diff --git a/openmp/libomptarget/DeviceRTL/include/Debug.h b/openmp/libomptarget/DeviceRTL/include/Debug.h
--- a/openmp/libomptarget/DeviceRTL/include/Debug.h
+++ b/openmp/libomptarget/DeviceRTL/include/Debug.h
@@ -13,6 +13,7 @@
 #define OMPTARGET_DEVICERTL_DEBUG_H
 
 #include "Configuration.h"
+#include "Utils.h"
 
 /// Assertion
 ///
@@ -25,9 +26,10 @@
 
 #define ASSERT(expr)                                                           \
   {                                                                            \
-    if (config::isDebugMode(config::DebugKind::Assertion) && !(expr))          \
+    if (config::isDebugMode(config::DebugKind::Assertion) && !(expr) &&        \
+        utils::SingletonFlag::testAndSet())                                    \
       __assert_fail(#expr, __FILE__, __LINE__, __PRETTY_FUNCTION__);           \
-    else                                                                       \
+    else if (!config::isDebugMode(config::DebugKind::Assertion))               \
       __assert_assume(expr);                                                   \
   }
 
diff --git a/openmp/libomptarget/DeviceRTL/include/Synchronization.h b/openmp/libomptarget/DeviceRTL/include/Synchronization.h
--- a/openmp/libomptarget/DeviceRTL/include/Synchronization.h
+++ b/openmp/libomptarget/DeviceRTL/include/Synchronization.h
@@ -74,6 +74,10 @@
 /// Atomically add \p V to \p *Addr with \p Ordering semantics.
 uint64_t add(uint64_t *Addr, uint64_t V, int Ordering);
 
+/// Atomically write \p V to \p *Addr with \p Ordering semantics and return the
+/// old value of \p *Addr.
+uint32_t exchange(uint32_t *Addr, uint32_t V, int Ordering);
+
 } // namespace atomic
 
 } // namespace _OMP
diff --git a/openmp/libomptarget/DeviceRTL/include/Utils.h b/openmp/libomptarget/DeviceRTL/include/Utils.h
--- a/openmp/libomptarget/DeviceRTL/include/Utils.h
+++ b/openmp/libomptarget/DeviceRTL/include/Utils.h
@@ -12,6 +12,7 @@
 #ifndef OMPTARGET_DEVICERTL_UTILS_H
 #define OMPTARGET_DEVICERTL_UTILS_H
 
+#include "Synchronization.h"
 #include "Types.h"
 
 namespace _OMP {
@@ -72,6 +73,26 @@
   return V - V % Align;
 }
 
+namespace {
+/// Helper class to perform an action only once.
+///
+/// Using this is probably costly even if it is not executed. It should be
+/// guarded such that release mode execution will not be impacted.
+template <uint32_t ID> struct SingletonFlagImpl {
+
+  /// Each SingletonFlag instantiation with the same ID has an internal flag.
+  /// This function will return true if the flag was not set before, otherwise
+  /// it will return false. In either case the flag is set afterwards.
+  static bool testAndSet() {
+    static uint32_t DoOnceFlag = 0;
+    return 1 != atomic::exchange(&DoOnceFlag, 1, __ATOMIC_ACQ_REL);
+  }
+};
+} // namespace
+
+/// Helper to hide the __COUNTER__ use away.
+#define SingletonFlag SingletonFlagImpl<__COUNTER__>
+
 #define OMP_LIKELY(EXPR) __builtin_expect((bool)(EXPR), true)
 #define OMP_UNLIKELY(EXPR) __builtin_expect((bool)(EXPR), false)
 
diff --git a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
--- a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
@@ -320,6 +320,10 @@
   return impl::atomicAdd(Addr, V, Ordering);
 }
 
+uint32_t atomic::exchange(uint32_t *Addr, uint32_t V, int Ordering) {
+  return impl::atomicExchange(Addr, V, Ordering);
+}
+
 extern "C" {
 void __kmpc_ordered(IdentTy *Loc, int32_t TId) { FunctionTracingRAII(); }