xgpui-windows 1.0.1

mod color_presentation;
mod wide_sdr;

use std::{
    slice,
    sync::{Arc, OnceLock},
};

use anyhow::{Context, Result};
use gpui_util::ResultExt;
use windows::{
    Win32::{
        Foundation::HWND,
        Graphics::{
            Direct3D::*,
            Direct3D11::*,
            DirectComposition::*,
            DirectWrite::*,
            Dxgi::{Common::*, *},
        },
    },
    core::{HSTRING, Interface},
};

use crate::directx_renderer::shader_resources::{RawShaderBytes, ShaderModule, ShaderTarget};
use crate::*;
use gpui::*;

pub(crate) const DISABLE_DIRECT_COMPOSITION: &str = "GPUI_DISABLE_DIRECT_COMPOSITION";
const RENDER_TARGET_FORMAT: DXGI_FORMAT = DXGI_FORMAT_B8G8R8A8_UNORM;
// This configuration is used for MSAA rendering on paths only, and it's guaranteed to be supported by DirectX 11.
const PATH_MULTISAMPLE_COUNT: u32 = 4;
const MAX_INSTANCE_BUFFER_SIZE: usize = 256 * 1024 * 1024;

pub(crate) struct FontInfo {
    pub gamma_ratios: [f32; 4],
    pub grayscale_enhanced_contrast: f32,
    pub subpixel_enhanced_contrast: f32,
    pub is_bgr: bool,
}

pub(crate) struct DirectXRenderer {
    hwnd: HWND,
    atlas: Arc<DirectXAtlas>,
    devices: Option<DirectXRendererDevices>,
    resources: Option<DirectXResources>,
    globals: DirectXGlobalElements,
    pipelines: DirectXRenderPipelines,
    custom_compositor: Option<crate::custom_draw::CustomCompositor>,
    custom_retirement: Vec<(ID3D11Query, Vec<Arc<dyn gpui::CustomContent>>)>,
    shaders: Option<crate::custom_draw::ShaderRenderer>,
    backdrops: Option<crate::custom_draw::Backdrops>,
    wide_sdr: Option<wide_sdr::Compositor>,
    color_mode: WindowColorMode,
    color_display: color_presentation::DisplayState,
    direct_composition: Option<DirectComposition>,
    font_info: &'static FontInfo,

    width: u32,
    height: u32,

    /// Whether we want to skip drwaing due to device lost events.
    ///
    /// In that case we want to discard the first frame that we draw as we got reset in the middle of a frame
    /// meaning we lost all the allocated gpu textures and scene resources.
    skip_draws: bool,
}

/// Direct3D objects
#[derive(Clone)]
pub(crate) struct DirectXRendererDevices {
    pub(crate) adapter: IDXGIAdapter1,
    pub(crate) dxgi_factory: IDXGIFactory6,
    pub(crate) device: ID3D11Device,
    pub(crate) device_context: ID3D11DeviceContext,
    dxgi_device: Option<IDXGIDevice>,
    annotation: Option<ID3DUserDefinedAnnotation>,
}

struct DirectXResources {
    // Direct3D rendering objects
    swap_chain: Option<IDXGISwapChain1>,
    render_target: Option<ID3D11Texture2D>,
    render_target_view: Option<ID3D11RenderTargetView>,

    // Path intermediate textures (with MSAA)
    path_intermediate_texture: ID3D11Texture2D,
    path_intermediate_srv: Option<ID3D11ShaderResourceView>,
    path_intermediate_msaa_texture: ID3D11Texture2D,
    path_intermediate_msaa_view: Option<ID3D11RenderTargetView>,

    // Cached viewport
    viewport: D3D11_VIEWPORT,
}

struct DirectXRenderPipelines {
    shadow_pipeline: PipelineState<Shadow>,
    quad_pipeline: PipelineState<Quad>,
    path_rasterization_pipeline: PipelineState<PathRasterizationSprite>,
    path_sprite_pipeline: PipelineState<PathSprite>,
    underline_pipeline: PipelineState<Underline>,
    mono_sprites: PipelineState<MonochromeSprite>,
    subpixel_sprites: PipelineState<SubpixelSprite>,
    poly_sprites: PipelineState<PolychromeSprite>,
}

struct DirectXGlobalElements {
    global_params_buffer: Option<ID3D11Buffer>,
    batch_params_buffer: Option<ID3D11Buffer>,
    sampler: Option<ID3D11SamplerState>,
}

struct Annotation<'a>(&'a ID3DUserDefinedAnnotation);

impl<'a> Annotation<'a> {
    fn new(annotation: &'a ID3DUserDefinedAnnotation, label: HSTRING) -> Self {
        unsafe { annotation.BeginEvent(&label) };
        Self(annotation)
    }
}

impl Drop for Annotation<'_> {
    fn drop(&mut self) {
        unsafe { self.0.EndEvent() };
    }
}

struct DirectComposition {
    comp_device: IDCompositionDevice,
    comp_target: IDCompositionTarget,
    comp_visual: IDCompositionVisual,
}

impl DirectXRendererDevices {
    pub(crate) fn new(
        directx_devices: &DirectXDevices,
        disable_direct_composition: bool,
    ) -> Result<Self> {
        let DirectXDevices {
            adapter,
            dxgi_factory,
            device,
            device_context,
        } = directx_devices;
        let dxgi_device = if disable_direct_composition {
            None
        } else {
            Some(device.cast().context("Creating DXGI device")?)
        };
        let annotation = device_context.cast().ok();

        Ok(Self {
            adapter: adapter.clone(),
            dxgi_factory: dxgi_factory.clone(),
            device: device.clone(),
            device_context: device_context.clone(),
            dxgi_device,
            annotation,
        })
    }
}

impl DirectXRenderer {
    pub(crate) fn new(
        hwnd: HWND,
        directx_devices: &DirectXDevices,
        disable_direct_composition: bool,
    ) -> Result<Self> {
        if disable_direct_composition {
            log::info!("Direct Composition is disabled.");
        }

        let devices = DirectXRendererDevices::new(directx_devices, disable_direct_composition)
            .context("Creating DirectX devices")?;
        let atlas = Arc::new(DirectXAtlas::new(&devices.device, &devices.device_context));

        let resources = DirectXResources::new(&devices, 1, 1, hwnd, disable_direct_composition)
            .context("Creating DirectX resources")?;
        let globals = DirectXGlobalElements::new(&devices.device)
            .context("Creating DirectX global elements")?;
        let pipelines = DirectXRenderPipelines::new(&devices.device)
            .context("Creating DirectX render pipelines")?;

        let direct_composition = if disable_direct_composition {
            None
        } else {
            let composition = DirectComposition::new(devices.dxgi_device.as_ref().unwrap(), hwnd)
                .context("Creating DirectComposition")?;
            composition
                .set_swap_chain(
                    resources
                        .swap_chain
                        .as_ref()
                        .context("missing swap chain")?,
                )
                .context("Setting swap chain for DirectComposition")?;
            Some(composition)
        };

        Ok(DirectXRenderer {
            hwnd,
            atlas,
            devices: Some(devices),
            resources: Some(resources),
            globals,
            pipelines,
            custom_compositor: None,
            custom_retirement: Vec::new(),
            backdrops: None,
            shaders: None,
            wide_sdr: None,
            color_mode: WindowColorMode::Default,
            color_display: Default::default(),
            direct_composition,
            font_info: Self::get_font_info(),
            width: 1,
            height: 1,
            skip_draws: false,
        })
    }

    pub(crate) fn sprite_atlas(&self) -> Arc<dyn PlatformAtlas> {
        self.atlas.clone()
    }

    fn pre_draw(&self, clear_color: &[f32; 4]) -> Result<()> {
        let resources = self.resources.as_ref().expect("resources missing");
        let device_context = &self
            .devices
            .as_ref()
            .expect("devices missing")
            .device_context;
        update_buffer(
            device_context,
            self.globals.global_params_buffer.as_ref().unwrap(),
            &[GlobalParams {
                gamma_ratios: self.font_info.gamma_ratios,
                viewport_size: [resources.viewport.Width, resources.viewport.Height],
                grayscale_enhanced_contrast: self.font_info.grayscale_enhanced_contrast,
                subpixel_enhanced_contrast: self.font_info.subpixel_enhanced_contrast,
                is_bgr: self.font_info.is_bgr as u32,
                _pad: [0; 3],
            }],
        )?;
        unsafe {
            device_context.ClearRenderTargetView(
                resources
                    .render_target_view
                    .as_ref()
                    .context("missing render target view")?,
                clear_color,
            );
            device_context
                .OMSetRenderTargets(Some(slice::from_ref(&resources.render_target_view)), None);
            device_context.RSSetViewports(Some(slice::from_ref(&resources.viewport)));
            device_context
                .VSSetConstantBuffers(0, Some(slice::from_ref(&self.globals.global_params_buffer)));
            device_context
                .VSSetConstantBuffers(1, Some(slice::from_ref(&self.globals.batch_params_buffer)));
            device_context
                .PSSetConstantBuffers(0, Some(slice::from_ref(&self.globals.global_params_buffer)));
        }
        Ok(())
    }

    #[inline]
    fn present(&mut self) -> Result<()> {
        let result = unsafe {
            self.resources
                .as_ref()
                .expect("resources missing")
                .swap_chain
                .as_ref()
                .context("headless target cannot present")?
                .Present(0, DXGI_PRESENT(0))
        };
        result.ok().context("Presenting swap chain failed")
    }

    pub(crate) fn handle_device_lost(&mut self, directx_devices: &DirectXDevices) -> Result<()> {
        let disable_direct_composition = self.direct_composition.is_none();
        try_to_recover_from_device_lost(|| {
            self.handle_device_lost_impl(directx_devices, disable_direct_composition)
                .context("DirectXRenderer handling device lost")
        })
    }

    fn handle_device_lost_impl(
        &mut self,
        directx_devices: &DirectXDevices,
        disable_direct_composition: bool,
    ) -> Result<()> {
        let requested_color = self.color_mode;
        self.wide_sdr = None;
        self.custom_compositor = None;
        self.shaders = None;
        self.backdrops = None;

        unsafe {
            #[cfg(debug_assertions)]
            if let Some(devices) = &self.devices {
                report_live_objects(&devices.device)
                    .context("Failed to report live objects after device lost")
                    .log_err();
            }

            self.resources.take();
            if let Some(devices) = &self.devices {
                devices.device_context.OMSetRenderTargets(None, None);
                devices.device_context.ClearState();
                devices.device_context.Flush();
                #[cfg(debug_assertions)]
                report_live_objects(&devices.device)
                    .context("Failed to report live objects after device lost")
                    .log_err();
            }

            self.direct_composition.take();
            if let Some(devices) = &self.devices {
                devices.device_context.Flush();
            }
            self.devices.take();
            // Device removal terminates old queue work; foreign queries cannot
            // be polled on the replacement context.
            self.custom_retirement.clear();
        }

        let devices = DirectXRendererDevices::new(directx_devices, disable_direct_composition)
            .context("Recreating DirectX devices")?;
        let resources = DirectXResources::new(
            &devices,
            self.width,
            self.height,
            self.hwnd,
            disable_direct_composition,
        )
        .context("Creating DirectX resources")?;
        let globals = DirectXGlobalElements::new(&devices.device)
            .context("Creating DirectXGlobalElements")?;
        let pipelines = DirectXRenderPipelines::new(&devices.device)
            .context("Creating DirectXRenderPipelines")?;

        let direct_composition = if disable_direct_composition {
            None
        } else {
            let composition =
                DirectComposition::new(devices.dxgi_device.as_ref().unwrap(), self.hwnd)?;
            composition.set_swap_chain(
                resources
                    .swap_chain
                    .as_ref()
                    .context("missing swap chain")?,
            )?;
            Some(composition)
        };

        self.atlas
            .handle_device_lost(&devices.device, &devices.device_context);

        unsafe {
            devices
                .device_context
                .OMSetRenderTargets(Some(slice::from_ref(&resources.render_target_view)), None);
        }
        self.devices = Some(devices);
        self.resources = Some(resources);
        self.globals = globals;
        self.pipelines = pipelines;
        self.direct_composition = direct_composition;
        self.color_mode = WindowColorMode::Default;
        self.skip_draws = true;
        if let Err(error) = self.set_color_mode(requested_color, WindowBackgroundAppearance::Opaque)
        {
            self.color_mode = requested_color;
            return Err(error);
        }
        Ok(())
    }

    pub(crate) fn draw(
        &mut self,
        scene: &Scene,
        background_appearance: WindowBackgroundAppearance,
    ) -> Result<()> {
        if self.skip_draws {
            // skip drawing this frame, we just recovered from a device lost event
            // and so likely do not have the textures anymore that are required for drawing
            return Ok(());
        }
        self.refresh_color_profile()?;
        self.ensure_color_ready()?;
        self.render(scene, background_appearance)?;
        if let Some(wide) = &self.wide_sdr {
            let devices = self.devices.as_ref().context("devices missing")?;
            let target = self
                .resources
                .as_ref()
                .context("resources missing")?
                .render_target_view
                .as_ref()
                .context("back buffer missing")?;
            wide.present(devices, target, self.color_display.white_scale)?;
        }
        self.present()
    }

