improve trace output

src/c_wrapper/error.h

@@ -97,6 +97,12 @@ struct __CLPost<T, decltype((void)(std::declval<T>().post()))> {
     }
 };
 
+template<typename T, class = void>
+struct is_out_arg : std::false_type {};
+
+template<typename T>
+struct is_out_arg<T, enable_if_t<rm_ref_t<T>::is_out> > : std::true_type {};
+
 template<typename T, class = void>
 struct __CLPrintOut {
     static PYOPENCL_INLINE void
@@ -106,7 +112,7 @@ struct __CLPrintOut {
 };
 
 template<typename T>
-struct __CLPrintOut<T, enable_if_t<rm_ref_t<T>::is_out> > {
+struct __CLPrintOut<T, enable_if_t<is_out_arg<T>::value> > {
     static inline void
     call(T v, std::ostream &stm)
     {
@@ -125,6 +131,9 @@ struct __CLPrint {
         } else {
             first = false;
         }
+        if (is_out_arg<T>::value) {
+            stm << "{out}";
+        }
         v.print(stm);
     }
 };
@@ -162,7 +171,7 @@ class CLArgPack : public ArgPack<CLArg, Types...> {
         std::cerr << name << "(";
         __CLCall<__CLPrint, sizeof...(Types) - 1,
                  decltype(*that)>::call(*that, std::cerr, true);
-        std::cerr << ") = (" << res;
+        std::cerr << ") = (ret: " << res;
         __CLCall<__CLPrintOut, sizeof...(Types) - 1,
                  decltype(*that)>::call(*that, std::cerr);
         std::cerr << ")" << std::endl;

src/c_wrapper/utils.h

@@ -137,9 +137,7 @@ struct CLGenericArgPrinter<std::nullptr_t, void> {
 };
 
 template<typename T, class = void>
-struct CLGenericArgOut {
-    constexpr static bool value = false;
-};
+struct CLGenericArgOut : std::false_type {};
 
 template<typename T, class = void>
 class CLArg {
@@ -411,16 +409,12 @@ template<typename T>
 using pyopencl_buf_ele_t = typename rm_ref_t<T>::element_type;
 
 template<typename T, class = void>
-struct is_pyopencl_buf {
-    constexpr static bool value = false;
-};
+struct is_pyopencl_buf : std::false_type {};
 
 template<typename T>
 struct is_pyopencl_buf<
     T, enable_if_t<std::is_base_of<pyopencl_buf<pyopencl_buf_ele_t<T> >,
-                                   rm_ref_t<T> >::value> > {
-    constexpr static bool value = true;
-};
+                                   rm_ref_t<T> >::value> > : std::true_type {};
 
 template<ArgType AT, typename T>
 struct _ToArgBuffer<AT, T, enable_if_t<is_pyopencl_buf<T>::value &&