Part of the Ewald energy was being returned even when that force group wasn't requested

platforms/cuda/src/CudaKernels.cpp

@@ -1495,12 +1495,13 @@ public:
             addEnergyKernel(addEnergyKernel), pmeEnergyBuffer(pmeEnergyBuffer), forceGroup(forceGroup) {
     }
     double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
-        if ((groups&(1<<forceGroup)) != 0)
+        if ((groups&(1<<forceGroup)) != 0) {
             cuStreamWaitEvent(cu.getCurrentStream(), event, 0);
-        if (includeEnergy) {
-            int bufferSize = pmeEnergyBuffer.getSize();
-            void* args[] = {&pmeEnergyBuffer.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &bufferSize};
-            cu.executeKernel(addEnergyKernel, args, bufferSize);
+            if (includeEnergy) {
+                int bufferSize = pmeEnergyBuffer.getSize();
+                void* args[] = {&pmeEnergyBuffer.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &bufferSize};
+                cu.executeKernel(addEnergyKernel, args, bufferSize);
+            }
         }
         return 0.0;
     }

platforms/opencl/src/OpenCLKernels.cpp

@@ -1502,9 +1502,9 @@ public:
             events[0] = event;
             event = cl::Event();
             cl.getQueue().enqueueWaitForEvents(events);
+            if (includeEnergy)
+                cl.executeKernel(addEnergyKernel, pmeEnergyBuffer.getSize());
         }
-        if (includeEnergy)
-            cl.executeKernel(addEnergyKernel, pmeEnergyBuffer.getSize());
         return 0.0;
     }
 private: