Add opt-in puffin feature to egui (#3298)

* Add opt-in `puffin` feature to egui

* fix web build

* Fix web for realz

crates/eframe/Cargo.toml

@@ -63,9 +63,9 @@ persistence = [
 
 ## Enable profiling with the [`puffin`](https://docs.rs/puffin) crate.
 ##
-## Only enabled on native, because of the low resolution (1ms) of time keeping in browsers.
+## Only enabled on native, because of the low resolution (1ms) of clocks in browsers.
 ## `eframe` will call `puffin::GlobalProfiler::lock().new_frame()` for you
-puffin = ["dep:puffin", "egui_glow?/puffin", "egui-wgpu?/puffin"]
+puffin = ["dep:puffin", "egui/puffin", "egui_glow?/puffin", "egui-wgpu?/puffin"]
 
 ## Enable screen reader support (requires `ctx.options_mut(|o| o.screen_reader = true);`) on web.
 ##

crates/egui-winit/Cargo.toml

@@ -34,7 +34,7 @@ clipboard = ["arboard", "smithay-clipboard"]
 links = ["webbrowser"]
 
 ## Enable profiling with the [`puffin`](https://docs.rs/puffin) crate.
-puffin = ["dep:puffin"]
+puffin = ["dep:puffin", "egui/puffin"]
 
 ## Allow serialization of [`WindowSettings`] using [`serde`](https://docs.rs/serde).
 serde = ["egui/serde", "dep:serde"]

crates/egui/Cargo.toml

@@ -54,6 +54,11 @@ mint = ["epaint/mint"]
 ## Enable persistence of memory (window positions etc).
 persistence = ["serde", "epaint/serde", "ron"]
 
+## Enable profiling with the [`puffin`](https://docs.rs/puffin) crate.
+##
+## Only enabled on native, because of the low resolution (1ms) of clocks in browsers.
+puffin = ["dep:puffin"]
+
 ## Allow serialization using [`serde`](https://docs.rs/serde).
 serde = ["dep:serde", "epaint/serde", "accesskit?/serde"]
 
@@ -79,5 +84,6 @@ accesskit = { version = "0.11", optional = true }
 document-features = { version = "0.2", optional = true }
 
 log = { version = "0.4", optional = true, features = ["std"] }
+puffin = { version = "0.16", optional = true }
 ron = { version = "0.8", optional = true }
 serde = { version = "1", optional = true, features = ["derive", "rc"] }

crates/egui/src/context.rs

@@ -207,6 +207,7 @@ impl ContextImpl {
 
         #[cfg(feature = "accesskit")]
         if self.is_accesskit_enabled {
+            crate::profile_scope!("accesskit");
             use crate::frame_state::AccessKitFrameState;
             let id = crate::accesskit_root_id();
             let mut builder = accesskit::NodeBuilder::new(accesskit::Role::Window);
@@ -224,24 +225,32 @@ impl ContextImpl {
 
     /// Load fonts unless already loaded.
     fn update_fonts_mut(&mut self) {
+        crate::profile_function!();
+
         let pixels_per_point = self.input.pixels_per_point();
         let max_texture_side = self.input.max_texture_side;
 
         if let Some(font_definitions) = self.memory.new_font_definitions.take() {
+            crate::profile_scope!("Fonts::new");
             let fonts = Fonts::new(pixels_per_point, max_texture_side, font_definitions);
             self.fonts = Some(fonts);
         }
 
         let fonts = self.fonts.get_or_insert_with(|| {
             let font_definitions = FontDefinitions::default();
+            crate::profile_scope!("Fonts::new");
             Fonts::new(pixels_per_point, max_texture_side, font_definitions)
         });
 
-        fonts.begin_frame(pixels_per_point, max_texture_side);
+        {
+            crate::profile_scope!("Fonts::begin_frame");
+            fonts.begin_frame(pixels_per_point, max_texture_side);
+        }
 
         if self.memory.options.preload_font_glyphs {
+            crate::profile_scope!("preload_font_glyphs");
             // Preload the most common characters for the most common fonts.
-            // This is not very important to do, but may a few GPU operations.
+            // This is not very important to do, but may save a few GPU operations.
             for font_id in self.memory.options.style.text_styles.values() {
                 fonts.lock().fonts.font(font_id).preload_common_characters();
             }
@@ -370,6 +379,7 @@ impl Context {
     /// ```
     #[must_use]
     pub fn run(&self, new_input: RawInput, run_ui: impl FnOnce(&Context)) -> FullOutput {
+        crate::profile_function!();
         self.begin_frame(new_input);
         run_ui(self);
         self.end_frame()
@@ -393,6 +403,7 @@ impl Context {
     /// // handle full_output
     /// ```
     pub fn begin_frame(&self, new_input: RawInput) {
+        crate::profile_function!();
         self.write(|ctx| ctx.begin_frame_mut(new_input));
     }
 }
@@ -1207,6 +1218,7 @@ impl Context {
     /// Call at the end of each frame.
     #[must_use]
     pub fn end_frame(&self) -> FullOutput {
+        crate::profile_function!();
         if self.input(|i| i.wants_repaint()) {
             self.request_repaint();
         }
@@ -1230,6 +1242,7 @@ impl Context {
 
