ONly device time for code objects

src/targets/gpu/driver/perf.cpp

@@ -64,16 +64,19 @@ time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
         (void)i;
         host_time += time<milliseconds>(run);
     }
-    // Measure device time
+    // Measure device time only for code_object ops which support it
+    double device_time = 0.0;
+    if (op.name() == "gpu::code_object")
+    {
         gctx.enable_perf_measurement();
         run();
-    double device_time = 0.0;
         for(auto i : range(n))
         {
             (void)i;
             run();
             device_time += gctx.get_elapsed_ms();
         }
+    }
     return std::make_pair(host_time / n, device_time / n);
 }