refactoring Array, Tuple

composable_kernel/include/tensor_description/multi_index.hpp

@@ -9,20 +9,12 @@ namespace ck {
 template <index_t N>
 using MultiIndex = Array<index_t, N>;
 
-#if 1 // debug
 template <typename... Xs>
 __host__ __device__ constexpr auto make_multi_index(Xs... xs)
 {
     return make_array<index_t>(xs...);
 }
 #else
-template <typename... Xs>
-__host__ __device__ constexpr auto make_multi_index(const Xs&... xs)
-{
-    return make_array(xs...);
-}
-#endif
-#else
 template <index_t N>
 using MultiIndex = StaticallyIndexedArray<index_t, N>;
 

composable_kernel/include/utility/array.hpp

@@ -12,7 +12,7 @@ struct Array
     using type      = Array;
     using data_type = TData;
 
-    TData mData[NSize] = {0};
+    TData mData[NSize];
 
     __host__ __device__ static constexpr index_t Size() { return NSize; }
 

composable_kernel/include/utility/tuple.hpp

@@ -20,7 +20,7 @@ struct TupleElement
     __host__ __device__ explicit constexpr TupleElement() : mData() {}
 
     template <typename T>
-    __host__ __device__ explicit constexpr TupleElement(T&& v) : mData(static_cast<T&&>(v))
+    __host__ __device__ explicit constexpr TupleElement(T&& v) : mData(std::forward<T>(v))
     {
     }
 
@@ -57,7 +57,7 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElement<TupleElementKey<Is>, Xs>
 
     template <typename... Ys>
     __host__ __device__ explicit constexpr TupleImpl(Ys&&... ys)
-        : TupleElement<TupleElementKey<Is>, Xs>(static_cast<Ys&&>(ys))...
+        : TupleElement<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
     {
     }
 
@@ -85,7 +85,7 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
         detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(Xs), 1>::type, Xs...>;
 
     template <typename... Ys>
-    __host__ __device__ explicit constexpr Tuple(Ys&&... ys) : base(static_cast<Ys&&>(ys)...)
+    __host__ __device__ explicit constexpr Tuple(Ys&&... ys) : base(std::forward<Ys>(ys)...)
     {
     }