u64)
            .unwrap_or(0)
    }

    /// CLOCK_THREAD_CPUTIME_ID nanoseconds — CPU time consumed by the
    /// (current) thread. For a slow frame window: wall ≫ cpu ⇒ the thread
    /// was descheduled/blocked between calls; wall ≈ cpu ⇒ pure CPU work.
    pub fn thread_cpu_ns() -> u64 {
        #[repr(C)]
        struct Ts {
            tv_sec: i64,
            tv_nsec: i64,
        }
        unsafe extern "C" {
            fn clock_gettime(clkid: i32, tp: *mut Ts) -> i32;
        }
        const CLOCK_THREAD_CPUTIME_ID: i32 = 3;
        let mut ts = Ts { tv_sec: 0, tv_nsec: 0 };
        if unsafe { clock_gettime(CLOCK_THREAD_CPUTIME_ID, &mut ts) } == 0 {
            ts.tv_sec as u64 * 1_000_000_000 + ts.tv_nsec as u64
        } else {
            0
        }
    }

    /// CLOCK_MONOTONIC nanoseconds — the same clock/epoch as the backend's
    /// VSYNC `mon=` stamp, so PRES and VSYNC can be correlated offline for
    /// frame-alignment (phase of present vs the vsync grid).
    pub fn monotonic_ns() -> u64 {
        #[repr(C)]
        struct Ts {
            tv_sec: i64,
            tv_nsec: i64,
        }
        unsafe extern "C" {
            fn clock_gettime(clkid: i32, tp: *mut Ts) -> i32;
        }
        const CLOCK_MONOTONIC: i32 = 1;
        let mut ts = Ts { tv_sec: 0, tv_nsec: 0 };
        if unsafe { clock_gettime(CLOCK_MONOTONIC, &mut ts) } == 0 {
            ts.tv_sec as u64 * 1_000_000_000 + ts.tv_nsec as u64
        } else {
            0
        }
    }

    // Log one slow stage (>= 3 ms) with its running-max and total so we can
    // see both the worst case and the accumulated cost per call site.
    pub fn report(stage: &str, slot: &AtomicU64, us: u64) {
        let _ = slot.fetch_max(us, Ordering::Relaxed);
        if us >= THRESHOLD_US.load(Ordering::Relaxed) {
            let f = FRAME.load(Ordering::Relaxed);
            log::info!("VK {} {}us f={} max={}us", stage, us, f, slot.load(Ordering::Relaxed));
        }
    }
}

#[cfg(target_family = "windows")]
mod dx12;
#[cfg(target_vendor = "apple")]
mod metal;
#[cfg(skia_wgpu_vulkan)]
mod vulkan;

/// See [`crate::attachment_color_space`].
pub(crate) fn attachment_color_space(texture: &wgpu::Texture) -> skia_safe::ColorSpace {
    crate::attachment_color_space(crate::TextureEncoding::from_format_is_srgb(
        texture.format().is_srgb(),
    ))
}

/// See [`crate::sampled_texture_color_space`].
#[cfg_attr(not(feature = "unstable-wgpu-30"), allow(dead_code))]
pub(crate) fn sampled_texture_color_space(texture: &wgpu::Texture) -> skia_safe::ColorSpace {
    crate::sampled_texture_color_space(crate::TextureEncoding::from_format_is_srgb(
        texture.format().is_srgb(),
    ))
}

/// Skia rendering surface backed by WGPU. Supports both on-screen rendering (with a
/// window surface) and offscreen rendering into caller-provided textures.
pub struct WGPUSurface {
    pub(crate) gr_context: RefCell<skia_safe::gpu::DirectContext>,
    wgpu: Rc<SharedWgpuState>,
    surface_config: RefCell<Option<wgpu::SurfaceConfiguration>>,
    surface: Option<wgpu::Surface<'static>>,
    textures_to_transition_for_sampling: RefCell<Vec<wgpu::Texture>>,
    pub(crate) backend: Backend,
    alpha_modes: Vec<wgpu::CompositeAlphaMode>,
    /// RENDER_AHEAD: the frame prepared by the previous `render()` call,
    /// presented at the start of the next one. `1` = frame index it was
    /// prepared under (for the PRES log line).
    pending_frame: RefCell<Option<(wgpu::SurfaceTexture, u64)>>,
}

