Merge branch 'v2.1' of https://github.com/embree/embree into v2.1

RenderKit · Jan 22, 2014 · 1fe8c45 · 1fe8c45
2 parents 1b85717 + 0d2fb8d
commit 1fe8c45
Show file tree

Hide file tree

Showing 14 changed files with 691 additions and 32 deletions.
diff --git a/common/cmake/icc_xeonphi.cmake b/common/cmake/icc_xeonphi.cmake
@@ -28,4 +28,4 @@ SET (CMAKE_CXX_FLAGS_RELWITHDEBINFO "-g -O3 -DNDEBUG -mCG_lrb_num_threads=4 -mCG
 
 SET(CMAKE_LINKER icpc)
 SET(CMAKE_CXX_LINK_EXECUTABLE  	    "<CMAKE_CXX_COMPILER> -mmic -static-intel -rdynamic -fPIC -L${CMAKE_BINARY_DIR} <LINK_FLAGS> -o <TARGET> ${CMAKE_GNULD_IMAGE_VERSION} <OBJECTS> <LINK_LIBRARIES>")
-SET(CMAKE_CXX_CREATE_SHARED_LIBRARY "<CMAKE_CXX_COMPILER> -mmic -static-intel <LANGUAGE_COMPILE_FLAGS> <CMAKE_SHARED_MODULE_CXX_FLAGS> <LINK_FLAGS> <CMAKE_SHARED_MODULE_CREATE_CXX_FLAGS> -o <TARGET> ${CMAKE_GNULD_IMAGE_VERSION} <OBJECTS> <LINK_LIBRARIES> -zmuldefs")
+SET(CMAKE_CXX_CREATE_SHARED_LIBRARY "<CMAKE_CXX_COMPILER> -mmic -static-intel  <LANGUAGE_COMPILE_FLAGS> <CMAKE_SHARED_MODULE_CXX_FLAGS> <LINK_FLAGS> <CMAKE_SHARED_MODULE_CREATE_CXX_FLAGS> -o <TARGET> ${CMAKE_GNULD_IMAGE_VERSION} <OBJECTS> <LINK_LIBRARIES> -zmuldefs")
diff --git a/common/cmake/ispc.props b/common/cmake/ispc.props
@@ -26,7 +26,11 @@
       <MathLibrary>0</MathLibrary>
       <Opt>1</Opt>
       <CommandLineTemplate>ispc.exe [AllOptions] [AdditionalOptions] [Inputs]</CommandLineTemplate>
-      <Outputs>$(IntDir)%(Filename).ispc*.obj</Outputs>
+      <Outputs Condition="'$(Configuration)'=='Debug' and '$(MSBuildProjectName)'!='embree'">$(IntDir)%(Filename).ispc.obj;$(IntDir)%(Filename).ispc_sse2.obj;$(IntDir)%(Filename).ispc_sse4.obj</Outputs>
+      <Outputs Condition="'$(Configuration)'=='Release' and '$(MSBuildProjectName)'!='embree'">$(IntDir)%(Filename).ispc.obj;$(IntDir)%(Filename).ispc_sse2.obj;$(IntDir)%(Filename).ispc_sse4.obj</Outputs>
+      <Outputs Condition="'$(Configuration)'=='ReleaseAVX' and '$(MSBuildProjectName)'!='embree'">$(IntDir)%(Filename).ispc.obj;$(IntDir)%(Filename).ispc_sse2.obj;$(IntDir)%(Filename).ispc_sse4.obj;$(IntDir)%(Filename).ispc_avx.obj</Outputs>
+      <Outputs Condition="'$(Configuration)'=='ReleaseAVX2' and '$(MSBuildProjectName)'!='embree'">$(IntDir)%(Filename).ispc.obj;$(IntDir)%(Filename).ispc_sse2.obj;$(IntDir)%(Filename).ispc_sse4.obj;$(IntDir)%(Filename).ispc_avx.obj</Outputs>
+      <Outputs Condition="'$(MSBuildProjectName)'=='embree'">$(IntDir)%(Filename).ispc.obj;$(IntDir)%(Filename).ispc_sse2.obj;$(IntDir)%(Filename).ispc_sse4.obj;$(IntDir)%(Filename).ispc_avx.obj;$(IntDir)%(Filename).ispc_avx2.obj</Outputs>
       <ExecutionDescription>Compiling %(Filename)%(Extension)  ...</ExecutionDescription>
     </ISPC>
   </ItemDefinitionGroup>

diff --git a/common/sys/sync/mutex.h b/common/sys/sync/mutex.h
@@ -57,7 +57,7 @@ namespace embree
 	  _mm_pause(); 
 	  _mm_pause();
 #else
-	  _mm_delay_32(128); // FIXME: exp falloff
+	  _mm_delay_32(256); // FIXME: exp falloff
 #endif
 	}
         __memory_barrier();

