gpui_wgpu/

wgpu_renderer.rs

use crate::{CompositorGpuHint, WgpuAtlas, WgpuContext};
use anyhow::{Context as _, Result};
use bytemuck::{Pod, Zeroable};
use gpui_engine::{
    AtlasTextureId, Path, PlatformAtlas, PrimitiveBatch, Scene, SceneRenderer,
    get_gamma_correction_ratios,
};
use gpui_platform::{Background, Bounds, DevicePixels, GpuSpecs, Point, ScaledPixels, Size};
use log::warn;
#[cfg(not(target_family = "wasm"))]
use raw_window_handle::{HasDisplayHandle, HasWindowHandle};
use std::cell::RefCell;
use std::num::NonZeroU64;
use std::ops::Range;
use std::rc::Rc;
use std::sync::{Arc, Mutex};

const MAX_INSTANCE_BUFFER_SIZE: u64 = 256 * 1024 * 1024;

const INSTANCE_TEXTURE_TEXEL_SIZE: u64 = 16;

/// Shader variant for backends with storage buffer support: the shared shader
/// logic plus the storage-buffer instance transport.
const STORAGE_BUFFER_SHADERS: &str = concat!(
    include_str!("shaders.wgsl"),
    include_str!("shaders_storage.wgsl"),
);

/// Shader variant for WebGL2, which has no storage buffers: the shared shader
/// logic plus the texture-based instance transport.
const WEBGL_SHADERS: &str = concat!(
    include_str!("shaders.wgsl"),
    include_str!("shaders_webgl.wgsl"),
);

/// Subpixel text rendering requires dual-source blending, which WebGL2 lacks, so
/// this variant only ever runs with the storage-buffer transport. The `enable`
/// directive must precede all declarations.
const SUBPIXEL_SHADERS: &str = concat!(
    "enable dual_source_blending;\n",
    include_str!("shaders.wgsl"),
    include_str!("shaders_storage.wgsl"),
    include_str!("shaders_subpixel.wgsl"),
);

fn least_common_multiple(left: u64, right: u64) -> u64 {
    let mut first = left;
    let mut second = right;
    while second != 0 {
        let remainder = first % second;
        first = second;
        second = remainder;
    }
    left / first * right
}

#[repr(C)]
#[derive(Clone, Copy, Pod, Zeroable)]
struct GlobalParams {
    viewport_size: [f32; 2],
    premultiplied_alpha: u32,
    pad: u32,
}

#[repr(C)]
#[derive(Clone, Copy, Pod, Zeroable)]
struct PodBounds {
    origin: [f32; 2],
    size: [f32; 2],
}

impl From<Bounds<ScaledPixels>> for PodBounds {
    fn from(bounds: Bounds<ScaledPixels>) -> Self {
        Self {
            origin: [bounds.origin.x.0, bounds.origin.y.0],
            size: [bounds.size.width.0, bounds.size.height.0],
        }
    }
}

#[repr(C)]
#[derive(Clone, Copy, Pod, Zeroable)]
struct SurfaceParams {
    bounds: PodBounds,
    content_mask: PodBounds,
}

#[repr(C)]
#[derive(Clone, Copy, Pod, Zeroable)]
struct GammaParams {
    gamma_ratios: [f32; 4],
    grayscale_enhanced_contrast: f32,
    subpixel_enhanced_contrast: f32,
    is_bgr: u32,
    _pad: u32,
}

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

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

pub struct WgpuSurfaceConfig {
    pub size: Size<DevicePixels>,
    pub transparent: bool,
    /// Preferred presentation mode. When `Some`, the renderer will use this
    /// mode if supported by the surface, falling back to `Fifo`.
    /// When `None`, defaults to `Fifo` (VSync).
    ///
    /// Mobile platforms may prefer `Mailbox` (triple-buffering) to avoid
    /// blocking in `get_current_texture()` during lifecycle transitions.
    pub preferred_present_mode: Option<wgpu::PresentMode>,
}

struct WgpuPipelines {
    quads: wgpu::RenderPipeline,
    shadows: wgpu::RenderPipeline,
    path_rasterization: wgpu::RenderPipeline,
    paths: wgpu::RenderPipeline,
    underlines: wgpu::RenderPipeline,
    mono_sprites: wgpu::RenderPipeline,
    subpixel_sprites: Option<wgpu::RenderPipeline>,
    poly_sprites: wgpu::RenderPipeline,
    #[allow(dead_code)]
    surfaces: wgpu::RenderPipeline,
}

/// One frame allocation of instance data, ready to bind.
struct InstanceBinding {
    bind_group: wgpu::BindGroup,
    /// Index of the allocation's first instance within the bound data. Always
    /// zero on the storage-buffer path, where the binding offset already
    /// positions the array; on the WebGL texture path the shader indexes the
    /// shared instance texture absolutely, so draws must offset their
    /// instance (or vertex) ranges by this value.
    first_instance: u32,
}

struct InstanceBindings {
    quads: InstanceBinding,
    shadows: InstanceBinding,
    underlines: InstanceBinding,
    monochrome_sprites: InstanceBinding,
    subpixel_sprites: InstanceBinding,
    polychrome_sprites: InstanceBinding,
}

struct WgpuBindGroupLayouts {
    globals: wgpu::BindGroupLayout,
    instances: wgpu::BindGroupLayout,
    texture: wgpu::BindGroupLayout,
    surfaces: wgpu::BindGroupLayout,
}

/// Shared GPU context reference, used to coordinate device recovery across multiple windows.
pub type GpuContext = Rc<RefCell<Option<WgpuContext>>>;

enum InstanceData {
    Storage(wgpu::Buffer),
    // WebGL2 has no storage buffers. A uint texture keeps the records available to both shader
    // stages while preserving integer and floating-point bit patterns exactly.
    Texture {
        texture: wgpu::Texture,
        view: wgpu::TextureView,
        width: u32,
        height: u32,
    },
}

/// GPU resources that must be dropped together during device recovery.
struct WgpuResources {
    device: Arc<wgpu::Device>,
    queue: Arc<wgpu::Queue>,
    surface: wgpu::Surface<'static>,
    pipelines: WgpuPipelines,
    bind_group_layouts: WgpuBindGroupLayouts,
    atlas_sampler: wgpu::Sampler,
    globals_buffer: wgpu::Buffer,
    globals_bind_group: wgpu::BindGroup,
    path_globals_bind_group: wgpu::BindGroup,
    instance_data: InstanceData,
    path_intermediate_texture: Option<wgpu::Texture>,
    path_intermediate_view: Option<wgpu::TextureView>,
    path_msaa_texture: Option<wgpu::Texture>,
    path_msaa_view: Option<wgpu::TextureView>,
}

impl WgpuResources {
    fn invalidate_intermediate_textures(&mut self) {
        self.path_intermediate_texture = None;
        self.path_intermediate_view = None;
        self.path_msaa_texture = None;
        self.path_msaa_view = None;
    }
}

pub struct WgpuRenderer {
    /// Shared GPU context for device recovery coordination (unused on WASM).
    #[allow(dead_code)]
    context: Option<GpuContext>,
    /// Compositor GPU hint for adapter selection (unused on WASM).
    #[allow(dead_code)]
    compositor_gpu: Option<CompositorGpuHint>,
    resources: Option<WgpuResources>,
    surface_config: wgpu::SurfaceConfiguration,
    atlas: Arc<WgpuAtlas>,
    path_globals_offset: u64,
    gamma_offset: u64,
    instance_data_capacity: u64,
    max_instance_data_size: u64,
    instance_data_alignment: u64,
    uses_webgl_instance_data: bool,
    rendering_params: RenderingParameters,
    is_bgr: bool,
    dual_source_blending: bool,
    adapter_info: wgpu::AdapterInfo,
    transparent_alpha_mode: wgpu::CompositeAlphaMode,
    opaque_alpha_mode: wgpu::CompositeAlphaMode,
    max_texture_size: u32,
    last_error: Arc<Mutex<Option<String>>>,
    failed_frame_count: u32,
    device_lost: std::sync::Arc<std::sync::atomic::AtomicBool>,
    surface_configured: bool,
    needs_redraw: bool,
}

impl WgpuRenderer {
    fn resources(&self) -> &WgpuResources {
        self.resources
            .as_ref()
            .expect("GPU resources not available")
    }

    fn resources_mut(&mut self) -> &mut WgpuResources {
        self.resources
            .as_mut()
            .expect("GPU resources not available")
    }

