Postpone call to get_current_texture (#3914)

This should help slightly with CPU/GPU parallelism when vsync is on.

I also return the time spent on vsync, which can help users figure out
how much CPU wall-time was used on non-vsync stuff

Cargo.lock

@@ -1239,6 +1239,7 @@ dependencies = [
  "puffin",
  "thiserror",
  "type-map",
+ "web-time",
  "wgpu",
  "winit",
 ]

Cargo.toml

@@ -53,6 +53,7 @@ glow = "0.13"
 puffin = "0.18"
 raw-window-handle = "0.6.0"
 thiserror = "1.0.37"
+web-time = "0.2" # Timekeeping for native and web
 wgpu = { version = "0.19.1", default-features = false, features = [
     # Make the renderer `Sync` even on wasm32, because it makes the code simpler:
     "fragile-send-sync-non-atomic-wasm",

crates/eframe/src/native/wgpu_integration.rs

@@ -644,7 +644,7 @@ impl WgpuWinitRunning {
             let clipped_primitives = egui_ctx.tessellate(shapes, pixels_per_point);
 
             let screenshot_requested = std::mem::take(&mut viewport.screenshot_requested);
-            let screenshot = painter.paint_and_update_textures(
+            let (_vsync_secs, screenshot) = painter.paint_and_update_textures(
                 viewport_id,
                 pixels_per_point,
                 app.clear_color(&egui_ctx.style().visuals),

crates/egui-wgpu/Cargo.toml

@@ -45,6 +45,7 @@ bytemuck = "1.7"
 log = { version = "0.4", features = ["std"] }
 thiserror.workspace = true
 type-map = "0.5.0"
+web-time.workspace = true
 wgpu = { workspace = true, features = ["wgsl"] }
 
 #! ### Optional dependencies

crates/egui-wgpu/src/winit.rs

@@ -504,7 +504,10 @@ impl Painter {
         })
     }
 
-    // Returns a vector with the frame's pixel data if it was requested.
+    /// Returns two things:
+    ///
+    /// The approximate number of seconds spent on vsync-waiting (if any),
+    /// and the captures captured screenshot if it was requested.
     pub fn paint_and_update_textures(
         &mut self,
         viewport_id: ViewportId,
@@ -513,29 +516,16 @@ impl Painter {
         clipped_primitives: &[epaint::ClippedPrimitive],
         textures_delta: &epaint::textures::TexturesDelta,
         capture: bool,
-    ) -> Option<epaint::ColorImage> {
+    ) -> (f32, Option<epaint::ColorImage>) {
         crate::profile_function!();
 
-        let render_state = self.render_state.as_mut()?;
-        let surface_state = self.surfaces.get(&viewport_id)?;
+        let mut vsync_sec = 0.0;
 
-        let output_frame = {
-            crate::profile_scope!("get_current_texture");
-            // This is what vsync-waiting happens, at least on Mac.
-            surface_state.surface.get_current_texture()
+        let Some(render_state) = self.render_state.as_mut() else {
+            return (vsync_sec, None);
         };
-
-        let output_frame = match output_frame {
-            Ok(frame) => frame,
-            Err(err) => match (*self.configuration.on_surface_error)(err) {
-                SurfaceErrorAction::RecreateSurface => {
-                    Self::configure_surface(surface_state, render_state, &self.configuration);
-                    return None;
-                }
-                SurfaceErrorAction::SkipFrame => {
-                    return None;
-                }
-            },
+        let Some(surface_state) = self.surfaces.get(&viewport_id) else {
+            return (vsync_sec, None);
         };
 
         let mut encoder =
@@ -580,6 +570,28 @@ impl Painter {
             }
         };
 
+        let output_frame = {
+            crate::profile_scope!("get_current_texture");
+            // This is what vsync-waiting happens on my Mac.
+            let start = web_time::Instant::now();
+            let output_frame = surface_state.surface.get_current_texture();
+            vsync_sec += start.elapsed().as_secs_f32();
+            output_frame
+        };
+
+        let output_frame = match output_frame {
+            Ok(frame) => frame,
+            Err(err) => match (*self.configuration.on_surface_error)(err) {
+                SurfaceErrorAction::RecreateSurface => {
+                    Self::configure_surface(surface_state, render_state, &self.configuration);
+                    return (vsync_sec, None);
+                }
+                SurfaceErrorAction::SkipFrame => {
+                    return (vsync_sec, None);
+                }
+            },
+        };
+
         {
             let renderer = render_state.renderer.read();
             let frame_view = if capture {
@@ -589,8 +601,11 @@ impl Painter {
                     render_state,
                 );
                 self.screen_capture_state
-                    .as_ref()?
-                    .texture
+                    .as_ref()
+                    .map_or_else(
+                        || &output_frame.texture,
+                        |capture_state| &capture_state.texture,
+                    )
                     .create_view(&wgpu::TextureViewDescriptor::default())
             } else {
                 output_frame
@@ -654,23 +669,33 @@ impl Painter {
         // Submit the commands: both the main buffer and user-defined ones.
         {
             crate::profile_scope!("Queue::submit");
+            // wgpu doesn't document where vsync can happen. Maybe here?
+            let start = web_time::Instant::now();
             render_state
                 .queue
                 .submit(user_cmd_bufs.into_iter().chain([encoded]));
+            vsync_sec += start.elapsed().as_secs_f32();
         };
 
         let screenshot = if capture {
-            let screen_capture_state = self.screen_capture_state.as_ref()?;
-            Self::read_screen_rgba(screen_capture_state, render_state, &output_frame)
+            self.screen_capture_state
+                .as_ref()
+                .and_then(|screen_capture_state| {
+                    Self::read_screen_rgba(screen_capture_state, render_state, &output_frame)
+                })
         } else {
             None
         };
 
         {
             crate::profile_scope!("present");
+            // wgpu doesn't document where vsync can happen. Maybe here?
+            let start = web_time::Instant::now();
             output_frame.present();
+            vsync_sec += start.elapsed().as_secs_f32();
         }
-        screenshot
+
+        (vsync_sec, screenshot)
     }
 
     pub fn gc_viewports(&mut self, active_viewports: &ViewportIdSet) {

crates/egui-winit/Cargo.toml

@@ -60,7 +60,7 @@ egui = { version = "0.25.0", path = "../egui", default-features = false, feature
 ] }
 log = { version = "0.4", features = ["std"] }
 raw-window-handle.workspace = true
-web-time = { version = "0.2" } # We use web-time so we can (maybe) compile for web
+web-time.workspace = true
 winit = { workspace = true, default-features = false, features = ["rwh_06"] }
 
 #! ### Optional dependencies