Revert to normal vector without allocators

Prohibit vector type in indices

src/dpl_shim.h

@@ -23,11 +23,6 @@ const static auto exe_policy = oneapi::dpl::execution::device_policy<>{
         oneapi::dpl::execution::make_device_policy(cl::sycl::default_selector{})
 };
 
-template<typename T> using Allocator = sycl::usm_allocator<T, sycl::usm::alloc::shared>;
-
-template<class T>
-constexpr Allocator<T> alloc_vec() { return {exe_policy.queue()}; };
-
 template<typename T>
 T *alloc_raw(size_t size) { return sycl::malloc_shared<T>(size, exe_policy.queue()); }
 
@@ -61,11 +56,6 @@ static constexpr auto exe_policy = std::execution::par_unseq;
 
 #ifdef USE_STD_PTR_ALLOC_DEALLOC
 
-template<typename T> using Allocator = std::allocator<T>;
-
-template<class T>
-constexpr Allocator<T> alloc_vec() { return {}; };
-
 template<typename T>
 T *alloc_raw(size_t size) { return (T *) aligned_alloc(ALIGNMENT, sizeof(T) * size); }
 

src/std-data/STDDataStream.cpp

@@ -18,7 +18,7 @@ template <class T>
 STDDataStream<T>::STDDataStream(const int ARRAY_SIZE, int device)
   noexcept : array_size{ARRAY_SIZE},
 #ifdef USE_VECTOR
-  a(ARRAY_SIZE, alloc_vec<T>()), b(ARRAY_SIZE, alloc_vec<T>()), c(ARRAY_SIZE, alloc_vec<T>())
+  a(ARRAY_SIZE), b(ARRAY_SIZE), c(ARRAY_SIZE)
 #else
   a(alloc_raw<T>(ARRAY_SIZE)), b(alloc_raw<T>(ARRAY_SIZE)), c(alloc_raw<T>(ARRAY_SIZE))
 #endif
@@ -69,28 +69,28 @@ void STDDataStream<T>::copy()
 {
   // c[i] = a[i]
   std::copy(exe_policy, BEGIN(a), END(a), BEGIN(c));
-  }
+}
 
 template <class T>
 void STDDataStream<T>::mul()
 {
   //  b[i] = scalar * c[i];
   std::transform(exe_policy, BEGIN(c), END(c), BEGIN(b), [scalar = startScalar](T ci){ return scalar*ci; });
-  }
+}
 
 template <class T>
 void STDDataStream<T>::add()
 {
   //  c[i] = a[i] + b[i];
   std::transform(exe_policy, BEGIN(a), END(a), BEGIN(b), BEGIN(c), std::plus<T>());
-  }
+}
 
 template <class T>
 void STDDataStream<T>::triad()
 {
   //  a[i] = b[i] + scalar * c[i];
   std::transform(exe_policy, BEGIN(b), END(b), BEGIN(c), BEGIN(a), [scalar = startScalar](T bi, T ci){ return bi+scalar*ci; });
-  }
+}
 
 template <class T>
 void STDDataStream<T>::nstream()
@@ -101,7 +101,7 @@ void STDDataStream<T>::nstream()
   //  2: a[i] += scalar * c[i];
   std::transform(exe_policy, BEGIN(a), END(a), BEGIN(b), BEGIN(a), [](T ai, T bi){ return ai + bi; });
   std::transform(exe_policy, BEGIN(a), END(a), BEGIN(c), BEGIN(a), [scalar = startScalar](T ai, T ci){ return ai + scalar*ci; });
-  }
+}
    
 
 template <class T>

src/std-data/STDDataStream.h

@@ -23,7 +23,7 @@ class STDDataStream : public Stream<T>
 
     // Device side pointers
 #ifdef USE_VECTOR
-    std::vector<T, Allocator<T>> a, b, c;
+    std::vector<T> a, b, c;
 #else
     T *a, *b, *c;
 #endif