    /// Creates a new WgpuRenderer from raw window handles.
    ///
    /// The `gpu_context` is a shared reference that coordinates GPU context across
    /// multiple windows. The first window to create a renderer will initialize the
    /// context; subsequent windows will share it.
    ///
    /// # Safety
    /// The caller must ensure that the window handle remains valid for the lifetime
    /// of the returned renderer.
    #[cfg(not(target_family = "wasm"))]
    pub fn new<W>(
        gpu_context: GpuContext,
        window: &W,
        config: WgpuSurfaceConfig,
        compositor_gpu: Option<CompositorGpuHint>,
    ) -> anyhow::Result<Self>
    where
        W: HasWindowHandle + HasDisplayHandle + std::fmt::Debug + Send + Sync + Clone + 'static,
    {
        let window_handle = window
            .window_handle()
            .map_err(|e| anyhow::anyhow!("Failed to get window handle: {e}"))?;

        let target = wgpu::SurfaceTargetUnsafe::RawHandle {
            // Fall back to the display handle already provided via InstanceDescriptor::display.
            raw_display_handle: None,
            raw_window_handle: window_handle.as_raw(),
        };

        // Use the existing context's instance if available, otherwise create a new one.
        // The surface must be created with the same instance that will be used for
        // adapter selection, otherwise wgpu will panic.
        let instance = gpu_context
            .borrow()
            .as_ref()
            .map(|ctx| ctx.instance.clone())
            .unwrap_or_else(|| WgpuContext::instance(Box::new(window.clone())));

        // Safety: The caller guarantees that the window handle is valid for the
        // lifetime of this renderer. In practice, the RawWindow struct is created
        // from the native window handles and the surface is dropped before the window.
        let surface = unsafe {
            instance
                .create_surface_unsafe(target)
                .map_err(|e| anyhow::anyhow!("Failed to create surface: {e}"))?
        };

        let mut ctx_ref = gpu_context.borrow_mut();
        let context = match ctx_ref.as_mut() {
            Some(context) => {
                context.check_compatible_with_surface(&surface)?;
                context
            }
            None => ctx_ref.insert(WgpuContext::new(instance, &surface, compositor_gpu)?),
        };

        let atlas = Arc::new(WgpuAtlas::from_context(context));

        Self::new_internal(
            Some(Rc::clone(&gpu_context)),
            context,
            surface,
            config,
            compositor_gpu,
            atlas,
        )
    }

    #[cfg(target_family = "wasm")]
    pub fn new_from_canvas(
        context: &WgpuContext,
        canvas: &web_sys::HtmlCanvasElement,
        config: WgpuSurfaceConfig,
    ) -> anyhow::Result<Self> {
        let surface = context
            .instance
            .create_surface(wgpu::SurfaceTarget::Canvas(canvas.clone()))
            .map_err(|e| anyhow::anyhow!("Failed to create surface: {e}"))?;
        Self::new_from_surface(context, surface, config)
    }

    #[cfg(target_family = "wasm")]
    #[allow(clippy::arc_with_non_send_sync)]
    pub fn new_from_surface(
        context: &WgpuContext,
        surface: wgpu::Surface<'static>,
        config: WgpuSurfaceConfig,
    ) -> anyhow::Result<Self> {
        let atlas = Arc::new(WgpuAtlas::from_context(context));
        Self::new_internal(None, context, surface, config, None, atlas)
    }

    fn new_internal(
        gpu_context: Option<GpuContext>,
        context: &WgpuContext,
        surface: wgpu::Surface<'static>,
        config: WgpuSurfaceConfig,
        compositor_gpu: Option<CompositorGpuHint>,
        atlas: Arc<WgpuAtlas>,
    ) -> anyhow::Result<Self> {
        let surface_caps = surface.get_capabilities(&context.adapter);
        let preferred_formats = [
            wgpu::TextureFormat::Bgra8Unorm,
            wgpu::TextureFormat::Rgba8Unorm,
        ];
        let surface_format = preferred_formats
            .iter()
            .find(|f| surface_caps.formats.contains(f))
            .copied()
            .or_else(|| surface_caps.formats.iter().find(|f| !f.is_srgb()).copied())
            .or_else(|| surface_caps.formats.first().copied())
            .ok_or_else(|| {
                anyhow::anyhow!(
                    "Surface reports no supported texture formats for adapter {:?}",
                    context.adapter.get_info().name
                )
            })?;

        let pick_alpha_mode =
            |preferences: &[wgpu::CompositeAlphaMode]| -> anyhow::Result<wgpu::CompositeAlphaMode> {
                preferences
                    .iter()
                    .find(|p| surface_caps.alpha_modes.contains(p))
                    .copied()
                    .or_else(|| surface_caps.alpha_modes.first().copied())
                    .ok_or_else(|| {
                        anyhow::anyhow!(
                            "Surface reports no supported alpha modes for adapter {:?}",
                            context.adapter.get_info().name
                        )
                    })
            };

        let transparent_alpha_mode = pick_alpha_mode(&[
            wgpu::CompositeAlphaMode::PreMultiplied,
            wgpu::CompositeAlphaMode::Inherit,
        ])?;

        let opaque_alpha_mode = pick_alpha_mode(&[
            wgpu::CompositeAlphaMode::Opaque,
            wgpu::CompositeAlphaMode::Inherit,
        ])?;

        let alpha_mode = if config.transparent {
            transparent_alpha_mode
        } else {
            opaque_alpha_mode
        };

        let device = Arc::clone(&context.device);
        let max_texture_size = device.limits().max_texture_dimension_2d;

        let requested_width = config.size.width.0 as u32;
        let requested_height = config.size.height.0 as u32;
        let clamped_width = requested_width.min(max_texture_size);
        let clamped_height = requested_height.min(max_texture_size);

        if clamped_width != requested_width || clamped_height != requested_height {
            warn!(
                "Requested surface size ({}, {}) exceeds maximum texture dimension {}. \
                 Clamping to ({}, {}). Window content may not fill the entire window.",
                requested_width, requested_height, max_texture_size, clamped_width, clamped_height
            );
        }

        let surface_config = wgpu::SurfaceConfiguration {
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
            format: surface_format,
            width: clamped_width.max(1),
            height: clamped_height.max(1),
            present_mode: config
                .preferred_present_mode
                .filter(|mode| surface_caps.present_modes.contains(mode))
                .unwrap_or(wgpu::PresentMode::Fifo),
            desired_maximum_frame_latency: 2,
            alpha_mode,
            view_formats: vec![],
        };
        // Configure the surface immediately. The adapter selection process already validated
        // that this adapter can successfully configure this surface.
        surface.configure(&context.device, &surface_config);

        let queue = Arc::clone(&context.queue);
        let rendering_params = RenderingParameters::new(&context.adapter, surface_format);
        let uses_webgl_instance_data = context.uses_webgl_instance_data();
        let dual_source_blending =
            context.supports_dual_source_blending() && !uses_webgl_instance_data;
        let bind_group_layouts = Self::create_bind_group_layouts(&device, uses_webgl_instance_data);
        let pipelines = Self::create_pipelines(
            &device,
            &bind_group_layouts,
            surface_format,
            alpha_mode,
            rendering_params.path_sample_count,
            dual_source_blending,
            uses_webgl_instance_data,
        );

        let atlas_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
            label: Some("atlas_sampler"),
            mag_filter: wgpu::FilterMode::Linear,
            min_filter: wgpu::FilterMode::Linear,
            ..Default::default()
        });

        let uniform_alignment = device.limits().min_uniform_buffer_offset_alignment as u64;
        let globals_size = std::mem::size_of::<GlobalParams>() as u64;
        let gamma_size = std::mem::size_of::<GammaParams>() as u64;
        let path_globals_offset = globals_size.next_multiple_of(uniform_alignment);
        let gamma_offset = (path_globals_offset + globals_size).next_multiple_of(uniform_alignment);

        let globals_buffer = device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("globals_buffer"),
            size: gamma_offset + gamma_size,
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });

        let (
            instance_data,
            instance_data_capacity,
            max_instance_data_size,
            instance_data_alignment,
        ) = if uses_webgl_instance_data {
            let max_texture_dimension = device.limits().max_texture_dimension_2d;
            let max_instance_data_size = (u64::from(max_texture_dimension).pow(2)
                * INSTANCE_TEXTURE_TEXEL_SIZE)
                .min(MAX_INSTANCE_BUFFER_SIZE);
            let initial_capacity = (2 * 1024 * 1024).min(max_instance_data_size);
            let (instance_data, capacity) =
                Self::create_instance_texture(&device, initial_capacity, max_texture_dimension);
            (
                instance_data,
                capacity,
                max_instance_data_size,
                INSTANCE_TEXTURE_TEXEL_SIZE,
            )
        } else {
            // Every frame allocation is exposed as one storage-buffer binding, so
            // its backing buffer must satisfy both the allocation and binding limits.
            let max_buffer_size = device
                .limits()
                .max_buffer_size
                .min(device.limits().max_storage_buffer_binding_size)
                .min(MAX_INSTANCE_BUFFER_SIZE);
            let initial_capacity = (2 * 1024 * 1024).min(max_buffer_size);
            let buffer = device.create_buffer(&wgpu::BufferDescriptor {
                label: Some("instance_buffer"),
                size: initial_capacity,
                usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
                mapped_at_creation: false,
            });
            (
                InstanceData::Storage(buffer),
                initial_capacity,
                max_buffer_size,
                device.limits().min_storage_buffer_offset_alignment as u64,
            )
        };

