Cycles: explicitly skip self-intersection

Remember the last intersected primitive and skip any intersections with the
same primitive.

Ref D12954

10 files changed, 61 insertions, 5 deletions

diff --git a/intern/cycles/kernel/integrator/intersect_closest.h b/intern/cycles/kernel/integrator/intersect_closest.h
index df710dc1d82..4c5265189fa 100644
--- a/intern/cycles/kernel/integrator/intersect_closest.h
+++ b/intern/cycles/kernel/integrator/intersect_closest.h
@@ -328,6 +328,12 @@ ccl_device void integrator_intersect_closest(KernelGlobals kg,
 
   /* Scene Intersection. */
   Intersection isect ccl_optional_struct_init;
+  isect.object = OBJECT_NONE;
+  isect.prim = PRIM_NONE;
+  ray.self.object = last_isect_object;
+  ray.self.prim = last_isect_prim;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   bool hit = scene_intersect(kg, &ray, visibility, &isect);
 
   /* TODO: remove this and do it in the various intersection functions instead. */
diff --git a/intern/cycles/kernel/integrator/intersect_shadow.h b/intern/cycles/kernel/integrator/intersect_shadow.h
index 90422445fad..1ba8724826b 100644
--- a/intern/cycles/kernel/integrator/intersect_shadow.h
+++ b/intern/cycles/kernel/integrator/intersect_shadow.h
@@ -156,7 +156,10 @@ ccl_device void integrator_intersect_shadow(KernelGlobals kg, IntegratorShadowSt
   /* Read ray from integrator state into local memory. */
   Ray ray ccl_optional_struct_init;
   integrator_state_read_shadow_ray(kg, state, &ray);
-
+  ray.self.object = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 0, object);
+  ray.self.prim = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 0, prim);
+  ray.self.light_object = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 1, object);
+  ray.self.light_prim = INTEGRATOR_STATE_ARRAY(state, shadow_isect, 1, prim);
   /* Compute visibility. */
   const uint visibility = integrate_intersect_shadow_visibility(kg, state);
 
diff --git a/intern/cycles/kernel/integrator/intersect_volume_stack.h b/intern/cycles/kernel/integrator/intersect_volume_stack.h
index 9fa5ff63ad2..aa7be879995 100644
--- a/intern/cycles/kernel/integrator/intersect_volume_stack.h
+++ b/intern/cycles/kernel/integrator/intersect_volume_stack.h
@@ -38,7 +38,10 @@ ccl_device void integrator_volume_stack_update_for_subsurface(KernelGlobals kg,
   Ray volume_ray ccl_optional_struct_init;
   volume_ray.P = from_P;
   volume_ray.D = normalize_len(to_P - from_P, &volume_ray.t);
-
+  volume_ray.self.object = INTEGRATOR_STATE(state, isect, object);
+  volume_ray.self.prim = INTEGRATOR_STATE(state, isect, prim);
+  volume_ray.self.light_object = OBJECT_NONE;
+  volume_ray.self.light_prim = PRIM_NONE;
   /* Store to avoid global fetches on every intersection step. */
   const uint volume_stack_size = kernel_data.volume_stack_size;
 
@@ -91,6 +94,10 @@ ccl_device void integrator_volume_stack_init(KernelGlobals kg, IntegratorState s
    * fewest hits. */
   volume_ray.D = make_float3(0.0f, 0.0f, 1.0f);
   volume_ray.t = FLT_MAX;
+  volume_ray.self.object = OBJECT_NONE;
+  volume_ray.self.prim = PRIM_NONE;
+  volume_ray.self.light_object = OBJECT_NONE;
+  volume_ray.self.light_prim = PRIM_NONE;
 
   int stack_index = 0, enclosed_index = 0;
 
