fix RTN logic

include/ck/utility/type_convert.hpp

@@ -32,19 +32,7 @@ inline __host__ __device__ constexpr float type_convert<float, bhalf_t>(bhalf_t
 }
 
 // convert fp32 to bfp16
-#ifndef FLASH_ATTENTION_INTERNAL_USE_RTN
-template <>
-inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
-{
-    union
-    {
-        float fp32;
-        uint32_t int32;
-    } u = {x};
-
-    return uint16_t(u.int32 >> 16);
-}
-#else
+#if FLASH_ATTENTION_INTERNAL_USE_RTN
 template <>
 inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
 {
@@ -87,6 +75,18 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float
 
     return uint16_t(u.int32 >> 16);
 }
+#else
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    return uint16_t(u.int32 >> 16);
+}
 #endif
 
 // convert bfp16 to fp16 via fp32