Merge pull request #174 from gboisse/master

Fixed instance rendering with bvh2

Author: yozhijk

RadeonRays/src/accelerator/bvh2.h

@@ -67,7 +67,7 @@ namespace RadeonRays
 
     protected:
         using RefArray = std::vector<std::uint32_t>;
-        using MetaDataArray = std::vector<std::pair<const Mesh *, std::size_t> >;
+        using MetaDataArray = std::vector<std::pair<const Shape *, std::size_t> >;
 
         // Constant values
         enum Constants
@@ -157,7 +157,7 @@ namespace RadeonRays
         static inline void SetPrimitive(
             Node &node,
             std::uint32_t index,
-            std::pair<const Mesh *, std::size_t> ref);
+            std::pair<const Shape *, std::size_t> ref);
 
         static inline bool IsInternal(const Node &node);
         static inline std::uint32_t GetChildIndex(const Node &node, std::uint8_t idx);
@@ -283,7 +283,7 @@ namespace RadeonRays
                 _mm_store_ps(&aabb_max[current_face].x, pmax);
                 _mm_store_ps(&aabb_centroid[current_face].x, centroid);
 
-                metadata[current_face] = std::make_pair(mesh, face_index);
+                metadata[current_face] = std::make_pair(shape, face_index);
             }
         }
 
@@ -339,11 +339,12 @@ namespace RadeonRays
     void Bvh2::SetPrimitive(
         Node &node,
         std::uint32_t index,
-        std::pair<const Mesh *, std::size_t> ref)
+        std::pair<const Shape *, std::size_t> ref)
     {
-        auto mesh = ref.first;
+        auto shape = ref.first;
         matrix worldmat, worldmatinv;
-        mesh->GetTransform(worldmat, worldmatinv);
+        shape->GetTransform(worldmat, worldmatinv);
+        auto mesh = static_cast<const Mesh *>(static_cast<const ShapeImpl *>(shape)->is_instance() ? static_cast<const Instance *>(shape)->GetBaseShape() : shape);
         auto face = mesh->GetFaceData()[ref.second];
         auto v0 = transform_point(mesh->GetVertexData()[face.idx[0]], worldmat);
         auto v1 = transform_point(mesh->GetVertexData()[face.idx[1]], worldmat);
@@ -357,7 +358,7 @@ namespace RadeonRays
         node.aabb_right_min_or_v2[0] = v2.x;
         node.aabb_right_min_or_v2[1] = v2.y;
         node.aabb_right_min_or_v2[2] = v2.z;
-        node.mesh_id = mesh->GetId();
+        node.mesh_id = shape->GetId();
         node.prim_id = static_cast<std::uint32_t>(ref.second);
     }
 

RadeonRays/src/device/calc_intersection_device.cpp

@@ -126,7 +126,7 @@ namespace RadeonRays
                         m_intersector_string = "fatbvh";
 #else
                         m_intersector.reset(new IntersectorLDS(m_device.get()));
-                        m_intersector_string = "bvh2";
+                        m_intersector_string = "fatbvh";
 #endif
                     }
                 }

RadeonRays/src/kernelcache/kernels_cl.h

@@ -2487,7 +2487,7 @@ static const char g_intersect_bvh2_lds_opencl[]= \
 " \n"\
 "#define GROUP_SIZE 64 \n"\
 "#define STACK_SIZE 32 \n"\
-"#define LDS_STACK_SIZE 8 \n"\
+"#define LDS_STACK_SIZE 16 \n"\
 " \n"\
 "// BVH node \n"\
 "typedef struct \n"\
@@ -2699,13 +2699,10 @@ static const char g_intersect_bvh2_lds_opencl[]= \
 "            const float3 invDir = safe_invdir(my_ray); \n"\
 "            const float3 oxInvDir = -my_ray.o.xyz * invDir; \n"\
 " \n"\
-"            // Intersection parametric distance \n"\
-"            float closest_t = my_ray.o.w; \n"\
-" \n"\
 "            // Current node address \n"\
 "            uint addr = 0; \n"\
-"            // Current closest address \n"\
-"            uint closest_addr = INVALID_ADDR; \n"\
+"            // Intersection parametric distance \n"\
+"            const float closest_t = my_ray.o.w; \n"\
 " \n"\
 "            uint stack_bottom = STACK_SIZE * index; \n"\
 "            uint sptr = stack_bottom; \n"\

RadeonRays/src/kernels/CL/intersect_bvh2_lds.cl

@@ -34,7 +34,7 @@ TYPE DEFINITIONS
 
 #define GROUP_SIZE 64
 #define STACK_SIZE 32
-#define LDS_STACK_SIZE 8
+#define LDS_STACK_SIZE 16
 
 // BVH node
 typedef struct
@@ -246,13 +246,10 @@ KERNEL void occluded_main(
             const float3 invDir = safe_invdir(my_ray);
             const float3 oxInvDir = -my_ray.o.xyz * invDir;
 
-            // Intersection parametric distance
-            float closest_t = my_ray.o.w;
-
             // Current node address
             uint addr = 0;
-            // Current closest address
-            uint closest_addr = INVALID_ADDR;
+            // Intersection parametric distance
+            const float closest_t = my_ray.o.w;
 
             uint stack_bottom = STACK_SIZE * index;
             uint sptr = stack_bottom;