        let globals_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("globals_bind_group"),
            layout: &bind_group_layouts.globals,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                        buffer: &globals_buffer,
                        offset: 0,
                        size: Some(NonZeroU64::new(globals_size).unwrap()),
                    }),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                        buffer: &globals_buffer,
                        offset: gamma_offset,
                        size: Some(NonZeroU64::new(gamma_size).unwrap()),
                    }),
                },
            ],
        });

        let path_globals_bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("path_globals_bind_group"),
            layout: &bind_group_layouts.globals,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                        buffer: &globals_buffer,
                        offset: path_globals_offset,
                        size: Some(NonZeroU64::new(globals_size).unwrap()),
                    }),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                        buffer: &globals_buffer,
                        offset: gamma_offset,
                        size: Some(NonZeroU64::new(gamma_size).unwrap()),
                    }),
                },
            ],
        });

        let adapter_info = context.adapter.get_info();

        let last_error: Arc<Mutex<Option<String>>> = Arc::new(Mutex::new(None));
        let last_error_clone = Arc::clone(&last_error);
        device.on_uncaptured_error(Arc::new(move |error| {
            let mut guard = last_error_clone.lock().unwrap();
            *guard = Some(error.to_string());
        }));

        let resources = WgpuResources {
            device,
            queue,
            surface,
            pipelines,
            bind_group_layouts,
            atlas_sampler,
            globals_buffer,
            globals_bind_group,
            path_globals_bind_group,
            instance_data,
            // Defer intermediate texture creation to first draw call via ensure_intermediate_textures().
            // This avoids panics when the device/surface is in an invalid state during initialization.
            path_intermediate_texture: None,
            path_intermediate_view: None,
            path_msaa_texture: None,
            path_msaa_view: None,
        };

        Ok(Self {
            context: gpu_context,
            compositor_gpu,
            resources: Some(resources),
            surface_config,
            atlas,
            path_globals_offset,
            gamma_offset,
            instance_data_capacity,
            max_instance_data_size,
            instance_data_alignment,
            uses_webgl_instance_data,
            rendering_params,
            is_bgr: false,
            dual_source_blending,
            adapter_info,
            transparent_alpha_mode,
            opaque_alpha_mode,
            max_texture_size,
            last_error,
            failed_frame_count: 0,
            device_lost: context.device_lost_flag(),
            surface_configured: true,
            needs_redraw: false,
        })
    }

    fn create_bind_group_layouts(
        device: &wgpu::Device,
        uses_webgl_instance_data: bool,
    ) -> WgpuBindGroupLayouts {
        let globals =
            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                label: Some("globals_layout"),
                entries: &[
                    wgpu::BindGroupLayoutEntry {
                        binding: 0,
                        visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                        ty: wgpu::BindingType::Buffer {
                            ty: wgpu::BufferBindingType::Uniform,
                            has_dynamic_offset: false,
                            min_binding_size: NonZeroU64::new(
                                std::mem::size_of::<GlobalParams>() as u64
                            ),
                        },
                        count: None,
                    },
                    wgpu::BindGroupLayoutEntry {
                        binding: 1,
                        visibility: wgpu::ShaderStages::FRAGMENT,
                        ty: wgpu::BindingType::Buffer {
                            ty: wgpu::BufferBindingType::Uniform,
                            has_dynamic_offset: false,
                            min_binding_size: NonZeroU64::new(
                                std::mem::size_of::<GammaParams>() as u64
                            ),
                        },
                        count: None,
                    },
                ],
            });

        let instance_data_entry = wgpu::BindGroupLayoutEntry {
            binding: 0,
            visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
            ty: if uses_webgl_instance_data {
                wgpu::BindingType::Texture {
                    sample_type: wgpu::TextureSampleType::Uint,
                    view_dimension: wgpu::TextureViewDimension::D2,
                    multisampled: false,
                }
            } else {
                wgpu::BindingType::Buffer {
                    ty: wgpu::BufferBindingType::Storage { read_only: true },
                    has_dynamic_offset: false,
                    min_binding_size: None,
                }
            },
            count: None,
        };

        let instances = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("instances_layout"),
            entries: &[instance_data_entry],
        });

        let texture = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("texture_layout"),
            entries: &[
                wgpu::BindGroupLayoutEntry {
                    binding: 0,
                    visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                    ty: wgpu::BindingType::Texture {
                        sample_type: wgpu::TextureSampleType::Float { filterable: true },
                        view_dimension: wgpu::TextureViewDimension::D2,
                        multisampled: false,
                    },
                    count: None,
                },
                wgpu::BindGroupLayoutEntry {
                    binding: 1,
                    visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                    ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Filtering),
                    count: None,
                },
            ],
        });

        let surfaces = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("surfaces_layout"),
            entries: &[
                wgpu::BindGroupLayoutEntry {
                    binding: 0,
                    visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                    ty: wgpu::BindingType::Buffer {
                        ty: wgpu::BufferBindingType::Uniform,
                        has_dynamic_offset: false,
                        min_binding_size: NonZeroU64::new(
                            std::mem::size_of::<SurfaceParams>() as u64
                        ),
                    },
                    count: None,
                },
                wgpu::BindGroupLayoutEntry {
                    binding: 1,
                    visibility: wgpu::ShaderStages::FRAGMENT,
                    ty: wgpu::BindingType::Texture {
                        sample_type: wgpu::TextureSampleType::Float { filterable: true },
                        view_dimension: wgpu::TextureViewDimension::D2,
                        multisampled: false,
                    },
                    count: None,
                },
                wgpu::BindGroupLayoutEntry {
                    binding: 2,
                    visibility: wgpu::ShaderStages::FRAGMENT,
                    ty: wgpu::BindingType::Texture {
                        sample_type: wgpu::TextureSampleType::Float { filterable: true },
                        view_dimension: wgpu::TextureViewDimension::D2,
                        multisampled: false,
                    },
                    count: None,
                },
                wgpu::BindGroupLayoutEntry {
                    binding: 3,
                    visibility: wgpu::ShaderStages::FRAGMENT,
                    ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Filtering),
                    count: None,
                },
            ],
        });

        WgpuBindGroupLayouts {
            globals,
            instances,
            texture,
            surfaces,
        }
    }

    fn create_instance_texture(
        device: &wgpu::Device,
        requested_capacity: u64,
        max_texture_dimension: u32,
    ) -> (InstanceData, u64) {
        let texel_count = requested_capacity.div_ceil(INSTANCE_TEXTURE_TEXEL_SIZE);
        let width = texel_count.min(u64::from(max_texture_dimension)).max(1) as u32;
        let height = texel_count
            .div_ceil(u64::from(width))
            .min(u64::from(max_texture_dimension))
            .max(1) as u32;
        let texture = device.create_texture(&wgpu::TextureDescriptor {
            label: Some("instance_texture"),
            size: wgpu::Extent3d {
                width,
                height,
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format: wgpu::TextureFormat::Rgba32Uint,
            usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
            view_formats: &[],
        });
        let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
        let capacity = u64::from(width) * u64::from(height) * INSTANCE_TEXTURE_TEXEL_SIZE;
        (
            InstanceData::Texture {
                texture,
                view,
                width,
                height,
            },
            capacity,
        )
    }

    fn create_pipelines(
        device: &wgpu::Device,
        layouts: &WgpuBindGroupLayouts,
        surface_format: wgpu::TextureFormat,
        alpha_mode: wgpu::CompositeAlphaMode,
        path_sample_count: u32,
        dual_source_blending: bool,
        uses_webgl_instance_data: bool,
    ) -> WgpuPipelines {
        // Diagnostic guard: verify the device actually has
        // DUAL_SOURCE_BLENDING. We have a crash report (ZED-5G1) where a
        // feature mismatch caused a wgpu-hal abort, but we haven't
        // identified the code path that produces the mismatch. This
        // guard prevents the crash and logs more evidence.
        // Remove this check once:
        // a) We find and fix the root cause, or
        // b) There are no reports of this warning appearing for some time.
        let device_has_feature = device
            .features()
            .contains(wgpu::Features::DUAL_SOURCE_BLENDING);
        if dual_source_blending && !device_has_feature {
            log::error!(
                "BUG: dual_source_blending flag is true but device does not \
                 have DUAL_SOURCE_BLENDING enabled (device features: {:?}). \
                 Falling back to mono text rendering. Please report this at \
                 https://github.com/zed-industries/zed/issues",
                device.features(),
            );
        }
        let dual_source_blending =
            dual_source_blending && device_has_feature && !uses_webgl_instance_data;

