diff --git a/src/c_wrapper/async.cpp b/src/c_wrapper/async.cpp
index abe9f9f9d3ced510b868a464ae1030b06dc7eb4a..69d76a16a5c8053add841367e2f88c24f019773e 100644
--- a/src/c_wrapper/async.cpp
+++ b/src/c_wrapper/async.cpp
@@ -1,4 +1,5 @@
 #include "async.h"
+#include "function.h"
 
 #include <queue>
 #include <thread>
@@ -14,7 +15,7 @@ private:
     std::mutex m_mutex;
     std::condition_variable m_cond;
 public:
-    T
+    PYOPENCL_INLINE T
     pop()
     {
         std::unique_lock<std::mutex> mlock(m_mutex);
@@ -25,10 +26,11 @@ public:
         m_queue.pop();
         return item;
     }
-    void
+    PYOPENCL_INLINE void
     push(const T &item)
     {
         {
+            // Sub scope for the lock
             std::unique_lock<std::mutex> mlock(m_mutex);
             m_queue.push(item);
         }
@@ -58,12 +60,12 @@ private:
         t.detach();
     }
 public:
-    void
+    PYOPENCL_INLINE void
     ensure_thread()
     {
         std::call_once(m_flag, &AsyncCaller::start_thread, this);
     }
-    void
+    PYOPENCL_INLINE void
     push(const std::function<void()> &func)
     {
         ensure_thread();
diff --git a/src/c_wrapper/async.h b/src/c_wrapper/async.h
index 321439258e22219bacefc70c60285baa06f83787..43fb60f59a7ef875375a4dd199e7cebdf31b8bef 100644
--- a/src/c_wrapper/async.h
+++ b/src/c_wrapper/async.h
@@ -5,7 +5,9 @@
 
 namespace pyopencl {
 
+// Start the helper thread
 void init_async();
+// Call @func in the helper thread
 void call_async(const std::function<void()> &func);
 
 }