Opt-in logging of egui-wgpu using puffin

Cargo.lock

@@ -1175,6 +1175,7 @@ dependencies = [
  "document-features",
  "egui",
  "pollster",
+ "puffin",
  "tracing",
  "type-map",
  "wgpu",

crates/eframe/Cargo.toml

@@ -47,7 +47,7 @@ persistence = [
 ##
 ## Only enabled on native, because of the low resolution (1ms) of time keeping in browsers.
 ## `eframe` will call `puffin::GlobalProfiler::lock().new_frame()` for you
-puffin = ["dep:puffin", "egui_glow?/puffin"]
+puffin = ["dep:puffin", "egui_glow?/puffin", "egui-wgpu?/puffin"]
 
 ## Enable screen reader support (requires `ctx.options().screen_reader = true;`)
 screen_reader = ["egui-winit/screen_reader", "tts"]

crates/egui-wgpu/Cargo.toml

@@ -28,6 +28,9 @@ all-features = true
 
 
 [features]
+## Enable profiling with the [`puffin`](https://docs.rs/puffin) crate.
+puffin = ["dep:puffin"]
+
 ## Enable [`winit`](https://docs.rs/winit) integration.
 winit = ["dep:pollster", "dep:winit"]
 
@@ -48,3 +51,7 @@ document-features = { version = "0.2", optional = true }
 
 pollster = { version = "0.2", optional = true }
 winit = { version = "0.27.2", optional = true }
+
+# Native:
+[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
+puffin = { version = "0.14", optional = true }

crates/egui-wgpu/src/lib.rs

@@ -95,3 +95,25 @@ pub fn preferred_framebuffer_format(formats: &[wgpu::TextureFormat]) -> wgpu::Te
     }
     formats[0] // take the first
 }
+
+// ---------------------------------------------------------------------------
+
+/// Profiling macro for feature "puffin"
+macro_rules! profile_function {
+    ($($arg: tt)*) => {
+        #[cfg(feature = "puffin")]
+        #[cfg(not(target_arch = "wasm32"))]
+        puffin::profile_function!($($arg)*);
+    };
+}
+pub(crate) use profile_function;
+
+/// Profiling macro for feature "puffin"
+macro_rules! profile_scope {
+    ($($arg: tt)*) => {
+        #[cfg(feature = "puffin")]
+        #[cfg(not(target_arch = "wasm32"))]
+        puffin::profile_scope!($($arg)*);
+    };
+}
+pub(crate) use profile_scope;

crates/egui-wgpu/src/renderer.rs

@@ -174,6 +174,8 @@ impl Renderer {
         output_depth_format: Option<wgpu::TextureFormat>,
         msaa_samples: u32,
     ) -> Self {
+        crate::profile_function!();
+
         let shader = wgpu::ShaderModuleDescriptor {
             label: Some("egui"),
             source: wgpu::ShaderSource::Wgsl(Cow::Borrowed(include_str!("egui.wgsl"))),
@@ -347,6 +349,8 @@ impl Renderer {
         paint_jobs: &[egui::epaint::ClippedPrimitive],
         screen_descriptor: &ScreenDescriptor,
     ) {
+        crate::profile_function!();
+
         let pixels_per_point = screen_descriptor.pixels_per_point;
         let size_in_pixels = screen_descriptor.size_in_pixels;
 
@@ -421,6 +425,8 @@ impl Renderer {
                     };
 
                     if callback.rect.is_positive() {
+                        crate::profile_scope!("callback");
+
                         needs_reset = true;
 
                         {
@@ -472,6 +478,8 @@ impl Renderer {
         id: egui::TextureId,
         image_delta: &egui::epaint::ImageDelta,
     ) {
+        crate::profile_function!();
+
         let width = image_delta.image.width() as u32;
         let height = image_delta.image.height() as u32;
 
@@ -653,6 +661,8 @@ impl Renderer {
         texture: &wgpu::TextureView,
         sampler_descriptor: wgpu::SamplerDescriptor<'_>,
     ) -> egui::TextureId {
+        crate::profile_function!();
+
         let sampler = device.create_sampler(&wgpu::SamplerDescriptor {
             compare: None,
             ..sampler_descriptor
@@ -692,6 +702,8 @@ impl Renderer {
         sampler_descriptor: wgpu::SamplerDescriptor<'_>,
         id: egui::TextureId,
     ) {
+        crate::profile_function!();
+
         let (_user_texture, user_texture_binding) = self
             .textures
             .get_mut(&id)
@@ -732,8 +744,12 @@ impl Renderer {
         paint_jobs: &[egui::epaint::ClippedPrimitive],
         screen_descriptor: &ScreenDescriptor,
     ) -> Vec<wgpu::CommandBuffer> {
+        crate::profile_function!();
+
         let screen_size_in_points = screen_descriptor.screen_size_in_points();
 