        let shader_source = if uses_webgl_instance_data {
            WEBGL_SHADERS
        } else {
            STORAGE_BUFFER_SHADERS
        };
        let shader_module = device.create_shader_module(wgpu::ShaderModuleDescriptor {
            label: Some("gpui_shaders"),
            source: wgpu::ShaderSource::Wgsl(shader_source.into()),
        });

        let subpixel_shader_module = if dual_source_blending {
            Some(device.create_shader_module(wgpu::ShaderModuleDescriptor {
                label: Some("gpui_subpixel_shaders"),
                source: wgpu::ShaderSource::Wgsl(SUBPIXEL_SHADERS.into()),
            }))
        } else {
            None
        };

        let blend_mode = match alpha_mode {
            wgpu::CompositeAlphaMode::PreMultiplied => {
                wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING
            }
            _ => wgpu::BlendState::ALPHA_BLENDING,
        };

        let color_target = wgpu::ColorTargetState {
            format: surface_format,
            blend: Some(blend_mode),
            write_mask: wgpu::ColorWrites::ALL,
        };

        let create_pipeline = |name: &str,
                               vs_entry: &str,
                               fs_entry: &str,
                               globals_layout: &wgpu::BindGroupLayout,
                               data_layout: &wgpu::BindGroupLayout,
                               texture_layout: Option<&wgpu::BindGroupLayout>,
                               topology: wgpu::PrimitiveTopology,
                               color_targets: &[Option<wgpu::ColorTargetState>],
                               sample_count: u32,
                               module: &wgpu::ShaderModule| {
            let mut bind_group_layouts = vec![Some(globals_layout), Some(data_layout)];
            bind_group_layouts.extend(texture_layout.map(Some));
            let pipeline_layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                label: Some(&format!("{name}_layout")),
                bind_group_layouts: &bind_group_layouts,
                immediate_size: 0,
            });

            device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
                label: Some(name),
                layout: Some(&pipeline_layout),
                vertex: wgpu::VertexState {
                    module,
                    entry_point: Some(vs_entry),
                    buffers: &[],
                    compilation_options: wgpu::PipelineCompilationOptions::default(),
                },
                fragment: Some(wgpu::FragmentState {
                    module,
                    entry_point: Some(fs_entry),
                    targets: color_targets,
                    compilation_options: wgpu::PipelineCompilationOptions::default(),
                }),
                primitive: wgpu::PrimitiveState {
                    topology,
                    strip_index_format: None,
                    front_face: wgpu::FrontFace::Ccw,
                    cull_mode: None,
                    polygon_mode: wgpu::PolygonMode::Fill,
                    unclipped_depth: false,
                    conservative: false,
                },
                depth_stencil: None,
                multisample: wgpu::MultisampleState {
                    count: sample_count,
                    mask: !0,
                    alpha_to_coverage_enabled: false,
                },
                multiview_mask: None,
                cache: None,
            })
        };

        let quads = create_pipeline(
            "quads",
            "vs_quad",
            "fs_quad",
            &layouts.globals,
            &layouts.instances,
            None,
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target.clone())],
            1,
            &shader_module,
        );

        let shadows = create_pipeline(
            "shadows",
            "vs_shadow",
            "fs_shadow",
            &layouts.globals,
            &layouts.instances,
            None,
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target.clone())],
            1,
            &shader_module,
        );

        let path_rasterization = create_pipeline(
            "path_rasterization",
            "vs_path_rasterization",
            "fs_path_rasterization",
            &layouts.globals,
            &layouts.instances,
            None,
            wgpu::PrimitiveTopology::TriangleList,
            &[Some(wgpu::ColorTargetState {
                format: surface_format,
                blend: Some(wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING),
                write_mask: wgpu::ColorWrites::ALL,
            })],
            path_sample_count,
            &shader_module,
        );

        let paths_blend = wgpu::BlendState {
            color: wgpu::BlendComponent {
                src_factor: wgpu::BlendFactor::One,
                dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
                operation: wgpu::BlendOperation::Add,
            },
            alpha: wgpu::BlendComponent {
                src_factor: wgpu::BlendFactor::One,
                dst_factor: wgpu::BlendFactor::One,
                operation: wgpu::BlendOperation::Add,
            },
        };

        let paths = create_pipeline(
            "paths",
            "vs_path",
            "fs_path",
            &layouts.globals,
            &layouts.instances,
            Some(&layouts.texture),
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(wgpu::ColorTargetState {
                format: surface_format,
                blend: Some(paths_blend),
                write_mask: wgpu::ColorWrites::ALL,
            })],
            1,
            &shader_module,
        );

        let underlines = create_pipeline(
            "underlines",
            "vs_underline",
            "fs_underline",
            &layouts.globals,
            &layouts.instances,
            None,
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target.clone())],
            1,
            &shader_module,
        );

        let mono_sprites = create_pipeline(
            "mono_sprites",
            "vs_mono_sprite",
            "fs_mono_sprite",
            &layouts.globals,
            &layouts.instances,
            Some(&layouts.texture),
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target.clone())],
            1,
            &shader_module,
        );

        let subpixel_sprites = if let Some(subpixel_module) = &subpixel_shader_module {
            let subpixel_blend = wgpu::BlendState {
                color: wgpu::BlendComponent {
                    src_factor: wgpu::BlendFactor::Src1,
                    dst_factor: wgpu::BlendFactor::OneMinusSrc1,
                    operation: wgpu::BlendOperation::Add,
                },
                alpha: wgpu::BlendComponent {
                    src_factor: wgpu::BlendFactor::One,
                    dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
                    operation: wgpu::BlendOperation::Add,
                },
            };

            Some(create_pipeline(
                "subpixel_sprites",
                "vs_subpixel_sprite",
                "fs_subpixel_sprite",
                &layouts.globals,
                &layouts.instances,
                Some(&layouts.texture),
                wgpu::PrimitiveTopology::TriangleStrip,
                &[Some(wgpu::ColorTargetState {
                    format: surface_format,
                    blend: Some(subpixel_blend),
                    write_mask: wgpu::ColorWrites::COLOR,
                })],
                1,
                subpixel_module,
            ))
        } else {
            None
        };

        let poly_sprites = create_pipeline(
            "poly_sprites",
            "vs_poly_sprite",
            "fs_poly_sprite",
            &layouts.globals,
            &layouts.instances,
            Some(&layouts.texture),
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target.clone())],
            1,
            &shader_module,
        );

        let surfaces = create_pipeline(
            "surfaces",
            "vs_surface",
            "fs_surface",
            &layouts.globals,
            &layouts.surfaces,
            None,
            wgpu::PrimitiveTopology::TriangleStrip,
            &[Some(color_target)],
            1,
            &shader_module,
        );

        WgpuPipelines {
            quads,
            shadows,
            path_rasterization,
            paths,
            underlines,
            mono_sprites,
            subpixel_sprites,
            poly_sprites,
            surfaces,
        }
    }

    fn create_path_intermediate(
        device: &wgpu::Device,
        format: wgpu::TextureFormat,
        width: u32,
        height: u32,
    ) -> (wgpu::Texture, wgpu::TextureView) {
        let texture = device.create_texture(&wgpu::TextureDescriptor {
            label: Some("path_intermediate"),
            size: wgpu::Extent3d {
                width: width.max(1),
                height: height.max(1),
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format,
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::TEXTURE_BINDING,
            view_formats: &[],
        });
        let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
        (texture, view)
    }

    fn create_msaa_if_needed(
        device: &wgpu::Device,
        format: wgpu::TextureFormat,
        width: u32,
        height: u32,
        sample_count: u32,
    ) -> Option<(wgpu::Texture, wgpu::TextureView)> {
        if sample_count <= 1 {
            return None;
        }
        let texture = device.create_texture(&wgpu::TextureDescriptor {
            label: Some("path_msaa"),
            size: wgpu::Extent3d {
                width: width.max(1),
                height: height.max(1),
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count,
            dimension: wgpu::TextureDimension::D2,
            format,
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
            view_formats: &[],
        });
        let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
        Some((texture, view))
    }

    pub fn update_drawable_size(&mut self, size: Size<DevicePixels>) {
        let width = size.width.0 as u32;
        let height = size.height.0 as u32;

        if width != self.surface_config.width || height != self.surface_config.height {
            let clamped_width = width.min(self.max_texture_size);
            let clamped_height = height.min(self.max_texture_size);

            if clamped_width != width || clamped_height != height {
                warn!(
                    "Requested surface size ({}, {}) exceeds maximum texture dimension {}. \
                     Clamping to ({}, {}). Window content may not fill the entire window.",
                    width, height, self.max_texture_size, clamped_width, clamped_height
                );
            }

            self.surface_config.width = clamped_width.max(1);
            self.surface_config.height = clamped_height.max(1);
            let surface_config = self.surface_config.clone();

            let Some(resources) = self.resources.as_mut() else {
                return;
            };