    /// Clear the render target for `background_appearance` and encode every
    /// primitive batch of `scene` into it, without presenting. Shared by
    /// [`draw`](Self::draw) (which then presents) and
    /// [`render_to_image`](Self::render_to_image) (which reads the target back
    /// instead), so the two cannot drift.
    fn render(&mut self, scene: &Scene, background: WindowBackgroundAppearance) -> Result<()> {
        if let Some(devices) = &self.devices {
            self.custom_retirement.retain(|(query, _)| {
                let mut complete = 0u32;
                let result = unsafe {
                    devices.device_context.GetData(
                        query,
                        Some((&mut complete as *mut u32).cast()),
                        4,
                        D3D11_ASYNC_GETDATA_DONOTFLUSH.0 as u32,
                    )
                };
                result.is_err() || complete == 0
            });
        }
        // Create the retirement fence before issuing any sampling work.
        let fence = if !scene.custom_draws.is_empty() {
            let devices = self.devices.as_ref().context("devices missing")?;
            let mut query = None;
            unsafe {
                devices.device.CreateQuery(
                    &D3D11_QUERY_DESC {
                        Query: D3D11_QUERY_EVENT,
                        MiscFlags: 0,
                    },
                    Some(&mut query),
                )?;
            }
            Some((
                query.context("custom retirement query missing")?,
                scene
                    .custom_draws
                    .iter()
                    .map(|d| d.content.clone())
                    .collect::<Vec<_>>(),
            ))
        } else {
            None
        };
        let result = self.render_inner(scene, background);
        if let Some((query, content)) = fence {
            if let Some(devices) = &self.devices {
                unsafe {
                    devices.device_context.End(&query);
                }
                self.custom_retirement.push((query, content));
            }
        }
        result
    }
    fn render_inner(
        &mut self,
        scene: &Scene,
        background_appearance: WindowBackgroundAppearance,
    ) -> Result<()> {
        if scene.backdrops.is_empty() && scene.layer_boundaries.is_empty() {
            if let Some(effects) = &mut self.backdrops {
                effects.clear();
            }
        }

        anyhow::ensure!(self.wide_sdr.is_some() || scene.shader_draws.iter().all(|d|
            d.program.output_encoding() == gpui::ShaderOutputEncoding::EncodedSrgb),
            "linear P3 shaders require WideGamutSdr");
        anyhow::ensure!(
            self.wide_sdr.is_none()
                || (scene.backdrops.is_empty()
                    && scene.layer_boundaries.is_empty()
                    && scene.shader_draws.iter().all(|d| d.path_mask.is_none())),
            "captured effects require SDR presentation"
        );
        if let Some(program) = scene
            .backdrops
            .first()
            .map(|d| &d.program)
            .or_else(|| scene.layer_boundaries.first().map(|b| &b.capture.program))
        {
            let devices = self.devices.as_ref().context("devices missing")?;
            if self.backdrops.is_none() {
                self.backdrops = Some(crate::custom_draw::Backdrops::new(
                    &devices.device,
                    program,
                )?);
            }
            self.backdrops.as_mut().unwrap().prepare(
                &devices.device,
                scene,
                size(
                    DevicePixels(self.width as i32),
                    DevicePixels(self.height as i32),
                ),
                &[
                    matches!(background_appearance, WindowBackgroundAppearance::Opaque) as u64,
                    self.font_info.gamma_ratios[0].to_bits() as u64,
                    self.font_info.gamma_ratios[1].to_bits() as u64,
                    self.font_info.gamma_ratios[2].to_bits() as u64,
                    self.font_info.gamma_ratios[3].to_bits() as u64,
                    self.font_info.grayscale_enhanced_contrast.to_bits() as u64,
                    self.font_info.subpixel_enhanced_contrast.to_bits() as u64,
                    self.font_info.is_bgr as u64,
                ],
            )?;
        }
        if let Some(wide) = self.wide_sdr.as_mut() {
            wide.resize(
                &self.devices.as_ref().context("devices missing")?.device,
                size(
                    DevicePixels(self.width as i32),
                    DevicePixels(self.height as i32),
                ),
            )?;
        }
        if !scene.shader_draws.is_empty() || self.shaders.is_some() {
            let devices = self.devices.as_ref().context("devices missing")?;
            if self.shaders.is_none() {
                self.shaders = Some(crate::custom_draw::ShaderRenderer::new(&devices.device)?);
            }
            self.shaders.as_mut().unwrap().prepare(
                &devices.device,
                &devices.device_context,
                scene,
            )?;
        }
        let custom_outputs = self.prepare_custom_draws(scene)?;
        self.pre_draw(&match background_appearance {
            WindowBackgroundAppearance::Opaque => [1.0f32; 4],
            _ => [0.0f32; 4],
        })?;
        self.upload_scene_buffers(scene)?;

        if let Some(mut wide) = self.wide_sdr.take() {
            let result = wide.render_scene(self, scene, &custom_outputs, background_appearance);
            self.wide_sdr = Some(wide);
            return result;
        }
        let mut executed_backdrops = Vec::new();
        let mut executed_groups = Vec::new();
        let mut skip_group_end = None;
        for batch in scene.batches() {
            if let Some(end) = skip_group_end {
                if !matches!(&batch,PrimitiveBatch::LayerBoundary(index) if *index==end) {
                    continue;
                }
                skip_group_end = None;
            }
            if let PrimitiveBatch::LayerBoundary(index) = &batch {
                let effects = self.backdrops.as_mut().unwrap();
                let (slot, begin) = effects.groups.boundary(*index);
                if begin {
                    effects.groups.start(slot);
                }
                if begin {
                    executed_groups.push(slot);
                    if effects.groups.hit(slot) {
                        skip_group_end = Some(effects.groups.end(slot));
                    }
                }
            }
            if let PrimitiveBatch::Backdrop(index) = &batch {
                executed_backdrops.push(*index);
            }
            self.render_batch(scene, &custom_outputs, batch)?;
        }
        if let Some(effects) = &mut self.backdrops {
            effects.groups.commit(&executed_groups);
            effects.cache.commit(&executed_backdrops);
        }
        Ok(())
    }

    fn render_batch(
        &mut self,
        scene: &Scene,
        custom_outputs: &[crate::custom_draw::PreparedDraw],
        batch: PrimitiveBatch,
    ) -> Result<()> {
        let annotation = self
            .devices
            .as_ref()
            .and_then(|d| d.annotation.clone())
            .filter(|a| unsafe { a.GetStatus().as_bool() });
        let _annotation = annotation
            .as_ref()
            .map(|a| Annotation::new(a, HSTRING::from(batch.label())));
        match batch {
            PrimitiveBatch::Shaders { program_id, range } => {
                if scene.shader_draws[range.start].path_mask.is_some() {
                    let paths: Vec<_> = scene.shader_draws[range.clone()]
                        .iter()
                        .map(|draw| draw.path_mask.as_ref().unwrap().as_ref().clone())
                        .collect();
                    self.draw_paths_to_intermediate(&paths)?;
                }
                let devices = self.devices.as_ref().context("devices missing")?;
                let resources = self.resources.as_ref().context("resources missing")?;
                self.shaders
                    .as_ref()
                    .context("shader renderer missing")?
                    .draw(
                        &devices.device_context,
                        resources
                            .render_target_view
                            .as_ref()
                            .context("render target missing")?,
                        size(
                            DevicePixels(resources.viewport.Width as i32),
                            DevicePixels(resources.viewport.Height as i32),
                        ),
                        program_id,
                        range.clone(),
                        if scene.shader_draws[range.start].path_mask.is_some() {
                            Some([resources.path_intermediate_srv.clone()])
                        } else {
                            scene.shader_draws[range.start]
                                .glyph_mask
                                .and_then(|tile| self.atlas.get_texture_view(tile.texture_id))
                        },
                    )
            }
            PrimitiveBatch::LayerBoundary(index) => {
                let devices = self.devices.as_ref().context("devices missing")?;
                let resources = self.resources.as_ref().context("resources missing")?;
                self.backdrops
                    .as_ref()
                    .context("effects missing")?
                    .group_boundary(
                        &devices.device_context,
                        resources.render_target.as_ref().context("target missing")?,
                        resources
                            .render_target_view
                            .as_ref()
                            .context("target view missing")?,
                        size(
                            DevicePixels(resources.viewport.Width as i32),
                            DevicePixels(resources.viewport.Height as i32),
                        ),
                        index,
                    )
            }
            PrimitiveBatch::Backdrop(index) => {
                let devices = self.devices.as_ref().context("devices missing")?;
                let resources = self.resources.as_ref().context("resources missing")?;
                self.backdrops
                    .as_mut()
                    .context("backdrop renderer missing")?
                    .draw(
                        &devices.device_context,
                        resources
                            .render_target
                            .as_ref()
                            .context("backdrop target missing")?,
                        resources
                            .render_target_view
                            .as_ref()
                            .context("backdrop target view missing")?,
                        size(
                            DevicePixels(resources.viewport.Width as i32),
                            DevicePixels(resources.viewport.Height as i32),
                        ),
                        index,
                    )
            }
            PrimitiveBatch::Shadows(range) => self.draw_shadows(range.start, range.len()),
            PrimitiveBatch::Quads(range) => self.draw_quads(range.start, range.len()),
            PrimitiveBatch::Paths(range) => {
                let paths = &scene.paths[range];
                self.draw_paths_to_intermediate(paths)?;
                self.draw_paths_from_intermediate(paths)
            }
            PrimitiveBatch::Underlines(range) => self.draw_underlines(range.start, range.len()),
            PrimitiveBatch::MonochromeSprites { texture_id, range } => {
                self.draw_monochrome_sprites(texture_id, range.start, range.len())
            }
            PrimitiveBatch::SubpixelSprites { texture_id, range } => {
                self.draw_subpixel_sprites(texture_id, range.start, range.len())
            }
            PrimitiveBatch::PolychromeSprites { texture_id, range } => {
                self.draw_polychrome_sprites(texture_id, range.start, range.len())
            }
            PrimitiveBatch::Custom(range) => {
                let devices = self.devices.as_ref().context("devices missing")?;
                let resources = self.resources.as_ref().context("resources missing")?;
                let target = resources
                    .render_target_view
                    .as_ref()
                    .context("render target missing")?;
                let viewport = size(
                    DevicePixels(resources.viewport.Width as i32),
                    DevicePixels(resources.viewport.Height as i32),
                );
                self.custom_compositor
                    .as_ref()
                    .context("custom compositor missing")?
                    .draw(
                        &devices.device_context,
                        target,
                        viewport,
                        &scene.custom_draws[range.clone()],
                        &custom_outputs[range],
                    )
            }
            PrimitiveBatch::Surfaces(range) => self.draw_surfaces(&scene.surfaces[range]),
        }
    }