impl WGPUSurface {
    pub fn new_with_surface(
        surface_target: impl Into<i_slint_core::graphics::wgpu_30::SurfaceTarget>,
        size: PhysicalWindowSize,
        requested_graphics_api: Option<RequestedGraphicsAPI>,
    ) -> Result<Self, PlatformError> {
        let (instance, adapter, device, queue, surface) =
            i_slint_core::graphics::wgpu_30::init_instance_adapter_device_queue_surface(
                surface_target,
                requested_graphics_api,
                i_slint_core::graphics::wgpu_30::default_backends_to_avoid(),
            )?;
        Self::init_with_parts(
            Rc::new(SharedWgpuState { instance, adapter, device, queue }),
            surface,
            size,
        )
    }

    fn init_with_parts(
        wgpu: Rc<SharedWgpuState>,
        surface: wgpu::Surface<'static>,
        size: PhysicalWindowSize,
    ) -> Result<Self, PlatformError> {
        let SharedWgpuState { adapter, device, queue, .. } = &*wgpu;
        #[cfg(target_vendor = "apple")]
        metal::set_layer_contents_gravity(&surface);

        // On iOS a window has no size until UIKit attaches it to a scene,
        // and wgpu rejects a zero-sized configure.
        // Start at 1x1; `resize_event` applies the real size.
        let mut surface_config = surface
            .get_default_config(adapter, size.width.max(1), size.height.max(1))
            .ok_or_else(|| PlatformError::from("WGPU surface is not compatible with adapter"))?;

        let swapchain_capabilities = surface.get_capabilities(adapter);
        let swapchain_format = swapchain_capabilities
            .formats
            .iter()
            .find(|f| {
                matches!(f, wgpu::TextureFormat::Rgba8Unorm | wgpu::TextureFormat::Bgra8Unorm)
            })
            .copied()
            .unwrap_or_else(|| swapchain_capabilities.formats[0]);
        surface_config.format = swapchain_format;

        // Prefer Mailbox / Immediate over Fifo/AutoVsync to prevent queue.present
        // from blocking the render thread on vsync backpressure (which causes
        // 17-55 ms tail stalls / dropped frames on 120Hz mobile GPUs).
        surface_config.present_mode = if swapchain_capabilities.present_modes.contains(&wgpu::PresentMode::Mailbox) {
            wgpu::PresentMode::Mailbox
        } else if swapchain_capabilities.present_modes.contains(&wgpu::PresentMode::Immediate) {
            wgpu::PresentMode::Immediate
        } else {
            wgpu::PresentMode::AutoVsync
        };

        let error_scope = device.push_error_scope(wgpu::ErrorFilter::Validation);
        surface.configure(device, &surface_config);
        if let Some(e) = spin_on::spin_on(error_scope.pop()) {
            return Err(PlatformError::from(format!("Error configuring WGPU surface: {e}")));
        }

        let backend = Backend::new(adapter, device)?;

        let gr_context = backend
            .make_context(adapter, device, queue)
            .ok_or_else(|| PlatformError::from("Failed to create Skia context from WGPU"))?;

        Ok(Self {
            gr_context: gr_context.into(),
            wgpu,
            surface_config: Some(surface_config).into(),
            surface: Some(surface),
            textures_to_transition_for_sampling: RefCell::new(Vec::new()),
            backend,
            alpha_modes: swapchain_capabilities.alpha_modes,
            pending_frame: RefCell::new(None),
        })
    }

    pub(crate) fn new_offscreen(
        instance: wgpu::Instance,
        adapter: wgpu::Adapter,
        device: wgpu::Device,
        queue: wgpu::Queue,
        backend: Backend,
        gr_context: skia_safe::gpu::DirectContext,
    ) -> Self {
        Self {
            gr_context: RefCell::new(gr_context),
            wgpu: Rc::new(SharedWgpuState { instance, adapter, device, queue }),
            surface_config: None.into(),
            surface: None,
            textures_to_transition_for_sampling: RefCell::new(Vec::new()),
            backend,
            alpha_modes: vec![],
            pending_frame: RefCell::new(None),
        }
    }

    /// RENDER_AHEAD: release a held frame before a reconfigure, where a
    /// stale-size frame would be wrong anyway. PRESENTS it (rather than
    /// just dropping it): wgpu's Presentation tracks the acquired image
    /// independently of the Rust Arc, so a bare drop would leave it
    /// "acquired" and the next get_current_texture would fail with
    /// AlreadyAcquired. The compositor handles the one-frame size change.
    fn discard_pending_frame(&self) {
        if let Some((pf, _)) = self.pending_frame.borrow_mut().take() {
            self.wgpu.queue.present(pf);
        }
    }