            // Wait for any in-flight GPU work to complete before destroying textures
            if let Err(e) = resources.device.poll(wgpu::PollType::Wait {
                submission_index: None,
                timeout: None,
            }) {
                warn!("Failed to poll device during resize: {e:?}");
            }

            // Destroy old textures before allocating new ones to avoid GPU memory spikes
            if let Some(ref texture) = resources.path_intermediate_texture {
                texture.destroy();
            }
            if let Some(ref texture) = resources.path_msaa_texture {
                texture.destroy();
            }

            resources
                .surface
                .configure(&resources.device, &surface_config);

            // Invalidate intermediate textures - they will be lazily recreated
            // in draw() after we confirm the surface is healthy. This avoids
            // panics when the device/surface is in an invalid state during resize.
            resources.invalidate_intermediate_textures();
        }
    }

    fn ensure_intermediate_textures(&mut self) {
        if self.resources().path_intermediate_texture.is_some() {
            return;
        }

        let format = self.surface_config.format;
        let width = self.surface_config.width;
        let height = self.surface_config.height;
        let path_sample_count = self.rendering_params.path_sample_count;
        let resources = self.resources_mut();

        let (t, v) = Self::create_path_intermediate(&resources.device, format, width, height);
        resources.path_intermediate_texture = Some(t);
        resources.path_intermediate_view = Some(v);

        let (path_msaa_texture, path_msaa_view) = Self::create_msaa_if_needed(
            &resources.device,
            format,
            width,
            height,
            path_sample_count,
        )
        .map(|(t, v)| (Some(t), Some(v)))
        .unwrap_or((None, None));
        resources.path_msaa_texture = path_msaa_texture;
        resources.path_msaa_view = path_msaa_view;
    }

    pub fn set_subpixel_layout(&mut self, is_bgr: bool) {
        self.is_bgr = is_bgr;
    }

    pub fn update_transparency(&mut self, transparent: bool) {
        let new_alpha_mode = if transparent {
            self.transparent_alpha_mode
        } else {
            self.opaque_alpha_mode
        };

        if new_alpha_mode != self.surface_config.alpha_mode {
            self.surface_config.alpha_mode = new_alpha_mode;
            let surface_config = self.surface_config.clone();
            let path_sample_count = self.rendering_params.path_sample_count;
            let dual_source_blending = self.dual_source_blending;
            let uses_webgl_instance_data = self.uses_webgl_instance_data;
            let Some(resources) = self.resources.as_mut() else {
                return;
            };
            resources
                .surface
                .configure(&resources.device, &surface_config);
            resources.pipelines = Self::create_pipelines(
                &resources.device,
                &resources.bind_group_layouts,
                surface_config.format,
                surface_config.alpha_mode,
                path_sample_count,
                dual_source_blending,
                uses_webgl_instance_data,
            );
        }
    }

    #[allow(dead_code)]
    pub fn viewport_size(&self) -> Size<DevicePixels> {
        Size {
            width: DevicePixels(self.surface_config.width as i32),
            height: DevicePixels(self.surface_config.height as i32),
        }
    }

    pub fn sprite_atlas(&self) -> &Arc<WgpuAtlas> {
        &self.atlas
    }

    pub fn supports_dual_source_blending(&self) -> bool {
        self.dual_source_blending
    }

    pub fn gpu_specs(&self) -> GpuSpecs {
        GpuSpecs {
            is_software_emulated: self.adapter_info.device_type == wgpu::DeviceType::Cpu,
            device_name: self.adapter_info.name.clone(),
            driver_name: self.adapter_info.driver.clone(),
            driver_info: self.adapter_info.driver_info.clone(),
        }
    }

    pub fn max_texture_size(&self) -> u32 {
        self.max_texture_size
    }

    pub fn draw(&mut self, scene: &Scene) -> bool {
        #[cfg(target_family = "wasm")]
        if self.device_lost() {
            if self.surface_configured {
                log::error!(
                    "Browser graphics context was lost; rendering has stopped. Reload the page to recover."
                );
                self.surface_configured = false;
            }
            return false;
        }

        // Bail out early if the surface has been unconfigured (e.g. during
        // Android background/rotation transitions).  Attempting to acquire
        // a texture from an unconfigured surface can block indefinitely on
        // some drivers (Adreno).
        if !self.surface_configured {
            return false;
        }

        let last_error = self.last_error.lock().unwrap().take();
        if let Some(error) = last_error {
            self.failed_frame_count += 1;
            log::error!(
                "GPU error during frame (failure {} of 10): {error}",
                self.failed_frame_count
            );

            // TBD. Does retrying more actually help?
            if self.failed_frame_count > 10 {
                panic!("Too many consecutive GPU errors. Last error: {error}");
            } else if self.failed_frame_count > 5 {
                if let Some(res) = self.resources.as_mut() {
                    res.invalidate_intermediate_textures();
                }
                self.atlas.clear();
                self.needs_redraw = true;
                self.failed_frame_count = 0;
                return false;
            }
        } else {
            self.failed_frame_count = 0;
        }

        self.atlas.before_frame();

        let frame = match self.resources().surface.get_current_texture() {
            wgpu::CurrentSurfaceTexture::Success(frame) => frame,
            wgpu::CurrentSurfaceTexture::Suboptimal(frame) => {
                // Textures must be destroyed before the surface can be reconfigured.
                drop(frame);
                let surface_config = self.surface_config.clone();
                let resources = self.resources_mut();
                resources
                    .surface
                    .configure(&resources.device, &surface_config);
                return false;
            }
            wgpu::CurrentSurfaceTexture::Lost | wgpu::CurrentSurfaceTexture::Outdated => {
                let surface_config = self.surface_config.clone();
                let resources = self.resources_mut();
                resources
                    .surface
                    .configure(&resources.device, &surface_config);
                return false;
            }
            wgpu::CurrentSurfaceTexture::Timeout | wgpu::CurrentSurfaceTexture::Occluded => {
                return false;
            }
            wgpu::CurrentSurfaceTexture::Validation => {
                *self.last_error.lock().unwrap() =
                    Some("Surface texture validation error".to_string());
                return false;
            }
        };

        // Now that we know the surface is healthy, ensure intermediate textures exist
        self.ensure_intermediate_textures();

        let frame_view = frame
            .texture
            .create_view(&wgpu::TextureViewDescriptor::default());

        let gamma_params = GammaParams {
            gamma_ratios: self.rendering_params.gamma_ratios,
            grayscale_enhanced_contrast: self.rendering_params.grayscale_enhanced_contrast,
            subpixel_enhanced_contrast: self.rendering_params.subpixel_enhanced_contrast,
            is_bgr: self.is_bgr as u32,
            _pad: 0,
        };

        let globals = GlobalParams {
            viewport_size: [
                self.surface_config.width as f32,
                self.surface_config.height as f32,
            ],
            premultiplied_alpha: if self.surface_config.alpha_mode
                == wgpu::CompositeAlphaMode::PreMultiplied
            {
                1
            } else {
                0
            },
            pad: 0,
        };

        let path_globals = GlobalParams {
            premultiplied_alpha: 0,
            ..globals
        };

        {
            let resources = self.resources();
            resources.queue.write_buffer(
                &resources.globals_buffer,
                0,
                bytemuck::bytes_of(&globals),
            );
            resources.queue.write_buffer(
                &resources.globals_buffer,
                self.path_globals_offset,
                bytemuck::bytes_of(&path_globals),
            );
            resources.queue.write_buffer(
                &resources.globals_buffer,
                self.gamma_offset,
                bytemuck::bytes_of(&gamma_params),
            );
        }

        if let Err(error) = self.record_frame(scene, &frame_view) {
            log::error!("{error:#}");
            self.resources().queue.submit(std::iter::empty());
            return false;
        }

        frame.present();
        true
    }

    fn record_frame(&mut self, scene: &Scene, frame_view: &wgpu::TextureView) -> Result<()> {
        let mut instance_offset = 0;
        let instance_bindings = self
            .write_instances(scene, &mut instance_offset)
            .with_context(|| {
                format!(
                    "scene too large: {} paths, {} shadows, {} quads, {} underlines, {} monochrome sprites, {} subpixel sprites, {} polychrome sprites",
                    scene.paths.len(),
                    scene.shadows.len(),
                    scene.quads.len(),
                    scene.underlines.len(),
                    scene.monochrome_sprites.len(),
                    scene.subpixel_sprites.len(),
                    scene.polychrome_sprites.len(),
                )
            })?;

        let mut encoder =
            self.resources()
                .device
                .create_command_encoder(&wgpu::CommandEncoderDescriptor {
                    label: Some("main_encoder"),
                });