    fn prepare_custom_draws(
        &mut self,
        scene: &Scene,
    ) -> Result<Vec<crate::custom_draw::PreparedDraw>> {
        use crate::custom_draw::*;
        if scene.custom_draws.is_empty() {
            return Ok(Vec::new());
        }
        let devices = self.devices.as_ref().context("devices missing")?.clone();
        let mut batch =
            PreparationBatch::new(devices.device.clone(), devices.device_context.clone());
        let mut compositor = match self.custom_compositor.take() {
            Some(compositor) => compositor,
            None => CustomCompositor::new(&devices.device)?,
        };
        let result = (|| {
            let mut prepared = Vec::with_capacity(scene.custom_draws.len());
            let preparation_frame = gpui::PreparationFrame::default();
            for draw in &scene.custom_draws {
                let prepared_draw = batch.prepare(|context| {
                    let extent = extent(draw)?;
                    let request =
                        PreparationRequest::new(draw.bounds.size.map(|v| px(v.0)), extent, 1.)?;
                    let output =
                        draw.content
                            .prepare_in_frame(context, &request, &preparation_frame)?;
                    let native = output.is::<Background>();
                    let typed_wide = output.is::<Dx11WideSdrOutput>();
                    let wide = typed_wide
                        || draw.content.image_color_mode() == gpui::ImageColorMode::LinearP3Wide;
                    anyhow::ensure!(
                        !wide || self.wide_sdr.is_some(),
                        "wide image requires WideGamutSdr presentation"
                    );
                    let source = if typed_wide {
                        anyhow::ensure!(
                            self.wide_sdr.is_some(),
                            "wide output requires WideGamutSdr presentation"
                        );
                        output
                            .downcast::<Dx11WideSdrOutput>()
                            .unwrap()
                            .validate(context, extent)?
                    } else if native {
                        let background = *output.downcast::<Background>().unwrap();
                        let bounds = Bounds::new(
                            point(ScaledPixels(0.), ScaledPixels(0.)),
                            size(
                                ScaledPixels(extent.width.0 as f32),
                                ScaledPixels(extent.height.0 as f32),
                            ),
                        );
                        let mut scene = Scene::default();
                        scene.insert_primitive(Quad {
                            bounds,
                            content_mask: ContentMask { bounds },
                            background,
                            ..Default::default()
                        });
                        scene.finish();
                        self.render_custom_mask(&scene, extent)?
                    } else {
                        output
                            .downcast::<Dx11CustomOutput>()
                            .map_err(|_| anyhow::anyhow!("producer returned a non-DirectX output"))?
                            .validate(context, extent)?
                    };
                    if draw.content.image_color_mode() != gpui::ImageColorMode::EncodedSrgb {
                        let mut description = D3D11_TEXTURE2D_DESC::default();
                        unsafe {
                            source.texture.GetDesc(&mut description);
                        }
                        anyhow::ensure!(
                            description.Format == DXGI_FORMAT_R16G16B16A16_FLOAT,
                            "linear P3 image metadata requires RGBA16F texture"
                        );
                    }
                    let coverage = if draw.coverage.is_some() {
                        let key = CoverageKey::new(draw)?;
                        if let Some(mask) = compositor.masks.get(&key) {
                            Some(mask.clone())
                        } else {
                            let mask =
                                self.render_custom_mask(&coverage_scene(draw, extent)?, extent)?;
                            let bytes = extent.width.0 as usize * extent.height.0 as usize * 4
                                + key.byte_len();
                            if compositor.mask_bytes + bytes > 64 * 1024 * 1024 {
                                compositor.masks.clear();
                                compositor.mask_bytes = 0;
                            }
                            if bytes <= 64 * 1024 * 1024 {
                                compositor.mask_bytes += bytes;
                                compositor.masks.insert(&key, mask.clone());
                            }
                            Some(mask)
                        }
                    } else {
                        None
                    };
                    Ok(PreparedDraw {
                        source,
                        coverage,
                        native,
                        wide,
                    })
                })?;
                prepared.push(prepared_draw);
            }
            batch.submit()?;
            Ok(prepared)
        })();
        self.custom_compositor = Some(compositor);
        result
    }

    fn render_custom_mask(
        &mut self,
        scene: &Scene,
        extent: Size<DevicePixels>,
    ) -> Result<crate::custom_draw::TextureBundle> {
        let devices = self.devices.as_ref().context("devices missing")?.clone();
        let target = crate::custom_draw::allocate_texture(&devices.device, extent)?;
        let (path_intermediate_texture, path_intermediate_srv) = create_path_intermediate_texture(
            &devices.device,
            extent.width.0 as u32,
            extent.height.0 as u32,
        )?;
        let (path_intermediate_msaa_texture, path_intermediate_msaa_view) =
            create_path_intermediate_msaa_texture_and_view(
                &devices.device,
                extent.width.0 as u32,
                extent.height.0 as u32,
            )?;
        let original = self.resources.take().context("resources missing")?;
        self.resources = Some(DirectXResources {
            swap_chain: original.swap_chain.clone(),
            render_target: Some(target.texture.clone()),
            render_target_view: Some(target.rtv.clone()),
            path_intermediate_texture,
            path_intermediate_srv,
            path_intermediate_msaa_texture,
            path_intermediate_msaa_view,
            viewport: crate::custom_draw::viewport(extent),
        });
        let wide = self.wide_sdr.take();
        let effects = self.backdrops.take();
        let result = self.render(scene, WindowBackgroundAppearance::Transparent);
        self.backdrops = effects;
        self.wide_sdr = wide;
        unsafe {
            devices
                .device_context
                .OMSetRenderTargets(Some(slice::from_ref(&original.render_target_view)), None);
            devices
                .device_context
                .RSSetViewports(Some(slice::from_ref(&original.viewport)));
        }
        self.resources = Some(original);
        unsafe {
            devices
                .device_context
                .PSSetShaderResources(0, Some(&[None, None]));
        }
        result?;
        Ok(target)
    }

    /// Render `scene` to an offscreen CPU image **without presenting** so
    /// the window need never be shown or visible (the macOS headless path
    /// goes through MetalRenderer; this is the Windows analogue). Draws into
    /// the existing render target, copies it into a `D3D11_USAGE_STAGING`
    /// texture, maps it, and converts BGRA to RGBA.
    #[cfg(any(test, feature = "test-support"))]
    pub(crate) fn render_to_image(
        &mut self,
        scene: &Scene,
        background_appearance: WindowBackgroundAppearance,
    ) -> Result<image::RgbaImage> {
        // A pending device-lost recovery (`skip_draws`) leaves the atlas holding
        // tile references from the previous device; drawing before the forced
        // re-render rebuilds them panics in `DirectXAtlasTextures::texture`.
        anyhow::ensure!(
            !self.skip_draws,
            "render_to_image unavailable while recovering from a lost device"
        );
        self.render(scene, background_appearance)?;

        let devices = self.devices.as_ref().context("devices missing")?;
        let device = &devices.device;
        let context = &devices.device_context;
        let resources = self.resources.as_ref().context("resources missing")?;
        let preview;
        let render_target = if let Some(wide) = &self.wide_sdr {
            preview = wide.preview(devices)?;
            &preview
        } else {
            resources
                .render_target
                .as_ref()
                .context("render target missing")?
        };

        // A CPU-readable copy of the render target.
        let mut desc = D3D11_TEXTURE2D_DESC::default();
        unsafe { render_target.GetDesc(&mut desc) };
        let width = desc.Width;
        let height = desc.Height;
        let staging_desc = D3D11_TEXTURE2D_DESC {
            Usage: D3D11_USAGE_STAGING,
            BindFlags: 0,
            CPUAccessFlags: D3D11_CPU_ACCESS_READ.0 as u32,
            MiscFlags: 0,
            MipLevels: 1,
            ArraySize: 1,
            SampleDesc: DXGI_SAMPLE_DESC {
                Count: 1,
                Quality: 0,
            },
            ..desc
        };
        let mut staging: Option<ID3D11Texture2D> = None;
        unsafe { device.CreateTexture2D(&staging_desc, None, Some(&mut staging))? };
        let staging = staging.context("creating staging texture")?;
        unsafe { context.CopyResource(&staging, render_target) };

        let mut mapped = D3D11_MAPPED_SUBRESOURCE::default();
        unsafe { context.Map(&staging, 0, D3D11_MAP_READ, 0, Some(&mut mapped))? };
        let row_bytes = (width as usize) * 4;
        let mut pixels = vec![0u8; row_bytes * height as usize];
        // SAFETY: `Map` succeeded, so `pData` points at `RowPitch * height`
        // readable bytes for as long as the mapping is held, and `RowPitch >=
        // row_bytes` (it only ever adds trailing padding). `pixels` is sized
        // `row_bytes * height`, so every copy stays in bounds on both sides,
        // and the regions cannot overlap (`pixels` is a fresh allocation).
        unsafe {
            let src = mapped.pData as *const u8;
            for row in 0..height as usize {
                let s = src.add(row * mapped.RowPitch as usize);
                let d = pixels.as_mut_ptr().add(row * row_bytes);
                std::ptr::copy_nonoverlapping(s, d, row_bytes);
            }
            context.Unmap(&staging, 0);
        }
        // The render target is BGRA; image::RgbaImage expects RGBA.
        for px in pixels.chunks_exact_mut(4) {
            px.swap(0, 2);
        }
        image::RgbaImage::from_raw(width, height, pixels)
            .context("Failed to build RgbaImage from staging readback")
    }

    pub(crate) fn resize(&mut self, new_size: Size<DevicePixels>) -> Result<()> {
        let width = new_size.width.0.max(1) as u32;
        let height = new_size.height.0.max(1) as u32;
        if self.width == width && self.height == height {
            return Ok(());
        }
        let buffer_format = self.buffer_format();
        if self.wide_sdr.is_some() {
            wide_sdr::validate_extent(size(
                DevicePixels(width as i32),
                DevicePixels(height as i32),
            ))?;
        }

        // Clear the render target before resizing
        let devices = self.devices.as_ref().context("devices missing")?;
        unsafe { devices.device_context.OMSetRenderTargets(None, None) };
        let resources = self.resources.as_mut().context("resources missing")?;
        if resources.swap_chain.is_none() {
            *resources = DirectXResources::offscreen(devices, width, height)?;
            self.width = width;
            self.height = height;
            return Ok(());
        }
        resources.render_target.take();
        resources.render_target_view.take();