    /// Transitions any imported wgpu textures to sampling state and flushes
    /// the Skia graphics context. Must be called after rendering to ensure
    /// Skia's GPU work is submitted.
    pub(crate) fn flush_and_submit(&self, gr_context: &mut skia_safe::gpu::DirectContext) {
        let textures_to_transition = self.textures_to_transition_for_sampling.take();
        if !textures_to_transition.is_empty() {
            let mut encoder =
                self.wgpu.device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("Skia texture transition encoder"),
                });
            encoder.transition_resources(
                std::iter::empty(),
                textures_to_transition.iter().map(|texture| wgpu::TextureTransition {
                    texture,
                    selector: None,
                    state: wgpu::TextureUses::RESOURCE,
                }),
            );
            self.wgpu.queue.submit(Some(encoder.finish()));
        }

        gr_context.submit(None);
    }
}

/// The wgpu stack a surface renders with.
///
/// [`SkiaSharedContext`] caches the one the first window created, so that later windows it can
/// serve reuse it instead of creating their own.
/// That cache holds a weak reference and every surface a strong one, so the resources go away
/// with the last surface using them.
/// Were they to outlive the surfaces, they'd be destroyed whenever the last renderer happens to
/// be dropped, and tearing a wgpu device down from a thread-local destructor aborts the process
/// on macOS, where the Metal backend needs an autorelease pool that is gone by then.
/// A window that is still shown owns its surface, so that case remains.
pub(crate) struct SharedWgpuState {
    instance: wgpu::Instance,
    adapter: wgpu::Adapter,
    device: wgpu::Device,
    queue: wgpu::Queue,
}

/// Whether `device` offers everything `settings` asks of it.
///
/// The instance and adapter parts of the settings are covered by
/// [`adapter_matches_graphics_api_request`].
#[cfg(feature = "unstable-wgpu-30")]
fn device_satisfies_settings(
    device: &wgpu::Device,
    settings: &i_slint_core::graphics::wgpu_30::api::WGPUSettings,
) -> bool {
    device.features().contains(settings.device_required_features)
        && settings.device_required_limits.check_limits(&device.limits())
}

fn adapter_matches_graphics_api_request(
    adapter: &wgpu::Adapter,
    requested_graphics_api: Option<&RequestedGraphicsAPI>,
) -> bool {
    let backend = adapter.get_info().backend;

    #[cfg_attr(slint_nightly_test, allow(non_exhaustive_omitted_patterns))]
    match requested_graphics_api {
        None => true,
        Some(RequestedGraphicsAPI::Metal) => backend == wgpu::Backend::Metal,
        Some(RequestedGraphicsAPI::Vulkan) => backend == wgpu::Backend::Vulkan,
        Some(RequestedGraphicsAPI::Direct3D) => backend == wgpu::Backend::Dx12,
        #[cfg(feature = "unstable-wgpu-30")]
        Some(RequestedGraphicsAPI::WGPU30(
            i_slint_core::graphics::wgpu_30::api::WGPUConfiguration::Automatic(settings),
        )) => {
            let backend_bit = match backend {
                wgpu::Backend::Vulkan => wgpu::Backends::VULKAN,
                wgpu::Backend::Metal => wgpu::Backends::METAL,
                wgpu::Backend::Dx12 => wgpu::Backends::DX12,
                wgpu::Backend::Gl => wgpu::Backends::GL,
                wgpu::Backend::BrowserWebGpu => wgpu::Backends::BROWSER_WEBGPU,
                _ => return false,
            };
            settings.backends.contains(backend_bit)
        }
        Some(_) => false,
    }
}

impl crate::Surface for WGPUSurface {
    fn new(
        shared_context: &SkiaSharedContext,
        window_handle: Arc<dyn raw_window_handle::HasWindowHandle + Send + Sync>,
        display_handle: Arc<dyn raw_window_handle::HasDisplayHandle + Send + Sync>,
        size: PhysicalWindowSize,
        requested_graphics_api: Option<RequestedGraphicsAPI>,
    ) -> Result<Self, PlatformError> {
        let make_target = || -> Box<dyn wgpu::DisplayAndWindowHandle + 'static> {
            Box::new(WindowAndDisplayHandle(window_handle.clone(), display_handle.clone()))
        };

        #[cfg(feature = "unstable-wgpu-30")]
        let manual_configuration = matches!(
            &requested_graphics_api,
            Some(RequestedGraphicsAPI::WGPU30(
                i_slint_core::graphics::wgpu_30::api::WGPUConfiguration::Manual { .. }
            ))
        );
        #[cfg(not(feature = "unstable-wgpu-30"))]
        let manual_configuration = false;

