diff --git a/src/traversal/intersection.impala b/src/traversal/intersection.impala
index 0dd1769..2a938f7 100644
--- a/src/traversal/intersection.impala
+++ b/src/traversal/intersection.impala
@@ -142,7 +142,7 @@ fn @bbox_half_area(a: BBox) -> f32 {
     ext.x * (ext.y + ext.z) + ext.y * ext.z
 }
 
-fn @intersect_ray_tri(math: Intrinsics, backface_culling: bool, ray: Ray, tri: Tri) -> Option[(f32, f32, f32)] {
+fn @intersect_ray_tri_cpu(math: Intrinsics, backface_culling: bool, ray: Ray, tri: Tri) -> Option[(f32, f32, f32)] {
     let c = vec3_sub(tri.v0, ray.org);
     let r = vec3_cross(ray.dir, c);
     let det = vec3_dot(tri.n, ray.dir);
@@ -172,6 +172,34 @@ fn @intersect_ray_tri(math: Intrinsics, backface_culling: bool, ray: Ray, tri: T
     }
 }
 
+fn @intersect_ray_tri_gpu(math: Intrinsics, backface_culling: bool, ray: Ray, tri: Tri) -> Option[(f32, f32, f32)] {
+    let c = vec3_sub(tri.v0, ray.org);
+    let r = vec3_cross(ray.dir, c);
+    let det = vec3_dot(tri.n, ray.dir);
+    let abs_det = math.fabsf(det);
+
+    let u = prodsign(vec3_dot(r, tri.e2), det);
+    let mut mask = u >= 0;
+
+    let v = prodsign(vec3_dot(r, tri.e1), det);
+    mask &= v >= 0;
+    mask &= u + v <= abs_det;
+
+    if likely(!mask) { return(Option[(f32, f32, f32)]::None) }
+
+    let t = prodsign(vec3_dot(c, tri.n), det);
+    mask &= if backface_culling { det < 0 } else { abs_det != 0 };
+    mask &= t >= abs_det * ray.tmin;
+    mask &= t <= abs_det * ray.tmax;
+
+    if mask {
+        let inv_det = 1 / abs_det;
+        make_option(t * inv_det, u * inv_det, v * inv_det)
+    } else {
+        Option[(f32, f32, f32)]::None
+    }
+}
+
 fn @intersect_ray_box(min_max: MinMax, ordered: bool, ray: Ray, bbox: BBox) -> (f32, f32) {
     let t0 = vec3_add(vec3_mul(ray.inv_dir, bbox.min), ray.inv_org);
     let t1 = vec3_add(vec3_mul(ray.inv_dir, bbox.max), ray.inv_org);
diff --git a/src/traversal/mapping_cpu.impala b/src/traversal/mapping_cpu.impala
index c11da16..c1e55a8 100644
--- a/src/traversal/mapping_cpu.impala
+++ b/src/traversal/mapping_cpu.impala
@@ -30,7 +30,7 @@ fn @make_cpu_tri4(tris: &[Tri4]) -> fn (i32) -> Prim {
             let e2  = make_vec3(tri_ptr.e2(0)(i), tri_ptr.e2(1)(i), tri_ptr.e2(2)(i));
             let n   = make_vec3(tri_ptr.n (0)(i), tri_ptr.n (1)(i), tri_ptr.n (2)(i));
             let tri = make_tri(v0, e1, e2, n);
-            match intersect_ray_tri(math, false /*backface_culling*/, ray, tri) {
+            match intersect_ray_tri_cpu(math, false /*backface_culling*/, ray, tri) {
                 Option[(f32, f32, f32)]::Some(t, u, v) => {
                     let prim_id = tri_ptr.prim_id(i) & 0x7FFFFFFF;
                     let geom_id = tri_ptr.geom_id(i);
diff --git a/src/traversal/mapping_gpu.impala b/src/traversal/mapping_gpu.impala
index bd4acb8..d913250 100644
--- a/src/traversal/mapping_gpu.impala
+++ b/src/traversal/mapping_gpu.impala
@@ -56,7 +56,7 @@ fn @make_gpu_bvh2_tri1(nodes: &[Node2], tris: &[Tri1], is_nvvm: bool) -> Bvh {
                     let e2  = make_vec3(tri2(0), tri2(1), tri2(2));
                     let n   = vec3_cross(e1, e2);
                     let tri = make_tri(v0, e1, e2, n);
-                    match intersect_ray_tri(math, false /*backface_culling*/, ray, tri) {
+                    match intersect_ray_tri_gpu(math, false /*backface_culling*/, ray, tri) {
                         Option[(f32, f32, f32)]::Some((t, u, v)) => {
                             make_option(make_hit(geom_id, prim_id & 0x7FFFFFFF, t, make_vec2(u, v)))
                         },