        // Resizing the swap chain requires a call to the underlying DXGI adapter, which can return the device removed error.
        // The app might have moved to a monitor that's attached to a different graphics device.
        // When a graphics device is removed or reset, the desktop resolution often changes, resulting in a window size change.
        // But here we just return the error, because we are handling device lost scenarios elsewhere.
        unsafe {
            resources
                .swap_chain
                .as_ref()
                .context("missing swap chain")?
                .ResizeBuffers(
                    BUFFER_COUNT as u32,
                    width,
                    height,
                    buffer_format,
                    DXGI_SWAP_CHAIN_FLAG(0),
                )
                .context("Failed to resize swap chain")?;
        }

        resources.recreate_resources(devices, width, height)?;

        unsafe {
            devices
                .device_context
                .OMSetRenderTargets(Some(slice::from_ref(&resources.render_target_view)), None);
        }

        self.width = width;
        self.height = height;
        self.resize_color_space()?;

        Ok(())
    }

    fn upload_scene_buffers(&mut self, scene: &Scene) -> Result<()> {
        let devices = self.devices.as_ref().context("devices missing")?;

        if !scene.shadows.is_empty() {
            self.pipelines.shadow_pipeline.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.shadows,
            )?;
        }

        if !scene.quads.is_empty() {
            self.pipelines.quad_pipeline.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.quads,
            )?;
        }

        if !scene.underlines.is_empty() {
            self.pipelines.underline_pipeline.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.underlines,
            )?;
        }

        if !scene.monochrome_sprites.is_empty() {
            self.pipelines.mono_sprites.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.monochrome_sprites,
            )?;
        }

        if !scene.subpixel_sprites.is_empty() {
            self.pipelines.subpixel_sprites.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.subpixel_sprites,
            )?;
        }

        if !scene.polychrome_sprites.is_empty() {
            self.pipelines.poly_sprites.update_buffer(
                &devices.device,
                &devices.device_context,
                &scene.polychrome_sprites,
            )?;
        }

        Ok(())
    }

    fn draw_shadows(&mut self, start: usize, len: usize) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        self.pipelines.shadow_pipeline.draw_range(
            &devices.device_context,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            start as u32,
            len as u32,
        )
    }

    fn draw_quads(&mut self, start: usize, len: usize) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        self.pipelines.quad_pipeline.draw_range(
            &devices.device_context,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            start as u32,
            len as u32,
        )
    }

    fn draw_paths_to_intermediate(&mut self, paths: &[Path<ScaledPixels>]) -> Result<()> {
        if paths.is_empty() {
            return Ok(());
        }

        let devices = self.devices.as_ref().context("devices missing")?;
        let resources = self.resources.as_ref().context("resources missing")?;
        // Clear intermediate MSAA texture
        unsafe {
            devices.device_context.ClearRenderTargetView(
                resources.path_intermediate_msaa_view.as_ref().unwrap(),
                &[0.0; 4],
            );
            // Set intermediate MSAA texture as render target
            devices.device_context.OMSetRenderTargets(
                Some(slice::from_ref(&resources.path_intermediate_msaa_view)),
                None,
            );
        }

        // Collect all vertices and sprites for a single draw call
        let mut vertices = Vec::new();

        for path in paths {
            vertices.extend(path.vertices.iter().map(|v| PathRasterizationSprite {
                xy_position: v.xy_position,
                st_position: v.st_position,
                color: path.color,
                bounds: path.clipped_bounds(),
            }));
        }

        self.pipelines.path_rasterization_pipeline.update_buffer(
            &devices.device,
            &devices.device_context,
            &vertices,
        )?;

        self.pipelines.path_rasterization_pipeline.draw(
            &devices.device_context,
            D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST,
            vertices.len() as u32,
            1,
        )?;

        // Resolve MSAA to non-MSAA intermediate texture
        unsafe {
            devices.device_context.ResolveSubresource(
                &resources.path_intermediate_texture,
                0,
                &resources.path_intermediate_msaa_texture,
                0,
                RENDER_TARGET_FORMAT,
            );
            // Restore main render target
            devices
                .device_context
                .OMSetRenderTargets(Some(slice::from_ref(&resources.render_target_view)), None);
        }

        Ok(())
    }

    fn draw_paths_from_intermediate(&mut self, paths: &[Path<ScaledPixels>]) -> Result<()> {
        let Some(first_path) = paths.first() else {
            return Ok(());
        };

        // When copying paths from the intermediate texture to the drawable,
        // each pixel must only be copied once, in case of transparent paths.
        //
        // If all paths have the same draw order, then their bounds are all
        // disjoint, so we can copy each path's bounds individually. If this
        // batch combines different draw orders, we perform a single copy
        // for a minimal spanning rect.
        let sprites = if paths.last().unwrap().order == first_path.order {
            paths
                .iter()
                .map(|path| PathSprite {
                    bounds: path.clipped_bounds(),
                })
                .collect::<Vec<_>>()
        } else {
            let mut bounds = first_path.clipped_bounds();
            for path in paths.iter().skip(1) {
                bounds = bounds.union(&path.clipped_bounds());
            }
            vec![PathSprite { bounds }]
        };

        let devices = self.devices.as_ref().context("devices missing")?;
        let resources = self.resources.as_ref().context("resources missing")?;
        self.pipelines.path_sprite_pipeline.update_buffer(
            &devices.device,
            &devices.device_context,
            &sprites,
        )?;

        // Draw the sprites with the path texture
        self.pipelines.path_sprite_pipeline.draw_with_texture(
            &devices.device_context,
            slice::from_ref(&resources.path_intermediate_srv),
            slice::from_ref(&self.globals.sampler),
            sprites.len() as u32,
        )
    }

    fn draw_underlines(&mut self, start: usize, len: usize) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        self.pipelines.underline_pipeline.draw_range(
            &devices.device_context,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            start as u32,
            len as u32,
        )
    }

    fn draw_monochrome_sprites(
        &mut self,
        texture_id: AtlasTextureId,
        start: usize,
        len: usize,
    ) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        let Some(texture_view) = self.atlas.get_texture_view(texture_id) else {
            return Ok(());
        };
        self.pipelines.mono_sprites.draw_range_with_texture(
            &devices.device_context,
            &texture_view,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            slice::from_ref(&self.globals.sampler),
            start as u32,
            len as u32,
        )
    }

    fn draw_subpixel_sprites(
        &mut self,
        texture_id: AtlasTextureId,
        start: usize,
        len: usize,
    ) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        let Some(texture_view) = self.atlas.get_texture_view(texture_id) else {
            return Ok(());
        };
        self.pipelines.subpixel_sprites.draw_range_with_texture(
            &devices.device_context,
            &texture_view,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            slice::from_ref(&self.globals.sampler),
            start as u32,
            len as u32,
        )
    }

    fn draw_polychrome_sprites(
        &mut self,
        texture_id: AtlasTextureId,
        start: usize,
        len: usize,
    ) -> Result<()> {
        if len == 0 {
            return Ok(());
        }
        let devices = self.devices.as_ref().context("devices missing")?;
        let Some(texture_view) = self.atlas.get_texture_view(texture_id) else {
            return Ok(());
        };
        self.pipelines.poly_sprites.draw_range_with_texture(
            &devices.device_context,
            &texture_view,
            self.globals
                .batch_params_buffer
                .as_ref()
                .context("batch params buffer missing")?,
            slice::from_ref(&self.globals.sampler),
            start as u32,
            len as u32,
        )
    }

    fn draw_surfaces(&mut self, surfaces: &[PaintSurface]) -> Result<()> {
        if surfaces.is_empty() {
            return Ok(());
        }
        Ok(())
    }

    pub(crate) fn gpu_specs(&self) -> Result<GpuSpecs> {
        let devices = self.devices.as_ref().context("devices missing")?;
        let desc = unsafe { devices.adapter.GetDesc1() }?;
        let is_software_emulated = (desc.Flags & DXGI_ADAPTER_FLAG_SOFTWARE.0 as u32) != 0;
        let device_name = String::from_utf16_lossy(&desc.Description)
            .trim_matches(char::from(0))
            .to_string();
        let driver_name = match desc.VendorId {
            0x10DE => "NVIDIA Corporation".to_string(),
            0x1002 => "AMD Corporation".to_string(),
            0x8086 => "Intel Corporation".to_string(),
            id => format!("Unknown Vendor (ID: {:#X})", id),
        };
        let driver_version = match desc.VendorId {
            0x10DE => nvidia::get_driver_version(),
            0x1002 => amd::get_driver_version(),
            // For Intel and other vendors, we use the DXGI API to get the driver version.
            _ => dxgi::get_driver_version(&devices.adapter),
        }
        .context("Failed to get gpu driver info")
        .log_err()
        .unwrap_or("Unknown Driver".to_string());
        Ok(GpuSpecs {
            is_software_emulated,
            device_name,
            driver_name,
            driver_info: driver_version,
        })
    }

    pub(crate) fn get_font_info() -> &'static FontInfo {
        static CACHED_FONT_INFO: OnceLock<FontInfo> = OnceLock::new();
        CACHED_FONT_INFO.get_or_init(|| unsafe {
            let factory: IDWriteFactory5 = DWriteCreateFactory(DWRITE_FACTORY_TYPE_SHARED).unwrap();
            let render_params: IDWriteRenderingParams1 =
                factory.CreateRenderingParams().unwrap().cast().unwrap();
            FontInfo {
                gamma_ratios: gpui::get_gamma_correction_ratios(render_params.GetGamma()),
                grayscale_enhanced_contrast: render_params.GetGrayscaleEnhancedContrast(),
                subpixel_enhanced_contrast: render_params.GetEnhancedContrast(),
                is_bgr: render_params.GetPixelGeometry() == DWRITE_PIXEL_GEOMETRY_BGR,
            }
        })
    }

    pub(crate) fn mark_drawable(&mut self) {
        self.skip_draws = false;
    }
}

impl DirectXResources {
    fn offscreen(devices: &DirectXRendererDevices, width: u32, height: u32) -> Result<Self> {
        let extent = size(DevicePixels(width as i32), DevicePixels(height as i32));
        let target = crate::custom_draw::allocate_texture(&devices.device, extent)?;
        let (path_intermediate_texture, path_intermediate_srv) =
            create_path_intermediate_texture(&devices.device, width, height)?;
        let (path_intermediate_msaa_texture, path_intermediate_msaa_view) =
            create_path_intermediate_msaa_texture_and_view(&devices.device, width, height)?;
        Ok(Self {
            swap_chain: None,
            render_target: Some(target.texture),
            render_target_view: Some(target.rtv),
            path_intermediate_texture,
            path_intermediate_srv,
            path_intermediate_msaa_texture,
            path_intermediate_msaa_view,
            viewport: crate::custom_draw::viewport(extent),
        })
    }
    pub fn new(
        devices: &DirectXRendererDevices,
        width: u32,
        height: u32,
        hwnd: HWND,
        disable_direct_composition: bool,
    ) -> Result<Self> {
        let swap_chain = if disable_direct_composition {
            create_swap_chain(&devices.dxgi_factory, &devices.device, hwnd, width, height)?
        } else {
            create_swap_chain_for_composition(
                &devices.dxgi_factory,
                &devices.device,
                width,
                height,
            )?
        };

        let (
            render_target,
            render_target_view,
            path_intermediate_texture,
            path_intermediate_srv,
            path_intermediate_msaa_texture,
            path_intermediate_msaa_view,
            viewport,
        ) = create_resources(devices, &swap_chain, width, height)?;
        set_rasterizer_state(&devices.device, &devices.device_context)?;

        Ok(Self {
            swap_chain: Some(swap_chain),
            render_target: Some(render_target),
            render_target_view,
            path_intermediate_texture,
            path_intermediate_msaa_texture,
            path_intermediate_msaa_view,
            path_intermediate_srv,
            viewport,
        })
    }