        #[cfg(feature = "unstable-wgpu-30")]
        #[cfg_attr(slint_nightly_test, allow(non_exhaustive_omitted_patterns))]
        let requested_settings = match &requested_graphics_api {
            Some(RequestedGraphicsAPI::WGPU30(
                i_slint_core::graphics::wgpu_30::api::WGPUConfiguration::Automatic(settings),
            )) => Some(settings.clone()),
            _ => None,
        };

        // try to reuse old / shared graphics primitives from previous windows with matching
        // settings
        if !manual_configuration {
            let shared_state = shared_context.0.wgpu_30_state.borrow().upgrade();
            if let Some(shared) = shared_state {
                #[cfg(feature = "unstable-wgpu-30")]
                let settings_compatible = requested_settings
                    .as_ref()
                    .is_none_or(|settings| device_satisfies_settings(&shared.device, settings));
                #[cfg(not(feature = "unstable-wgpu-30"))]
                let settings_compatible = true;
                if settings_compatible
                    && adapter_matches_graphics_api_request(
                        &shared.adapter,
                        requested_graphics_api.as_ref(),
                    )
                    && let Ok(surface) = shared.instance.create_surface(make_target())
                    && shared.adapter.is_surface_supported(&surface)
                {
                    match Self::init_with_parts(shared, surface, size) {
                        Ok(surface) => return Ok(surface),
                        Err(err) => {
                            // The shared device may be lost. Drop it and start over.
                            i_slint_core::debug_log!(
                                "Failed to reuse the shared WGPU device: {err} . Re-initializing"
                            );
                            shared_context.0.wgpu_30_state.take();
                        }
                    }
                }
            }
        }

        let (instance, adapter, device, queue, surface) =
            i_slint_core::graphics::wgpu_30::init_instance_adapter_device_queue_surface(
                make_target(),
                requested_graphics_api,
                i_slint_core::graphics::wgpu_30::default_backends_to_avoid(),
            )?;
        let wgpu = Rc::new(SharedWgpuState { instance, adapter, device, queue });
        let new_surface = Self::init_with_parts(wgpu.clone(), surface, size)?;
        if !manual_configuration {
            *shared_context.0.wgpu_30_state.borrow_mut() = Rc::downgrade(&wgpu);
        }
        Ok(new_surface)
    }

    fn name(&self) -> &'static str {
        if self.surface.is_some() { "wgpu" } else { "wgpu-texture" }
    }

    fn resize_event(&self, size: PhysicalWindowSize) -> Result<(), PlatformError> {
        let mut surface_config_opt = self.surface_config.borrow_mut();
        let (Some(surface_config), Some(surface)) = (surface_config_opt.as_mut(), &self.surface)
        else {
            return Ok(());
        };
        self.discard_pending_frame();

        // Skip reconfigure if size hasn't changed — DRM/KMS surfaces don't
        // support being reconfigured.
        if surface_config.width == size.width && surface_config.height == size.height {
            return Ok(());
        }

        {
            let gr_context = &mut self.gr_context.borrow_mut();
            // This is brute force, but for the lack of access to the fences this seems to work: Avoid any pending work so that
            // IDXGISwapChain::ResizeBuffers doesn't complain that the surface is still in use.
            gr_context.flush_submit_and_sync_cpu();
        }

        surface_config.width = size.width;
        surface_config.height = size.height;

        surface.configure(&self.wgpu.device, surface_config);
        Ok(())
    }

    fn render(
        &self,
        _window: &Window,
        _size: PhysicalWindowSize,
        callback: &dyn Fn(
            &skia_safe::Canvas,
            Option<&mut skia_safe::gpu::DirectContext>,
            u8,
        ) -> Option<DirtyRegion>,
        pre_present_callback: &RefCell<Option<Box<dyn FnMut()>>>,
    ) -> Result<DrawOutcome, PlatformError> {
        let (Some(surface), Some(surface_config)) = (&self.surface, &*self.surface_config.borrow())
        else {
            return Err("WGPUSurface::render() called on offscreen surface".into());
        };

        let gr_context = &mut self.gr_context.borrow_mut();

        // Jank-hunt: Vulkan stage timing (see vkprobe).
        let __f = vkprobe::FRAME.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        vkprobe::scn_reset();

