Adjusted the Kokkos implementation to fix view initialisation, and store local copies of views for lambda scoping

KOKKOSStream.cpp

@@ -7,7 +7,7 @@
 
 #include "KOKKOSStream.hpp"
 
-using Kokkos::parallel_for;
+using namespace Kokkos;
 
 template <class T>
 KOKKOSStream<T>::KOKKOSStream(
@@ -16,21 +16,21 @@ KOKKOSStream<T>::KOKKOSStream(
 {
   Kokkos::initialize();
 
-  new(d_a) Kokkos::View<double*, DEVICE>("d_a", ARRAY_SIZE);
-  new(d_b) Kokkos::View<double*, DEVICE>("d_b", ARRAY_SIZE);
-  new(d_c) Kokkos::View<double*, DEVICE>("d_c", ARRAY_SIZE);
-  new(hm_a) Kokkos::View<double*>::HostMirror();
-  new(hm_b) Kokkos::View<double*>::HostMirror();
-  new(hm_c) Kokkos::View<double*>::HostMirror();
-  hm_a = Kokkos::create_mirror_view(d_a);
-  hm_b = Kokkos::create_mirror_view(d_b);
-  hm_c = Kokkos::create_mirror_view(d_c);
+  d_a = new View<double*, DEVICE>("d_a", ARRAY_SIZE);
+  d_b = new View<double*, DEVICE>("d_b", ARRAY_SIZE);
+  d_c = new View<double*, DEVICE>("d_c", ARRAY_SIZE);
+  hm_a = new View<double*, DEVICE>::HostMirror();
+  hm_b = new View<double*, DEVICE>::HostMirror();
+  hm_c = new View<double*, DEVICE>::HostMirror();
+  *hm_a = create_mirror_view(*d_a);
+  *hm_b = create_mirror_view(*d_b);
+  *hm_c = create_mirror_view(*d_c);
 }
 
 template <class T>
 KOKKOSStream<T>::~KOKKOSStream()
 {
-  Kokkos::finalize();
+  finalize();
 }
 
 template <class T>
@@ -39,65 +39,81 @@ void KOKKOSStream<T>::write_arrays(
 {
   for(int ii = 0; ii < array_size; ++ii)
   {
-    hm_a(ii) = a[ii];
-    hm_b(ii) = b[ii];
-    hm_c(ii) = c[ii];
+    (*hm_a)(ii) = a[ii];
+    (*hm_b)(ii) = b[ii];
+    (*hm_c)(ii) = c[ii];
   }
-  Kokkos::deep_copy(hm_a, d_a);
-  Kokkos::deep_copy(hm_b, d_b);
-  Kokkos::deep_copy(hm_c, d_c);
+  deep_copy(*hm_a, *d_a);
+  deep_copy(*hm_b, *d_b);
+  deep_copy(*hm_c, *d_c);
 }
 
 template <class T>
 void KOKKOSStream<T>::read_arrays(
         std::vector<T>& a, std::vector<T>& b, std::vector<T>& c)
 {
-  Kokkos::deep_copy(d_a, hm_a);
-  Kokkos::deep_copy(d_a, hm_b);
-  Kokkos::deep_copy(d_a, hm_c);
+  deep_copy(*d_a, *hm_a);
+  deep_copy(*d_b, *hm_b);
+  deep_copy(*d_c, *hm_c);
   for(int ii = 0; ii < array_size; ++ii)
   {
-    a[ii] = hm_a(ii);
-    b[ii] = hm_b(ii);
-    c[ii] = hm_c(ii);
+    a[ii] = (*hm_a)(ii);
+    b[ii] = (*hm_b)(ii);
+    c[ii] = (*hm_c)(ii);
   }
 }
 
 template <class T>
 void KOKKOSStream<T>::copy()
 {
-  Kokkos::parallel_for(array_size, KOKKOS_LAMBDA (const int index) 
+  View<double*, DEVICE> a(*d_a);
+  View<double*, DEVICE> b(*d_b);
+  View<double*, DEVICE> c(*d_c);
+
+  parallel_for(array_size, KOKKOS_LAMBDA (const int index) 
   {
-    d_c[index] = d_a[index];
+    c[index] = a[index];
   });
 }
 
 template <class T>
 void KOKKOSStream<T>::mul()
 {
+  View<double*, DEVICE> a(*d_a);
+  View<double*, DEVICE> b(*d_b);
+  View<double*, DEVICE> c(*d_c);
+
   const T scalar = 3.0;
   parallel_for(array_size, KOKKOS_LAMBDA (const int index) 
   {
-    d_b[index] = scalar*d_c[index];
+    b[index] = scalar*c[index];
   });
 }
 
 template <class T>
 void KOKKOSStream<T>::add()
 {
+  View<double*, DEVICE> a(*d_a);
+  View<double*, DEVICE> b(*d_b);
+  View<double*, DEVICE> c(*d_c);
+
   parallel_for(array_size, KOKKOS_LAMBDA (const int index) 
   {
-    d_c[index] = d_a[index] + d_b[index];
+    c[index] = a[index] + b[index];
   });
 }
 
 template <class T>
 void KOKKOSStream<T>::triad()
 {
+  View<double*, DEVICE> a(*d_a);
+  View<double*, DEVICE> b(*d_b);
+  View<double*, DEVICE> c(*d_c);
+
   const T scalar = 3.0;
   parallel_for(array_size, KOKKOS_LAMBDA (const int index) 
   {
-    d_a[index] = d_b[index] + scalar*d_c[index];
+    a[index] = b[index] + scalar*c[index];
   });
 }
 
@@ -118,6 +134,6 @@ std::string getDeviceDriver(const int device)
   return "Kokkos";
 }
 
-template class KOKKOSStream<float>;
+//template class KOKKOSStream<float>;
 template class KOKKOSStream<double>;
 

KOKKOSStream.hpp

@@ -17,8 +17,11 @@
 
 #define IMPLEMENTATION_STRING "KOKKOS"
 
-#define DEVICE Kokkos::OpenMP
-
+#ifdef KOKKOS_TARGET_CPU
+  #define DEVICE Kokkos::OpenMP
+#else
+  #define DEVICE Kokkos::Cuda
+#endif
 
 template <class T>
 class KOKKOSStream : public Stream<T>
@@ -28,12 +31,12 @@ class KOKKOSStream : public Stream<T>
     unsigned int array_size;
 
     // Device side pointers to arrays
-    Kokkos::View<double*, DEVICE> d_a;
-    Kokkos::View<double*, DEVICE> d_b;
-    Kokkos::View<double*, DEVICE> d_c;
-    Kokkos::View<double*>::HostMirror hm_a;
-    Kokkos::View<double*>::HostMirror hm_b;
-    Kokkos::View<double*>::HostMirror hm_c;
+    Kokkos::View<double*, DEVICE>* d_a;
+    Kokkos::View<double*, DEVICE>* d_b;
+    Kokkos::View<double*, DEVICE>* d_c;
+    Kokkos::View<double*>::HostMirror* hm_a;
+    Kokkos::View<double*>::HostMirror* hm_b;
+    Kokkos::View<double*>::HostMirror* hm_c;
 
   public:
 

main.cpp

@@ -59,9 +59,11 @@ int main(int argc, char *argv[])
 
   // TODO: Fix SYCL to allow multiple template specializations
 #ifndef SYCL
+#ifndef KOKKOS
   if (use_float)
     run<float>();
   else
+#endif
 #endif
     run<double>();