diff --git a/kernels/common/scene.cpp b/kernels/common/scene.cpp
@@ -34,6 +34,8 @@ namespace embree
       numTriangleMeshes(0), numTriangleMeshes2(0), numUserGeometries(0),
       flat_triangle_source_1(this,1), flat_triangle_source_2(this,2)
   {
+    geometries.reserve(128);
+
 #if defined(__MIC__)
 
     accels.accel0 = NULL; 
@@ -209,9 +211,8 @@ namespace embree
 
   unsigned Scene::add(Geometry* geometry) 
   {
-#if !defined(__MIC__)
     Lock<AtomicMutex> lock(geometriesMutex);
-#endif
+
     if (usedIDs.size()) {
       int id = usedIDs.back(); 
       usedIDs.pop_back();
@@ -225,9 +226,8 @@ namespace embree
 
   void Scene::remove(Geometry* geometry) 
   {
-#if !defined(__MIC__)
     Lock<AtomicMutex> lock(geometriesMutex);
-#endif
+
     usedIDs.push_back(geometry->id);
     geometries[geometry->id] = NULL;
     delete geometry;
@@ -243,9 +243,7 @@ namespace embree
 
   void Scene::build () 
   {
-#if !defined(__MIC__)
     Lock<MutexSys> lock(mutex);
-#endif
 
     if ((isStatic() && isBuild()) || !ready()) {
       recordError(RTC_INVALID_OPERATION);

diff --git a/kernels/common/scene.h b/kernels/common/scene.h
@@ -81,13 +81,10 @@ namespace embree
     __forceinline const Geometry* get(size_t i) const { assert(i < geometries.size()); return geometries[i]; }
     __forceinline       Geometry* get_locked(size_t i)  { 
 
-#if !defined(__MIC__)
-      Lock<AtomicMutex> lock(geometriesMutex); 
-#endif
-
+      Lock<AtomicMutex> lock(geometriesMutex);
       Geometry *g = geometries[i]; 
-
-      assert(i < geometries.size()); return g; 
+      assert(i < geometries.size()); 
+      return g; 
     }
 
     /* get triangle mesh by ID */
@@ -209,7 +206,6 @@ namespace embree
 
 
   public:
-    AtomicMutex geometriesMutex;
     std::vector<int> usedIDs;
     std::vector<Geometry*> geometries; //!< list of all user geometries
 
@@ -222,6 +218,7 @@ namespace embree
     bool needVertices;
     bool is_build;
     MutexSys mutex;
+    AtomicMutex geometriesMutex;
 
   public:
     atomic_t numTriangleMeshes;        //!< number of enabled triangle meshes

diff --git a/kernels/xeon/CMakeLists.txt b/kernels/xeon/CMakeLists.txt
@@ -122,6 +122,7 @@ IF (TARGET_AVX)
    bvh4i/bvh4i_intersector1.cpp   
    bvh4i/bvh4i_intersector4_chunk.cpp   
    bvh4i/bvh4i_intersector8_chunk.cpp   
+   bvh4i/bvh4i_intersector8_hybrid.cpp
 
    bvh4i/bvh4i_builder_fast.cpp
    bvh4i/bvh4i_builder_binner.cpp
@@ -154,6 +155,7 @@ IF (TARGET_AVX2)
     bvh4/bvh4_intersector8_chunk.cpp
     bvh4/bvh4_intersector8_hybrid.cpp
     bvh4i/bvh4i_intersector8_chunk_avx2.cpp  
+    bvh4i/bvh4i_intersector8_hybrid.cpp
 
     bvh4i/bvh4i_intersector1.cpp   
     bvh4i/bvh4i_intersector4_chunk.cpp   

diff --git a/kernels/xeon/bvh4/bvh4.cpp b/kernels/xeon/bvh4/bvh4.cpp
@@ -253,7 +253,7 @@ namespace embree
     BVH4* accel = new BVH4(SceneTriangle4::type);
 
     Accel::Intersectors intersectors;
-    if      (g_traverser == "default") intersectors = BVH4Triangle4IntersectorsChunk(accel);
+    if      (g_traverser == "default") intersectors = BVH4Triangle4IntersectorsHybrid(accel);
     else if (g_traverser == "chunk"  ) intersectors = BVH4Triangle4IntersectorsChunk(accel);
     else if (g_traverser == "hybrid" ) intersectors = BVH4Triangle4IntersectorsHybrid(accel);
     else throw std::runtime_error("unknown traverser "+g_traverser+" for BVH4<Triangle4>");
@@ -278,7 +278,7 @@ namespace embree
     BVH4* accel = new BVH4(SceneTriangle8::type);
 
