Disable PME stream when using CPU PME (#3148)

platforms/cuda/src/CudaKernels.cpp

@@ -725,7 +725,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
             }
             char deviceName[100];
             cuDeviceGetName(deviceName, 100, cu.getDevice());
-            usePmeStream = (!cu.getPlatformData().disablePmeStream && string(deviceName) != "GeForce GTX 980"); // Using a separate stream is slower on GTX 980
+            usePmeStream = (!cu.getPlatformData().disablePmeStream && !cu.getPlatformData().useCpuPme && string(deviceName) != "GeForce GTX 980"); // Using a separate stream is slower on GTX 980
             map<string, string> pmeDefines;
             pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
             pmeDefines["NUM_ATOMS"] = cu.intToString(numParticles);

platforms/opencl/src/OpenCLKernels.cpp

@@ -803,7 +803,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                     dispersionFft = new OpenCLFFT3D(cl, dispersionGridSizeX, dispersionGridSizeY, dispersionGridSizeZ, true);
                 string vendor = cl.getDevice().getInfo<CL_DEVICE_VENDOR>();
                 bool isNvidia = (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA");
-                usePmeQueue = (!cl.getPlatformData().disablePmeStream && cl.getSupports64BitGlobalAtomics() && isNvidia);
+                usePmeQueue = (!cl.getPlatformData().disablePmeStream && !cl.getPlatformData().useCpuPme && cl.getSupports64BitGlobalAtomics() && isNvidia);
                 if (usePmeQueue) {
                     pmeDefines["USE_PME_STREAM"] = "1";
                     pmeQueue = cl::CommandQueue(cl.getContext(), cl.getDevice());