[WIP] use C++11 alignas for struct alignment

.gitignore

@@ -9,4 +9,5 @@ kernels/config.h
 kernels/hash.h
 include/embree4/rtcore_config.h
 build
+build*
 out

common/math/vec2fa.h

@@ -18,10 +18,8 @@ namespace embree
   /// SSE Vec2fa Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec2fa
+  struct alignas(16) Vec2fa
   {
-    ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 2 };
     union {

common/math/vec2fa_sycl.h

@@ -15,10 +15,8 @@ namespace embree
   /// SSE Vec2fa Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec2fa
+  struct alignas(16) Vec2fa
   {
-    //ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 2 };
     struct { float x,y; };

common/math/vec3ba.h

@@ -18,10 +18,8 @@ namespace embree
   /// SSE Vec3ba Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3ba
+  struct alignas(16) Vec3ba
   {
-    ALIGNED_STRUCT_(16);
-
     union {
       __m128 m128;
       struct { int x,y,z; };

common/math/vec3ba_sycl.h

@@ -13,10 +13,8 @@ namespace embree
   /// SSE Vec3ba Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3ba
+  struct alignas(16) Vec3ba
   {
-    //ALIGNED_STRUCT_(16);
-
     struct { bool x,y,z; };
 
     typedef bool Scalar;

common/math/vec3fa.h

@@ -18,10 +18,8 @@ namespace embree
   /// SSE Vec3fa Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3fa
+  struct alignas(16) Vec3fa
   {
-    ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 3 };
     union {
@@ -421,10 +419,8 @@ namespace embree
   /// SSE Vec3fx Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3fx
+  struct alignas(16) Vec3fx
   {
-    ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 3 };
     union {

common/math/vec3fa_sycl.h

@@ -13,10 +13,8 @@ namespace embree
   /// SSE Vec3fa Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3fa
+  struct alignas(16) Vec3fa
   {
-    //ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 3 };
     struct { float x,y,z, do_not_use; };
@@ -303,10 +301,8 @@ namespace embree
   /// SSE Vec3fx Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3fx
+  struct alignas(16) Vec3fx
   {
-    //ALIGNED_STRUCT_(16);
-
     typedef float Scalar;
     enum { N = 3 };
     struct { float x,y,z; union { int a; unsigned u; float w; }; };

common/math/vec3ia.h

@@ -18,10 +18,8 @@ namespace embree
   /// SSE Vec3ia Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3ia
+  struct alignas(16) Vec3ia
   {
-    ALIGNED_STRUCT_(16);
-
     union {
       __m128i m128;
       struct { int x,y,z; };

common/math/vec3ia_sycl.h

@@ -13,10 +13,8 @@ namespace embree
   /// SSE Vec3ia Type
   ////////////////////////////////////////////////////////////////////////////////
 
-  struct __aligned(16) Vec3ia
+  struct alignas(16) Vec3ia
   {
-    ALIGNED_STRUCT_(16);
-
     struct { int x,y,z; };
 
     typedef int Scalar;

common/tasking/taskschedulerinternal.h

@@ -23,9 +23,8 @@ namespace embree
      hide is also in the API namespace when requested. */
   RTC_NAMESPACE_BEGIN
 
-  struct TaskScheduler : public RefCount
+  struct alignas(64) TaskScheduler : public RefCount
   {
-    ALIGNED_STRUCT_(64);
     friend class Device;
 
     static const size_t TASK_STACK_SIZE = 4*1024;           //!< task structure stack
@@ -172,10 +171,8 @@ namespace embree
     };
 
     /*! thread local structure for each thread */
-    struct Thread
+    struct alignas(64) Thread
     {
-      ALIGNED_STRUCT_(64);
-
       Thread (size_t threadIndex, const Ref<TaskScheduler>& scheduler)
       : threadIndex(threadIndex), task(nullptr), scheduler(scheduler) {}
 

kernels/bvh/bvh_builder_subdiv.cpp

@@ -28,10 +28,8 @@ namespace embree
     typedef FastAllocator::CachedAllocator Allocator;
 
     template<int N>
-    struct BVHNSubdivPatch1BuilderSAH : public Builder
+    struct alignas(64) BVHNSubdivPatch1BuilderSAH : public Builder
     {
-      ALIGNED_STRUCT_(64);
-
       typedef BVHN<N> BVH;
       typedef typename BVH::NodeRef NodeRef;
 
