Cycles: improve ray tracing precision near triangle edges

Detect cases where a ray-intersection would miss the current triangle, which if
the intersection is strictly watertight, implies that a neighboring triangle would
incorrectly be hit instead.

When that is detected, apply a ray-offset. The idea being that we only want to
introduce potential error from ray offsets if we really need to.

This work for BVH2 and Embree, as we are able to match the ray-interesction
bit-for-bit, though doing so for Embree requires ugly hacks. Tiny differences
like fused-multiply-add or dot product intrinstics in matrix inversion and ray
intersection needed to be matched exactly, so this is fragile.

Unfortunately we're not able to do the same for OptiX or MetalRT, since those
implementations are unknown (and possibly impossible to match as hardware
instructions). Still artifacts are much reduced, though not eliminated.

Ref T97259

Differential Revision: https://developer.blender.org/D15559

2 files changed, 80 insertions, 2 deletions

diff --git a/intern/cycles/kernel/bvh/util.h b/intern/cycles/kernel/bvh/util.h
index b67c9394bea..a57703a8b8c 100644
--- a/intern/cycles/kernel/bvh/util.h
+++ b/intern/cycles/kernel/bvh/util.h
@@ -33,6 +33,30 @@ ccl_device_forceinline float intersection_t_offset(const float t)
   return __uint_as_float(bits);
 }
 
+/* Ray offset to avoid self intersection.
+ *
+ * This function can be used to compute a modified ray start position for rays
+ * leaving from a surface. This is from:
+ * "A Fast and Robust Method for Avoiding Self-Intersection"
+ * Ray Tracing Gems, chapter 6.
+ */
+ccl_device_inline float3 ray_offset(const float3 P, const float3 Ng)
+{
+  const float int_scale = 256.0f;
+  const int3 of_i = make_int3(
+      (int)(int_scale * Ng.x), (int)(int_scale * Ng.y), (int)(int_scale * Ng.z));
+
+  const float3 p_i = make_float3(
+      __int_as_float(__float_as_int(P.x) + ((P.x < 0) ? -of_i.x : of_i.x)),
+      __int_as_float(__float_as_int(P.y) + ((P.y < 0) ? -of_i.y : of_i.y)),
+      __int_as_float(__float_as_int(P.z) + ((P.z < 0) ? -of_i.z : of_i.z)));
+  const float origin = 1.0f / 32.0f;
+  const float float_scale = 1.0f / 65536.0f;
+  return make_float3(fabsf(P.x) < origin ? P.x + float_scale * Ng.x : p_i.x,
+                     fabsf(P.y) < origin ? P.y + float_scale * Ng.y : p_i.y,
+                     fabsf(P.z) < origin ? P.z + float_scale * Ng.z : p_i.z);
+}
+
 #ifndef __KERNEL_GPU__
 ccl_device int intersections_compare(const void *a, const void *b)
 {
diff --git a/intern/cycles/kernel/integrator/shade_surface.h b/intern/cycles/kernel/integrator/shade_surface.h
index 70b20a93b6a..19b8946e865 100644
--- a/intern/cycles/kernel/integrator/shade_surface.h
+++ b/intern/cycles/kernel/integrator/shade_surface.h
@@ -31,6 +31,52 @@ ccl_device_forceinline void integrate_surface_shader_setup(KernelGlobals kg,
   shader_setup_from_ray(kg, sd, &ray, &isect);
 }
 
+ccl_device_forceinline float3 integrate_surface_ray_offset(KernelGlobals kg,
+                                                           const ccl_private ShaderData *sd,
+                                                           const float3 ray_P,
+                                                           const float3 ray_D)
+{
+  /* No ray offset needed for other primitive types. */
+  if (!(sd->type & PRIMITIVE_TRIANGLE)) {
+    return ray_P;
+  }
+
+  /* Self intersection tests already account for the case where a ray hits the
+   * same primitive. However precision issues can still cause neighboring
+   * triangles to be hit. Here we test if the ray-triangle intersection with
+   * the same primitive would miss, implying that a neighbouring triangle would
+   * be hit instead.
+   *
+   * This relies on triangle intersection to be watertight, and the object inverse
+   * object transform to match the one used by ray intersection exactly.
+   *
+   * Potential improvements:
+   * - It appears this happens when either barycentric coordinates are small,
+   *   or dot(sd->Ng, ray_D)  is small. Detect such cases and skip test?
+   * - Instead of ray offset, can we tweak P to lie within the triangle?
+   */
+  const uint tri_vindex = kernel_data_fetch(tri_vindex, sd->prim).w;
+  const packed_float3 tri_a = kernel_data_fetch(tri_verts, tri_vindex + 0),
+                      tri_b = kernel_data_fetch(tri_verts, tri_vindex + 1),
+                      tri_c = kernel_data_fetch(tri_verts, tri_vindex + 2);
+
+  float3 local_ray_P = ray_P;
+  float3 local_ray_D = ray_D;
+
+  if (!(sd->object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
+    const Transform itfm = object_get_inverse_transform(kg, sd);
+    local_ray_P = transform_point(&itfm, local_ray_P);
+    local_ray_D = transform_direction(&itfm, local_ray_D);
+  }
+
+  if (ray_triangle_intersect_self(local_ray_P, local_ray_D, tri_a, tri_b, tri_c)) {
+    return ray_P;
+  }
+  else {
+    return ray_offset(ray_P, sd->Ng);
+  }
+}
+
 #ifdef __HOLDOUT__
 ccl_device_forceinline bool integrate_surface_holdout(KernelGlobals kg,
                                                       ConstIntegratorState state,
@@ -200,6 +246,10 @@ ccl_device_forceinline void integrate_surface_direct_light(KernelGlobals kg,
 #  endif
   }
 
+  if (ray.self.object != OBJECT_NONE) {
+    ray.P = integrate_surface_ray_offset(kg, sd, ray.P, ray.D);
+  }
+
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
   // Save memory by storing the light and object indices in the shadow_isect
@@ -327,8 +377,9 @@ ccl_device_forceinline int integrate_surface_bsdf_bssrdf_bounce(
   }
   else {
     /* Setup ray with changed origin and direction. */
-    INTEGRATOR_STATE_WRITE(state, ray, P) = sd->P;
-    INTEGRATOR_STATE_WRITE(state, ray, D) = normalize(bsdf_omega_in);
+    const float3 D = normalize(bsdf_omega_in);
+    INTEGRATOR_STATE_WRITE(state, ray, P) = integrate_surface_ray_offset(kg, sd, sd->P, D);
+    INTEGRATOR_STATE_WRITE(state, ray, D) = D;
     INTEGRATOR_STATE_WRITE(state, ray, tmin) = 0.0f;
     INTEGRATOR_STATE_WRITE(state, ray, tmax) = FLT_MAX;
 #ifdef __RAY_DIFFERENTIALS__
@@ -422,6 +473,9 @@ ccl_device_forceinline void integrate_surface_ao(KernelGlobals kg,
   Ray ray ccl_optional_struct_init;
   ray.P = shadow_ray_offset(kg, sd, ao_D, &skip_self);
   ray.D = ao_D;
+  if (skip_self) {
+    ray.P = integrate_surface_ray_offset(kg, sd, ray.P, ray.D);
+  }
   ray.tmin = 0.0f;
   ray.tmax = kernel_data.integrator.ao_bounces_distance;
   ray.time = sd->time;