        {
            let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                label: Some("main_pass"),
                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                    view: frame_view,
                    resolve_target: None,
                    ops: wgpu::Operations {
                        load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                        store: wgpu::StoreOp::Store,
                    },
                    depth_slice: None,
                })],
                depth_stencil_attachment: None,
                ..Default::default()
            });

            for batch in scene.batches() {
                match batch {
                    PrimitiveBatch::Quads(range) => self.draw_instances(
                        &instance_bindings.quads,
                        &self.resources().pipelines.quads,
                        instance_range(range),
                        &mut pass,
                    ),
                    PrimitiveBatch::Shadows(range) => self.draw_instances(
                        &instance_bindings.shadows,
                        &self.resources().pipelines.shadows,
                        instance_range(range),
                        &mut pass,
                    ),
                    PrimitiveBatch::Paths(range) => {
                        let paths = &scene.paths[range];
                        if paths.is_empty() {
                            continue;
                        }

                        drop(pass);
                        let rasterized = self.draw_paths_to_intermediate(
                            &mut encoder,
                            paths,
                            &mut instance_offset,
                        )?;

                        pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                            label: Some("main_pass_continued"),
                            color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                                view: frame_view,
                                resolve_target: None,
                                ops: wgpu::Operations {
                                    load: wgpu::LoadOp::Load,
                                    store: wgpu::StoreOp::Store,
                                },
                                depth_slice: None,
                            })],
                            depth_stencil_attachment: None,
                            ..Default::default()
                        });

                        if rasterized {
                            self.draw_paths_from_intermediate(
                                paths,
                                &mut instance_offset,
                                &mut pass,
                            )?;
                        }
                    }
                    PrimitiveBatch::Underlines(range) => self.draw_instances(
                        &instance_bindings.underlines,
                        &self.resources().pipelines.underlines,
                        instance_range(range),
                        &mut pass,
                    ),
                    PrimitiveBatch::MonochromeSprites { texture_id, range } => self.draw_sprites(
                        &instance_bindings.monochrome_sprites,
                        texture_id,
                        &self.resources().pipelines.mono_sprites,
                        instance_range(range),
                        &mut pass,
                    ),
                    PrimitiveBatch::SubpixelSprites { texture_id, range } => {
                        let resources = self.resources();
                        self.draw_sprites(
                            &instance_bindings.subpixel_sprites,
                            texture_id,
                            resources
                                .pipelines
                                .subpixel_sprites
                                .as_ref()
                                .unwrap_or(&resources.pipelines.mono_sprites),
                            instance_range(range),
                            &mut pass,
                        );
                    }
                    PrimitiveBatch::PolychromeSprites { texture_id, range } => self.draw_sprites(
                        &instance_bindings.polychrome_sprites,
                        texture_id,
                        &self.resources().pipelines.poly_sprites,
                        instance_range(range),
                        &mut pass,
                    ),
                    // Surfaces are macOS-only for video playback and are not
                    // implemented by the WGPU renderer.
                    PrimitiveBatch::Surfaces(_surfaces) => {}
                }
            }
        }

        self.resources()
            .queue
            .submit(std::iter::once(encoder.finish()));
        Ok(())
    }

    fn write_instances(
        &mut self,
        scene: &Scene,
        instance_offset: &mut u64,
    ) -> Result<InstanceBindings> {
        Ok(InstanceBindings {
            quads: self.write_instance_binding(
                "quads_bind_group",
                instance_offset,
                &scene.quads,
            )?,
            shadows: self.write_instance_binding(
                "shadows_bind_group",
                instance_offset,
                &scene.shadows,
            )?,
            underlines: self.write_instance_binding(
                "underlines_bind_group",
                instance_offset,
                &scene.underlines,
            )?,
            monochrome_sprites: self.write_instance_binding(
                "monochrome_sprites_bind_group",
                instance_offset,
                &scene.monochrome_sprites,
            )?,
            subpixel_sprites: self.write_instance_binding(
                "subpixel_sprites_bind_group",
                instance_offset,
                &scene.subpixel_sprites,
            )?,
            polychrome_sprites: self.write_instance_binding(
                "polychrome_sprites_bind_group",
                instance_offset,
                &scene.polychrome_sprites,
            )?,
        })
    }

    fn create_texture_bind_group(
        &self,
        label: &str,
        texture_view: &wgpu::TextureView,
    ) -> wgpu::BindGroup {
        let resources = self.resources();
        resources
            .device
            .create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(label),
                layout: &resources.bind_group_layouts.texture,
                entries: &[
                    wgpu::BindGroupEntry {
                        binding: 0,
                        resource: wgpu::BindingResource::TextureView(texture_view),
                    },
                    wgpu::BindGroupEntry {
                        binding: 1,
                        resource: wgpu::BindingResource::Sampler(&resources.atlas_sampler),
                    },
                ],
            })
    }

    fn draw_instances(
        &self,
        instances: &InstanceBinding,
        pipeline: &wgpu::RenderPipeline,
        range: Range<u32>,
        pass: &mut wgpu::RenderPass<'_>,
    ) {
        if range.is_empty() {
            return;
        }
        pass.set_pipeline(pipeline);
        pass.set_bind_group(0, &self.resources().globals_bind_group, &[]);
        pass.set_bind_group(1, &instances.bind_group, &[]);
        pass.draw(
            0..4,
            instances.first_instance + range.start..instances.first_instance + range.end,
        );
    }

    fn draw_sprites(
        &self,
        sprite_instances: &InstanceBinding,
        texture_id: AtlasTextureId,
        pipeline: &wgpu::RenderPipeline,
        range: Range<u32>,
        pass: &mut wgpu::RenderPass<'_>,
    ) {
        if range.is_empty() {
            return;
        }
        let texture_info = self.atlas.get_texture_info(texture_id);
        let texture =
            self.create_texture_bind_group("atlas_texture_bind_group", &texture_info.view);
        pass.set_pipeline(pipeline);
        pass.set_bind_group(0, &self.resources().globals_bind_group, &[]);
        pass.set_bind_group(1, &sprite_instances.bind_group, &[]);
        pass.set_bind_group(2, &texture, &[]);
        pass.draw(
            0..4,
            sprite_instances.first_instance + range.start
                ..sprite_instances.first_instance + range.end,
        );
    }

    unsafe fn instance_bytes<T>(instances: &[T]) -> &[u8] {
        unsafe {
            std::slice::from_raw_parts(
                instances.as_ptr() as *const u8,
                std::mem::size_of_val(instances),
            )
        }
    }

    fn draw_paths_from_intermediate(
        &mut self,
        paths: &[Path<ScaledPixels>],
        instance_offset: &mut u64,
        pass: &mut wgpu::RenderPass<'_>,
    ) -> Result<()> {
        let first_path = &paths[0];
        let sprites: Vec<PathSprite> = if paths.last().map(|p| &p.order) == Some(&first_path.order)
        {
            paths
                .iter()
                .map(|p| PathSprite {
                    bounds: p.clipped_bounds(),
                })
                .collect()
        } 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 Some(path_intermediate_view) = self.resources().path_intermediate_view.clone() else {
            return Ok(());
        };
        let instances =
            self.write_instance_binding("path_sprites_bind_group", instance_offset, &sprites)?;
        let texture = self.create_texture_bind_group(
            "path_intermediate_texture_bind_group",
            &path_intermediate_view,
        );
        let resources = self.resources();
        pass.set_pipeline(&resources.pipelines.paths);
        pass.set_bind_group(0, &resources.globals_bind_group, &[]);
        pass.set_bind_group(1, &instances.bind_group, &[]);
        pass.set_bind_group(2, &texture, &[]);
        pass.draw(
            0..4,
            instances.first_instance..instances.first_instance + sprites.len() as u32,
        );
        Ok(())
    }

    fn draw_paths_to_intermediate(
        &mut self,
        encoder: &mut wgpu::CommandEncoder,
        paths: &[Path<ScaledPixels>],
        instance_offset: &mut u64,
    ) -> Result<bool> {
        let mut vertices = Vec::new();
        for path in paths {
            let bounds = path.clipped_bounds();
            vertices.extend(path.vertices.iter().map(|v| PathRasterizationVertex {
                xy_position: v.xy_position,
                st_position: v.st_position,
                color: path.color,
                bounds,
            }));
        }

        if vertices.is_empty() {
            return Ok(false);
        }

        let vertex_binding = self.write_instance_binding(
            "path_rasterization_bind_group",
            instance_offset,
            &vertices,
        )?;

        let resources = self.resources();
        let Some(path_intermediate_view) = resources.path_intermediate_view.as_ref() else {
            return Ok(false);
        };

        let (target_view, resolve_target) = if let Some(ref msaa_view) = resources.path_msaa_view {
            (msaa_view, Some(path_intermediate_view))
        } else {
            (path_intermediate_view, None)
        };