@@ -211,10 +209,8 @@ namespace embree
     };
 
     template<int N>
-    struct BVHNSubdivPatch1MBlurBuilderSAH : public Builder
+    struct alignas(64) BVHNSubdivPatch1MBlurBuilderSAH : public Builder
     {
-      ALIGNED_STRUCT_(64);
-
       typedef BVHN<N> BVH;
       typedef typename BVH::NodeRef NodeRef;
       typedef typename BVH::NodeRecordMB4D NodeRecordMB4D;

kernels/common/buffer.h

@@ -35,7 +35,6 @@ namespace embree
       device->refInc();
 
       ptr = alloc(ptr_in, shared, EmbreeMemoryType::SHARED);
-
 #if defined(EMBREE_SYCL_SUPPORT)
       dshared = true;
       dptr = ptr;
@@ -57,14 +56,11 @@ namespace embree
 #endif
       {
         ptr = alloc(ptr_in, shared, EmbreeMemoryType::SHARED);
-      }
-
 #if defined(EMBREE_SYCL_SUPPORT)
-      if (!dptr) {
         dshared = true;
         dptr = ptr;
-      }
 #endif
+      }
     }
 
     /*! Buffer destruction */

kernels/common/device.cpp

@@ -137,6 +137,7 @@ namespace embree
 
   Device::~Device ()
   {
+    std::cout << "destructor device" << std::endl;
     setCacheSize(0);
     exitTaskingSystem();
   }

kernels/common/geometry.h

@@ -126,10 +126,8 @@ namespace embree
   };
 
   /*! Base class all geometries are derived from */
-  class Geometry : public RefCount
+  class alignas(16) Geometry : public RefCount
   {
-    ALIGNED_CLASS_(16);
-
     friend class Scene;
   public:
 

kernels/common/scene.h

@@ -35,10 +35,8 @@ namespace embree
   struct TaskGroup;
 
   /*! Base class all scenes are derived from */
-  class Scene : public AccelN
+  class alignas(16) Scene : public AccelN
   {
-    ALIGNED_CLASS_(std::alignment_of<Scene>::value);
-
   public:
     template<typename Ty, bool mblur = false>
       class Iterator

kernels/common/scene_instance.h

@@ -13,7 +13,6 @@ namespace embree
   /*! Instanced acceleration structure */
   struct Instance : public Geometry
   {
-    //ALIGNED_STRUCT_(16);
     static const Geometry::GTypeMask geom_type = Geometry::MTY_INSTANCE;
 
   public:

kernels/common/scene_instance_array.h

@@ -13,7 +13,6 @@ namespace embree
   /*! Instanced acceleration structure */
   struct InstanceArray : public Geometry
   {
-    //ALIGNED_STRUCT_(16);
     static const Geometry::GTypeMask geom_type = Geometry::MTY_INSTANCE_ARRAY;
 
   public:

kernels/common/scene_line_segments.cpp

@@ -224,7 +224,6 @@ namespace embree
         flags[i] |= hasRight * RTC_CURVE_FLAG_NEIGHBOR_RIGHT;
         hasLeft = hasRight;
       }
-      std::cout << "flags.buffer->commit()" << std::endl;
       flags.buffer->commit();
     }
     segments.clearLocalModified();

kernels/subdiv/catmullclark_ring.h

@@ -10,15 +10,13 @@
 
 namespace embree
 {
-  struct __aligned(64) FinalQuad {
+  struct alignas(64) FinalQuad {
     Vec3fa vtx[4];
   };
 
