diff --git a/extras/Hadrons/Modules/MScalarSUN/EMT.hpp b/extras/Hadrons/Modules/MScalarSUN/EMT.hpp
index fbc9c62a..025b7936 100644
--- a/extras/Hadrons/Modules/MScalarSUN/EMT.hpp
+++ b/extras/Hadrons/Modules/MScalarSUN/EMT.hpp
@@ -104,6 +104,7 @@ std::vector<std::string> TEMT<SImpl>::getInput(void)
             in.push_back(varName(par().improvement, mu, nu));
         }
     }
+    in.push_back(varName(par().kinetic, "sum"));
     in.push_back(varName(par().phiPow, 2));
     in.push_back(varName(par().phiPow, 4));
 
@@ -133,7 +134,6 @@ void TEMT<SImpl>::setup(void)
     {
         envCreateLat(ComplexField, varName(getName(), mu, nu));
     }
-    envTmpLat(ComplexField, "sumkin");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -158,15 +158,8 @@ void TEMT<SImpl>::execute(void)
     const unsigned int N       = SImpl::Group::Dimension;
     auto               &trphi2 = envGet(ComplexField, varName(par().phiPow, 2));
     auto               &trphi4 = envGet(ComplexField, varName(par().phiPow, 4));
+    auto               &sumkin = envGet(ComplexField, varName(par().kinetic, "sum"));
 
-    envGetTmp(ComplexField, sumkin);
-    sumkin = zero;
-    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
-    {
-        auto &trkin = envGet(ComplexField, varName(par().kinetic, mu, mu));
-
-        sumkin += trkin;
-    }
     for (unsigned int mu = 0; mu < env().getNd(); ++mu)
     for (unsigned int nu = mu; nu < env().getNd(); ++nu)
     {
diff --git a/extras/Hadrons/Modules/MScalarSUN/TrKinetic.hpp b/extras/Hadrons/Modules/MScalarSUN/TrKinetic.hpp
index a8c7f521..59aa27b8 100644
--- a/extras/Hadrons/Modules/MScalarSUN/TrKinetic.hpp
+++ b/extras/Hadrons/Modules/MScalarSUN/TrKinetic.hpp
@@ -110,6 +110,7 @@ std::vector<std::string> TTrKinetic<SImpl>::getOutput(void)
     {
         out.push_back(varName(getName(), mu, nu));
     }
+    out.push_back(varName(getName(), "sum"));
     
     return out;
 }
@@ -123,6 +124,7 @@ void TTrKinetic<SImpl>::setup(void)
     {
         envCreateLat(ComplexField, varName(getName(), mu, nu));
     }
+    envCreateLat(ComplexField, varName(getName(), "sum"));
     envTmp(std::vector<Field>, "der", 1, env().getNd(), env().getGrid());
 }
 
@@ -134,9 +136,11 @@ void TTrKinetic<SImpl>::execute(void)
                  << " derivative" << std::endl; 
 
     std::vector<Result> result;
-    auto                &phi = envGet(Field, par().field);
+    auto                &phi    = envGet(Field, par().field);
+    auto                &sumkin = envGet(ComplexField, varName(getName(), "sum"));
 
     envGetTmp(std::vector<Field>, der);
+    sumkin = zero;
     for (unsigned int mu = 0; mu < env().getNd(); ++mu)
     {
         dmu(der[mu], phi, mu, par().type);
@@ -147,8 +151,17 @@ void TTrKinetic<SImpl>::execute(void)
         auto &out = envGet(ComplexField, varName(getName(), mu, nu));
 
         out = -trace(der[mu]*der[nu]);
-        if (!par().output.empty())
+        if (mu == nu)
         {
+            sumkin += out;
+        }
+    }
+    if (!par().output.empty())
+    {
+        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+        for (unsigned int nu = mu; nu < env().getNd(); ++nu)
+        {
+            auto &out = envGet(ComplexField, varName(getName(), mu, nu));
             Result r;
 
             r.op    = "tr(d_" + std::to_string(mu) + "phi*d_" 
@@ -156,6 +169,13 @@ void TTrKinetic<SImpl>::execute(void)
             r.value = TensorRemove(sum(out));
             result.push_back(r);
         }
+        {
+            Result r;
+
+            r.op    = "sum_mu tr(d_mu phi*d_mu phi)";
+            r.value = TensorRemove(sum(sumkin));
+            result.push_back(r);
+        }
     }
     if (result.size() > 0)
     {