optimized to use all pointcloud types

Cargo.toml

@@ -3,8 +3,16 @@ name = "sherpa_ros2_pointcloud_downsampling"
 version = "0.1.0"
 edition = "2024"
 
+[profile.dev]
+opt-level = 1
+
+[profile.release]
+opt-level = 3
+lto = true            # Link Time Optimization (makes final binary faster)
+codegen-units = 1     # Better for speed (slower compile)
+
 [dependencies]
 byteorder = "1.5.0"
 futures = "0.3.31"
 itertools = "0.14.0"
-r2r = "0.9.5"
+r2r = "0.9.5"

src/main.rs

@@ -1,47 +1,83 @@
-use byteorder::{LittleEndian, BigEndian, ByteOrder};
+use byteorder::{BigEndian, ByteOrder, LittleEndian};
 use futures::{executor::LocalPool, future, stream::StreamExt, task::LocalSpawnExt};
 use itertools::Itertools;
 use r2r;
-use r2r::sensor_msgs::msg::{PointField, PointCloud2};
+use r2r::sensor_msgs::msg::{PointCloud2, PointField};
 use std::f32::consts::PI;
 use std::ops::{Add, Sub};
 use std::time::{Duration, Instant};
 
+trait GetOffset {
+    fn get_offset(&self, field: &str) -> Option<usize>;
+}
+
+impl GetOffset for Vec<PointField> {
+    fn get_offset(&self, field: &str) -> Option<usize> {
+        for point_field in self {
+            if field == point_field.name {
+                return Some(point_field.offset as usize);
+            }
+        }
+        None
+    }
+}
+
 trait Downsample {
     fn downsample(self) -> Self;
 }
 
 impl Downsample for PointCloud2 {
     fn downsample(self) -> Self {
-        let data: Vec<u8> = self.data
-            .chunks_exact(self.point_step as usize)
-            .map(|chunk| PointF32::from_bytes(chunk))
-            .filter(|point| point.is_valid())
-            .tuple_windows::<(_, _, _)>()
-            .filter(|(last, current, next)| {
-                let v1: VectorF32 = last.vec_to(current);
-                let v2: VectorF32 = current.vec_to(next);
+        let offset_x: usize = self.fields.get_offset("x").unwrap();
+        let offset_y: usize = self.fields.get_offset("y").unwrap();
+        let offset_z: usize = self.fields.get_offset("z").unwrap();
+
+        let mut data: Vec<u8> = Vec::with_capacity(self.data.len());
+        self.data
+            .chunks_exact(self.point_step as usize)
+            .enumerate()
+            .map(|(point_nr, chunk)| {
+                (
+                    point_nr,
+                    match self.is_bigendian {
+                        true => VectorF32::new(
+                            BigEndian::read_f32(&chunk[offset_x..offset_x + 4]),
+                            BigEndian::read_f32(&chunk[offset_y..offset_y + 4]),
+                            BigEndian::read_f32(&chunk[offset_z..offset_z + 4]),
+                        ),
+                        false => VectorF32::new(
+                            LittleEndian::read_f32(&chunk[offset_x..offset_x + 4]),
+                            LittleEndian::read_f32(&chunk[offset_y..offset_y + 4]),
+                            LittleEndian::read_f32(&chunk[offset_z..offset_z + 4]),
+                        ),
+                    },
+                )
+            })
+            .filter(|(_, point)| point.is_valid())
+            .tuple_windows::<(_, _, _)>()
+            .filter(|((_, last), (_, current), (_, next))| {
+                let v1: VectorF32 = *last - *current;
+                let v2: VectorF32 = *current - *next;
+
+                let angle_deg: f32 =
+                    (v1.dot(&v2).abs() / (v1.abs() * v2.abs())).acos() * 180f32 / PI;
 
-                let angle_deg: f32 = (v1.dot(&v2).abs() / (v1.abs() * v2.abs())).acos() * 180f32 / PI;
-                
                 angle_deg > 30f32
             })
-            .flat_map(|(_, point, _)| point.to_bytes_vec())
-            .collect();
+            .for_each(|(_, (point_nr, _), _)| {
+                let point_idx: usize = point_nr * self.point_step as usize;
+
+                data.extend_from_slice(&self.data[point_idx..point_idx + self.point_step as usize])
+            });
 