         #[cfg(feature = "accesskit")]
         {
+            crate::profile_scope!("accesskit");
             let state = self.frame_state_mut(|fs| fs.accesskit_state.take());
             if let Some(state) = state {
                 let has_focus = self.input(|i| i.raw.focused);
@@ -1269,11 +1282,13 @@ impl Context {
     }
 
     fn drain_paint_lists(&self) -> Vec<ClippedShape> {
+        crate::profile_function!();
         self.write(|ctx| ctx.graphics.drain(ctx.memory.areas.order()).collect())
     }
 
     /// Tessellate the given shapes into triangle meshes.
     pub fn tessellate(&self, shapes: Vec<ClippedShape>) -> Vec<ClippedPrimitive> {
+        crate::profile_function!();
         // A tempting optimization is to reuse the tessellation from last frame if the
         // shapes are the same, but just comparing the shapes takes about 50% of the time
         // it takes to tessellate them, so it is not a worth optimization.
@@ -1293,13 +1308,16 @@ impl Context {
             };
 
             let paint_stats = PaintStats::from_shapes(&shapes);
-            let clipped_primitives = tessellator::tessellate_shapes(
-                pixels_per_point,
-                tessellation_options,
-                font_tex_size,
-                prepared_discs,
-                shapes,
-            );
+            let clipped_primitives = {
+                crate::profile_scope!("tessellator::tessellate_shapes");
+                tessellator::tessellate_shapes(
+                    pixels_per_point,
+                    tessellation_options,
+                    font_tex_size,
+                    prepared_discs,
+                    shapes,
+                )
+            };
             ctx.paint_stats = paint_stats.with_clipped_primitives(&clipped_primitives);
             clipped_primitives
         })

crates/egui/src/frame_state.rs

@@ -76,6 +76,7 @@ impl Default for FrameState {
 
 impl FrameState {
     pub(crate) fn begin_frame(&mut self, input: &InputState) {
+        crate::profile_function!();
         let Self {
             used_ids,
             available_rect,

crates/egui/src/input_state.rs

@@ -149,6 +149,7 @@ impl InputState {
         mut new: RawInput,
         requested_repaint_last_frame: bool,
     ) -> InputState {
+        crate::profile_function!();
         let time = new.time.unwrap_or(self.time + new.predicted_dt as f64);
         let unstable_dt = (time - self.time) as f32;
 

crates/egui/src/lib.rs

@@ -579,3 +579,32 @@ pub fn __run_test_ui(mut add_contents: impl FnMut(&mut Ui)) {
 pub fn accesskit_root_id() -> Id {
     Id::new("accesskit_root")
 }
+
+// ---------------------------------------------------------------------------
+
+mod profiling_scopes {
+    #![allow(unused_macros)]
+    #![allow(unused_imports)]
+
+    /// Profiling macro for feature "puffin"
+    macro_rules! profile_function {
+        ($($arg: tt)*) => {
+            #[cfg(not(target_arch = "wasm32"))] // Disabled on web because of the coarse 1ms clock resolution there.
+            #[cfg(feature = "puffin")]
+            puffin::profile_function!($($arg)*);
+        };
+    }
+    pub(crate) use profile_function;
+
+    /// Profiling macro for feature "puffin"
+    macro_rules! profile_scope {
+        ($($arg: tt)*) => {
+            #[cfg(not(target_arch = "wasm32"))] // Disabled on web because of the coarse 1ms clock resolution there.
+            #[cfg(feature = "puffin")]
+            puffin::profile_scope!($($arg)*);
+        };
+    }
+    pub(crate) use profile_scope;
+}
+
+pub(crate) use profiling_scopes::*;

crates/egui/src/memory.rs

@@ -493,6 +493,7 @@ impl Memory {
         prev_input: &crate::input_state::InputState,
         new_input: &crate::data::input::RawInput,
     ) {
+        crate::profile_function!();
         self.interaction.begin_frame(prev_input, new_input);
 
         if !prev_input.pointer.any_down() {

crates/egui/src/util/id_type_map.rs

@@ -515,6 +515,7 @@ struct PersistedMap(Vec<(u64, SerializedElement)>);
 #[cfg(feature = "persistence")]
 impl PersistedMap {
     fn from_map(map: &IdTypeMap) -> Self {
+        crate::profile_function!();
         // filter out the elements which cannot be serialized:
         Self(
             map.0
@@ -525,6 +526,7 @@ impl PersistedMap {
     }
 
     fn into_map(self) -> IdTypeMap {
+        crate::profile_function!();
         IdTypeMap(
             self.0
                 .into_iter()
@@ -542,6 +544,7 @@ impl serde::Serialize for IdTypeMap {
     where
         S: serde::Serializer,
     {
+        crate::profile_scope!("IdTypeMap::serialize");
         PersistedMap::from_map(self).serialize(serializer)
     }
 }
@@ -552,6 +555,7 @@ impl<'de> serde::Deserialize<'de> for IdTypeMap {
     where
         D: serde::Deserializer<'de>,
     {
+        crate::profile_scope!("IdTypeMap::deserialize");
         <PersistedMap>::deserialize(deserializer).map(PersistedMap::into_map)
     }
 }

crates/egui_glow/Cargo.toml

@@ -37,7 +37,7 @@ clipboard = ["egui-winit?/clipboard"]
 links = ["egui-winit?/links"]
 
 ## Enable profiling with the [`puffin`](https://docs.rs/puffin) crate.
-puffin = ["dep:puffin", "egui-winit?/puffin"]
+puffin = ["dep:puffin", "egui-winit?/puffin", "egui/puffin"]
 
 ## Enable [`winit`](https://docs.rs/winit) integration.
 winit = ["egui-winit"]