more profiler things

6 files changed, 83 insertions, 11 deletions

diff --git a/client/Cargo.toml b/client/Cargo.toml
index c98374a..823160e 100644
--- a/client/Cargo.toml
+++ b/client/Cargo.toml
@@ -22,3 +22,4 @@ egui = { version = "0.30.0", features = ["bytemuck"] }
 bytemuck = "1.21.0"
 xdg = "2.5.2"
 nnnoiseless = "0.5.1"
+humansize = "2.1.3"
diff --git a/client/src/interfaces/mod.rs b/client/src/interfaces/mod.rs
index 09db60f..eb0eb86 100644
--- a/client/src/interfaces/mod.rs
+++ b/client/src/interfaces/mod.rs
@@ -20,7 +20,7 @@ pub mod profiler;
 use crate::{download::Downloader, network::Network, scene_prepare::ScenePreparer};
 use egui::Widget;
 use prefabindex::PrefabIndexInterface;
-use profiler::Profiler;
+use profiler::{Profiler, TimingProfiler};
 use std::sync::{Arc, Mutex};
 use weareshared::resources::PrefabIndex;
 
@@ -35,6 +35,7 @@ pub struct InterfaceData {
     pub network: Arc<Network>,
     pub downloader: Arc<Downloader>,
     pub prefab_index: Arc<PrefabIndex>,
+    pub render_timing: Arc<Mutex<TimingProfiler>>,
 }
 
 pub fn ui_selector(idata: Arc<InterfaceData>) -> impl Fn(&egui::Context) -> bool {
diff --git a/client/src/interfaces/profiler.rs b/client/src/interfaces/profiler.rs
index 5c8737b..85b5858 100644
--- a/client/src/interfaces/profiler.rs
+++ b/client/src/interfaces/profiler.rs
@@ -15,8 +15,8 @@
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 */
 use super::InterfaceData;
-use egui::Widget;
-use std::sync::Arc;
+use egui::{Grid, Widget};
+use std::{sync::Arc, time::Instant};
 
 pub struct Profiler {
     pub idata: Arc<InterfaceData>,
@@ -28,6 +28,52 @@ impl Widget for &mut Profiler {
         ui.add(&*self.idata.scene_prepare);
         ui.heading("Download");
         ui.add(&*self.idata.downloader);
+        ui.heading("Render");
+        ui.add(&*self.idata.render_timing.lock().unwrap());
         ui.response()
     }
 }
+
+pub struct TimingProfiler {
+    last_cp: Instant,
+    cur_cp: &'static str,
+    checkpoints: Vec<(&'static str, f32)>,
+}
+
+impl Default for TimingProfiler {
+    fn default() -> Self {
+        Self {
+            last_cp: Instant::now(),
+            checkpoints: Default::default(),
+            cur_cp: "none",
+        }
+    }
+}
+impl TimingProfiler {
+    pub fn begin(&mut self, name: &'static str) {
+        self.checkpoints.clear();
+        self.last_cp = Instant::now();
+        self.cur_cp = name;
+    }
+    pub fn checkpoint(&mut self, name: &'static str) {
+        let now = Instant::now();
+        let dur = (now - self.last_cp).as_secs_f32();
+        self.last_cp = now;
+        self.checkpoints.push((self.cur_cp, dur));
+        self.cur_cp = name;
+    }
+}
+impl Widget for &TimingProfiler {
+    fn ui(self, ui: &mut egui::Ui) -> egui::Response {
+        Grid::new("tp")
+            .num_columns(2)
+            .show(ui, |ui| {
+                for (name, dur) in &self.checkpoints {
+                    ui.label(*name);
+                    ui.label(format!("{:.02}ms", dur * 1000.));
+                    ui.end_row();
+                }
+            })
+            .response
+    }
+}
diff --git a/client/src/renderer.rs b/client/src/renderer.rs
index b4c2c05..2777526 100644
--- a/client/src/renderer.rs
+++ b/client/src/renderer.rs
@@ -15,14 +15,15 @@
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 */
 use crate::{
-    camera::Camera, download::Downloader, scene_prepare::ScenePreparer,
-    scene_render::ScenePipeline, state::InputState, ui::UiRenderer,
+    camera::Camera, download::Downloader, interfaces::profiler::TimingProfiler,
+    scene_prepare::ScenePreparer, scene_render::ScenePipeline, state::InputState, ui::UiRenderer,
 };
 use anyhow::{Result, anyhow};
 use log::{info, warn};
 use pollster::FutureExt;
 use std::{
-    sync::Arc,
+    mem::swap,
+    sync::{Arc, Mutex},
     thread::{sleep, spawn},
     time::Duration,
 };
@@ -45,6 +46,8 @@ pub struct Renderer<'a> {
     pub scene_prepare: Arc<ScenePreparer>,
     surface_needs_reconfigure: bool,
     depth: TextureView,
+    pub timing: TimingProfiler,
+    pub timing_submit: Arc<Mutex<TimingProfiler>>,
 }
 impl<'a> Renderer<'a> {
     pub fn new(window: &'a Window, downloader: Arc<Downloader>) -> Result<Self> {
@@ -138,6 +141,8 @@ impl<'a> Renderer<'a> {
             surface_configuration,
             ui_renderer,
             surface_needs_reconfigure: false,
+            timing: Default::default(),
+            timing_submit: Default::default(),
         })
     }
 
