__half2 should walk like CUDA and talk like CUDA

hipamd/include/hip/hcc_detail/hip_fp16.h

@@ -1126,8 +1126,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data ==
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __half2_raw{__builtin_convertvector(-r, _Float16_2)};
             }
             inline
             __device__
@@ -1135,8 +1134,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data !=
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __builtin_convertvector(-r, _Float16_2);
             }
             inline
             __device__
@@ -1144,8 +1142,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data <=
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __builtin_convertvector(-r, _Float16_2);
             }
             inline
             __device__
@@ -1153,8 +1150,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data >=
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __builtin_convertvector(-r, _Float16_2);
             }
             inline
             __device__
@@ -1162,8 +1158,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data <
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __builtin_convertvector(-r, _Float16_2);
             }
             inline
             __device__
@@ -1171,8 +1166,7 @@ THE SOFTWARE.
             {
                 auto r = static_cast<__half2_raw>(x).data >
                     static_cast<__half2_raw>(y).data;
-                return __half2_raw{_Float16_2{
-                    static_cast<_Float16>(r.x), static_cast<_Float16>(r.y)}};
+                return __builtin_convertvector(-r, _Float16_2);
             }
             inline
             __device__