Added benchmarks and fixed a parallelism bug

crates/resolved-cli/src/main.rs

@@ -109,6 +109,7 @@ fn render(
         },
         appearance: file.appearance(),
         read_radiance: false,
+        profile: timings.is_some(),
     };
 
     let start = Instant::now();
@@ -121,6 +122,7 @@ fn render(
     let loading = start.elapsed();
 
     let mut frame_times = Vec::new();
+    let mut pass_times = Vec::new();
     let mut frame = None;
     for _ in 0..timings.unwrap_or(1).max(1) {
         let start = Instant::now();
@@ -129,6 +131,7 @@ fn render(
             .and_then(|p| p.wait())
             .map_err(|e| e.to_string())?;
         frame_times.push(start.elapsed().as_secs_f64() * 1e3);
+        pass_times.extend(rendered.timings);
         frame = Some(rendered);
     }
     let frame = frame.expect("at least one frame");
@@ -153,6 +156,23 @@ fn render(
                 warm.len()
             );
         }
+        // GPU time per pass, median over the frames after the first.
+        if pass_times.len() > 1 {
+            let median = |f: fn(&resolved::gpu::profile::PassTimings) -> f64| {
+                let mut v: Vec<f64> = pass_times[1..].iter().map(f).collect();
+                v.sort_by(f64::total_cmp);
+                v[v.len() / 2]
+            };
+            println!(
+                "gpu      bodies {:.3} ms, mesh {:.3} ms, sensor {:.3} ms, first to last {:.3} ms (medians)",
+                median(|t| t.bodies),
+                median(|t| t.mesh),
+                median(|t| t.sensor),
+                median(|t| t.total)
+            );
+        } else if pass_times.is_empty() {
+            println!("gpu      no pass timings: this adapter has no timestamp queries");
+        }
     }
     let detector = &renderer.camera().detector;
     if let Some(out) = out {