diff --git a/ockl/src/gaaf.cl b/ockl/src/gaaf.cl
index cb5ed822..2821aeaa 100644
--- a/ockl/src/gaaf.cl
+++ b/ockl/src/gaaf.cl
@@ -17,16 +17,28 @@
 
 extern void __llvm_amdgcn_global_atomic_fadd_p1f32_f32(__global float *, float) __asm("llvm.amdgcn.global.atomic.fadd.p1f32.f32");
 
+__attribute__((target("atomic-fadd-insts"))) static void
+global_atomic_fadd(__global float *p, float v)
+{
+    __llvm_amdgcn_global_atomic_fadd_p1f32_f32(p, v);
+}
+
+static void
+generic_atomic_fadd(float *p, float v)
+{
+    atomic_uint *t = (atomic_uint *)p;
+    uint e = AL(t, memory_order_relaxed, memory_scope_device);
+    while (!AC(t, &e, AS_UINT(v + AS_FLOAT(e)), memory_order_relaxed, memory_order_relaxed, memory_scope_device))
+        ;
+}
+
 void
 __ockl_atomic_add_noret_f32(float *p, float v)
 {
     if (__oclc_ISA_version == 9008 && !__ockl_is_local_addr(p) && !__ockl_is_private_addr(p)) {
-        __llvm_amdgcn_global_atomic_fadd_p1f32_f32((__global float *)p, v);
+        global_atomic_fadd((__global float *)p, v);
     } else {
-        atomic_uint *t = (atomic_uint *)p;
-        uint e = AL(t, memory_order_relaxed, memory_scope_device);
-        while (!AC(t, &e, AS_UINT(v + AS_FLOAT(e)), memory_order_relaxed, memory_order_relaxed, memory_scope_device))
-            ;
+        generic_atomic_fadd(p, v);
     }
 }