diff --git a/intern/cycles/kernel/integrator/shade_shadow.h b/intern/cycles/kernel/integrator/shade_shadow.h
index a68fcaa7a64..10ec48c8637 100644
--- a/intern/cycles/kernel/integrator/shade_shadow.h
+++ b/intern/cycles/kernel/integrator/shade_shadow.h
@@ -83,7 +83,10 @@ ccl_device_inline void integrate_transparent_volume_shadow(KernelGlobals kg,
   /* Setup shader data. */
   Ray ray ccl_optional_struct_init;
   integrator_state_read_shadow_ray(kg, state, &ray);
-
+  ray.self.object = OBJECT_NONE;
+  ray.self.prim = PRIM_NONE;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   /* Modify ray position and length to match current segment. */
   const float start_t = (hit == 0) ? 0.0f :
                                      INTEGRATOR_STATE_ARRAY(state, shadow_isect, hit - 1, t);
diff --git a/intern/cycles/kernel/integrator/shade_surface.h b/intern/cycles/kernel/integrator/shade_surface.h
index 9f6077e5d66..3ca9e773591 100644
--- a/intern/cycles/kernel/integrator/shade_surface.h
+++ b/intern/cycles/kernel/integrator/shade_surface.h
@@ -182,6 +182,11 @@ ccl_device_forceinline void integrate_surface_direct_light(KernelGlobals kg,
 
   /* 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
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);
@@ -364,6 +369,10 @@ ccl_device_forceinline void integrate_surface_ao(KernelGlobals kg,
   ray.D = ao_D;
   ray.t = kernel_data.integrator.ao_bounces_distance;
   ray.time = sd->time;
+  ray.self.object = sd->object;
+  ray.self.prim = sd->prim;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
   ray.dP = differential_zero_compact();
   ray.dD = differential_zero_compact();
 
@@ -375,6 +384,10 @@ ccl_device_forceinline void integrate_surface_ao(KernelGlobals kg,
 
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);
diff --git a/intern/cycles/kernel/integrator/shade_volume.h b/intern/cycles/kernel/integrator/shade_volume.h
index 00fa256d894..107d5ec1795 100644
--- a/intern/cycles/kernel/integrator/shade_volume.h
+++ b/intern/cycles/kernel/integrator/shade_volume.h
@@ -791,6 +791,10 @@ ccl_device_forceinline void integrate_volume_direct_light(
 
   /* Write shadow ray and associated state to global memory. */
   integrator_state_write_shadow_ray(kg, shadow_state, &ray);
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, object) = ray.self.object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 0, prim) = ray.self.prim;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, object) = ray.self.light_object;
+  INTEGRATOR_STATE_ARRAY_WRITE(shadow_state, shadow_isect, 1, prim) = ray.self.light_prim;
 
   /* Copy state from main path to shadow path. */
   const uint16_t bounce = INTEGRATOR_STATE(state, path, bounce);
@@ -878,6 +882,9 @@ ccl_device_forceinline bool integrate_volume_phase_scatter(
   INTEGRATOR_STATE_WRITE(state, ray, dP) = differential_make_compact(sd->dP);
   INTEGRATOR_STATE_WRITE(state, ray, dD) = differential_make_compact(phase_domega_in);
 #  endif
+  // Save memory by storing last hit prim and object in isect
+  INTEGRATOR_STATE_WRITE(state, isect, prim) = sd->prim;
+  INTEGRATOR_STATE_WRITE(state, isect, object) = sd->object;
 
   /* Update throughput. */
   const float3 throughput = INTEGRATOR_STATE(state, path, throughput);
diff --git a/intern/cycles/kernel/integrator/shadow_state_template.h b/intern/cycles/kernel/integrator/shadow_state_template.h
index 625a429d3db..86fcabdcd82 100644
--- a/intern/cycles/kernel/integrator/shadow_state_template.h
+++ b/intern/cycles/kernel/integrator/shadow_state_template.h
@@ -61,6 +61,7 @@ KERNEL_STRUCT_MEMBER(shadow_ray, packed_float3, D, KERNEL_FEATURE_PATH_TRACING)
 KERNEL_STRUCT_MEMBER(shadow_ray, float, t, KERNEL_FEATURE_PATH_TRACING)
 KERNEL_STRUCT_MEMBER(shadow_ray, float, time, KERNEL_FEATURE_PATH_TRACING)
 KERNEL_STRUCT_MEMBER(shadow_ray, float, dP, KERNEL_FEATURE_PATH_TRACING)
+KERNEL_STRUCT_MEMBER(shadow_ray, int, object, KERNEL_FEATURE_PATH_TRACING)
 KERNEL_STRUCT_END(shadow_ray)
 
