55493a823e
# Objective - Allow prepare windows to run off of the main thread on all platforms. - Fixes https://github.com/bevyengine/bevy/issues/9964 on all platforms. ## Solution - Running `prepare_windows` on the main thread on apple platforms is only mandatory to create surface, which is only needed during window creation. Split that part into its own system that happens before `prepare_windows` - Tested on macOS and iOS --- ## Changelog - Allow prepare windows to run off main thread on all platforms.
469 lines
20 KiB
Rust
469 lines
20 KiB
Rust
use crate::{
|
|
render_resource::{
|
|
BindGroupEntries, PipelineCache, SpecializedRenderPipelines, SurfaceTexture, TextureView,
|
|
},
|
|
renderer::{RenderAdapter, RenderDevice, RenderInstance},
|
|
texture::TextureFormatPixelInfo,
|
|
Extract, ExtractSchedule, Render, RenderApp, RenderSet,
|
|
};
|
|
use bevy_app::{App, Plugin};
|
|
use bevy_ecs::prelude::*;
|
|
use bevy_utils::{default, tracing::debug, EntityHashMap, HashSet};
|
|
use bevy_window::{
|
|
CompositeAlphaMode, PresentMode, PrimaryWindow, RawHandleWrapper, Window, WindowClosed,
|
|
};
|
|
use std::{
|
|
ops::{Deref, DerefMut},
|
|
sync::PoisonError,
|
|
};
|
|
use wgpu::{
|
|
BufferUsages, SurfaceTargetUnsafe, TextureFormat, TextureUsages, TextureViewDescriptor,
|
|
};
|
|
|
|
pub mod screenshot;
|
|
|
|
use screenshot::{
|
|
ScreenshotManager, ScreenshotPlugin, ScreenshotPreparedState, ScreenshotToScreenPipeline,
|
|
};
|
|
|
|
use super::Msaa;
|
|
|
|
/// Token to ensure a system runs on the main thread.
|
|
#[derive(Resource, Default)]
|
|
pub struct NonSendMarker;
|
|
|
|
pub struct WindowRenderPlugin;
|
|
|
|
impl Plugin for WindowRenderPlugin {
|
|
fn build(&self, app: &mut App) {
|
|
app.add_plugins(ScreenshotPlugin);
|
|
|
|
if let Ok(render_app) = app.get_sub_app_mut(RenderApp) {
|
|
render_app
|
|
.init_resource::<ExtractedWindows>()
|
|
.init_resource::<WindowSurfaces>()
|
|
.add_systems(ExtractSchedule, extract_windows)
|
|
.add_systems(Render, prepare_windows.in_set(RenderSet::PrepareAssets))
|
|
.add_systems(Render, create_surfaces.in_set(RenderSet::ManageViews));
|
|
}
|
|
}
|
|
|
|
fn finish(&self, app: &mut App) {
|
|
if let Ok(render_app) = app.get_sub_app_mut(RenderApp) {
|
|
render_app.init_resource::<ScreenshotToScreenPipeline>();
|
|
}
|
|
}
|
|
}
|
|
|
|
pub struct ExtractedWindow {
|
|
/// An entity that contains the components in [`Window`].
|
|
pub entity: Entity,
|
|
pub handle: RawHandleWrapper,
|
|
pub physical_width: u32,
|
|
pub physical_height: u32,
|
|
pub present_mode: PresentMode,
|
|
/// Note: this will not always be the swap chain texture view. When taking a screenshot,
|
|
/// this will point to an alternative texture instead to allow for copying the render result
|
|
/// to CPU memory.
