diff --git a/scene/src/lib.rs b/scene/src/lib.rs
index b330dd0..7b950db 100644
--- a/scene/src/lib.rs
+++ b/scene/src/lib.rs
@@ -18,6 +18,7 @@ struct TriangleNormals {
 }
 
 impl TriangleNormals {
+    #[inline]
     fn lerp(&self, u: f32, v: f32) -> Vec3 {
         ((1.0 - (u + v)) * self.n0 + u * self.n1 + v * self.n2).normalize()
     }
@@ -31,6 +32,7 @@ struct TriangleTexcoords {
 }
 
 impl TriangleTexcoords {
+    #[inline]
     fn lerp(&self, u: f32, v: f32) -> Vec2 {
         (1.0 - (u + v)) * self.uv0 + u * self.uv1 + v * self.uv2
     }
diff --git a/tracing/src/pathtracer.rs b/tracing/src/pathtracer.rs
index c40c967..ca32d9f 100644
--- a/tracing/src/pathtracer.rs
+++ b/tracing/src/pathtracer.rs
@@ -1,5 +1,3 @@
-use std::ops::RangeInclusive;
-
 use crate::{
     image_buffer::ImageBuffer,
     material::{material_brdf, material_sample, IncomingRay, OutgoingRay},
@@ -8,7 +6,7 @@ use crate::{
 };
 use geometry::ray::Ray;
 use glam::{UVec2, Vec3};
-use kdtree::{intersection::KdIntersection, KdNode};
+use kdtree::KdNode;
 use rand::rngs::SmallRng;
 use scene::{camera::Pinhole, Scene};
 
@@ -19,14 +17,6 @@ pub struct Pathtracer {
 }
 
 impl Pathtracer {
-    fn intersect(&self, ray: &Ray, t_range: RangeInclusive<f32>) -> Option<KdIntersection> {
-        self.kdtree.intersect(&self.scene.geometries, ray, t_range)
-    }
-
-    fn intersect_any(&self, ray: &Ray, t_range: RangeInclusive<f32>) -> bool {
-        self.intersect(ray, t_range).is_some()
-    }
-
     fn trace_ray(
         &self,
         mut ray_logger: RayLoggerWithIterationAndPixel,
@@ -40,7 +30,9 @@ impl Pathtracer {
             return accumulated_radiance;
         }
 
-        let intersection = self.intersect(ray, 0.0..=f32::MAX);
+        let intersection = self
+            .kdtree
+            .intersect(&self.scene.geometries, ray, 0.0..=f32::MAX);
         if intersection.is_none() {
             ray_logger.log_infinite(ray, accumulated_bounces).unwrap();
             return accumulated_radiance + accumulated_transport * self.scene.environment;
@@ -69,7 +61,10 @@ impl Pathtracer {
             .iter()
             .map(|light| {
                 let shadow_ray = Ray::between(point_above, sample_light(light, rng));
-                if self.intersect_any(&shadow_ray, 0.0..=1.0) {
+                let intersection =
+                    self.kdtree
+                        .intersect(&self.scene.geometries, &shadow_ray, 0.0..=1.0);
+                if intersection.is_some() {
                     return Vec3::ZERO;
                 }
                 let wo = shadow_ray.direction.normalize();
@@ -89,27 +84,24 @@ impl Pathtracer {
             &IncomingRay { wi, n, uv },
             rng,
         );
-        let next_ray = Ray::new(
-            if sample.wo.dot(n) >= 0.0 {
-                point_above
-            } else {
-                point_below
-            },
-            sample.wo,
-        );
-
         if sample.pdf <= 0.01 {
             return accumulated_radiance;
         }
 
-        let cosine_term = sample.wo.dot(n).abs();
+        let cosine_term = sample.wo.dot(n);
         let accumulated_transport =
-            accumulated_transport * sample.brdf * (cosine_term / sample.pdf);
-
-        if accumulated_transport.length() <= 0.01 {
+            accumulated_transport * sample.brdf * (cosine_term.abs() / sample.pdf);
+        if accumulated_transport.length_squared() <= 1.0e-4 {
             return accumulated_radiance;
         }
 
+        let next_start_point = if cosine_term >= 0.0 {
+            point_above
+        } else {
+            point_below
+        };
+        let next_ray = Ray::new(next_start_point, sample.wo);
+
         self.trace_ray(
             ray_logger,
             rng,
diff --git a/tracing/src/sampling.rs b/tracing/src/sampling.rs
index 2645181..dc17232 100644
--- a/tracing/src/sampling.rs
+++ b/tracing/src/sampling.rs
@@ -10,8 +10,9 @@ pub fn uniform_sample_unit_sphere(rng: &mut SmallRng) -> Vec3 {
     let z = rng.gen_range(-1.0..1.0);
     let a = rng.gen_range(0.0..std::f32::consts::TAU);
     let r = (1.0f32 - z * z).sqrt();
-    let x = r * a.cos();
-    let y = r * a.sin();
+    let (a_sin, a_cos) = a.sin_cos();
+    let x = r * a_cos;
+    let y = r * a_sin;
     Vec3::new(x, y, z)
 }
 
@@ -24,8 +25,8 @@ pub fn uniform_sample_unit_sphere(rng: &mut SmallRng) -> Vec3 {
 // }
 
 fn concentric_sample_unit_disk(rng: &mut SmallRng) -> Vec2 {
-    let x = rng.gen_range(-1.0..=1.0);
-    let y = rng.gen_range(-1.0..=1.0);
+    let x = rng.gen_range(-1.0f32..=1.0f32);
+    let y = rng.gen_range(-1.0f32..=1.0f32);
     if x == 0.0 && y == 0.0 {
         return Vec2::ZERO;
     }
@@ -37,8 +38,8 @@ fn concentric_sample_unit_disk(rng: &mut SmallRng) -> Vec2 {
         (x, y) => (-y, 6.0 - x / y),
     };
 
-    let theta = theta * std::f32::consts::FRAC_PI_4;
-    Vec2::new(r * theta.cos(), r * theta.sin())
+    let (theta_sin, theta_cos) = (theta * std::f32::consts::FRAC_PI_4).sin_cos();
+    Vec2::new(r * theta_cos, r * theta_sin)
 }
 
 pub fn cosine_sample_hemisphere(rng: &mut SmallRng) -> Vec3 {