Embree shape support fix performance issue

Author: Speierers

include/mitsuba/render/kdtree.h

@@ -30,7 +30,7 @@
  * Temporary scratch space that is used to cache intersection information
  * (# of floats)
  */
-#define MTS_KD_INTERSECTION_CACHE_SIZE 7
+#define MTS_KD_INTERSECTION_CACHE_SIZE 6
 
 NAMESPACE_BEGIN(mitsuba)
 
@@ -2427,7 +2427,7 @@ class MTS_EXPORT_RENDER ShapeKDTree : public TShapeKDTree<BoundingBox<Point<scal
         else if (ShadowRay)
             hit = shape->ray_test(ray, active);
         else
-            std::tie(hit, t) = shape->ray_intersect(ray, cache + 3, active);
+            std::tie(hit, t) = shape->ray_intersect(ray, cache + 2, active);
 
         if (!ShadowRay && any(hit)) {
             Float shape_index_v = reinterpret_array<Float>(UInt(shape_index));
@@ -2441,16 +2441,13 @@ class MTS_EXPORT_RENDER ShapeKDTree : public TShapeKDTree<BoundingBox<Point<scal
                 masked(cache[1], hit) = prim_index_v;
             }
 
-            // Indicates that all lanes have a valid cache
-            cache[2] = 1.f;
-
             if (is_mesh) {
                 if constexpr (!is_array_v<Float>) {
-                    cache[3] = u;
-                    cache[4] = v;
+                    cache[2] = u;
+                    cache[3] = v;
                 } else {
-                    masked(cache[3], hit) = u;
-                    masked(cache[4], hit) = v;
+                    masked(cache[2], hit) = u;
+                    masked(cache[3], hit) = v;
                 }
             }
         }

include/mitsuba/render/shape.h

@@ -162,9 +162,7 @@ class MTS_EXPORT_RENDER Shape : public Object {
      * uninitialized.
      *
      * \param cache
-     *     Cached information about the previously computed intersection. The
-     *     first entry of the cache indicates which lanes in the entries are
-     *     valid. For invalid lanes, the information needs to be recomputed.
+     *     Cached information about the previously computed intersection.
      */
     virtual void fill_surface_interaction(const Ray3f &ray, const Float *cache,
                                           SurfaceInteraction3f &si, Mask active = true) const;
@@ -348,7 +346,6 @@ NAMESPACE_END(mitsuba)
 ENOKI_CALL_SUPPORT_TEMPLATE_BEGIN(mitsuba::Shape)
     ENOKI_CALL_SUPPORT_METHOD(normal_derivative)
     ENOKI_CALL_SUPPORT_METHOD(fill_surface_interaction)
-    ENOKI_CALL_SUPPORT_METHOD(is_mesh)
     ENOKI_CALL_SUPPORT_GETTER_TYPE(emitter, m_emitter, const typename Class::Emitter *)
     ENOKI_CALL_SUPPORT_GETTER_TYPE(sensor, m_sensor, const typename Class::Sensor *)
     ENOKI_CALL_SUPPORT_GETTER_TYPE(bsdf, m_bsdf, const typename Class::BSDF *)

src/librender/mesh.cpp

@@ -265,10 +265,6 @@ MTS_VARIANT void Mesh<Float, Spectrum>::fill_surface_interaction(const Ray3f & /
                                                                  Mask active) const {
     MTS_MASK_ARGUMENT(active);
 
-    // Only fill surface interaction for lanes with valid cache
-    Mask invalid_cache = neq(*cache++, 1.f);
-    active &= !invalid_cache;
-
     // Barycentric coordinates within triangle
     Float b1 = cache[0],
           b2 = cache[1];

src/librender/scene_embree.inl

@@ -75,7 +75,7 @@ Scene<Float, Spectrum>::ray_intersect_cpu(const Ray3f &ray, Mask active) const {
                 si.prim_index = prim_index;
 
