diff --git a/openmp/libomptarget/deviceRTLs/nvptx/src/target_impl.h b/openmp/libomptarget/deviceRTLs/nvptx/src/target_impl.h
--- a/openmp/libomptarget/deviceRTLs/nvptx/src/target_impl.h
+++ b/openmp/libomptarget/deviceRTLs/nvptx/src/target_impl.h
@@ -130,6 +130,15 @@
 DEVICE unsigned GetWarpId();
 DEVICE unsigned GetLaneId();
 
+// Forward declaration of atomics. Although they're template functions, we
+// already have definitions for different types in CUDA internal headers with
+// the right mangled names.
+template <typename T> T atomicAdd(T *address, T val);
+template <typename T> T atomicInc(T *address, T val);
+template <typename T> T atomicMax(T *address, T val);
+template <typename T> T atomicExch(T *address, T val);
+template <typename T> T atomicCAS(T *address, T compare, T val);
+
 // Atomics
 template <typename T> INLINE T __kmpc_atomic_add(T *address, T val) {
   return atomicAdd(address, val);