        // RENDER_AHEAD: present the frame prepared by the PREVIOUS render()
        // call FIRST, so the present lands at (vsync wake + ~1.1 ms) instead
        // of (vsync wake + full render). The frame prepared below is held
        // (when ahead is active) until the next call presents it (one slot
        // of constant latency).
        //
        // This block must run in BOTH modes: an unlocked render following a
        // held frame must PRESENT the held frame (not just drop it) —
        // wgpu's Presentation tracks the acquired image independently of the
        // Rust Arc, so a bare drop leaves it "acquired" and the next
        // get_current_texture fails with AlreadyAcquired.
        let __ahead = vkprobe::RENDER_AHEAD.load(std::sync::atomic::Ordering::Relaxed)
            && vkprobe::RENDER_AHEAD_ACTIVE.load(std::sync::atomic::Ordering::Relaxed);
        if let Some((pf, pf_idx)) = self.pending_frame.borrow_mut().take() {
            if let Some(cb) = pre_present_callback.borrow_mut().as_mut() {
                cb();
            }
            let t_pre = vkprobe::now_us();
            self.wgpu.queue.present(pf);
            let pre_us = vkprobe::now_us().saturating_sub(t_pre);
            vkprobe::report("PRE", &vkprobe::PRE, pre_us);
            let mon = vkprobe::monotonic_ns();
            if mon > 0 {
                log::info!("PRES f={} mon={} pre_us={}", pf_idx, mon, pre_us);
            }
        }
        let __cpu0 = vkprobe::thread_cpu_ns();
        let __t_acq = vkprobe::now_us();
        let frame = match surface.get_current_texture() {
            wgpu::CurrentSurfaceTexture::Success(t) => t,
            wgpu::CurrentSurfaceTexture::Occluded => return Ok(DrawOutcome::Occluded),
            wgpu::CurrentSurfaceTexture::Timeout => return Ok(DrawOutcome::Timeout),
            wgpu::CurrentSurfaceTexture::Validation => {
                return Err("WGPU surface validation error in get_current_texture".into());
            }
            stale @ (wgpu::CurrentSurfaceTexture::Outdated
            | wgpu::CurrentSurfaceTexture::Suboptimal(_)
            | wgpu::CurrentSurfaceTexture::Lost) => {
                // `Suboptimal` carries a live `SurfaceTexture`; matched with `_` it is not bound,
                // so the value returned by `get_current_texture()` keeps it alive across the
                // `surface.configure()` below — which wgpu forbids ("`SurfaceOutput` must be
                // dropped before a new `Surface` is made"), panicking on the first frame on
                // Wayland. Drop it first. (`Outdated`/`Lost` carry nothing → no-op.)
                drop(stale);
                surface.configure(&self.wgpu.device, surface_config);
                match surface.get_current_texture() {
                    wgpu::CurrentSurfaceTexture::Success(t) => t,
                    _ => return Ok(DrawOutcome::Occluded),
                }
            }
        };
        vkprobe::report("ACQ", &vkprobe::ACQ, vkprobe::now_us().saturating_sub(__t_acq));

        if !vkprobe::REORDER_NOCLR.load(std::sync::atomic::Ordering::Relaxed) {
            // Skia renders through the raw backend queue, invisible to wgpu's usage
            // tracking, so `Queue::present` would treat the frame as never written
            // and clear it. Clear it through wgpu instead before Skia draws; the
            // in-order queue keeps Skia's later-committed work on top.
            let __t_clr = vkprobe::now_us();
            let view = frame.texture.create_view(&wgpu::TextureViewDescriptor::default());
            let mut encoder =
                self.wgpu.device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("Slint frame init"),
                });
            encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                label: Some("Slint frame init"),
                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                    view: &view,
                    depth_slice: None,
                    resolve_target: None,
                    ops: wgpu::Operations {
                        load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                        store: wgpu::StoreOp::Store,
                    },
                })],
                depth_stencil_attachment: None,
                timestamp_writes: None,
                occlusion_query_set: None,
                multiview_mask: None,
            });
            self.wgpu.queue.submit(Some(encoder.finish()));
            vkprobe::report("CLR", &vkprobe::CLR, vkprobe::now_us().saturating_sub(__t_clr));
        }

        let __t_skd = vkprobe::now_us();
        let skia_surface = self.backend.make_swapchain_surface(gr_context, &frame.texture);

        let mut skia_surface = skia_surface
            .ok_or_else(|| PlatformError::from("Failed to create Skia surface from WGPU"))?;

        callback(skia_surface.canvas(), Some(gr_context), 0);

        self.backend.release_swapchain_surface(gr_context, &mut skia_surface);
        vkprobe::report("SKD", &vkprobe::SKD, vkprobe::now_us().saturating_sub(__t_skd));
        vkprobe::scn_report(__f);

        let __t_skf = vkprobe::now_us();
        self.flush_and_submit(gr_context);
        vkprobe::report("SKF", &vkprobe::SKF, vkprobe::now_us().saturating_sub(__t_skf));