        {
            let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                label: Some("path_rasterization_pass"),
                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
                    view: target_view,
                    resolve_target,
                    ops: wgpu::Operations {
                        load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                        store: wgpu::StoreOp::Store,
                    },
                    depth_slice: None,
                })],
                depth_stencil_attachment: None,
                ..Default::default()
            });

            pass.set_pipeline(&resources.pipelines.path_rasterization);
            pass.set_bind_group(0, &resources.path_globals_bind_group, &[]);
            pass.set_bind_group(1, &vertex_binding.bind_group, &[]);
            // The path rasterization shader loads records by vertex index
            // rather than instance index, so the allocation's base shifts the
            // vertex range here.
            pass.draw(
                vertex_binding.first_instance
                    ..vertex_binding.first_instance + vertices.len() as u32,
                0..1,
            );
        }

        Ok(true)
    }

    fn write_instance_binding<T>(
        &mut self,
        label: &str,
        instance_offset: &mut u64,
        instances: &[T],
    ) -> Result<InstanceBinding> {
        let data = unsafe { Self::instance_bytes(instances) };
        // wgpu rejects zero-sized bindings, so empty primitive arrays still
        // reserve the 16-byte minimum.
        let size = (data.len() as u64).max(16);
        let stride = (std::mem::size_of::<T>() as u64).max(1);
        let (alignment, allocation_size) = if self.uses_webgl_instance_data {
            // The texture transport has no binding offset: the shader indexes
            // the instance texture absolutely, so each allocation must start on
            // a whole instance (a stride multiple) and a whole texel, and must
            // end on a texel boundary so the zero padding of its final partial
            // texel cannot overlap the next allocation.
            (
                least_common_multiple(self.instance_data_alignment, stride),
                size.next_multiple_of(INSTANCE_TEXTURE_TEXEL_SIZE),
            )
        } else {
            (self.instance_data_alignment.max(1), size)
        };
        let mut offset = (*instance_offset).next_multiple_of(alignment);
        if offset + allocation_size > self.instance_data_capacity {
            self.grow_instance_data(allocation_size)?;
            offset = 0;
        }
        *instance_offset = offset + allocation_size;

        let first_instance = if self.uses_webgl_instance_data {
            u32::try_from(offset / stride).context("instance index exceeds u32 range")?
        } else {
            0
        };

        let resources = self.resources();
        if !data.is_empty() {
            match &resources.instance_data {
                InstanceData::Storage(buffer) => resources.queue.write_buffer(buffer, offset, data),
                InstanceData::Texture { .. } => {
                    Self::write_instance_texture(resources, offset, data)
                }
            }
        }
        let bind_group = resources
            .device
            .create_bind_group(&wgpu::BindGroupDescriptor {
                label: Some(label),
                layout: &resources.bind_group_layouts.instances,
                entries: &[wgpu::BindGroupEntry {
                    binding: 0,
                    resource: match &resources.instance_data {
                        InstanceData::Storage(buffer) => {
                            wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                                buffer,
                                offset,
                                size: NonZeroU64::new(size),
                            })
                        }
                        InstanceData::Texture { view, .. } => {
                            wgpu::BindingResource::TextureView(view)
                        }
                    },
                }],
            });
        Ok(InstanceBinding {
            bind_group,
            first_instance,
        })
    }

    fn write_instance_texture(resources: &WgpuResources, offset: u64, data: &[u8]) {
        let InstanceData::Texture {
            texture,
            width,
            height,
            ..
        } = &resources.instance_data
        else {
            return;
        };
        let mut byte_offset = 0usize;
        let mut texel_offset = offset / INSTANCE_TEXTURE_TEXEL_SIZE;
        while byte_offset < data.len() {
            let x = (texel_offset % u64::from(*width)) as u32;
            let y = (texel_offset / u64::from(*width)) as u32;
            if y >= *height {
                // The capacity check in write_instance_binding should make this
                // unreachable. Truncating silently would leave stale bytes in the
                // texture and draw garbage for the remaining instances.
                debug_assert!(
                    false,
                    "instance texture write out of bounds: row {y} >= height {}",
                    *height
                );
                log::error!(
                    "instance texture write out of bounds; dropping {} bytes of instance data",
                    data.len() - byte_offset
                );
                return;
            }
            let available_texels = u64::from(*width - x);
            let remaining_bytes = data.len() - byte_offset;
            let complete_texels = remaining_bytes as u64 / INSTANCE_TEXTURE_TEXEL_SIZE;
            let texels = complete_texels.min(available_texels);
            if texels > 0 {
                let byte_count = (texels * INSTANCE_TEXTURE_TEXEL_SIZE) as usize;
                resources.queue.write_texture(
                    wgpu::TexelCopyTextureInfo {
                        texture,
                        mip_level: 0,
                        origin: wgpu::Origin3d { x, y, z: 0 },
                        aspect: wgpu::TextureAspect::All,
                    },
                    &data[byte_offset..byte_offset + byte_count],
                    wgpu::TexelCopyBufferLayout {
                        offset: 0,
                        bytes_per_row: Some(byte_count as u32),
                        rows_per_image: None,
                    },
                    wgpu::Extent3d {
                        width: texels as u32,
                        height: 1,
                        depth_or_array_layers: 1,
                    },
                );
                byte_offset += byte_count;
                texel_offset += texels;
                continue;
            }

            let mut final_texel = [0; INSTANCE_TEXTURE_TEXEL_SIZE as usize];
            final_texel[..remaining_bytes].copy_from_slice(&data[byte_offset..]);
            resources.queue.write_texture(
                wgpu::TexelCopyTextureInfo {
                    texture,
                    mip_level: 0,
                    origin: wgpu::Origin3d { x, y, z: 0 },
                    aspect: wgpu::TextureAspect::All,
                },
                &final_texel,
                wgpu::TexelCopyBufferLayout {
                    offset: 0,
                    bytes_per_row: Some(INSTANCE_TEXTURE_TEXEL_SIZE as u32),
                    rows_per_image: None,
                },
                wgpu::Extent3d {
                    width: 1,
                    height: 1,
                    depth_or_array_layers: 1,
                },
            );
            break;
        }
    }

    fn grow_instance_data(&mut self, required: u64) -> Result<()> {
        let capacity = (self.instance_data_capacity * 2)
            .max(required.next_power_of_two())
            .min(self.max_instance_data_size);
        anyhow::ensure!(
            capacity >= required,
            "instance data needs {required} bytes, above the maximum of {}",
            self.max_instance_data_size
        );
        anyhow::ensure!(
            capacity > self.instance_data_capacity,
            "frame instance data exceeds the {}-byte maximum",
            self.max_instance_data_size
        );
        log::debug!(
            "instance data grown from {} to {capacity}",
            self.instance_data_capacity
        );
        // Bind groups created earlier in the frame keep the previous buffer or
        // texture alive, so allocations written before the grow remain valid;
        // only subsequent writes land in the new allocation.
        let uses_webgl_instance_data = self.uses_webgl_instance_data;
        let resources = self.resources_mut();
        if uses_webgl_instance_data {
            let max_texture_dimension = resources.device.limits().max_texture_dimension_2d;
            let (instance_data, actual_capacity) =
                Self::create_instance_texture(&resources.device, capacity, max_texture_dimension);
            resources.instance_data = instance_data;
            self.instance_data_capacity = actual_capacity;
        } else {
            resources.instance_data =
                InstanceData::Storage(resources.device.create_buffer(&wgpu::BufferDescriptor {
                    label: Some("instance_buffer"),
                    size: capacity,
                    usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
                    mapped_at_creation: false,
                }));
            self.instance_data_capacity = capacity;
        }
        Ok(())
    }

    /// Mark the surface as unconfigured so rendering is skipped until a new
    /// surface is provided via [`replace_surface`](Self::replace_surface).
    ///
    /// This does **not** drop the renderer — the device, queue, atlas, and
    /// pipelines stay alive.  Use this when the native window is destroyed
    /// (e.g. Android `TerminateWindow`) but you intend to re-create the
    /// surface later without losing cached atlas textures.
    pub fn unconfigure_surface(&mut self) {
        self.surface_configured = false;
        // Drop intermediate textures since they reference the old surface size.
        if let Some(res) = self.resources.as_mut() {
            res.invalidate_intermediate_textures();
        }
    }