   template<typename Vertex, typename Vertex_t = Vertex>
-    struct __aligned(64) CatmullClark1RingT
+    struct alignas(64) CatmullClark1RingT
   {
-    ALIGNED_STRUCT_(64);
-
     int border_index;                                   //!< edge index where border starts
     unsigned int face_valence;                          //!< number of adjacent quad faces
     unsigned int edge_valence;                          //!< number of adjacent edges (2*face_valence)
@@ -518,10 +516,8 @@ namespace embree
   typedef CatmullClark1RingT<Vec3fa,Vec3fa_t> CatmullClark1Ring3fa;
 
   template<typename Vertex, typename Vertex_t = Vertex>
-    struct __aligned(64) GeneralCatmullClark1RingT
+    struct alignas(64) GeneralCatmullClark1RingT
   {
-    ALIGNED_STRUCT_(64);
-
     typedef CatmullClark1RingT<Vertex,Vertex_t> CatmullClark1Ring;
 
     struct Face 

kernels/subdiv/tessellation_cache.h

@@ -41,10 +41,8 @@ namespace embree
  ////////////////////////////////////////////////////////////////////////////////
  ////////////////////////////////////////////////////////////////////////////////
 
- struct __aligned(64) ThreadWorkState 
+ struct alignas(64) ThreadWorkState 
  {
-   ALIGNED_STRUCT_(64);
-
    std::atomic<size_t> counter;
    ThreadWorkState* next;
    bool allocated;
@@ -56,7 +54,7 @@ namespace embree
    }   
  };
 
- class __aligned(64) SharedLazyTessellationCache 
+ class alignas(64) SharedLazyTessellationCache 
  {
  public:
 

kernels/sycl/rthwif_embree.cpp

@@ -695,7 +695,6 @@ __forceinline void trav_loop(intel_ray_query_t& query, Ray& ray, Scene* scene0,
 
 SYCL_EXTERNAL __attribute__((always_inline)) void rtcIntersectRTHW(sycl::global_ptr<RTCTraversableTy> htraversable, sycl::private_ptr<RTCRayQueryContext> ucontext, sycl::private_ptr<RTCRayHit> rayhit_i, sycl::private_ptr<RTCIntersectArguments> args)
 {
-#if 0
   Scene* scene = (Scene*) htraversable.get();
 
   Scene* scenes[RTC_MAX_INSTANCE_LEVEL_COUNT];
@@ -824,12 +823,10 @@ SYCL_EXTERNAL __attribute__((always_inline)) void rtcIntersectRTHW(sycl::global_
   }
 
   intel_ray_query_abandon(query);
-#endif
 }
 
 SYCL_EXTERNAL __attribute__((always_inline)) void rtcOccludedRTHW(sycl::global_ptr<RTCTraversableTy> htraversable, sycl::private_ptr<RTCRayQueryContext> ucontext, sycl::private_ptr<RTCRay> ray_i, sycl::private_ptr<RTCOccludedArguments> args)
 {
-#if 0
   Scene* scene = (Scene*) htraversable.get();
 
   Scene* scenes[RTC_MAX_INSTANCE_LEVEL_COUNT];
@@ -880,7 +877,6 @@ SYCL_EXTERNAL __attribute__((always_inline)) void rtcOccludedRTHW(sycl::global_p
     ray_i->tfar = -INFINITY;
 
   intel_ray_query_abandon(query);
-#endif
 }
 
 RTC_NAMESPACE_END;

tutorials/closest_point/closest_point_device.cpp

@@ -62,9 +62,8 @@ unsigned int g_line_index_buffer[g_num_point_queries] = {
 // ======================================================================== //
 //                         User defined instancing                          //
 // ======================================================================== //
-struct Instance
+struct alignas(16) Instance
 {
-  ALIGNED_STRUCT_(16)
   RTCGeometry geometry;
   RTCScene object;
   int userID;
@@ -227,9 +226,8 @@ void updateInstance (RTCScene scene, Instance* instance)
 //                      triangle mesh geometry                              //
 // ======================================================================== //
 
-struct TriangleMesh
+struct alignas(16) TriangleMesh
 {
-  ALIGNED_STRUCT_(16)
   Vertex* vertices;
   Triangle* triangles;
   unsigned int num_vertices;