        // Skia drew via the raw queue behind wgpu's back, and `Queue::present` skips its own
        // transition when the tracker already says PRESENT, so nothing would order the
        // presentation after Skia's work: the clear submission above signals its semaphore
        // before Skia even starts. Referencing the frame texture in a submission of our own
        // makes present wait for it. (With REORDER_NOCLR this submit is the ONLY
        // wgpu submission of the frame and happens AFTER Skia's, so the in-order
        // vkQueue + present's semaphore wait still order presentation after the draw.)
        let __t_ord = vkprobe::now_us();
        let mut encoder =
            self.wgpu.device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
                label: Some("Skia present ordering encoder"),
            });
        encoder.transition_resources(
            std::iter::empty(),
            std::iter::once(wgpu::TextureTransition {
                texture: &frame.texture,
                selector: None,
                state: wgpu::TextureUses::PRESENT,
            }),
        );
        self.wgpu.queue.submit(Some(encoder.finish()));
        vkprobe::report("ORD", &vkprobe::ORD, vkprobe::now_us().saturating_sub(__t_ord));

        if let Some(pre_present_callback) = pre_present_callback.borrow_mut().as_mut() {
            pre_present_callback();
        }

        let __t_pre = vkprobe::now_us();
        let __pre_us = if __ahead {
            // Render-ahead (locked/animating renders): do NOT present yet;
            // hold the frame for the next render() call (one slot of
            // constant latency). The PREVIOUS frame was already presented at
            // the start of this call. (The very first frame of the app's
            // lifetime is an unlocked render → presented immediately, so the
            // screen never stays black.)
            self.pending_frame.borrow_mut().replace((frame, __f));
            0
        } else {
            // Unlocked render: present immediately. (Any frame left in the
            // held slot was already presented at the top of this call.)
            // PRESENT GATE: while animating, land this present at a fixed
            // vsync-grid phase (next edge + 1.5 ms) so frame-to-frame
            // cadence is exactly one period (see `present_gate`).
            vkprobe::present_gate();
            self.wgpu.queue.present(frame);
            vkprobe::now_us().saturating_sub(__t_pre)
        };
        if __pre_us > 0 {
            vkprobe::report("PRE", &vkprobe::PRE, __pre_us);

            // Frame-alignment probe: monotonic ns at the moment present()
            // returned, tagged with the frame counter. Correlated offline
            // against VSYNC mon= stamps: phase = pres_mon - previous_vsync_mon.
            let __mon = vkprobe::monotonic_ns();
            if __mon > 0 {
                log::info!(
                    "PRES f={} mon={} pre_us={} gate_us={}",
                    __f,
                    __mon,
                    __pre_us,
                    vkprobe::GATE_US.load(std::sync::atomic::Ordering::Relaxed)
                );
            }
        }

        // Jank-hunt: for slow frames, CPU-vs-wall split for the WHOLE render
        // window. Correlated offline with the VKF line (time inside vk calls
        // per frame) from the Vulkan shim:
        //   wall ≈ cpu               → pure CPU work (Skia/wgpu on the UI thread)
        //   wall ≫ cpu, vk-time high → blocked inside the driver (fence etc.)
        //   wall ≫ cpu, vk-time low  → descheduled/between calls
        let __wall = vkprobe::now_us().saturating_sub(__t_acq);
        let __cpu = vkprobe::thread_cpu_ns().saturating_sub(__cpu0) / 1000;
        if __wall >= 8_000 {
            log::info!("CPUW f={} wall_us={} cpu_us={}", __f, __wall, __cpu);
        }

        Ok(DrawOutcome::Success)
    }

    fn bits_per_pixel(&self) -> Result<u8, PlatformError> {
        if let Some(surface_config) = &*self.surface_config.borrow() {
            Ok(match surface_config.format {
                wgpu_30::TextureFormat::Rgba8Unorm
                | wgpu_30::TextureFormat::Rgba8UnormSrgb
                | wgpu_30::TextureFormat::Bgra8Unorm
                | wgpu_30::TextureFormat::Bgra8UnormSrgb => 32,
                fmt => return Err(format!("Unsupported surface format {:#?}", fmt).into()),
            })
        } else {
            // All supported render-target formats (Rgba8Unorm, Bgra8Unorm, and sRGB variants) are 32bpp.
            Ok(32)
        }
    }