src/std-indices/STDIndicesStream.cpp

@@ -22,11 +22,21 @@
 #define END(x) ((x) + array_size)
 #endif
 
+#ifdef USE_VECTOR
+#if (defined(__NVCOMPILER) || defined(__NVCOMPILER_LLVM__))
+#error "std::vector *is* supported in NVHPC if we capture `this`, however, oneDPL (via SYCL2020) only works correctly with explicit *value* captures."
+#endif
+
+#if defined(USE_ONEDPL)
+#error "std::vector is unspported: oneDPL (via SYCL2020) only works correctly with explicit *value* captures"
+#endif
+#endif
+
 template <class T>
 STDIndicesStream<T>::STDIndicesStream(const int ARRAY_SIZE, int device)
 noexcept : array_size{ARRAY_SIZE}, range(0, array_size),
 #ifdef USE_VECTOR
-  a(ARRAY_SIZE, alloc_vec<T>()), b(ARRAY_SIZE, alloc_vec<T>()), c(ARRAY_SIZE, alloc_vec<T>())
+  a(ARRAY_SIZE), b(ARRAY_SIZE), c(ARRAY_SIZE)
 #else
   a(alloc_raw<T>(ARRAY_SIZE)), b(alloc_raw<T>(ARRAY_SIZE)), c(alloc_raw<T>(ARRAY_SIZE))
 #endif
@@ -77,7 +87,7 @@ void STDIndicesStream<T>::copy()
 {
   // c[i] = a[i]
   std::copy(exe_policy, BEGIN(a), END(a), BEGIN(c));
-  }
+}
 
 template <class T>
 void STDIndicesStream<T>::mul()
@@ -86,7 +96,7 @@ void STDIndicesStream<T>::mul()
   std::transform(exe_policy, range.begin(), range.end(), BEGIN(b), [c = this->c, scalar = startScalar](int i) {
     return scalar * c[i];
   });
-  }
+}
 
 template <class T>
 void STDIndicesStream<T>::add()
@@ -95,7 +105,7 @@ void STDIndicesStream<T>::add()
   std::transform(exe_policy, range.begin(), range.end(), BEGIN(c), [a = this->a, b = this->b](int i) {
     return a[i] + b[i];
   });
-  }
+}
 
 template <class T>
 void STDIndicesStream<T>::triad()
@@ -104,7 +114,7 @@ void STDIndicesStream<T>::triad()
   std::transform(exe_policy, range.begin(), range.end(), BEGIN(a), [b = this->b, c = this->c, scalar = startScalar](int i) {
     return b[i] + scalar * c[i];
   });
-  }
+}
 
 template <class T>
 void STDIndicesStream<T>::nstream()
@@ -116,7 +126,7 @@ void STDIndicesStream<T>::nstream()
   std::transform(exe_policy, range.begin(), range.end(), BEGIN(a), [a = this->a, b = this->b, c = this->c, scalar = startScalar](int i) {
     return a[i] + b[i] + scalar * c[i];
   });
-  }
+}
    
 
 template <class T>

src/std-indices/STDIndicesStream.h

@@ -78,7 +78,7 @@ class STDIndicesStream : public Stream<T>
 
     // Device side pointers
 #ifdef USE_VECTOR
-    std::vector<T, Allocator<T>> a, b, c;
+    std::vector<T> a, b, c;
 #else
     T *a, *b, *c;
 #endif

src/std-ranges/STDRangesStream.cpp

@@ -26,7 +26,7 @@ template <class T>
 STDRangesStream<T>::STDRangesStream(const int ARRAY_SIZE, int device)
 noexcept : array_size{ARRAY_SIZE},
 #ifdef USE_VECTOR
-  a(ARRAY_SIZE, alloc_vec<T>()), b(ARRAY_SIZE, alloc_vec<T>()), c(ARRAY_SIZE, alloc_vec<T>())
+  a(ARRAY_SIZE), b(ARRAY_SIZE), c(ARRAY_SIZE)
 #else
   a(alloc_raw<T>(ARRAY_SIZE)), b(alloc_raw<T>(ARRAY_SIZE)), c(alloc_raw<T>(ARRAY_SIZE))
 #endif
@@ -89,7 +89,7 @@ void STDRangesStream<T>::copy()
       c[i] = a[i];
     }
   );
-  }
+}
 
 template <class T>
 void STDRangesStream<T>::mul()
@@ -103,7 +103,7 @@ void STDRangesStream<T>::mul()
       b[i] = scalar * c[i];
     }
   );
-  }
+}
 
 template <class T>
 void STDRangesStream<T>::add()
@@ -115,7 +115,7 @@ void STDRangesStream<T>::add()
       c[i] = a[i] + b[i];
     }
   );
-  }
+}
 
 template <class T>
 void STDRangesStream<T>::triad()
@@ -129,7 +129,7 @@ void STDRangesStream<T>::triad()
       a[i] = b[i] + scalar * c[i];
     }
   );
-  }
+}
 
 template <class T>
 void STDRangesStream<T>::nstream()
@@ -143,7 +143,7 @@ void STDRangesStream<T>::nstream()
       a[i] += b[i] + scalar * c[i];
     }
   );
-  }
+}
 
 template <class T>
 T STDRangesStream<T>::dot()

src/std-ranges/STDRangesStream.hpp

@@ -22,7 +22,7 @@ class STDRangesStream : public Stream<T>
 
     // Device side pointers
 #ifdef USE_VECTOR
-    std::vector<T, Allocator<T>> a, b, c;
+    std::vector<T> a, b, c;
 #else
     T *a, *b, *c;
 #endif