    #[inline]
    fn recreate_resources(
        &mut self,
        devices: &DirectXRendererDevices,
        width: u32,
        height: u32,
    ) -> Result<()> {
        let (
            render_target,
            render_target_view,
            path_intermediate_texture,
            path_intermediate_srv,
            path_intermediate_msaa_texture,
            path_intermediate_msaa_view,
            viewport,
        ) = create_resources(
            devices,
            self.swap_chain.as_ref().context("missing swap chain")?,
            width,
            height,
        )?;
        self.render_target = Some(render_target);
        self.render_target_view = render_target_view;
        self.path_intermediate_texture = path_intermediate_texture;
        self.path_intermediate_msaa_texture = path_intermediate_msaa_texture;
        self.path_intermediate_msaa_view = path_intermediate_msaa_view;
        self.path_intermediate_srv = path_intermediate_srv;
        self.viewport = viewport;
        Ok(())
    }
}

impl DirectXRenderPipelines {
    pub fn new(device: &ID3D11Device) -> Result<Self> {
        let shadow_pipeline = PipelineState::new(
            device,
            "shadow_pipeline",
            ShaderModule::Shadow,
            4,
            create_blend_state(device)?,
        )?;
        let quad_pipeline = PipelineState::new(
            device,
            "quad_pipeline",
            ShaderModule::Quad,
            64,
            create_blend_state(device)?,
        )?;
        let path_rasterization_pipeline = PipelineState::new(
            device,
            "path_rasterization_pipeline",
            ShaderModule::PathRasterization,
            32,
            create_blend_state_for_path_rasterization(device)?,
        )?;
        let path_sprite_pipeline = PipelineState::new(
            device,
            "path_sprite_pipeline",
            ShaderModule::PathSprite,
            4,
            create_blend_state_for_path_sprite(device)?,
        )?;
        let underline_pipeline = PipelineState::new(
            device,
            "underline_pipeline",
            ShaderModule::Underline,
            4,
            create_blend_state(device)?,
        )?;
        let mono_sprites = PipelineState::new(
            device,
            "monochrome_sprite_pipeline",
            ShaderModule::MonochromeSprite,
            512,
            create_blend_state(device)?,
        )?;
        let subpixel_sprites = PipelineState::new(
            device,
            "subpixel_sprite_pipeline",
            ShaderModule::SubpixelSprite,
            512,
            create_blend_state_for_subpixel_rendering(device)?,
        )?;
        let poly_sprites = PipelineState::new(
            device,
            "polychrome_sprite_pipeline",
            ShaderModule::PolychromeSprite,
            16,
            create_blend_state(device)?,
        )?;

        Ok(Self {
            shadow_pipeline,
            quad_pipeline,
            path_rasterization_pipeline,
            path_sprite_pipeline,
            underline_pipeline,
            mono_sprites,
            subpixel_sprites,
            poly_sprites,
        })
    }
}

impl DirectComposition {
    pub fn new(dxgi_device: &IDXGIDevice, hwnd: HWND) -> Result<Self> {
        let comp_device = get_comp_device(dxgi_device)?;
        let comp_target = unsafe { comp_device.CreateTargetForHwnd(hwnd, true) }?;
        let comp_visual = unsafe { comp_device.CreateVisual() }?;

        Ok(Self {
            comp_device,
            comp_target,
            comp_visual,
        })
    }

    pub fn set_swap_chain(&self, swap_chain: &IDXGISwapChain1) -> Result<()> {
        unsafe {
            self.comp_visual.SetContent(swap_chain)?;
            self.comp_target.SetRoot(&self.comp_visual)?;
            self.comp_device.Commit()?;
        }
        Ok(())
    }
}

impl DirectXGlobalElements {
    pub fn new(device: &ID3D11Device) -> Result<Self> {
        let global_params_buffer = create_constant_buffer::<GlobalParams>(device)?;
        let batch_params_buffer = create_constant_buffer::<BatchParams>(device)?;

        let sampler = unsafe {
            let desc = D3D11_SAMPLER_DESC {
                Filter: D3D11_FILTER_MIN_MAG_MIP_LINEAR,
                AddressU: D3D11_TEXTURE_ADDRESS_WRAP,
                AddressV: D3D11_TEXTURE_ADDRESS_WRAP,
                AddressW: D3D11_TEXTURE_ADDRESS_WRAP,
                MipLODBias: 0.0,
                MaxAnisotropy: 1,
                ComparisonFunc: D3D11_COMPARISON_ALWAYS,
                BorderColor: [0.0; 4],
                MinLOD: 0.0,
                MaxLOD: D3D11_FLOAT32_MAX,
            };
            let mut output = None;
            device.CreateSamplerState(&desc, Some(&mut output))?;
            output
        };

        Ok(Self {
            global_params_buffer,
            batch_params_buffer,
            sampler,
        })
    }
}

#[derive(Debug, Default)]
#[repr(C)]
struct GlobalParams {
    gamma_ratios: [f32; 4],
    viewport_size: [f32; 2],
    grayscale_enhanced_contrast: f32,
    subpixel_enhanced_contrast: f32,
    is_bgr: u32,
    _pad: [u32; 3],
}

#[derive(Clone, Copy, Debug, Default)]
#[repr(C, align(16))]
struct BatchParams {
    start_index: u32,
    _padding: [u32; 3],
}

const _: () = assert!(std::mem::size_of::<BatchParams>() == 16);

struct PipelineState<T> {
    label: &'static str,
    vertex: ID3D11VertexShader,
    fragment: ID3D11PixelShader,
    buffer: ID3D11Buffer,
    buffer_size: usize,
    view: Option<ID3D11ShaderResourceView>,
    blend_state: ID3D11BlendState,
    _marker: std::marker::PhantomData<T>,
}

impl<T> PipelineState<T> {
    fn new(
        device: &ID3D11Device,
        label: &'static str,
        shader_module: ShaderModule,
        buffer_size: usize,
        blend_state: ID3D11BlendState,
    ) -> Result<Self> {
        let vertex = {
            let raw_shader = RawShaderBytes::new(shader_module, ShaderTarget::Vertex)?;
            create_vertex_shader(device, raw_shader.as_bytes())?
        };
        let fragment = {
            let raw_shader = RawShaderBytes::new(shader_module, ShaderTarget::Fragment)?;
            create_fragment_shader(device, raw_shader.as_bytes())?
        };
        let buffer = create_buffer(device, std::mem::size_of::<T>(), buffer_size)?;
        let view = create_buffer_view(device, &buffer)?;

        Ok(PipelineState {
            label,
            vertex,
            fragment,
            buffer,
            buffer_size,
            view,
            blend_state,
            _marker: std::marker::PhantomData,
        })
    }

    fn update_buffer(
        &mut self,
        device: &ID3D11Device,
        device_context: &ID3D11DeviceContext,
        data: &[T],
    ) -> Result<()> {
        if self.buffer_size < data.len() {
            let element_size = std::mem::size_of::<T>();
            let required_size = std::mem::size_of_val(data);
            anyhow::ensure!(
                required_size <= MAX_INSTANCE_BUFFER_SIZE,
                "{} buffer needs {required_size} bytes, above the maximum of {MAX_INSTANCE_BUFFER_SIZE}",
                self.label
            );
            let new_buffer_size = data
                .len()
                .next_power_of_two()
                .min(MAX_INSTANCE_BUFFER_SIZE / element_size);
            log::debug!(
                "Updating {} buffer size from {} to {}",
                self.label,
                self.buffer_size,
                new_buffer_size
            );
            let buffer = create_buffer(device, std::mem::size_of::<T>(), new_buffer_size)?;
            let view = create_buffer_view(device, &buffer)?;
            self.buffer = buffer;
            self.view = view;
            self.buffer_size = new_buffer_size;
        }
        update_buffer(device_context, &self.buffer, data)
    }

    fn draw(
        &self,
        device_context: &ID3D11DeviceContext,
        topology: D3D_PRIMITIVE_TOPOLOGY,
        vertex_count: u32,
        instance_count: u32,
    ) -> Result<()> {
        set_pipeline_state(
            device_context,
            slice::from_ref(&self.view),
            topology,
            &self.vertex,
            &self.fragment,
            &self.blend_state,
        );
        unsafe {
            device_context.DrawInstanced(vertex_count, instance_count, 0, 0);
        }
        Ok(())
    }

    fn draw_with_texture(
        &self,
        device_context: &ID3D11DeviceContext,
        texture: &[Option<ID3D11ShaderResourceView>],
        sampler: &[Option<ID3D11SamplerState>],
        instance_count: u32,
    ) -> Result<()> {
        set_pipeline_state(
            device_context,
            slice::from_ref(&self.view),
            D3D_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP,
            &self.vertex,
            &self.fragment,
            &self.blend_state,
        );
        unsafe {
            device_context.PSSetSamplers(0, Some(sampler));
            device_context.VSSetShaderResources(0, Some(texture));
            device_context.PSSetShaderResources(0, Some(texture));

            device_context.DrawInstanced(4, instance_count, 0, 0);
        }
        Ok(())
    }

    fn draw_range(
        &self,
        device_context: &ID3D11DeviceContext,
        batch_params_buffer: &ID3D11Buffer,
        first_instance: u32,
        instance_count: u32,
    ) -> Result<()> {
        anyhow::ensure!(
            first_instance as usize + instance_count as usize <= self.buffer_size,
            "DirectX instance range exceeds the {} buffer",
            self.label
        );
        update_batch_start(device_context, batch_params_buffer, first_instance)?;
        set_pipeline_state(
            device_context,
            slice::from_ref(&self.view),
            D3D_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP,
            &self.vertex,
            &self.fragment,
            &self.blend_state,
        );
        unsafe {
            device_context.DrawInstanced(4, instance_count, 0, 0);
        }
        Ok(())
    }

    fn draw_range_with_texture(
        &self,
        device_context: &ID3D11DeviceContext,
        texture: &[Option<ID3D11ShaderResourceView>],
        batch_params_buffer: &ID3D11Buffer,
        sampler: &[Option<ID3D11SamplerState>],
        first_instance: u32,
        instance_count: u32,
    ) -> Result<()> {
        anyhow::ensure!(
            first_instance as usize + instance_count as usize <= self.buffer_size,
            "DirectX instance range exceeds the {} buffer",
            self.label
        );
        update_batch_start(device_context, batch_params_buffer, first_instance)?;
        set_pipeline_state(
            device_context,
            slice::from_ref(&self.view),
            D3D_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP,
            &self.vertex,
            &self.fragment,
            &self.blend_state,
        );
        unsafe {
            device_context.PSSetSamplers(0, Some(sampler));
            device_context.VSSetShaderResources(0, Some(texture));
            device_context.PSSetShaderResources(0, Some(texture));
            device_context.DrawInstanced(4, instance_count, 0, 0);
        }
        Ok(())
    }
}

#[derive(Clone, Copy)]
#[repr(C)]
struct PathRasterizationSprite {
    xy_position: Point<ScaledPixels>,
    st_position: Point<f32>,
    color: Background,
    bounds: Bounds<ScaledPixels>,
}

#[derive(Clone, Copy)]
#[repr(C)]
struct PathSprite {
    bounds: Bounds<ScaledPixels>,
}

impl Drop for DirectXRenderer {
    fn drop(&mut self) {
        #[cfg(debug_assertions)]
        if let Some(devices) = &self.devices {
            report_live_objects(&devices.device).ok();
        }
    }
}