|
|
pub swap_chain_texture_view: Option<TextureView>,
|
|
pub swap_chain_texture: Option<SurfaceTexture>,
|
|
pub swap_chain_texture_format: Option<TextureFormat>,
|
|
pub screenshot_memory: Option<ScreenshotPreparedState>,
|
|
pub size_changed: bool,
|
|
pub present_mode_changed: bool,
|
|
pub alpha_mode: CompositeAlphaMode,
|
|
pub screenshot_func: Option<screenshot::ScreenshotFn>,
|
|
}
|
|
|
|
impl ExtractedWindow {
|
|
fn set_swapchain_texture(&mut self, frame: wgpu::SurfaceTexture) {
|
|
let texture_view_descriptor = TextureViewDescriptor {
|
|
format: Some(frame.texture.format().add_srgb_suffix()),
|
|
..default()
|
|
};
|
|
self.swap_chain_texture_view = Some(TextureView::from(
|
|
frame.texture.create_view(&texture_view_descriptor),
|
|
));
|
|
self.swap_chain_texture = Some(SurfaceTexture::from(frame));
|
|
}
|
|
}
|
|
|
|
#[derive(Default, Resource)]
|
|
pub struct ExtractedWindows {
|
|
pub primary: Option<Entity>,
|
|
pub windows: EntityHashMap<Entity, ExtractedWindow>,
|
|
}
|
|
|
|
impl Deref for ExtractedWindows {
|
|
type Target = EntityHashMap<Entity, ExtractedWindow>;
|
|
|
|
fn deref(&self) -> &Self::Target {
|
|
&self.windows
|
|
}
|
|
}
|
|
|
|
impl DerefMut for ExtractedWindows {
|
|
fn deref_mut(&mut self) -> &mut Self::Target {
|
|
&mut self.windows
|
|
}
|
|
}
|
|
|
|
fn extract_windows(
|
|
mut extracted_windows: ResMut<ExtractedWindows>,
|
|
screenshot_manager: Extract<Res<ScreenshotManager>>,
|
|
mut closed: Extract<EventReader<WindowClosed>>,
|
|
windows: Extract<Query<(Entity, &Window, &RawHandleWrapper, Option<&PrimaryWindow>)>>,
|
|
mut removed: Extract<RemovedComponents<RawHandleWrapper>>,
|
|
mut window_surfaces: ResMut<WindowSurfaces>,
|
|
) {
|
|
for (entity, window, handle, primary) in windows.iter() {
|
|
if primary.is_some() {
|
|
extracted_windows.primary = Some(entity);
|
|
}
|
|
|
|
let (new_width, new_height) = (
|
|
window.resolution.physical_width().max(1),
|
|
window.resolution.physical_height().max(1),
|
|
);
|
|
|
|
let extracted_window = extracted_windows.entry(entity).or_insert(ExtractedWindow {
|
|
entity,
|
|
handle: handle.clone(),
|
|
physical_width: new_width,
|
|
physical_height: new_height,
|
|
present_mode: window.present_mode,
|
|
swap_chain_texture: None,
|
|
swap_chain_texture_view: None,
|
|
size_changed: false,
|
|
swap_chain_texture_format: None,
|
|
present_mode_changed: false,
|
|
alpha_mode: window.composite_alpha_mode,
|
|
screenshot_func: None,
|
|
screenshot_memory: None,
|
|
});
|
|
|
|
// NOTE: Drop the swap chain frame here
|
|
extracted_window.swap_chain_texture_view = None;
|
|
extracted_window.size_changed = new_width != extracted_window.physical_width
|
|
|| new_height != extracted_window.physical_height;
|
|
extracted_window.present_mode_changed =
|
|
window.present_mode != extracted_window.present_mode;
|
|
|
|
if extracted_window.size_changed {
|
|
debug!(
|
|
"Window size changed from {}x{} to {}x{}",
|
|
extracted_window.physical_width,
|
|
extracted_window.physical_height,
|
|
new_width,
|
|
new_height
|
|
);
|
|
extracted_window.physical_width = new_width;
|
|
extracted_window.physical_height = new_height;
|
|
}
|
|
|
|
if extracted_window.present_mode_changed {
|
|
debug!(
|
|
"Window Present Mode changed from {:?} to {:?}",
|
|
extracted_window.present_mode, window.present_mode
|
|
);
|
|
extracted_window.present_mode = window.present_mode;
|
|
}
|
|
}
|
|
|
|
for closed_window in closed.read() {
|
|
extracted_windows.remove(&closed_window.window);
|
|
window_surfaces.remove(&closed_window.window);
|
|
}
|
|
for removed_window in removed.read() {
|
|
extracted_windows.remove(&removed_window);
|
|
window_surfaces.remove(&removed_window);
|
|
}
|
|
// This lock will never block because `callbacks` is `pub(crate)` and this is the singular callsite where it's locked.
|
|
// Even if a user had multiple copies of this system, since the system has a mutable resource access the two systems would never run
|
|
// at the same time
|
|
// TODO: since this is guaranteed, should the lock be replaced with an UnsafeCell to remove the overhead, or is it minor enough to be ignored?
|
|
for (window, screenshot_func) in screenshot_manager
|
|
.callbacks
|
|
.lock()
|
|
.unwrap_or_else(PoisonError::into_inner)
|
|
.drain()
|
|
{
|
|
if let Some(window) = extracted_windows.get_mut(&window) {
|
|
window.screenshot_func = Some(screenshot_func);
|
|
}
|
|
}
|
|
}
|
|
|
|
struct SurfaceData {
|
|
// TODO: what lifetime should this be?
