Change memset kernel to use memcpy instead of placement new. Simplify indexers.

[ROCm/clr commit: d5c8de3f41]

projects/clr/hipamd/include/hip/hcc_detail/hip_runtime.h

@@ -389,18 +389,9 @@ template<typename std::common_type<
 class Coordinates {
     using R = decltype(f(0));
 
-    struct X {
-        __device__ operator R() const { return f(0); }
-        __host__ operator R() const { return R{}; }
-    };
-    struct Y {
-        __device__ operator R() const { return f(1); }
-        __host__ operator R() const { return R{}; }
-    };
-    struct Z {
-        __device__ operator R() const { return f(2); }
-        __host__ operator R() const { return R{}; }
-    };
+    struct X { __device__ operator R() const { return f(0); } };
+    struct Y { __device__ operator R() const { return f(1); } };
+    struct Z { __device__ operator R() const { return f(2); } };
 public:
     static constexpr X x{};
     static constexpr Y y{};

projects/clr/hipamd/src/hip_memory.cpp

@@ -1311,7 +1311,10 @@ namespace
 
         size_t idx = blockIdx.x * block_dim + threadIdx.x;
         while (idx < n) {
-            new (&f[idx]) T{value};
+            __builtin_memcpy(
+                reinterpret_cast<void*>(&f[idx]),
+                reinterpret_cast<const void*>(&value),
+                sizeof(T));
             idx += grid_dim;
         }
     }