@@ -172,6 +177,7 @@ impl<'a> Renderer<'a> {
         camera: &Camera,
         input_state: &mut InputState,
     ) -> Result<()> {
+        self.timing.begin("prepare");
         if self.surface_needs_reconfigure {
             self.surface
                 .configure(&self.device, &self.surface_configuration);
@@ -192,6 +198,7 @@ impl<'a> Renderer<'a> {
 
         let projection = camera.to_matrix();
 
+        self.timing.checkpoint("draw scene");
         self.scene_pipeline.draw(
             &mut commands,
             &target_view,
@@ -201,6 +208,7 @@ impl<'a> Renderer<'a> {
             projection,
         );
 
+        self.timing.checkpoint("draw ui");
         self.ui_renderer.draw(
             &mut commands,
             &target_view,
@@ -210,11 +218,18 @@ impl<'a> Renderer<'a> {
             &self.surface_configuration,
         );
 
+        self.timing.checkpoint("submit");
         let i = self.queue.submit(Some(commands.finish()));
+        self.timing.checkpoint("poll");
         self.device.poll(MaintainBase::WaitForSubmissionIndex(i));
 
+        self.timing.checkpoint("present");
         target.present();
 
+        self.timing.checkpoint("");
+
+        let mut ts = self.timing_submit.lock().unwrap();
+        swap(&mut *ts, &mut self.timing);
         Ok(())
     }
 }
diff --git a/client/src/scene_prepare.rs b/client/src/scene_prepare.rs
index cee55f7..5c5102d 100644
--- a/client/src/scene_prepare.rs
+++ b/client/src/scene_prepare.rs
@@ -17,6 +17,7 @@
 use crate::download::Downloader;
 use anyhow::Result;
 use egui::{Grid, Widget};
+use humansize::DECIMAL;
 use image::ImageReader;
 use log::debug;
 use std::{
@@ -45,6 +46,7 @@ pub struct DemandMap<K, V> {
 struct DemandMapState<K, V> {
     values: HashMap<K, V>,
     needed: HashSet<K>,
+    size_metric: usize,
 }
 impl<K: Hash + Eq + Clone, V: Clone> DemandMap<K, V> {
     pub fn new() -> Self {
@@ -52,6 +54,7 @@ impl<K: Hash + Eq + Clone, V: Clone> DemandMap<K, V> {
             inner: DemandMapState {
                 needed: HashSet::new(),
                 values: HashMap::new(),
+                size_metric: 0,
             }
             .into(),
         }
@@ -59,10 +62,11 @@ impl<K: Hash + Eq + Clone, V: Clone> DemandMap<K, V> {
     pub fn needed(&self) -> Vec<K> {
         self.inner.read().unwrap().needed.iter().cloned().collect()
     }
-    pub fn insert(&self, key: K, value: V) {
+    pub fn insert(&self, key: K, value: V, size: usize) {
         let mut s = self.inner.write().unwrap();
         s.needed.remove(&key);
         s.values.insert(key, value);
+        s.size_metric += size;
     }
     pub fn try_get(&self, key: K) -> Option<V> {
         let mut s = self.inner.write().unwrap();
@@ -127,7 +131,7 @@ impl ScenePreparer {
                     }
                 }
                 if rprefab.0.len() == prefab.mesh.len() {
-                    self.prefabs.insert(pres.clone(), Arc::new(rprefab));
+                    self.prefabs.insert(pres.clone(), Arc::new(rprefab), 0);