|
|
surface: wgpu::Surface<'static>,
|
|
format: TextureFormat,
|
|
}
|
|
|
|
#[derive(Resource, Default)]
|
|
pub struct WindowSurfaces {
|
|
surfaces: EntityHashMap<Entity, SurfaceData>,
|
|
/// List of windows that we have already called the initial `configure_surface` for
|
|
configured_windows: HashSet<Entity>,
|
|
}
|
|
|
|
impl WindowSurfaces {
|
|
fn remove(&mut self, window: &Entity) {
|
|
self.surfaces.remove(window);
|
|
self.configured_windows.remove(window);
|
|
}
|
|
}
|
|
|
|
/// (re)configures window surfaces, and obtains a swapchain texture for rendering.
|
|
///
|
|
/// NOTE: `get_current_texture` in `prepare_windows` can take a long time if the GPU workload is
|
|
/// the performance bottleneck. This can be seen in profiles as multiple prepare-set systems all
|
|
/// taking an unusually long time to complete, and all finishing at about the same time as the
|
|
/// `prepare_windows` system. Improvements in bevy are planned to avoid this happening when it
|
|
/// should not but it will still happen as it is easy for a user to create a large GPU workload
|
|
/// relative to the GPU performance and/or CPU workload.
|
|
/// This can be caused by many reasons, but several of them are:
|
|
/// - GPU workload is more than your current GPU can manage
|
|
/// - Error / performance bug in your custom shaders
|
|
/// - wgpu was unable to detect a proper GPU hardware-accelerated device given the chosen
|
|
/// [`Backends`](crate::settings::Backends), [`WgpuLimits`](crate::settings::WgpuLimits),
|
|
/// and/or [`WgpuFeatures`](crate::settings::WgpuFeatures). For example, on Windows currently
|
|
/// `DirectX 11` is not supported by wgpu 0.12 and so if your GPU/drivers do not support Vulkan,
|
|
/// it may be that a software renderer called "Microsoft Basic Render Driver" using `DirectX 12`
|
|
/// will be chosen and performance will be very poor. This is visible in a log message that is
|
|
/// output during renderer initialization. Future versions of wgpu will support `DirectX 11`, but
|
|
/// another alternative is to try to use [`ANGLE`](https://github.com/gfx-rs/wgpu#angle) and
|
|
/// [`Backends::GL`](crate::settings::Backends::GL) if your GPU/drivers support `OpenGL 4.3` / `OpenGL ES 3.0` or
|
|
/// later.
|
|
#[allow(clippy::too_many_arguments)]
|
|
pub fn prepare_windows(
|
|
mut windows: ResMut<ExtractedWindows>,
|
|
mut window_surfaces: ResMut<WindowSurfaces>,
|
|
render_device: Res<RenderDevice>,
|
|
render_adapter: Res<RenderAdapter>,
|
|
screenshot_pipeline: Res<ScreenshotToScreenPipeline>,
|
|
pipeline_cache: Res<PipelineCache>,
|
|
mut pipelines: ResMut<SpecializedRenderPipelines<ScreenshotToScreenPipeline>>,
|
|
mut msaa: ResMut<Msaa>,
|
|
#[cfg(target_os = "linux")] render_instance: Res<RenderInstance>,
|
|
) {
|
|
for window in windows.windows.values_mut() {
|
|
let window_surfaces = window_surfaces.deref_mut();
|
|
let Some(surface_data) = window_surfaces.surfaces.get(&window.entity) else {
|
|
continue;
|
|
};
|
|
|
|
let surface_configuration = wgpu::SurfaceConfiguration {
|
|
format: surface_data.format,
|
|
width: window.physical_width,
|
|
height: window.physical_height,
|
|
usage: TextureUsages::RENDER_ATTACHMENT,
|
|
present_mode: match window.present_mode {
|
|
PresentMode::Fifo => wgpu::PresentMode::Fifo,
|
|
PresentMode::FifoRelaxed => wgpu::PresentMode::FifoRelaxed,
|
|
PresentMode::Mailbox => wgpu::PresentMode::Mailbox,
|
|
PresentMode::Immediate => wgpu::PresentMode::Immediate,
|
|
PresentMode::AutoVsync => wgpu::PresentMode::AutoVsync,
|
|
PresentMode::AutoNoVsync => wgpu::PresentMode::AutoNoVsync,
|
|
},
|
|
// TODO: Expose this as a setting somewhere
|
|
// 2 is wgpu's default/what we've been using so far.