    #[cfg(feature = "unstable-wgpu-30")]
    fn with_graphics_api(&self, callback: &mut dyn FnMut(GraphicsAPI<'_>)) {
        let api = i_slint_core::graphics::create_graphics_api_wgpu_30(
            self.wgpu.instance.clone(),
            self.wgpu.device.clone(),
            self.wgpu.queue.clone(),
        );
        callback(api)
    }

    #[cfg(any(feature = "unstable-wgpu-29", feature = "unstable-wgpu-30"))]
    fn import_wgpu_texture(
        &self,
        canvas: &skia_safe::Canvas,
        any_wgpu_texture: &i_slint_core::graphics::WGPUTexture,
    ) -> Option<skia_safe::Image> {
        let texture: wgpu_30::Texture = match any_wgpu_texture {
            #[cfg(feature = "unstable-wgpu-29")]
            i_slint_core::graphics::WGPUTexture::WGPU29Texture(..) => return None,
            #[cfg(feature = "unstable-wgpu-30")]
            i_slint_core::graphics::WGPUTexture::WGPU30Texture(texture) => texture.clone(),
        };

        // Skia won't submit commands right away, so remember the texture and transition before
        // submitting.
        self.textures_to_transition_for_sampling.borrow_mut().push(texture.clone());

        self.backend.import_texture(canvas, texture)
    }

    fn set_transparent(&self, transparent: bool) -> Result<(), PlatformError> {
        // `Opaque` discards the scene's alpha; pick a translucent mode if offered.
        // Metal (CAMetalLayer) only offers `PostMultiplied`, so it must be a fallback.
        use wgpu::CompositeAlphaMode::{Opaque, PostMultiplied, PreMultiplied};
        let wanted: &[wgpu::CompositeAlphaMode] =
            if transparent { &[PreMultiplied, PostMultiplied] } else { &[Opaque] };
        let Some(mode) = wanted.iter().copied().find(|m| self.alpha_modes.contains(m)) else {
            return Ok(());
        };

        let mut surface_config_opt = self.surface_config.borrow_mut();
        let (Some(surface_config), Some(surface)) = (surface_config_opt.as_mut(), &self.surface)
        else {
            return Ok(());
        };
        if surface_config.alpha_mode != mode {
            surface_config.alpha_mode = mode;
            self.discard_pending_frame();
            surface.configure(&self.wgpu.device, surface_config);
        }
        Ok(())
    }

    fn as_any(&self) -> &dyn core::any::Any {
        self
    }
}

struct WindowAndDisplayHandle(
    Arc<dyn raw_window_handle::HasWindowHandle + Send + Sync>,
    Arc<dyn raw_window_handle::HasDisplayHandle + Send + Sync>,
);

impl raw_window_handle::HasWindowHandle for WindowAndDisplayHandle {
    fn window_handle(
        &self,
    ) -> Result<raw_window_handle::WindowHandle<'_>, raw_window_handle::HandleError> {
        self.0.window_handle()
    }
}

impl raw_window_handle::HasDisplayHandle for WindowAndDisplayHandle {
    fn display_handle(
        &self,
    ) -> Result<raw_window_handle::DisplayHandle<'_>, raw_window_handle::HandleError> {
        self.1.display_handle()
    }
}

pub(crate) enum Backend {
    #[cfg(target_vendor = "apple")]
    Metal,
    #[cfg(target_family = "windows")]
    Dx12,
    #[cfg(skia_wgpu_vulkan)]
    Vulkan {
        /// The family Skia has to hand the swapchain image back to, see
        /// [`Backend::release_swapchain_surface`].
        queue_family_index: u32,
    },
}

impl Backend {
    pub(crate) fn new(
        adapter: &wgpu::Adapter,
        _device: &wgpu::Device,
    ) -> Result<Self, PlatformError> {
        match adapter.get_info().backend {
            wgpu_30::Backend::Noop => {
                Err(PlatformError::from("Cannot use WGPU Noop backend with Skia"))
            }
            #[cfg(skia_wgpu_vulkan)]
            wgpu_30::Backend::Vulkan => Ok(Self::Vulkan {
                // SAFETY: `_device` is a Vulkan device, as the adapter's backend just said.
                queue_family_index: unsafe { vulkan::queue_family_index(_device) }.ok_or_else(
                    || {
                        PlatformError::from(
                            "Cannot query the queue family of the WGPU Vulkan device",
                        )
                    },
                )?,
            }),
            #[cfg(target_vendor = "apple")]
            wgpu_30::Backend::Metal => Ok(Self::Metal),
            #[cfg(target_family = "windows")]
            wgpu_30::Backend::Dx12 => Ok(Self::Dx12),
            other => Err(PlatformError::from(format!(
                "Unsupported WGPU backend for use with Skia: {}",
                other
            ))),
        }
    }

