Merge pull request #27 from ingowald/main

added missing omp::freeBVH, and atomic_min/max<double>

Author: iwald-nvidia

cuBQL/builder/omp.h

@@ -11,6 +11,10 @@ namespace cuBQL {
     void refit(BinaryBVH<T,D>    &bvh,
                const box_t<T,D>  *boxes,
                Context *ctx);
+    
+    template<typename T, int D>
+    void freeBVH(BinaryBVH<T,D> &bvh,
+                 Context        *ctx);
   }
 }
 

cuBQL/builder/omp/AtomicBox.h

@@ -60,6 +60,42 @@ namespace cuBQL {
 #endif
     }
 
+    /*! iw - note: this implementation of atomic min/max via atomic
+        compare-exchange (CAS); which is cetainly not optimal on any
+        sort of modern GPU - but it works in any C++-21 compliant
+        compiler, so it's what we do for now */
+    inline void atomic_min(double *ptr, double value)
+    {
+#ifdef __NVCOMPILER
+# if 1
+      double &mem = *ptr;
+      if (mem <= value) return;
+      while (1) {
+        double wasBefore;
+#pragma omp atomic capture
+        { wasBefore = mem; mem = value; }
+        if (wasBefore >= value) break;
+        value = wasBefore;
+      }
+# else
+      double current = *(volatile double *)ptr;
+      while (current > value) {
+        bool wasChanged
+          = ((std::atomic<long long int>*)ptr)
+          ->compare_exchange_weak((long long int&)current,(long long int&)value);
+        if (wasChanged) break;
+      }
+# endif
+#else
+      double &x = *ptr;
+#pragma omp atomic compare 
+      if (x > value) { x = value; }
+//       double t;
+// #pragma omp atomic capture
+//       { t = *ptr; *ptr = std::min(t,value); }
+#endif
+    }
+    
     /*! iw - note: this implementation of atomic min/max via atomic
         compare-exchange (CAS); which is cetainly not optimal on any
         sort of modern GPU - but it works in any C++-21 compliant
@@ -96,6 +132,42 @@ namespace cuBQL {
 #endif
     }
 
+    /*! iw - note: this implementation of atomic min/max via atomic
+        compare-exchange (CAS); which is cetainly not optimal on any
+        sort of modern GPU - but it works in any C++-21 compliant
+        compiler, so it's what we do for now */
+    inline void atomic_max(double *ptr, double value)
+    { 
+#ifdef __NVCOMPILER
+# if 1
+      double &mem = *ptr;
+      if (mem >= value) return;
+      while (1) {
+        double wasBefore;
+#pragma omp atomic capture
+        { wasBefore = mem; mem = value; }
+        if (wasBefore <= value) break;
+        value = wasBefore;
+      }
+# else
+      double current = *(volatile double *)ptr;
+      while (current < value) {
+        bool wasChanged
+          = ((std::atomic<long long int>*)ptr)
+          ->compare_exchange_weak((long long int&)current,(long long int&)value);
+        if (wasChanged) break;
+      }
+# endif
+#else
+      double &x = *ptr;
+#pragma omp atomic compare 
+      if (x < value) { x = value; }
+        //       double t;
+// #pragma omp atomic capture
+//       { t = *ptr; *ptr = std::max(t,value); }
+#endif
+    }
+    
     template<typename T, int D>
     inline void v_atomic_min(vec_t<T,D> *ptr, vec_t<T,D> v);
     template<typename T, int D>

cuBQL/builder/omp/spatialMedian.h

@@ -423,6 +423,15 @@ namespace cuBQL {
       cuBQL::omp::refit(bvh,boxes,ctx);
     }
 
+    template<typename T, int D>
+    void freeBVH(BinaryBVH<T,D>    &bvh,
+                 Context           *ctx)
+    {
+      ctx->free(bvh.primIDs);
+      ctx->free(bvh.nodes);
+      bvh.primIDs = 0;
+    }
+    
   }
 
   template<typename T, int D>