|
|
// 1 is the minimum, but may cause lower framerates due to the cpu waiting for the gpu to finish
|
|
// all work for the previous frame before starting work on the next frame, which then means the gpu
|
|
// has to wait for the cpu to finish to start on the next frame.
|
|
desired_maximum_frame_latency: 2,
|
|
alpha_mode: match window.alpha_mode {
|
|
CompositeAlphaMode::Auto => wgpu::CompositeAlphaMode::Auto,
|
|
CompositeAlphaMode::Opaque => wgpu::CompositeAlphaMode::Opaque,
|
|
CompositeAlphaMode::PreMultiplied => wgpu::CompositeAlphaMode::PreMultiplied,
|
|
CompositeAlphaMode::PostMultiplied => wgpu::CompositeAlphaMode::PostMultiplied,
|
|
CompositeAlphaMode::Inherit => wgpu::CompositeAlphaMode::Inherit,
|
|
},
|
|
view_formats: if !surface_data.format.is_srgb() {
|
|
vec![surface_data.format.add_srgb_suffix()]
|
|
} else {
|
|
vec![]
|
|
},
|
|
};
|
|
|
|
// This is an ugly hack to work around drivers that don't support MSAA.
|
|
// This should be removed once https://github.com/bevyengine/bevy/issues/7194 lands and we're doing proper
|
|
// feature detection for MSAA.
|
|
// When removed, we can also remove the `.after(prepare_windows)` of `prepare_core_3d_depth_textures` and `prepare_prepass_textures`
|
|
let sample_flags = render_adapter
|
|
.get_texture_format_features(surface_configuration.format)
|
|
.flags;
|
|
|
|
if !sample_flags.sample_count_supported(msaa.samples()) {
|
|
let fallback = if sample_flags.sample_count_supported(Msaa::default().samples()) {
|
|
Msaa::default()
|
|
} else {
|
|
Msaa::Off
|
|
};
|
|
|
|
let fallback_str = if fallback == Msaa::Off {
|
|
"disabling MSAA".to_owned()
|
|
} else {
|
|
format!("MSAA {}x", fallback.samples())
|
|
};
|
|
|
|
bevy_log::warn!(
|
|
"MSAA {}x is not supported on this device. Falling back to {}.",
|
|
msaa.samples(),
|
|
fallback_str,
|
|
);
|
|
*msaa = fallback;
|
|
}
|
|
|
|
// A recurring issue is hitting `wgpu::SurfaceError::Timeout` on certain Linux
|
|
// mesa driver implementations. This seems to be a quirk of some drivers.
|
|
// We'd rather keep panicking when not on Linux mesa, because in those case,
|
|
// the `Timeout` is still probably the symptom of a degraded unrecoverable
|
|
// application state.
|
|
// see https://github.com/bevyengine/bevy/pull/5957
|
|
// and https://github.com/gfx-rs/wgpu/issues/1218
|
|
#[cfg(target_os = "linux")]
|
|
let may_erroneously_timeout = || {
|
|
render_instance
|
|
.enumerate_adapters(wgpu::Backends::VULKAN)
|
|
.iter()
|
|
.any(|adapter| {
|
|
let name = adapter.get_info().name;
|
|
name.starts_with("Radeon")
|
|
|| name.starts_with("AMD")
|
|
|| name.starts_with("Intel")
|
|
})
|
|
};
|
|
|
|
let not_already_configured = window_surfaces.configured_windows.insert(window.entity);
|
|
|
|
let surface = &surface_data.surface;
|
|
if not_already_configured || window.size_changed || window.present_mode_changed {
|
|
render_device.configure_surface(surface, &surface_configuration);
|
|
let frame = surface
|
|
.get_current_texture()
|
|
.expect("Error configuring surface");
|
|
window.set_swapchain_texture(frame);
|
|
} else {
|
|
match surface.get_current_texture() {
|
|
Ok(frame) => {
|
|
window.set_swapchain_texture(frame);
|
|
}
|
|
Err(wgpu::SurfaceError::Outdated) => {
|
|
render_device.configure_surface(surface, &surface_configuration);
|
|
let frame = surface
|
|
.get_current_texture()
|
|
.expect("Error reconfiguring surface");
|
|
window.set_swapchain_texture(frame);
|
|
}
|
|
#[cfg(target_os = "linux")]
|
|
Err(wgpu::SurfaceError::Timeout) if may_erroneously_timeout() => {
|
|
bevy_utils::tracing::trace!(
|
|
"Couldn't get swap chain texture. This is probably a quirk \
|
|
of your Linux GPU driver, so it can be safely ignored."