                     debug!("prefab created ({pres})");
                     num_done += 1;
                 }
@@ -149,6 +153,7 @@ impl ScenePreparer {
                 self.index_buffers.insert(
                     pres.clone(),
                     (Arc::new(buffer), (buf.len() / size_of::<u32>()) as u32),
+                    buf.len(),
                 );
                 debug!(
                     "index buffer created (len={}, took {:?}) {pres}",
@@ -173,6 +178,7 @@ impl ScenePreparer {
                 self.vertex_buffers.insert(
                     pres.clone(),
                     (Arc::new(buffer), (buf.len() / size_of::<f32>()) as u32),
+                    buf.len(),
                 );
                 debug!(
                     "vertex attribute buffer created (len={}, took {:?}) {pres}",
@@ -198,7 +204,7 @@ impl ScenePreparer {
                     dims.0,
                     dims.1,
                 );
-                self.textures.insert(pres.clone(), tex_bg);
+                self.textures.insert(pres.clone(), tex_bg, image.len());
                 debug!(
                     "texture created (res={}x{}, took {:?})",
                     dims.0,
@@ -218,7 +224,7 @@ impl ScenePreparer {
                 1,
                 1,
             );
-            self.placeholder_textures.insert(variant, tex_bg);
+            self.placeholder_textures.insert(variant, tex_bg, 4);
             num_done += 1;
         }
         for pres in self.mesh_parts.needed() {
@@ -294,6 +300,7 @@ impl ScenePreparer {
                                 tex_normal,
                                 double_sided: part.g_double_sided.is_some(),
                             }),
+                            0,
                         );
                         num_done += 1;
                     }
@@ -361,6 +368,7 @@ impl<K, V> Widget for &DemandMap<K, V> {
         let state = self.inner.read().unwrap();
         ui.label(state.needed.len().to_string());
         ui.label(state.values.len().to_string());
+        ui.label(humansize::format_size(state.size_metric, DECIMAL));
         ui.end_row();
         ui.response()
     }
@@ -369,7 +377,7 @@ impl<K, V> Widget for &DemandMap<K, V> {
 impl Widget for &ScenePreparer {
     fn ui(self, ui: &mut egui::Ui) -> egui::Response {
         Grid::new("sp")
-            .num_columns(3)
+            .num_columns(4)
             .show(ui, |ui| {
                 ui.label("prefabs");
                 self.prefabs.ui(ui);
diff --git a/client/src/state.rs b/client/src/state.rs
index 439b7ca..2912e14 100644
--- a/client/src/state.rs
+++ b/client/src/state.rs
@@ -108,6 +108,7 @@ impl<'a> State<'a> {
             downloader: self.downloader.clone(),
             network: self.network.clone(),
             prefab_index: self.prefab_index.clone(),
+            render_timing: self.renderer.timing_submit.clone(),
         });
         self.renderer
             .ui_renderer