diff --git a/src/TNL/Algorithms/AtomicOperations.h b/src/TNL/Algorithms/AtomicOperations.h
index 4be725d482f4ebb255c42c1d984b58a2ed4975e1..679b14eb5c1c1662da979d311ad0101ee924c035 100644
--- a/src/TNL/Algorithms/AtomicOperations.h
+++ b/src/TNL/Algorithms/AtomicOperations.h
@@ -12,7 +12,9 @@
 
 #pragma once
 
+#ifdef HAVE_CUDA
 #include <cuda.h>
+#endif
 #include <TNL/Devices/Sequential.h>
 #include <TNL/Devices/Host.h>
 #include <TNL/Devices/Cuda.h>
@@ -80,7 +82,7 @@ struct AtomicOperations< Devices::Cuda >
       TNL_ASSERT_TRUE( false, "Atomic add for long int is not supported on CUDA." );
 #endif // HAVE_CUDA
    }
-   
+
    __cuda_callable__
    static void add( short int& v, const short int& a )
    {