 /*********************** Shadow Intersection result **************************/
diff --git a/intern/cycles/kernel/integrator/subsurface.h b/intern/cycles/kernel/integrator/subsurface.h
index 59b0cd2596c..5a09fc51d75 100644
--- a/intern/cycles/kernel/integrator/subsurface.h
+++ b/intern/cycles/kernel/integrator/subsurface.h
@@ -57,7 +57,6 @@ ccl_device int subsurface_bounce(KernelGlobals kg,
 
   /* Pass along object info, reusing isect to save memory. */
   INTEGRATOR_STATE_WRITE(state, subsurface, Ng) = sd->Ng;
-  INTEGRATOR_STATE_WRITE(state, isect, object) = sd->object;
 
   uint32_t path_flag = (INTEGRATOR_STATE(state, path, flag) & ~PATH_RAY_CAMERA) |
                        ((sc->type == CLOSURE_BSSRDF_BURLEY_ID) ? PATH_RAY_SUBSURFACE_DISK :
diff --git a/intern/cycles/kernel/integrator/subsurface_disk.h b/intern/cycles/kernel/integrator/subsurface_disk.h
index cc6f5048cda..f5641d1fa5e 100644
--- a/intern/cycles/kernel/integrator/subsurface_disk.h
+++ b/intern/cycles/kernel/integrator/subsurface_disk.h
@@ -99,6 +99,10 @@ ccl_device_inline bool subsurface_disk(KernelGlobals kg,
   ray.dP = ray_dP;
   ray.dD = differential_zero_compact();
   ray.time = time;
+  ray.self.object = OBJECT_NONE;
+  ray.self.prim = PRIM_NONE;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = OBJECT_NONE;
 
   /* Intersect with the same object. if multiple intersections are found it
    * will use at most BSSRDF_MAX_HITS hits, a random subset of all hits. */
diff --git a/intern/cycles/kernel/integrator/subsurface_random_walk.h b/intern/cycles/kernel/integrator/subsurface_random_walk.h
index 7a8b467e199..43676fccfe5 100644
--- a/intern/cycles/kernel/integrator/subsurface_random_walk.h
+++ b/intern/cycles/kernel/integrator/subsurface_random_walk.h
@@ -195,6 +195,7 @@ ccl_device_inline bool subsurface_random_walk(KernelGlobals kg,
   const float time = INTEGRATOR_STATE(state, ray, time);
   const float3 Ng = INTEGRATOR_STATE(state, subsurface, Ng);
   const int object = INTEGRATOR_STATE(state, isect, object);
+  const int prim = INTEGRATOR_STATE(state, isect, prim);
 
   /* Sample diffuse surface scatter into the object. */
   float3 D;
@@ -211,6 +212,10 @@ ccl_device_inline bool subsurface_random_walk(KernelGlobals kg,
   ray.time = time;
   ray.dP = ray_dP;
   ray.dD = differential_zero_compact();
+  ray.self.object = object;
+  ray.self.prim = prim;
+  ray.self.light_object = OBJECT_NONE;
+  ray.self.light_prim = PRIM_NONE;
 
 #ifndef __KERNEL_GPU_RAYTRACING__
   /* Compute or fetch object transforms. */
@@ -377,7 +382,15 @@ ccl_device_inline bool subsurface_random_walk(KernelGlobals kg,
      * If yes, we will later use backwards guided sampling in order to have a decent
      * chance of connecting to it.
      * TODO: Maybe use less than 10 times the mean free path? */
-    ray.t = (bounce == 0) ? max(t, 10.0f / (min3(sigma_t))) : t;
+    if (bounce == 0) {
+      ray.t = max(t, 10.0f / (min3(sigma_t)));
+    }
+    else {
+      ray.t = t;
+      /* After the first bounce the object can intersect the same surface again */
+      ray.self.object = OBJECT_NONE;
+      ray.self.prim = PRIM_NONE;
+    }
     scene_intersect_local(kg, &ray, &ss_isect, object, NULL, 1);
     hit = (ss_isect.num_hits > 0);