|
|
);
|
|
}
|
|
Err(err) => {
|
|
panic!("Couldn't get swap chain texture, operation unrecoverable: {err}");
|
|
}
|
|
}
|
|
};
|
|
window.swap_chain_texture_format = Some(surface_data.format);
|
|
|
|
if window.screenshot_func.is_some() {
|
|
let texture = render_device.create_texture(&wgpu::TextureDescriptor {
|
|
label: Some("screenshot-capture-rendertarget"),
|
|
size: wgpu::Extent3d {
|
|
width: surface_configuration.width,
|
|
height: surface_configuration.height,
|
|
depth_or_array_layers: 1,
|
|
},
|
|
mip_level_count: 1,
|
|
sample_count: 1,
|
|
dimension: wgpu::TextureDimension::D2,
|
|
format: surface_configuration.format.add_srgb_suffix(),
|
|
usage: TextureUsages::RENDER_ATTACHMENT
|
|
| TextureUsages::COPY_SRC
|
|
| TextureUsages::TEXTURE_BINDING,
|
|
view_formats: &[],
|
|
});
|
|
let texture_view = texture.create_view(&Default::default());
|
|
let buffer = render_device.create_buffer(&wgpu::BufferDescriptor {
|
|
label: Some("screenshot-transfer-buffer"),
|
|
size: screenshot::get_aligned_size(
|
|
window.physical_width,
|
|
window.physical_height,
|
|
surface_data.format.pixel_size() as u32,
|
|
) as u64,
|
|
usage: BufferUsages::MAP_READ | BufferUsages::COPY_DST,
|
|
mapped_at_creation: false,
|
|
});
|
|
let bind_group = render_device.create_bind_group(
|
|
"screenshot-to-screen-bind-group",
|
|
&screenshot_pipeline.bind_group_layout,
|
|
&BindGroupEntries::single(&texture_view),
|
|
);
|
|
let pipeline_id = pipelines.specialize(
|
|
&pipeline_cache,
|
|
&screenshot_pipeline,
|
|
surface_configuration.format,
|
|
);
|
|
window.swap_chain_texture_view = Some(texture_view);
|
|
window.screenshot_memory = Some(ScreenshotPreparedState {
|
|
texture,
|
|
buffer,
|
|
bind_group,
|
|
pipeline_id,
|
|
});
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Creates window surfaces.
|
|
pub fn create_surfaces(
|
|
// By accessing a NonSend resource, we tell the scheduler to put this system on the main thread,
|
|
// which is necessary for some OS's
|
|
#[cfg(any(target_os = "macos", target_os = "ios"))] _marker: Option<NonSend<NonSendMarker>>,
|
|
windows: Res<ExtractedWindows>,
|
|
mut window_surfaces: ResMut<WindowSurfaces>,
|
|
render_instance: Res<RenderInstance>,
|
|
render_adapter: Res<RenderAdapter>,
|
|
) {
|
|
for window in windows.windows.values() {
|
|
window_surfaces
|
|
.surfaces
|
|
.entry(window.entity)
|
|
.or_insert_with(|| {
|
|
let surface_target = SurfaceTargetUnsafe::RawHandle {
|
|
raw_display_handle: window.handle.display_handle,
|
|
raw_window_handle: window.handle.window_handle,
|
|
};
|
|
// SAFETY: The window handles in ExtractedWindows will always be valid objects to create surfaces on
|
|
let surface = unsafe {
|
|
// NOTE: On some OSes this MUST be called from the main thread.
|
|
// As of wgpu 0.15, only fallible if the given window is a HTML canvas and obtaining a WebGPU or WebGL2 context fails.
|
|
render_instance
|
|
.create_surface_unsafe(surface_target)
|
|
.expect("Failed to create wgpu surface")
|
|
};
|
|
let caps = surface.get_capabilities(&render_adapter);
|
|
let formats = caps.formats;
|
|
// For future HDR output support, we'll need to request a format that supports HDR,
|
|
// but as of wgpu 0.15 that is not yet supported.
|
|
// Prefer sRGB formats for surfaces, but fall back to first available format if no sRGB formats are available.
|
|
let mut format = *formats.first().expect("No supported formats for surface");
|
|
for available_format in formats {
|
|
// Rgba8UnormSrgb and Bgra8UnormSrgb and the only sRGB formats wgpu exposes that we can use for surfaces.
|
|
if available_format == TextureFormat::Rgba8UnormSrgb
|
|
|| available_format == TextureFormat::Bgra8UnormSrgb
|
|
{
|
|
format = available_format;
|
|
break;
|
|
}
|
|
}
|
|
|
|
SurfaceData { surface, format }
|
|
});
|
|
}
|
|
}
|