    pub(crate) fn make_context(
        &self,
        _adapter: &wgpu::Adapter,
        device: &wgpu::Device,
        queue: &wgpu::Queue,
    ) -> Option<skia_safe::gpu::DirectContext> {
        match self {
            #[cfg(target_vendor = "apple")]
            Self::Metal => metal::make_metal_context(device, queue),
            #[cfg(target_family = "windows")]
            Self::Dx12 => unsafe { dx12::make_dx12_context(&_adapter, &device, &queue) },
            #[cfg(skia_wgpu_vulkan)]
            Self::Vulkan { .. } => unsafe { vulkan::make_vulkan_context(device, queue) },
        }
    }

    pub(crate) fn make_surface(
        &self,
        gr_context: &mut skia_safe::gpu::DirectContext,
        texture: &wgpu::Texture,
    ) -> Option<skia_safe::Surface> {
        match self {
            #[cfg(target_vendor = "apple")]
            Self::Metal => unsafe { metal::make_metal_surface(gr_context, texture) },
            #[cfg(target_family = "windows")]
            Self::Dx12 => unsafe { dx12::make_dx12_surface(gr_context, texture) },
            #[cfg(skia_wgpu_vulkan)]
            Self::Vulkan { .. } => unsafe {
                vulkan::make_vulkan_surface(
                    gr_context,
                    texture,
                    skia_safe::gpu::vk::ImageLayout::COLOR_ATTACHMENT_OPTIMAL,
                )
            },
        }
    }

    /// Like [`Self::make_surface`], but for the swapchain image handed out by
    /// [`wgpu::Surface::get_current_texture`].
    ///
    /// wgpu returns surface textures to their `PRESENT` state at the end of every submission,
    /// so that is the layout Skia finds the image in, not the color-attachment one a plain
    /// render target is left in. Pair every call with [`Self::release_swapchain_surface`].
    pub(crate) fn make_swapchain_surface(
        &self,
        gr_context: &mut skia_safe::gpu::DirectContext,
        texture: &wgpu::Texture,
    ) -> Option<skia_safe::Surface> {
        match self {
            #[cfg(target_vendor = "apple")]
            Self::Metal => unsafe { metal::make_metal_surface(gr_context, texture) },
            #[cfg(target_family = "windows")]
            Self::Dx12 => unsafe { dx12::make_dx12_surface(gr_context, texture) },
            #[cfg(skia_wgpu_vulkan)]
            Self::Vulkan { .. } => unsafe {
                vulkan::make_vulkan_surface(
                    gr_context,
                    texture,
                    skia_safe::gpu::vk::ImageLayout::PRESENT_SRC_KHR,
                )
            },
        }
    }

    /// Hands a surface made by [`Self::make_swapchain_surface`] back in the state
    /// [`wgpu::Queue::present`] expects to find it in.
    pub(crate) fn release_swapchain_surface(
        &self,
        _gr_context: &mut skia_safe::gpu::DirectContext,
        _skia_surface: &mut skia_safe::Surface,
    ) {
        match self {
            // Metal and D3D12 have no image layout for Skia to hand back.
            #[cfg(target_vendor = "apple")]
            Self::Metal => {}
            #[cfg(target_family = "windows")]
            Self::Dx12 => {}
            #[cfg(skia_wgpu_vulkan)]
            Self::Vulkan { queue_family_index } => vulkan::release_vulkan_swapchain_surface(
                _gr_context,
                _skia_surface,
                *queue_family_index,
            ),
        }
    }

    #[cfg_attr(not(feature = "unstable-wgpu-30"), allow(dead_code))]
    pub(crate) fn import_texture(
        &self,
        canvas: &skia_safe::Canvas,
        texture: wgpu::Texture,
    ) -> Option<skia_safe::Image> {
        match self {
            #[cfg(target_vendor = "apple")]
            Self::Metal => unsafe { metal::import_metal_texture(canvas, texture) },
            #[cfg(target_family = "windows")]
            Self::Dx12 => unsafe { dx12::import_dx12_texture(canvas, texture) },
            #[cfg(skia_wgpu_vulkan)]
            Self::Vulkan { .. } => unsafe { vulkan::import_vulkan_texture(canvas, texture) },
        }
    }
}