Merge pull request #1296 from peastman/bugs

Bug fixes

platforms/cuda/src/CudaContext.cpp

@@ -76,7 +76,7 @@ bool CudaContext::hasInitializedCuda = false;
 CudaContext::CudaContext(const System& system, int deviceIndex, bool useBlockingSync, const string& precision, const string& compiler,
         const string& tempDir, const std::string& hostCompiler, CudaPlatform::PlatformData& platformData) : system(system), currentStream(0),
         time(0.0), platformData(platformData), stepCount(0), computeForceCount(0), stepsSinceReorder(99999), contextIsValid(false), atomsWereReordered(false), hasCompilerKernel(false),
-        pinnedBuffer(NULL), posq(NULL), posqCorrection(NULL), velm(NULL), force(NULL), energyBuffer(NULL), integration(NULL), expression(NULL), bonded(NULL), nonbonded(NULL), thread(NULL) {
+        pinnedBuffer(NULL), posq(NULL), posqCorrection(NULL), velm(NULL), force(NULL), energyBuffer(NULL), atomIndexDevice(NULL), integration(NULL), expression(NULL), bonded(NULL), nonbonded(NULL), thread(NULL) {
     this->compiler = "\""+compiler+"\"";
     if (platformData.context != NULL) {
         try {
@@ -339,6 +339,8 @@ CudaContext::~CudaContext() {
         delete force;
     if (energyBuffer != NULL)
         delete energyBuffer;
+    if (atomIndexDevice != NULL)
+        delete atomIndexDevice;
     if (integration != NULL)
         delete integration;
     if (expression != NULL)

platforms/cuda/src/CudaKernels.cpp

@@ -1643,6 +1643,9 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
         defines["USE_EWALD"] = "1";
         if (cu.getContextIndex() == 0) {
             ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
+            char deviceName[100];
+            cuDeviceGetName(deviceName, 100, cu.getDevice());
+            usePmeStream = (string(deviceName) != "GeForce GTX 980"); // Using a separate stream is slower on GTX 980
             pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
             pmeDefines["NUM_ATOMS"] = cu.intToString(numParticles);
             pmeDefines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
@@ -1654,6 +1657,8 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
             pmeDefines["M_PI"] = cu.doubleToString(M_PI);
             if (cu.getUseDoublePrecision())
                 pmeDefines["USE_DOUBLE_PRECISION"] = "1";
+            if (usePmeStream)
+                pmeDefines["USE_PME_STREAM"] = "1";
             CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaKernelSources::pme, pmeDefines);
             if (cu.getPlatformData().useCpuPme) {
                 // Create the CPU PME kernel.
@@ -1713,11 +1718,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
                 
                 // Prepare for doing PME on its own stream.
                 
-                char deviceName[100];
-                cuDeviceGetName(deviceName, 100, cu.getDevice());
-                usePmeStream = (string(deviceName) != "GeForce GTX 980"); // Using a separate stream is slower on GTX 980
                 if (usePmeStream) {
-                    pmeDefines["USE_PME_STREAM"] = "1";
                     cuStreamCreate(&pmeStream, CU_STREAM_NON_BLOCKING);
                     if (useCudaFFT) {
                         cufftSetStream(fftForward, pmeStream);