#[inline]
fn get_comp_device(dxgi_device: &IDXGIDevice) -> Result<IDCompositionDevice> {
    Ok(unsafe { DCompositionCreateDevice(dxgi_device)? })
}

fn create_swap_chain_for_composition(
    dxgi_factory: &IDXGIFactory6,
    device: &ID3D11Device,
    width: u32,
    height: u32,
) -> Result<IDXGISwapChain1> {
    let desc = DXGI_SWAP_CHAIN_DESC1 {
        Width: width,
        Height: height,
        Format: RENDER_TARGET_FORMAT,
        Stereo: false.into(),
        SampleDesc: DXGI_SAMPLE_DESC {
            Count: 1,
            Quality: 0,
        },
        BufferUsage: DXGI_USAGE_RENDER_TARGET_OUTPUT,
        BufferCount: BUFFER_COUNT as u32,
        // Composition SwapChains only support the DXGI_SCALING_STRETCH Scaling.
        Scaling: DXGI_SCALING_STRETCH,
        SwapEffect: DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL,
        AlphaMode: DXGI_ALPHA_MODE_PREMULTIPLIED,
        Flags: 0,
    };
    Ok(unsafe { dxgi_factory.CreateSwapChainForComposition(device, &desc, None)? })
}

fn create_swap_chain(
    dxgi_factory: &IDXGIFactory6,
    device: &ID3D11Device,
    hwnd: HWND,
    width: u32,
    height: u32,
) -> Result<IDXGISwapChain1> {
    use windows::Win32::Graphics::Dxgi::DXGI_MWA_NO_ALT_ENTER;

    let desc = DXGI_SWAP_CHAIN_DESC1 {
        Width: width,
        Height: height,
        Format: RENDER_TARGET_FORMAT,
        Stereo: false.into(),
        SampleDesc: DXGI_SAMPLE_DESC {
            Count: 1,
            Quality: 0,
        },
        BufferUsage: DXGI_USAGE_RENDER_TARGET_OUTPUT,
        BufferCount: BUFFER_COUNT as u32,
        Scaling: DXGI_SCALING_NONE,
        SwapEffect: DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL,
        AlphaMode: DXGI_ALPHA_MODE_IGNORE,
        Flags: 0,
    };
    let swap_chain =
        unsafe { dxgi_factory.CreateSwapChainForHwnd(device, hwnd, &desc, None, None) }?;
    unsafe { dxgi_factory.MakeWindowAssociation(hwnd, DXGI_MWA_NO_ALT_ENTER) }?;
    Ok(swap_chain)
}

#[inline]
fn create_resources(
    devices: &DirectXRendererDevices,
    swap_chain: &IDXGISwapChain1,
    width: u32,
    height: u32,
) -> Result<(
    ID3D11Texture2D,
    Option<ID3D11RenderTargetView>,
    ID3D11Texture2D,
    Option<ID3D11ShaderResourceView>,
    ID3D11Texture2D,
    Option<ID3D11RenderTargetView>,
    D3D11_VIEWPORT,
)> {
    let (render_target, render_target_view) =
        create_render_target_and_its_view(swap_chain, &devices.device)?;
    let (path_intermediate_texture, path_intermediate_srv) =
        create_path_intermediate_texture(&devices.device, width, height)?;
    let (path_intermediate_msaa_texture, path_intermediate_msaa_view) =
        create_path_intermediate_msaa_texture_and_view(&devices.device, width, height)?;
    let viewport = D3D11_VIEWPORT {
        TopLeftX: 0.0,
        TopLeftY: 0.0,
        Width: width as f32,
        Height: height as f32,
        MinDepth: 0.0,
        MaxDepth: 1.0,
    };
    Ok((
        render_target,
        render_target_view,
        path_intermediate_texture,
        path_intermediate_srv,
        path_intermediate_msaa_texture,
        path_intermediate_msaa_view,
        viewport,
    ))
}

#[inline]
fn create_render_target_and_its_view(
    swap_chain: &IDXGISwapChain1,
    device: &ID3D11Device,
) -> Result<(ID3D11Texture2D, Option<ID3D11RenderTargetView>)> {
    let render_target: ID3D11Texture2D = unsafe { swap_chain.GetBuffer(0) }?;
    let mut render_target_view = None;
    unsafe { device.CreateRenderTargetView(&render_target, None, Some(&mut render_target_view))? };
    Ok((render_target, render_target_view))
}

#[inline]
fn create_path_intermediate_texture(
    device: &ID3D11Device,
    width: u32,
    height: u32,
) -> Result<(ID3D11Texture2D, Option<ID3D11ShaderResourceView>)> {
    let texture = unsafe {
        let mut output = None;
        let desc = D3D11_TEXTURE2D_DESC {
            Width: width,
            Height: height,
            MipLevels: 1,
            ArraySize: 1,
            Format: RENDER_TARGET_FORMAT,
            SampleDesc: DXGI_SAMPLE_DESC {
                Count: 1,
                Quality: 0,
            },
            Usage: D3D11_USAGE_DEFAULT,
            BindFlags: (D3D11_BIND_RENDER_TARGET.0 | D3D11_BIND_SHADER_RESOURCE.0) as u32,
            CPUAccessFlags: 0,
            MiscFlags: 0,
        };
        device.CreateTexture2D(&desc, None, Some(&mut output))?;
        output.unwrap()
    };

    let mut shader_resource_view = None;
    unsafe { device.CreateShaderResourceView(&texture, None, Some(&mut shader_resource_view))? };

    Ok((texture, Some(shader_resource_view.unwrap())))
}

#[inline]
fn create_path_intermediate_msaa_texture_and_view(
    device: &ID3D11Device,
    width: u32,
    height: u32,
) -> Result<(ID3D11Texture2D, Option<ID3D11RenderTargetView>)> {
    let msaa_texture = unsafe {
        let mut output = None;
        let desc = D3D11_TEXTURE2D_DESC {
            Width: width,
            Height: height,
            MipLevels: 1,
            ArraySize: 1,
            Format: RENDER_TARGET_FORMAT,
            SampleDesc: DXGI_SAMPLE_DESC {
                Count: PATH_MULTISAMPLE_COUNT,
                Quality: D3D11_STANDARD_MULTISAMPLE_PATTERN.0 as u32,
            },
            Usage: D3D11_USAGE_DEFAULT,
            BindFlags: D3D11_BIND_RENDER_TARGET.0 as u32,
            CPUAccessFlags: 0,
            MiscFlags: 0,
        };
        device.CreateTexture2D(&desc, None, Some(&mut output))?;
        output.unwrap()
    };
    let mut msaa_view = None;
    unsafe { device.CreateRenderTargetView(&msaa_texture, None, Some(&mut msaa_view))? };
    Ok((msaa_texture, Some(msaa_view.unwrap())))
}

#[inline]
fn set_rasterizer_state(device: &ID3D11Device, device_context: &ID3D11DeviceContext) -> Result<()> {
    let desc = D3D11_RASTERIZER_DESC {
        FillMode: D3D11_FILL_SOLID,
        CullMode: D3D11_CULL_NONE,
        FrontCounterClockwise: false.into(),
        DepthBias: 0,
        DepthBiasClamp: 0.0,
        SlopeScaledDepthBias: 0.0,
        DepthClipEnable: true.into(),
        ScissorEnable: false.into(),
        MultisampleEnable: true.into(),
        AntialiasedLineEnable: false.into(),
    };
    let rasterizer_state = unsafe {
        let mut state = None;
        device.CreateRasterizerState(&desc, Some(&mut state))?;
        state.unwrap()
    };
    unsafe { device_context.RSSetState(&rasterizer_state) };
    Ok(())
}

// https://learn.microsoft.com/en-us/windows/win32/api/d3d11/ns-d3d11-d3d11_blend_desc
#[inline]
fn create_blend_state(device: &ID3D11Device) -> Result<ID3D11BlendState> {
    let mut desc = D3D11_BLEND_DESC::default();
    desc.RenderTarget[0].BlendEnable = true.into();
    desc.RenderTarget[0].BlendOp = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].SrcBlend = D3D11_BLEND_SRC_ALPHA;
    desc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_ONE;
    desc.RenderTarget[0].DestBlend = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL.0 as u8;
    unsafe {
        let mut state = None;
        device.CreateBlendState(&desc, Some(&mut state))?;
        Ok(state.unwrap())
    }
}

#[inline]
fn create_blend_state_for_subpixel_rendering(device: &ID3D11Device) -> Result<ID3D11BlendState> {
    let mut desc = D3D11_BLEND_DESC::default();
    desc.RenderTarget[0].BlendEnable = true.into();
    desc.RenderTarget[0].BlendOp = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].SrcBlend = D3D11_BLEND_SRC1_COLOR;
    desc.RenderTarget[0].DestBlend = D3D11_BLEND_INV_SRC1_COLOR;
    // It does not make sense to draw transparent subpixel-rendered text, since it cannot be meaningfully alpha-blended onto anything else.
    desc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_ONE;
    desc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_ZERO;
    desc.RenderTarget[0].RenderTargetWriteMask =
        D3D11_COLOR_WRITE_ENABLE_ALL.0 as u8 & !D3D11_COLOR_WRITE_ENABLE_ALPHA.0 as u8;

    unsafe {
        let mut state = None;
        device.CreateBlendState(&desc, Some(&mut state))?;
        Ok(state.unwrap())
    }
}

#[inline]
fn create_blend_state_for_path_rasterization(device: &ID3D11Device) -> Result<ID3D11BlendState> {
    // If the feature level is set to greater than D3D_FEATURE_LEVEL_9_3, the display
    // device performs the blend in linear space, which is ideal.
    let mut desc = D3D11_BLEND_DESC::default();
    desc.RenderTarget[0].BlendEnable = true.into();
    desc.RenderTarget[0].BlendOp = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].SrcBlend = D3D11_BLEND_ONE;
    desc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_ONE;
    desc.RenderTarget[0].DestBlend = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL.0 as u8;
    unsafe {
        let mut state = None;
        device.CreateBlendState(&desc, Some(&mut state))?;
        Ok(state.unwrap())
    }
}

#[inline]
fn create_blend_state_for_path_sprite(device: &ID3D11Device) -> Result<ID3D11BlendState> {
    // If the feature level is set to greater than D3D_FEATURE_LEVEL_9_3, the display
    // device performs the blend in linear space, which is ideal.
    let mut desc = D3D11_BLEND_DESC::default();
    desc.RenderTarget[0].BlendEnable = true.into();
    desc.RenderTarget[0].BlendOp = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
    desc.RenderTarget[0].SrcBlend = D3D11_BLEND_ONE;
    desc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_ONE;
    desc.RenderTarget[0].DestBlend = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
    desc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALL.0 as u8;
    unsafe {
        let mut state = None;
        device.CreateBlendState(&desc, Some(&mut state))?;
        Ok(state.unwrap())
    }
}

#[inline]
fn create_vertex_shader(device: &ID3D11Device, bytes: &[u8]) -> Result<ID3D11VertexShader> {
    unsafe {
        let mut shader = None;
        device.CreateVertexShader(bytes, None, Some(&mut shader))?;
        Ok(shader.unwrap())
    }
}

#[inline]
fn create_fragment_shader(device: &ID3D11Device, bytes: &[u8]) -> Result<ID3D11PixelShader> {
    unsafe {
        let mut shader = None;
        device.CreatePixelShader(bytes, None, Some(&mut shader))?;
        Ok(shader.unwrap())
    }
}