    /// Replace the wgpu surface with a new one (e.g. after Android destroys
    /// and recreates the native window).  Keeps the device, queue, atlas, and
    /// all pipelines intact so cached `AtlasTextureId`s remain valid.
    ///
    /// The `instance` **must** be the same [`wgpu::Instance`] that was used to
    /// create the adapter and device (i.e. from the [`WgpuContext`]).  Using a
    /// different instance will cause a "Device does not exist" panic because
    /// the wgpu device is bound to its originating instance.
    #[cfg(not(target_family = "wasm"))]
    pub fn replace_surface<W: HasWindowHandle>(
        &mut self,
        window: &W,
        config: WgpuSurfaceConfig,
        instance: &wgpu::Instance,
    ) -> anyhow::Result<()> {
        let window_handle = window
            .window_handle()
            .map_err(|e| anyhow::anyhow!("Failed to get window handle: {e}"))?;

        let surface = create_surface(instance, window_handle.as_raw())?;

        let width = (config.size.width.0 as u32).max(1);
        let height = (config.size.height.0 as u32).max(1);

        let alpha_mode = if config.transparent {
            self.transparent_alpha_mode
        } else {
            self.opaque_alpha_mode
        };

        self.surface_config.width = width;
        self.surface_config.height = height;
        self.surface_config.alpha_mode = alpha_mode;
        if let Some(mode) = config.preferred_present_mode {
            self.surface_config.present_mode = mode;
        }

        {
            let res = self
                .resources
                .as_mut()
                .expect("GPU resources not available");
            surface.configure(&res.device, &self.surface_config);
            res.surface = surface;

            // Invalidate intermediate textures — they'll be recreated lazily.
            res.invalidate_intermediate_textures();
        }

        self.surface_configured = true;

        Ok(())
    }

    pub fn destroy(&mut self) {
        // Release surface-bound GPU resources eagerly so the underlying native
        // window can be destroyed before the renderer itself is dropped.
        self.resources.take();
    }

    /// Returns true if the GPU device was lost and recovery is needed.
    pub fn device_lost(&self) -> bool {
        self.device_lost.load(std::sync::atomic::Ordering::SeqCst)
    }

    /// Returns true if a redraw is needed because GPU state was cleared.
    /// Calling this method clears the flag.
    pub fn needs_redraw(&mut self) -> bool {
        std::mem::take(&mut self.needs_redraw)
    }

    /// Recovers from a lost GPU device by recreating the renderer with a new context.
    ///
    /// Call this after detecting `device_lost()` returns true.
    ///
    /// This method coordinates recovery across multiple windows:
    /// - The first window to call this will recreate the shared context
    /// - Subsequent windows will adopt the already-recovered context
    #[cfg(not(target_family = "wasm"))]
    pub fn recover<W>(&mut self, window: &W) -> anyhow::Result<()>
    where
        W: HasWindowHandle + HasDisplayHandle + std::fmt::Debug + Send + Sync + Clone + 'static,
    {
        let gpu_context = self.context.as_ref().expect("recover requires gpu_context");

        // Check if another window already recovered the context
        let needs_new_context = gpu_context
            .borrow()
            .as_ref()
            .is_none_or(|ctx| ctx.device_lost());

        let window_handle = window
            .window_handle()
            .map_err(|e| anyhow::anyhow!("Failed to get window handle: {e}"))?;

        let surface = if needs_new_context {
            log::warn!("GPU device lost, recreating context...");

            // Drop old resources to release Arc<Device>/Arc<Queue> and GPU resources
            self.resources = None;
            *gpu_context.borrow_mut() = None;

            // Wait briefly for the GPU driver to stabilize, then try to
            // recreate the context without software renderers. If this fails
            // the caller should request another frame and retry — the real GPU
            // may need more time to come back (e.g. after suspend/resume).
            std::thread::sleep(std::time::Duration::from_millis(350));

            let instance = WgpuContext::instance(Box::new(window.clone()));
            let surface = create_surface(&instance, window_handle.as_raw())?;
            let new_context =
                WgpuContext::new_rejecting_software(instance, &surface, self.compositor_gpu)?;
            *gpu_context.borrow_mut() = Some(new_context);
            surface
        } else {
            let ctx_ref = gpu_context.borrow();
            let instance = &ctx_ref.as_ref().unwrap().instance;
            create_surface(instance, window_handle.as_raw())?
        };

        let config = WgpuSurfaceConfig {
            size: gpui_platform::Size {
                width: gpui_platform::DevicePixels(self.surface_config.width as i32),
                height: gpui_platform::DevicePixels(self.surface_config.height as i32),
            },
            transparent: self.surface_config.alpha_mode != wgpu::CompositeAlphaMode::Opaque,
            preferred_present_mode: Some(self.surface_config.present_mode),
        };
        let gpu_context = Rc::clone(gpu_context);
        let ctx_ref = gpu_context.borrow();
        let context = ctx_ref.as_ref().expect("context should exist");

        self.resources = None;
        self.atlas.handle_device_lost(context);

        *self = Self::new_internal(
            Some(gpu_context.clone()),
            context,
            surface,
            config,
            self.compositor_gpu,
            self.atlas.clone(),
        )?;

        log::info!("GPU recovery complete");
        Ok(())
    }
}

fn instance_range(range: Range<usize>) -> Range<u32> {
    range.start as u32..range.end as u32
}

#[cfg(not(target_family = "wasm"))]
fn create_surface(
    instance: &wgpu::Instance,
    raw_window_handle: raw_window_handle::RawWindowHandle,
) -> anyhow::Result<wgpu::Surface<'static>> {
    unsafe {
        instance
            .create_surface_unsafe(wgpu::SurfaceTargetUnsafe::RawHandle {
                // Fall back to the display handle already provided via InstanceDescriptor::display.
                raw_display_handle: None,
                raw_window_handle,
            })
            .map_err(|e| anyhow::anyhow!("{e}"))
    }
}

struct RenderingParameters {
    path_sample_count: u32,
    gamma_ratios: [f32; 4],
    grayscale_enhanced_contrast: f32,
    subpixel_enhanced_contrast: f32,
}

impl RenderingParameters {
    fn new(adapter: &wgpu::Adapter, surface_format: wgpu::TextureFormat) -> Self {
        use std::env;

        let format_features = adapter.get_texture_format_features(surface_format);
        let path_sample_count = [4, 2, 1]
            .into_iter()
            .find(|&n| format_features.flags.sample_count_supported(n))
            .unwrap_or(1);

        let gamma = env::var("ZED_FONTS_GAMMA")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or(1.8_f32)
            .clamp(1.0, 2.2);
        let gamma_ratios = get_gamma_correction_ratios(gamma);

        let grayscale_enhanced_contrast = env::var("ZED_FONTS_GRAYSCALE_ENHANCED_CONTRAST")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or(1.0_f32)
            .max(0.0);

        let subpixel_enhanced_contrast = env::var("ZED_FONTS_SUBPIXEL_ENHANCED_CONTRAST")
            .ok()
            .and_then(|v| v.parse().ok())
            .unwrap_or(0.5_f32)
            .max(0.0);

        Self {
            path_sample_count,
            gamma_ratios,
            grayscale_enhanced_contrast,
            subpixel_enhanced_contrast,
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use gpui_engine::{MonochromeSprite, PolychromeSprite, Quad, SubpixelSprite, Underline};
    use gpui_platform::Shadow;

    #[test]
    fn webgl_shader_is_valid_wgsl_without_storage_buffers() {
        assert!(!WEBGL_SHADERS.contains("var<storage"));
        validate_wgsl(WEBGL_SHADERS, naga::valid::Capabilities::empty());
    }

    #[test]
    fn storage_buffer_shader_is_valid_wgsl() {
        validate_wgsl(STORAGE_BUFFER_SHADERS, naga::valid::Capabilities::empty());
    }

    #[test]
    fn subpixel_shader_is_valid_wgsl() {
        validate_wgsl(
            SUBPIXEL_SHADERS,
            naga::valid::Capabilities::DUAL_SOURCE_BLENDING,
        );
    }

    fn validate_wgsl(source: &str, capabilities: naga::valid::Capabilities) {
        let module = naga::front::wgsl::parse_str(source).expect("shader should parse");
        naga::valid::Validator::new(naga::valid::ValidationFlags::all(), capabilities)
            .validate(&module)
            .expect("shader should validate");
    }

    #[test]
    fn webgl_record_sizes_match_shader_word_strides() {
        assert_eq!(std::mem::size_of::<Quad>(), 40 * 4);
        assert_eq!(std::mem::size_of::<Shadow>(), 28 * 4);
        assert_eq!(std::mem::size_of::<PathRasterizationVertex>(), 26 * 4);
        assert_eq!(std::mem::size_of::<PathSprite>(), 4 * 4);
        assert_eq!(std::mem::size_of::<Underline>(), 16 * 4);
        assert_eq!(std::mem::size_of::<MonochromeSprite>(), 28 * 4);
        assert_eq!(std::mem::size_of::<SubpixelSprite>(), 28 * 4);
        assert_eq!(std::mem::size_of::<PolychromeSprite>(), 24 * 4);
    }
}

impl SceneRenderer for WgpuRenderer {
    fn draw(&mut self, scene: &Scene) -> bool {
        WgpuRenderer::draw(self, scene)
    }

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