                 // Create the cache for the Mesh shape
-                Float cache[3] = { (si.shape->is_mesh() ? 1.f : 0.f), rh.hit.u, rh.hit.v };
+                Float cache[2] = { rh.hit.u, rh.hit.v };
 
                 // Ask shape to fill in the rest
                 si.shape->fill_surface_interaction(ray, cache, si);
@@ -130,10 +130,7 @@ Scene<Float, Spectrum>::ray_intersect_cpu(const Ray3f &ray, Mask active) const {
                 si.prim_index = prim_index;
 
                 // Create the cache for the Mesh shapes
-                Float cache[3] = {
-                    select(si.shape->is_mesh(), Float(1.f), Float(0.f)),
-                    load<Float>(rh.hit.u), load<Float>(rh.hit.v)
-                };
+                Float cache[2] = { load<Float>(rh.hit.u), load<Float>(rh.hit.v) };
 
                 // Ask shape(s) to fill in the rest
                 si.shape->fill_surface_interaction(ray, cache, si, active);

src/librender/shape.cpp

@@ -276,9 +276,7 @@ Shape<Float, Spectrum>::ray_intersect(const Ray3f &ray, Mask active) const {
 
     SurfaceInteraction3f si = zero<SurfaceInteraction3f>();
     Float cache[MTS_KD_INTERSECTION_CACHE_SIZE];
-    cache[0] = Float(1.f); // Indicates that all lanes have a valid cache
-
-    auto [success, t] = ray_intersect(ray, cache + 1, active);
+    auto [success, t] = ray_intersect(ray, cache, active);
     active &= success;
     si.t = select(active, t,  math::Infinity<Float>);
 

src/shapes/disk.cpp

@@ -182,17 +182,16 @@ class Disk final : public Shape<Float, Spectrum> {
                                   SurfaceInteraction3f &si_out, Mask active) const override {
         MTS_MASK_ARGUMENT(active);
 
-        // Load and/or recompute cache if necessary
-        Mask invalid_cache = neq(*cache++, 1.f);
+#if !defined(MTS_ENABLE_EMBREE)
         Float local_x = cache[0];
         Float local_y = cache[1];
-        if (any(invalid_cache)) {
-            Ray3f ray_    = m_world_to_object.transform_affine(ray);
-            Float t       = -ray_.o.z() / ray_.d.z();
-            Point3f local = ray_(t);
-            masked(local_x, invalid_cache) = local.x();
-            masked(local_y, invalid_cache) = local.y();
-        }
+#else
+        Ray3f ray_    = m_world_to_object.transform_affine(ray);
+        Float t       = -ray_.o.z() / ray_.d.z();
+        Point3f local = ray_(t);
+        Float local_x = local.x();
+        Float local_y = local.y();
+#endif
 
         SurfaceInteraction3f si(si_out);
 

src/shapes/rectangle.cpp

@@ -212,17 +212,16 @@ class Rectangle final : public Shape<Float, Spectrum> {
                                   SurfaceInteraction3f &si_out, Mask active) const override {
         MTS_MASK_ARGUMENT(active);
 
-        // Load and/or recompute cache if necessary
-        Mask invalid_cache = neq(*cache++, 1.f);
+#if !defined(MTS_ENABLE_EMBREE)
         Float local_x = cache[0];
         Float local_y = cache[1];
-        if (any(invalid_cache)) {
-            Ray3f ray     = m_world_to_object.transform_affine(ray_);
-            Float t       = -ray.o.z() * ray.d_rcp.z();
-            Point3f local = ray(t);
-            masked(local_x, invalid_cache) = local.x();
-            masked(local_y, invalid_cache) = local.y();
-        }
+#else
+        Ray3f ray     = m_world_to_object.transform_affine(ray_);
+        Float t       = -ray.o.z() * ray.d_rcp.z();
+        Point3f local = ray(t);
+        Float local_x = local.x();
+        Float local_y = local.y();
+#endif
 
         SurfaceInteraction3f si(si_out);