         PointCloud2 {
             header: self.header,
             height: 1,
-            width: data.len() as u32 / 16,
-            fields: vec![
-                PointField { name: "x".to_string(), offset: 0, datatype: 7, count: 1 },
-                PointField { name: "y".to_string(), offset: 4, datatype: 7, count: 1 },
-                PointField { name: "z".to_string(), offset: 8, datatype: 7, count: 1 },
-                PointField { name: "intensity".to_string(), offset: 12, datatype: 7, count: 1 },
-            ],
+            width: data.len() as u32 / self.point_step,
+            fields: self.fields,
             is_bigendian: false,
-            point_step: 16,
-            row_step: 16,
+            point_step: self.point_step,
+            row_step: self.point_step,
             data,
             is_dense: true,
         }
@@ -55,43 +91,11 @@ struct VectorF32 {
     z: f32,
 }
 
-#[derive(Debug, Clone, Copy)]
-struct PointF32 {
-    position: VectorF32,
-    intensity: f32, 
-}
-
-impl PointF32 {
-    fn from_bytes(chunk: &[u8]) -> Self {
-        Self {
-            position: VectorF32 {
-                x: LittleEndian::read_f32(&chunk[0..4]),
-                y: LittleEndian::read_f32(&chunk[4..8]),
-                z: LittleEndian::read_f32(&chunk[8..12]),
-            },
-            intensity: LittleEndian::read_f32(&chunk[12..16]),
-        }
-    }
-
-    fn to_bytes_vec(&self) -> Vec<u8> {
-        let mut buf = [0u8; 16];
-        buf[0..4].copy_from_slice(&self.position.x.to_le_bytes());
-        buf[4..8].copy_from_slice(&self.position.y.to_le_bytes());
-        buf[8..12].copy_from_slice(&self.position.z.to_le_bytes());
-        buf[12..16].copy_from_slice(&self.intensity.to_le_bytes());
-        buf.to_vec()
-    }
-
-    fn is_valid(&self) -> bool {
-        self.position.is_valid() && !self.intensity.is_nan()
-    }
-
-    fn vec_to(&self, target: &Self) -> VectorF32 {
-        target.position - self.position
-    }
-}
-
 impl VectorF32 {
+    fn new(x: f32, y: f32, z: f32) -> Self {
+        Self { x, y, z }
+    }
+
     fn is_valid(&self) -> bool {
         !(self.x.is_nan() || self.y.is_nan() || self.z.is_nan())
     }
@@ -110,7 +114,7 @@ impl Add for VectorF32 {
 
     fn add(self, rhs: Self) -> Self::Output {
         Self::Output {
-            x: self.x + rhs.x, 
+            x: self.x + rhs.x,
             y: self.y + rhs.y,
             z: self.z + rhs.z,
         }
@@ -122,41 +126,44 @@ impl Sub for VectorF32 {
 
     fn sub(self, rhs: Self) -> Self::Output {
         Self::Output {
-            x: self.x - rhs.x, 
+            x: self.x - rhs.x,
             y: self.y - rhs.y,
             z: self.z - rhs.z,
         }
     }
 }
 
-
 fn main() {
     let ctx = r2r::Context::create().unwrap();
     let mut node = r2r::Node::create(ctx, "pointcloud_downsampling_node", "").unwrap();
-    let subscriber = node.subscribe::<PointCloud2>("/velodyne_points", r2r::QosProfile::default()).unwrap();
-    let publisher = node.create_publisher::<PointCloud2>("/filtered_points", r2r::QosProfile::default()).unwrap();
-
+    let subscriber = node
+        .subscribe::<PointCloud2>("/velodyne_points", r2r::QosProfile::default())
+        .unwrap();
+    let publisher = node
+        .create_publisher::<PointCloud2>("/filtered_points", r2r::QosProfile::default())
+        .unwrap();
 
     // Set up a simple task executor.
     let mut pool = LocalPool::new();
     let spawner = pool.spawner();
 
-    spawner.spawn_local(async move {
-        subscriber
-            .for_each(|msg| {
+    spawner
+        .spawn_local(async move {
+            subscriber
+                .for_each(|msg| {
+                    let start = Instant::now();
+                    let downsampled = msg.downsample();
+                    println!("Duration: {} us", start.elapsed().as_micros());
+                    publisher.publish(&downsampled);
 
-                let start = Instant::now();
-                let downsampled = msg.downsample();
-                println!("Duration: {} ms", start.elapsed().as_millis());
-                publisher.publish(&downsampled);
-
-                future::ready(())
-            })
-            .await
-    }).unwrap();
+                    future::ready(())
+                })
+                .await
+        })
+        .unwrap();
 
     loop {
         node.spin_once(Duration::from_millis(10));
         pool.run_until_stalled();
     }
-}
+}