Workaround for driver bugs that affect GTX 980

platforms/cuda/include/CudaKernels.h

@@ -632,7 +632,7 @@ private:
     std::vector<std::pair<int, int> > exceptionAtoms;
     double ewaldSelfEnergy, dispersionCoefficient, alpha;
     int interpolateForceThreads;
-    bool hasCoulomb, hasLJ;
+    bool hasCoulomb, hasLJ, usePmeStream;
     static const int PmeOrder = 5;
 };
 

platforms/cuda/src/CudaKernels.cpp

@@ -1457,9 +1457,11 @@ CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() {
     if (hasInitializedFFT) {
         cufftDestroy(fftForward);
         cufftDestroy(fftBackward);
+        if (usePmeStream) {
             cuStreamDestroy(pmeStream);
             cuEventDestroy(pmeSyncEvent);
         }
+    }
 }
 
 /**
@@ -1670,6 +1672,8 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
                 
                 // Prepare for doing PME on its own stream.
                 
+                usePmeStream = (cu.getComputeCapability() < 5.0); // A driver bug causes this to be very slow on GTX 980.
+                if (usePmeStream) {
                     cuStreamCreate(&pmeStream, CU_STREAM_NON_BLOCKING);
                     cufftSetStream(fftForward, pmeStream);
                     cufftSetStream(fftBackward, pmeStream);
@@ -1679,6 +1683,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
                         recipForceGroup = force.getForceGroup();
                     cu.addPreComputation(new SyncStreamPreComputation(pmeStream, pmeSyncEvent, recipForceGroup));
                     cu.addPostComputation(new SyncStreamPostComputation(pmeSyncEvent, recipForceGroup));
+                }
                 hasInitializedFFT = true;
 
                 // Initialize the b-spline moduli.
@@ -1795,6 +1800,7 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
         cu.executeKernel(ewaldForcesKernel, forcesArgs, cu.getNumAtoms());
     }
     if (directPmeGrid != NULL && includeReciprocal) {
+        if (usePmeStream)
             cu.setCurrentStream(pmeStream);
         void* gridIndexArgs[] = {&cu.getPosq().getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
         cu.executeKernel(pmeGridIndexKernel, gridIndexArgs, cu.getNumAtoms());
@@ -1832,9 +1838,11 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
         void* interpolateArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &directPmeGrid->getDevicePointer(),
                 cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer(), &pmeAtomGridIndex->getDevicePointer()};
         cu.executeKernel(pmeInterpolateForceKernel, interpolateArgs, cu.getNumAtoms(), 128);
+        if (usePmeStream) {
             cuEventRecord(pmeSyncEvent, pmeStream);
             cu.restoreDefaultStream();
         }
+    }
     double energy = (includeReciprocal ? ewaldSelfEnergy : 0.0);
     if (dispersionCoefficient != 0.0 && includeDirect) {
         double4 boxSize = cu.getPeriodicBoxSize();

platforms/opencl/src/OpenCLKernels.cpp

@@ -1611,6 +1611,12 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                 fft = new OpenCLFFT3D(cl, gridSizeX, gridSizeY, gridSizeZ);
                 string vendor = cl.getDevice().getInfo<CL_DEVICE_VENDOR>();
                 usePmeQueue = (vendor.size() >= 6 && vendor.substr(0, 6) == "NVIDIA");
+                if (cl.getDevice().getInfo<CL_DEVICE_EXTENSIONS>().find("cl_nv_device_attribute_query") != string::npos) {
+                    cl_uint computeCapabilityMajor;
+                    clGetDeviceInfo(cl.getDevice()(), 0x4000, sizeof(cl_uint), &computeCapabilityMajor, NULL); // CL_DEVICE_COMPUTE_CAPABILITY_MAJOR_NV
+                    if (computeCapabilityMajor == 5)
+                        usePmeQueue = false; // Workaround for driver bug that affects GTX 980.
+                }
                 if (usePmeQueue) {
                     pmeQueue = cl::CommandQueue(cl.getContext(), cl.getDevice());
                     int recipForceGroup = force.getReciprocalSpaceForceGroup();