From bb7d473edf3fcf8f19e3dfc4bb861545d6167a1a Mon Sep 17 00:00:00 2001
From: guillaume <guillaume@adaptive-ml.com>
Date: Mon, 25 May 2026 15:14:27 +0200
Subject: [PATCH 1/2] perf: apply init-and-skip pattern to ciou and diou
 distance

Mirrors the iou_distance_slice refactor from #85: pre-fill result with ONE
and skip on no-overlap instead of writing ONE per-cell. Adds ciou+diou
benchmarks so CodSpeed can measure the win.
---
 powerboxesrs/benches/bench_iou.rs | 40 +++++++++++++++++++++++++++++++
 powerboxesrs/src/ciou.rs          |  3 +--
 powerboxesrs/src/diou.rs          |  3 +--
 3 files changed, 42 insertions(+), 4 deletions(-)

diff --git a/powerboxesrs/benches/bench_iou.rs b/powerboxesrs/benches/bench_iou.rs
index 8cca8fa..fe9911c 100644
--- a/powerboxesrs/benches/bench_iou.rs
+++ b/powerboxesrs/benches/bench_iou.rs
@@ -1,5 +1,7 @@
 use codspeed_criterion_compat::{black_box, criterion_group, criterion_main, Criterion};
 use ndarray::Array2;
+use powerboxesrs::ciou::ciou_distance;
+use powerboxesrs::diou::diou_distance;
 use powerboxesrs::giou::{
     giou_distance, parallel_giou_distance, parallel_rotated_giou_distance, rotated_giou_distance,
 };
@@ -80,6 +82,42 @@ pub fn parallel_giou_distance_benchmark(c: &mut Criterion) {
     });
 }
 
+pub fn ciou_distance_benchmark(c: &mut Criterion) {
+    let mut boxes1 = Array2::<f64>::zeros((100, 4));
+    for i in 0..100 {
+        for j in 0..4 {
+            if j < 2 {
+                boxes1[[i, j]] = 0.0;
+            } else {
+                boxes1[[i, j]] = 10.0;
+            }
+        }
+    }
+    let boxes2 = boxes1.clone();
+
+    c.bench_function("ciou distance benchmark", |b| {
+        b.iter(|| ciou_distance(black_box(&boxes1), black_box(&boxes2)))
+    });
+}
+
+pub fn diou_distance_benchmark(c: &mut Criterion) {
+    let mut boxes1 = Array2::<f64>::zeros((100, 4));
+    for i in 0..100 {
+        for j in 0..4 {
+            if j < 2 {
+                boxes1[[i, j]] = 0.0;
+            } else {
+                boxes1[[i, j]] = 10.0;
+            }
+        }
+    }
+    let boxes2 = boxes1.clone();
+
+    c.bench_function("diou distance benchmark", |b| {
+        b.iter(|| diou_distance(black_box(&boxes1), black_box(&boxes2)))
+    });
+}
+
 /// Generate 100x5 rotated boxes with varied angles for benchmarking
 fn make_rotated_boxes(n: usize) -> Array2<f64> {
     let mut boxes = Array2::<f64>::zeros((n, 5));
@@ -154,6 +192,8 @@ criterion_group!(
     parallel_iou_distance_benchmark,
     giou_distance_benchmark,
     parallel_giou_distance_benchmark,
+    ciou_distance_benchmark,
+    diou_distance_benchmark,
     rotated_iou_distance_benchmark,
     parallel_rotated_iou_distance_benchmark,
     rotated_giou_distance_benchmark,
diff --git a/powerboxesrs/src/ciou.rs b/powerboxesrs/src/ciou.rs
index e702bfc..03560c6 100644
--- a/powerboxesrs/src/ciou.rs
+++ b/powerboxesrs/src/ciou.rs
@@ -28,7 +28,7 @@ where
 {
     let two = N::one() + N::one();
     let four_over_pi_sq: f64 = 4.0 / (std::f64::consts::PI * std::f64::consts::PI);
-    let mut result = vec![0.0f64; n1 * n2];
+    let mut result = vec![utils::ONE; n1 * n2];
     let areas1 = boxes::box_areas_slice(boxes1, n1);
     let areas2 = boxes::box_areas_slice(boxes2, n2);
 
@@ -44,7 +44,6 @@ where
             let x2 = utils::min(a1_x2, a2_x2);
             let y2 = utils::min(a1_y2, a2_y2);
             if x2 < x1 || y2 < y1 {
-                result[i * n2 + j] = utils::ONE;
                 continue;
             }
             let intersection = (x2 - x1) * (y2 - y1);
diff --git a/powerboxesrs/src/diou.rs b/powerboxesrs/src/diou.rs
index f907276..01acf78 100644
--- a/powerboxesrs/src/diou.rs
+++ b/powerboxesrs/src/diou.rs
@@ -24,7 +24,7 @@ where
     N: Num + PartialOrd + ToPrimitive + Float,
 {
     let two = N::one() + N::one();
-    let mut result = vec![0.0f64; n1 * n2];
+    let mut result = vec![utils::ONE; n1 * n2];
     let areas1 = boxes::box_areas_slice(boxes1, n1);
     let areas2 = boxes::box_areas_slice(boxes2, n2);
 
@@ -40,7 +40,6 @@ where
             let x2 = utils::min(a1_x2, a2_x2);
             let y2 = utils::min(a1_y2, a2_y2);
             if x2 < x1 || y2 < y1 {
-                result[i * n2 + j] = utils::ONE;
                 continue;
             }
             let intersection = (x2 - x1) * (y2 - y1);

From c3ec7d11bcf7750ede5f2f0dc6f66c40937b17f9 Mon Sep 17 00:00:00 2001
From: guillaume <guillaume@adaptive-ml.com>
Date: Sat, 13 Jun 2026 17:28:17 +0200
Subject: [PATCH 2/2] refactor: drop redundant zero-init in ciou/diou
 benchmarks

---
 powerboxesrs/benches/bench_iou.rs | 16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/powerboxesrs/benches/bench_iou.rs b/powerboxesrs/benches/bench_iou.rs
index fe9911c..adbb6e9 100644
--- a/powerboxesrs/benches/bench_iou.rs
+++ b/powerboxesrs/benches/bench_iou.rs
@@ -85,12 +85,8 @@ pub fn parallel_giou_distance_benchmark(c: &mut Criterion) {
 pub fn ciou_distance_benchmark(c: &mut Criterion) {
     let mut boxes1 = Array2::<f64>::zeros((100, 4));
     for i in 0..100 {
-        for j in 0..4 {
-            if j < 2 {
-                boxes1[[i, j]] = 0.0;
-            } else {
-                boxes1[[i, j]] = 10.0;
-            }
+        for j in 2..4 {
+            boxes1[[i, j]] = 10.0;
         }
     }
     let boxes2 = boxes1.clone();
@@ -103,12 +99,8 @@ pub fn ciou_distance_benchmark(c: &mut Criterion) {
 pub fn diou_distance_benchmark(c: &mut Criterion) {
     let mut boxes1 = Array2::<f64>::zeros((100, 4));
     for i in 0..100 {
-        for j in 0..4 {
-            if j < 2 {
-                boxes1[[i, j]] = 0.0;
-            } else {
-                boxes1[[i, j]] = 10.0;
-            }
+        for j in 2..4 {
+            boxes1[[i, j]] = 10.0;
         }
     }
     let boxes2 = boxes1.clone();