#[inline]
fn create_constant_buffer<T>(device: &ID3D11Device) -> Result<Option<ID3D11Buffer>> {
    const { assert!(std::mem::size_of::<T>() != 0 && std::mem::size_of::<T>().is_multiple_of(16)) };
    let desc = D3D11_BUFFER_DESC {
        ByteWidth: std::mem::size_of::<T>() as u32,
        Usage: D3D11_USAGE_DYNAMIC,
        BindFlags: D3D11_BIND_CONSTANT_BUFFER.0 as u32,
        CPUAccessFlags: D3D11_CPU_ACCESS_WRITE.0 as u32,
        MiscFlags: 0,
        StructureByteStride: 0,
    };
    let mut buffer = None;
    unsafe { device.CreateBuffer(&desc, None, Some(&mut buffer)) }?;
    Ok(buffer)
}

#[inline]
fn create_buffer(
    device: &ID3D11Device,
    element_size: usize,
    buffer_size: usize,
) -> Result<ID3D11Buffer> {
    let desc = D3D11_BUFFER_DESC {
        ByteWidth: (element_size * buffer_size) as u32,
        Usage: D3D11_USAGE_DYNAMIC,
        BindFlags: D3D11_BIND_SHADER_RESOURCE.0 as u32,
        CPUAccessFlags: D3D11_CPU_ACCESS_WRITE.0 as u32,
        MiscFlags: D3D11_RESOURCE_MISC_BUFFER_STRUCTURED.0 as u32,
        StructureByteStride: element_size as u32,
    };
    let mut buffer = None;
    unsafe { device.CreateBuffer(&desc, None, Some(&mut buffer)) }?;
    Ok(buffer.unwrap())
}

#[inline]
fn create_buffer_view(
    device: &ID3D11Device,
    buffer: &ID3D11Buffer,
) -> Result<Option<ID3D11ShaderResourceView>> {
    let mut view = None;
    unsafe { device.CreateShaderResourceView(buffer, None, Some(&mut view)) }?;
    Ok(view)
}

#[inline]
fn update_buffer<T>(
    device_context: &ID3D11DeviceContext,
    buffer: &ID3D11Buffer,
    data: &[T],
) -> Result<()> {
    unsafe {
        let mut dest = std::mem::zeroed();
        device_context.Map(buffer, 0, D3D11_MAP_WRITE_DISCARD, 0, Some(&mut dest))?;
        std::ptr::copy_nonoverlapping(data.as_ptr(), dest.pData as _, data.len());
        device_context.Unmap(buffer, 0);
    }
    Ok(())
}

#[inline]
fn update_batch_start(
    device_context: &ID3D11DeviceContext,
    buffer: &ID3D11Buffer,
    first_instance: u32,
) -> Result<()> {
    update_buffer(
        device_context,
        buffer,
        &[BatchParams {
            start_index: first_instance,
            _padding: [0; 3],
        }],
    )
}

#[inline]
fn set_pipeline_state(
    device_context: &ID3D11DeviceContext,
    buffer_view: &[Option<ID3D11ShaderResourceView>],
    topology: D3D_PRIMITIVE_TOPOLOGY,
    vertex_shader: &ID3D11VertexShader,
    fragment_shader: &ID3D11PixelShader,
    blend_state: &ID3D11BlendState,
) {
    unsafe {
        device_context.VSSetShaderResources(1, Some(buffer_view));
        device_context.PSSetShaderResources(1, Some(buffer_view));
        device_context.IASetPrimitiveTopology(topology);
        device_context.VSSetShader(vertex_shader, None);
        device_context.PSSetShader(fragment_shader, None);
        device_context.OMSetBlendState(blend_state, None, 0xFFFFFFFF);
    }
}

#[cfg(debug_assertions)]
fn report_live_objects(device: &ID3D11Device) -> Result<()> {
    let debug_device: ID3D11Debug = device.cast()?;
    unsafe {
        debug_device.ReportLiveDeviceObjects(D3D11_RLDO_DETAIL)?;
    }
    Ok(())
}

const BUFFER_COUNT: usize = 3;

pub(crate) mod shader_resources {
    use anyhow::Result;

    #[cfg(debug_assertions)]
    use windows::{
        Win32::Graphics::Direct3D::{
            Fxc::{D3DCOMPILE_DEBUG, D3DCOMPILE_SKIP_OPTIMIZATION, D3DCompileFromFile},
            ID3DBlob,
        },
        core::{HSTRING, PCSTR},
    };

    #[derive(Copy, Clone, Debug, Eq, PartialEq)]
    pub(crate) enum ShaderModule {
        Quad,
        Shadow,
        Underline,
        PathRasterization,
        PathSprite,
        MonochromeSprite,
        SubpixelSprite,
        PolychromeSprite,
        EmojiRasterization,
    }

    #[derive(Copy, Clone, Debug, Eq, PartialEq)]
    pub(crate) enum ShaderTarget {
        Vertex,
        Fragment,
    }

    pub(crate) struct RawShaderBytes<'t> {
        inner: &'t [u8],

        #[cfg(debug_assertions)]
        _blob: ID3DBlob,
    }

    impl<'t> RawShaderBytes<'t> {
        pub(crate) fn new(module: ShaderModule, target: ShaderTarget) -> Result<Self> {
            #[cfg(not(debug_assertions))]
            {
                Ok(Self::from_bytes(module, target))
            }
            #[cfg(debug_assertions)]
            {
                let blob = build_shader_blob(module, target)?;
                let inner = unsafe {
                    std::slice::from_raw_parts(
                        blob.GetBufferPointer() as *const u8,
                        blob.GetBufferSize(),
                    )
                };
                Ok(Self { inner, _blob: blob })
            }
        }

        pub(crate) fn as_bytes(&'t self) -> &'t [u8] {
            self.inner
        }

        #[cfg(not(debug_assertions))]
        fn from_bytes(module: ShaderModule, target: ShaderTarget) -> Self {
            let bytes = match module {
                ShaderModule::Quad => match target {
                    ShaderTarget::Vertex => QUAD_VERTEX_BYTES,
                    ShaderTarget::Fragment => QUAD_FRAGMENT_BYTES,
                },
                ShaderModule::Shadow => match target {
                    ShaderTarget::Vertex => SHADOW_VERTEX_BYTES,
                    ShaderTarget::Fragment => SHADOW_FRAGMENT_BYTES,
                },
                ShaderModule::Underline => match target {
                    ShaderTarget::Vertex => UNDERLINE_VERTEX_BYTES,
                    ShaderTarget::Fragment => UNDERLINE_FRAGMENT_BYTES,
                },
                ShaderModule::PathRasterization => match target {
                    ShaderTarget::Vertex => PATH_RASTERIZATION_VERTEX_BYTES,
                    ShaderTarget::Fragment => PATH_RASTERIZATION_FRAGMENT_BYTES,
                },
                ShaderModule::PathSprite => match target {
                    ShaderTarget::Vertex => PATH_SPRITE_VERTEX_BYTES,
                    ShaderTarget::Fragment => PATH_SPRITE_FRAGMENT_BYTES,
                },
                ShaderModule::MonochromeSprite => match target {
                    ShaderTarget::Vertex => MONOCHROME_SPRITE_VERTEX_BYTES,
                    ShaderTarget::Fragment => MONOCHROME_SPRITE_FRAGMENT_BYTES,
                },
                ShaderModule::SubpixelSprite => match target {
                    ShaderTarget::Vertex => SUBPIXEL_SPRITE_VERTEX_BYTES,
                    ShaderTarget::Fragment => SUBPIXEL_SPRITE_FRAGMENT_BYTES,
                },
                ShaderModule::PolychromeSprite => match target {
                    ShaderTarget::Vertex => POLYCHROME_SPRITE_VERTEX_BYTES,
                    ShaderTarget::Fragment => POLYCHROME_SPRITE_FRAGMENT_BYTES,
                },
                ShaderModule::EmojiRasterization => match target {
                    ShaderTarget::Vertex => EMOJI_RASTERIZATION_VERTEX_BYTES,
                    ShaderTarget::Fragment => EMOJI_RASTERIZATION_FRAGMENT_BYTES,
                },
            };
            Self { inner: bytes }
        }
    }

    #[cfg(debug_assertions)]
    pub(super) fn build_shader_blob(entry: ShaderModule, target: ShaderTarget) -> Result<ID3DBlob> {
        unsafe {
            use windows::Win32::Graphics::{
                Direct3D::ID3DInclude, Hlsl::D3D_COMPILE_STANDARD_FILE_INCLUDE,
            };

            let shader_name = if matches!(entry, ShaderModule::EmojiRasterization) {
                "color_text_raster.hlsl"
            } else {
                "shaders.hlsl"
            };

            let entry = format!(
                "{}_{}\0",
                entry.as_str(),
                match target {
                    ShaderTarget::Vertex => "vertex",
                    ShaderTarget::Fragment => "fragment",
                }
            );
            let target = match target {
                ShaderTarget::Vertex => "vs_4_1\0",
                ShaderTarget::Fragment => "ps_4_1\0",
            };

            let mut compile_blob = None;
            let mut error_blob = None;
            let shader_path = std::path::PathBuf::from(env!("CARGO_MANIFEST_DIR"))
                .join(&format!("src/{}", shader_name))
                .canonicalize()?;

            let entry_point = PCSTR::from_raw(entry.as_ptr());
            let target_cstr = PCSTR::from_raw(target.as_ptr());

            // really dirty trick because winapi bindings are unhappy otherwise
            let include_handler = &std::mem::transmute::<usize, ID3DInclude>(
                D3D_COMPILE_STANDARD_FILE_INCLUDE as usize,
            );

            let ret = D3DCompileFromFile(
                &HSTRING::from(shader_path.to_str().unwrap()),
                None,
                include_handler,
                entry_point,
                target_cstr,
                D3DCOMPILE_DEBUG | D3DCOMPILE_SKIP_OPTIMIZATION,
                0,
                &mut compile_blob,
                Some(&mut error_blob),
            );
            if ret.is_err() {
                let Some(error_blob) = error_blob else {
                    return Err(anyhow::anyhow!("{ret:?}"));
                };

                let error_string =
                    std::ffi::CStr::from_ptr(error_blob.GetBufferPointer() as *const i8)
                        .to_string_lossy();
                log::error!("Shader compile error: {}", error_string);
                return Err(anyhow::anyhow!("Compile error: {}", error_string));
            }
            Ok(compile_blob.unwrap())
        }
    }

    #[cfg(not(debug_assertions))]
    include!(concat!(env!("OUT_DIR"), "/shaders_bytes.rs"));

    #[cfg(debug_assertions)]
    impl ShaderModule {
        pub fn as_str(self) -> &'static str {
            match self {
                ShaderModule::Quad => "quad",
                ShaderModule::Shadow => "shadow",
                ShaderModule::Underline => "underline",
                ShaderModule::PathRasterization => "path_rasterization",
                ShaderModule::PathSprite => "path_sprite",
                ShaderModule::MonochromeSprite => "monochrome_sprite",
                ShaderModule::SubpixelSprite => "subpixel_sprite",
                ShaderModule::PolychromeSprite => "polychrome_sprite",
                ShaderModule::EmojiRasterization => "emoji_rasterization",
            }
        }
    }
}

mod nvidia {
    use std::{
        ffi::CStr,
        os::raw::{c_char, c_int, c_uint},
    };

    use anyhow::Result;
    use windows::{Win32::System::LibraryLoader::GetProcAddress, core::s};

    use crate::with_dll_library;

    // https://github.com/NVIDIA/nvapi/blob/7cb76fce2f52de818b3da497af646af1ec16ce27/nvapi_lite_common.h#L180
    const NVAPI_SHORT_STRING_MAX: usize = 64;

