Fallback to per-element converts for CUDA

Author: AlexeySachkov

sycl/include/sycl/types.hpp

@@ -678,6 +678,10 @@ template <typename Type, int NumElements> class vec {
 
     // Whole vector conversion can only be done, if:
     constexpr bool canUseNativeVectorConvert =
+#ifdef __NVPTX__
+        // - we are on CUDA, see intel/llvm#11840
+        false &&
+#endif // __NVPTX__
         // - both vectors are represented using native vector types;
         NativeVec && vec<convertT, NumElements>::NativeVec &&
         // - it is not a signed to unsigned (or vice versa) conversion