+        {
+            crate::profile_scope!("uniforms");
             // Update uniform buffer
             queue.write_buffer(
                 &self.uniform_buffer,
@@ -743,9 +759,11 @@ impl Renderer {
                     _padding: Default::default(),
                 }]),
             );
+        }
 
         // Determine how many vertices & indices need to be rendered.
-        let (vertex_count, index_count) =
+        let (vertex_count, index_count) = {
+            crate::profile_scope!("count_vertices_indices");
             paint_jobs.iter().fold((0, 0), |acc, clipped_primitive| {
                 match &clipped_primitive.primitive {
                     Primitive::Mesh(mesh) => {
@@ -753,9 +771,11 @@ impl Renderer {
                     }
                     Primitive::Callback(_) => acc,
                 }
-            });
-        // Resize index buffer if needed.
+            })
+        };
+
         {
+            // Resize index buffer if needed:
             self.index_buffer.slices.clear();
             let required_size = (std::mem::size_of::<u32>() * index_count) as u64;
             if self.index_buffer.capacity < required_size {
@@ -764,8 +784,9 @@ impl Renderer {
                 self.index_buffer.buffer = create_index_buffer(device, self.index_buffer.capacity);
             }
         }
-        // Resize vertex buffer if needed.
+
         {
+            // Resize vertex buffer if needed:
             self.vertex_buffer.slices.clear();
             let required_size = (std::mem::size_of::<Vertex>() * vertex_count) as u64;
             if self.vertex_buffer.capacity < required_size {
@@ -778,6 +799,8 @@ impl Renderer {
 
         // Upload index & vertex data and call user callbacks
         let mut user_cmd_bufs = Vec::new(); // collect user command buffers
+
+        crate::profile_scope!("primitives");
         for egui::ClippedPrimitive { primitive, .. } in paint_jobs.iter() {
             match primitive {
                 Primitive::Mesh(mesh) => {
@@ -806,6 +829,7 @@ impl Renderer {
                         continue;
                     };
 
+                    crate::profile_scope!("callback");
                     user_cmd_bufs.extend((cbfn.prepare)(
                         device,
                         queue,
@@ -841,6 +865,7 @@ fn create_sampler(options: egui::TextureOptions, device: &wgpu::Device) -> wgpu:
 }
 
 fn create_vertex_buffer(device: &wgpu::Device, size: u64) -> wgpu::Buffer {
+    crate::profile_function!();
     device.create_buffer(&wgpu::BufferDescriptor {
         label: Some("egui_vertex_buffer"),
         usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
@@ -850,6 +875,7 @@ fn create_vertex_buffer(device: &wgpu::Device, size: u64) -> wgpu::Buffer {
 }
 
 fn create_index_buffer(device: &wgpu::Device, size: u64) -> wgpu::Buffer {
+    crate::profile_function!();
     device.create_buffer(&wgpu::BufferDescriptor {
         label: Some("egui_index_buffer"),
         usage: wgpu::BufferUsages::INDEX | wgpu::BufferUsages::COPY_DST,

crates/egui-wgpu/src/winit.rs

@@ -109,6 +109,8 @@ impl Painter {
     }
 
     fn configure_surface(&mut self, width_in_pixels: u32, height_in_pixels: u32) {
+        crate::profile_function!();
+
         let render_state = self
             .render_state
             .as_ref()
@@ -227,6 +229,8 @@ impl Painter {
         clipped_primitives: &[egui::ClippedPrimitive],
         textures_delta: &egui::TexturesDelta,
     ) {
+        crate::profile_function!();
+
         let render_state = match self.render_state.as_mut() {
             Some(rs) => rs,
             None => return,
@@ -237,7 +241,13 @@ impl Painter {
         };
         let (width, height) = (surface_state.width, surface_state.height);
 
-        let output_frame = match surface_state.surface.get_current_texture() {
+        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 output_frame = match output_frame {
             Ok(frame) => frame,
             #[allow(clippy::single_match_else)]
             Err(e) => match (*self.configuration.on_surface_error)(e) {
@@ -326,16 +336,25 @@ impl Painter {
             }
         }
 
+        let encoded = {
+            crate::profile_scope!("CommandEncoder::finish");
+            encoder.finish()
+        };
+
         // Submit the commands: both the main buffer and user-defined ones.
-        render_state.queue.submit(
-            user_cmd_bufs
-                .into_iter()
-                .chain(std::iter::once(encoder.finish())),
-        );
+        {
+            crate::profile_scope!("Queue::submit");
+            render_state
+                .queue
+                .submit(user_cmd_bufs.into_iter().chain(std::iter::once(encoded)));
+        };
 
         // Redraw egui
+        {
+            crate::profile_scope!("present");
             output_frame.present();
         }
+    }
 
     #[allow(clippy::unused_self)]
     pub fn destroy(&mut self) {