    // https://github.com/NVIDIA/nvapi/blob/7cb76fce2f52de818b3da497af646af1ec16ce27/nvapi_lite_common.h#L235
    #[allow(non_camel_case_types)]
    type NvAPI_ShortString = [c_char; NVAPI_SHORT_STRING_MAX];

    // https://github.com/NVIDIA/nvapi/blob/7cb76fce2f52de818b3da497af646af1ec16ce27/nvapi_lite_common.h#L447
    #[allow(non_camel_case_types)]
    type NvAPI_SYS_GetDriverAndBranchVersion_t = unsafe extern "C" fn(
        driver_version: *mut c_uint,
        build_branch_string: *mut NvAPI_ShortString,
    ) -> c_int;

    pub(super) fn get_driver_version() -> Result<String> {
        #[cfg(target_pointer_width = "64")]
        let nvidia_dll_name = s!("nvapi64.dll");
        #[cfg(target_pointer_width = "32")]
        let nvidia_dll_name = s!("nvapi.dll");

        with_dll_library(nvidia_dll_name, |nvidia_dll| unsafe {
            let nvapi_query_addr = GetProcAddress(nvidia_dll, s!("nvapi_QueryInterface"))
                .ok_or_else(|| anyhow::anyhow!("Failed to get nvapi_QueryInterface address"))?;
            let nvapi_query: extern "C" fn(u32) -> *mut () = std::mem::transmute(nvapi_query_addr);

            // https://github.com/NVIDIA/nvapi/blob/7cb76fce2f52de818b3da497af646af1ec16ce27/nvapi_interface.h#L41
            let nvapi_get_driver_version_ptr = nvapi_query(0x2926aaad);
            if nvapi_get_driver_version_ptr.is_null() {
                anyhow::bail!("Failed to get NVIDIA driver version function pointer");
            }
            let nvapi_get_driver_version: NvAPI_SYS_GetDriverAndBranchVersion_t =
                std::mem::transmute(nvapi_get_driver_version_ptr);

            let mut driver_version: c_uint = 0;
            let mut build_branch_string: NvAPI_ShortString = [0; NVAPI_SHORT_STRING_MAX];
            let result = nvapi_get_driver_version(
                &mut driver_version as *mut c_uint,
                &mut build_branch_string as *mut NvAPI_ShortString,
            );

            if result != 0 {
                anyhow::bail!(
                    "Failed to get NVIDIA driver version, error code: {}",
                    result
                );
            }
            let major = driver_version / 100;
            let minor = driver_version % 100;
            let branch_string = CStr::from_ptr(build_branch_string.as_ptr());
            Ok(format!(
                "{}.{} {}",
                major,
                minor,
                branch_string.to_string_lossy()
            ))
        })
    }
}

mod amd {
    use std::os::raw::{c_char, c_int, c_void};

    use anyhow::Result;
    use windows::{Win32::System::LibraryLoader::GetProcAddress, core::s};

    use crate::with_dll_library;

    // https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK/blob/5d8812d703d0335741b6f7ffc37838eeb8b967f7/ags_lib/inc/amd_ags.h#L145
    const AGS_CURRENT_VERSION: i32 = (6 << 22) | (3 << 12);

    // https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK/blob/5d8812d703d0335741b6f7ffc37838eeb8b967f7/ags_lib/inc/amd_ags.h#L204
    // This is an opaque type, using struct to represent it properly for FFI
    #[repr(C)]
    struct AGSContext {
        _private: [u8; 0],
    }

    #[repr(C)]
    pub struct AGSGPUInfo {
        pub driver_version: *const c_char,
        pub radeon_software_version: *const c_char,
        pub num_devices: c_int,
        pub devices: *mut c_void,
    }

    // https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK/blob/5d8812d703d0335741b6f7ffc37838eeb8b967f7/ags_lib/inc/amd_ags.h#L429
    #[allow(non_camel_case_types)]
    type agsInitialize_t = unsafe extern "C" fn(
        version: c_int,
        config: *const c_void,
        context: *mut *mut AGSContext,
        gpu_info: *mut AGSGPUInfo,
    ) -> c_int;

    // https://github.com/GPUOpen-LibrariesAndSDKs/AGS_SDK/blob/5d8812d703d0335741b6f7ffc37838eeb8b967f7/ags_lib/inc/amd_ags.h#L436
    #[allow(non_camel_case_types)]
    type agsDeInitialize_t = unsafe extern "C" fn(context: *mut AGSContext) -> c_int;

    pub(super) fn get_driver_version() -> Result<String> {
        #[cfg(target_pointer_width = "64")]
        let amd_dll_name = s!("amd_ags_x64.dll");
        #[cfg(target_pointer_width = "32")]
        let amd_dll_name = s!("amd_ags_x86.dll");

        with_dll_library(amd_dll_name, |amd_dll| unsafe {
            let ags_initialize_addr = GetProcAddress(amd_dll, s!("agsInitialize"))
                .ok_or_else(|| anyhow::anyhow!("Failed to get agsInitialize address"))?;
            let ags_deinitialize_addr = GetProcAddress(amd_dll, s!("agsDeInitialize"))
                .ok_or_else(|| anyhow::anyhow!("Failed to get agsDeInitialize address"))?;

            let ags_initialize: agsInitialize_t = std::mem::transmute(ags_initialize_addr);
            let ags_deinitialize: agsDeInitialize_t = std::mem::transmute(ags_deinitialize_addr);

            let mut context: *mut AGSContext = std::ptr::null_mut();
            let mut gpu_info: AGSGPUInfo = AGSGPUInfo {
                driver_version: std::ptr::null(),
                radeon_software_version: std::ptr::null(),
                num_devices: 0,
                devices: std::ptr::null_mut(),
            };

            let result = ags_initialize(
                AGS_CURRENT_VERSION,
                std::ptr::null(),
                &mut context,
                &mut gpu_info,
            );
            if result != 0 {
                anyhow::bail!("Failed to initialize AMD AGS, error code: {}", result);
            }

            // Vulkan actually returns this as the driver version
            let software_version = if !gpu_info.radeon_software_version.is_null() {
                std::ffi::CStr::from_ptr(gpu_info.radeon_software_version)
                    .to_string_lossy()
                    .into_owned()
            } else {
                "Unknown Radeon Software Version".to_string()
            };

            let driver_version = if !gpu_info.driver_version.is_null() {
                std::ffi::CStr::from_ptr(gpu_info.driver_version)
                    .to_string_lossy()
                    .into_owned()
            } else {
                "Unknown Radeon Driver Version".to_string()
            };

            ags_deinitialize(context);
            Ok(format!("{} ({})", software_version, driver_version))
        })
    }
}

mod dxgi {
    use windows::{
        Win32::Graphics::Dxgi::{IDXGIAdapter1, IDXGIDevice},
        core::Interface,
    };

    pub(super) fn get_driver_version(adapter: &IDXGIAdapter1) -> anyhow::Result<String> {
        let number = unsafe { adapter.CheckInterfaceSupport(&IDXGIDevice::IID as _) }?;
        Ok(format!(
            "{}.{}.{}.{}",
            number >> 48,
            (number >> 32) & 0xFFFF,
            (number >> 16) & 0xFFFF,
            number & 0xFFFF
        ))
    }
}

/// Native D3D11 offscreen renderer using WARP, for deterministic test/perf apps.
#[cfg(any(test, feature = "test-support"))]
pub struct Dx11HeadlessRenderer(DirectXRenderer);
#[cfg(any(test, feature = "test-support"))]
impl Dx11HeadlessRenderer {
    /// Bytes in retained subtree textures, excluding driver/in-flight references.
    pub fn layer_texture_bytes(&self) -> u64 {
        self.0
            .backdrops
            .as_ref()
            .map(|effects| effects.groups.texture_bytes())
            .unwrap_or(0)
    }

    /// Number of retained effect textures; excludes driver and in-flight references.
    pub fn effect_texture_count(&self) -> usize {
        self.0
            .backdrops
            .as_ref()
            .map(|effects| effects.texture_count())
            .unwrap_or(0)
    }

    /// Cumulative backdrop filter cache statistics.
    pub fn backdrop_cache_stats(&self) -> gpui::BackdropCacheStats {
        self.0
            .backdrops
            .as_ref()
            .map(|e| e.cache.stats)
            .unwrap_or_default()
    }

    /// Cumulative captured-subtree cache statistics.
    pub fn layer_cache_stats(&self) -> gpui::LayerCacheStats {
        self.0
            .backdrops
            .as_ref()
            .map(|e| e.groups.stats)
            .unwrap_or_default()
    }

    /// Configure native wide-SDR composition for offscreen validation.
    pub fn set_color_mode(&mut self, mode: WindowColorMode) -> Result<gpui::WindowColorStatus> {
        self.0
            .set_color_mode(mode, WindowBackgroundAppearance::Opaque)
    }
    pub fn new() -> Result<Self> {
        let (mut device, mut immediate) = (None, None);
        unsafe {
            D3D11CreateDevice(
                None,
                D3D_DRIVER_TYPE_WARP,
                windows::Win32::Foundation::HMODULE::default(),
                D3D11_CREATE_DEVICE_BGRA_SUPPORT,
                Some(&[D3D_FEATURE_LEVEL_11_0]),
                D3D11_SDK_VERSION,
                Some(&mut device),
                None,
                Some(&mut immediate),
            )?;
        }
        let device = device.context("missing WARP device")?;
        let device_context = immediate.context("missing WARP context")?;
        let dxgi: IDXGIDevice = device.cast()?;
        let adapter: IDXGIAdapter1 = unsafe { dxgi.GetAdapter()? }.cast()?;
        let dxgi_factory: IDXGIFactory6 = unsafe { adapter.GetParent()? };
        let devices = DirectXRendererDevices {
            adapter,
            dxgi_factory,
            device,
            device_context,
            dxgi_device: None,
            annotation: None,
        };
        let atlas = Arc::new(DirectXAtlas::new(&devices.device, &devices.device_context));
        set_rasterizer_state(&devices.device, &devices.device_context)?;
        let resources = DirectXResources::offscreen(&devices, 1, 1)?;
        let globals = DirectXGlobalElements::new(&devices.device)?;
        let pipelines = DirectXRenderPipelines::new(&devices.device)?;
        Ok(Self(DirectXRenderer {
            hwnd: HWND::default(),
            atlas,
            devices: Some(devices),
            resources: Some(resources),
            globals,
            pipelines,
            custom_compositor: None,
            custom_retirement: Vec::new(),
            backdrops: None,
            shaders: None,
            wide_sdr: None,
            color_mode: WindowColorMode::Default,
            color_display: Default::default(),
            direct_composition: None,
            font_info: DirectXRenderer::get_font_info(),
            width: 1,
            height: 1,
            skip_draws: false,
        }))
    }
}
#[cfg(any(test, feature = "test-support"))]
impl PlatformHeadlessRenderer for Dx11HeadlessRenderer {
    fn render_scene_to_image(
        &mut self,
        scene: &Scene,
        size: Size<DevicePixels>,
    ) -> Result<image::RgbaImage> {
        self.0.resize(size)?;
        self.0
            .render_to_image(scene, WindowBackgroundAppearance::Transparent)
    }
    fn render_scene(&mut self, scene: &Scene, size: Size<DevicePixels>) -> Result<()> {
        self.0.resize(size)?;
        self.0
            .render(scene, WindowBackgroundAppearance::Transparent)
    }
    fn sprite_atlas(&self) -> Arc<dyn PlatformAtlas> {
        self.0.sprite_atlas()
    }
}