     Accel::Intersectors intersectors;
-    if      (g_traverser == "default") intersectors = BVH4Triangle8IntersectorsChunk(accel);
+    if      (g_traverser == "default") intersectors = BVH4Triangle8IntersectorsHybrid(accel);
     else if (g_traverser == "chunk"  ) intersectors = BVH4Triangle8IntersectorsChunk(accel);
     else if (g_traverser == "hybrid" ) intersectors = BVH4Triangle8IntersectorsHybrid(accel);
     else throw std::runtime_error("unknown traverser "+g_traverser+" for BVH4<Triangle8>");
@@ -314,7 +314,7 @@ namespace embree
     BVH4* accel = new BVH4(SceneTriangle4v::type);
 
     Accel::Intersectors intersectors;
-    if      (g_traverser == "default") intersectors = BVH4Triangle4vIntersectorsChunk(accel);
+    if      (g_traverser == "default") intersectors = BVH4Triangle4vIntersectorsHybrid(accel);
     else if (g_traverser == "chunk"  ) intersectors = BVH4Triangle4vIntersectorsChunk(accel);
     else if (g_traverser == "hybrid" ) intersectors = BVH4Triangle4vIntersectorsHybrid(accel);
     else throw std::runtime_error("unknown traverser "+g_traverser+" for BVH4<Triangle4>");

diff --git a/kernels/xeon/bvh4/bvh4_intersector8_hybrid.cpp b/kernels/xeon/bvh4/bvh4_intersector8_hybrid.cpp
@@ -20,7 +20,7 @@
 #include "geometry/triangle8_intersector8_moeller.h"
 #include "geometry/triangle4v_intersector8_pluecker.h"
 
-#define SWITCH_THRESHOLD 3
+#define SWITCH_THRESHOLD 6
 
 namespace embree
 {

diff --git a/kernels/xeon/bvh4i/bvh4i.cpp b/kernels/xeon/bvh4i/bvh4i.cpp
@@ -51,6 +51,10 @@ namespace embree
   DECLARE_SYMBOL(Accel::Intersector8,BVH4iTriangle1vIntersector8ChunkPluecker);
   DECLARE_SYMBOL(Accel::Intersector8,BVH4iTriangle4vIntersector8ChunkPluecker);
   DECLARE_SYMBOL(Accel::Intersector8,BVH4iVirtualIntersector8Chunk);
+
+  DECLARE_SYMBOL(Accel::Intersector8,BVH4iTriangle4Intersector8HybridMoeller);
+
+
 #if defined(__TARGET_AVX2__)
   extern Accel::Intersector8 BVH4iTriangle1Intersector8ChunkAVX2;
 #endif
@@ -90,6 +94,8 @@ namespace embree
     SELECT_SYMBOL_AVX_AVX2(features,BVH4iTriangle1vIntersector8ChunkPluecker);
     SELECT_SYMBOL_AVX_AVX2(features,BVH4iTriangle4vIntersector8ChunkPluecker);
     SELECT_SYMBOL_AVX_AVX2(features,BVH4iVirtualIntersector8Chunk);
+
+    SELECT_SYMBOL_AVX_AVX2(features,BVH4iTriangle4Intersector8HybridMoeller);
   }
 
   Accel::Intersectors BVH4iTriangle1Intersectors(BVH4i* bvh)
@@ -109,7 +115,8 @@ namespace embree
     intersectors.ptr = bvh;
     intersectors.intersector1 = BVH4iTriangle4Intersector1Moeller;
     intersectors.intersector4 = BVH4iTriangle4Intersector4ChunkMoeller;
-    intersectors.intersector8 = BVH4iTriangle4Intersector8ChunkMoeller;
+    //intersectors.intersector8 = BVH4iTriangle4Intersector8ChunkMoeller;
+    intersectors.intersector8 = BVH4iTriangle4Intersector8HybridMoeller;
     intersectors.intersector16 = NULL;
     return intersectors;
   }

diff --git a/kernels/xeon/bvh4i/bvh4i_builder_morton.cpp b/kernels/xeon/bvh4i/bvh4i_builder_morton.cpp
@@ -21,11 +21,7 @@
 
 #define BVH_NODE_PREALLOC_FACTOR 1.1f
 
-//#define PROFILE
-
-#if defined(__USE_STAT_COUNTERS__)
 #define PROFILE
-#endif
 
 #define DBG(x)
 
@@ -61,14 +57,14 @@ namespace embree
       double dt_min = pos_inf;
       double dt_avg = 0.0f;
       double dt_max = neg_inf;
-      for (size_t i=0; i<200; i++) 
+      for (size_t i=0; i<20; i++) 
       {
         TaskScheduler::executeTask(threadIndex,threadCount,_build_parallel_morton,this,TaskScheduler::getNumThreads(),"build_parallel_morton");
         dt_min = min(dt_min,dt);
         dt_avg = dt_avg + dt;
         dt_max = max(dt_max,dt);
       }
-      dt_avg /= double(200);
+      dt_avg /= double(20);
 
       std::cout << "[DONE]" << std::endl;
       std::cout << "  min = " << 1000.0f*dt_min << "ms (" << source->size()/dt_min*1E-6 << " Mtris/s)" << std::endl;