Created platform property for specifying whether to use the CPU based PME

platforms/cuda/include/CudaPlatform.h

@@ -81,6 +81,13 @@ public:
         static const std::string key = "CudaPrecision";
         return key;
     }
+    /**
+     * This is the name of the parameter for selecting whether to use the CPU based PME calculation.
+     */
+    static const std::string& UseCpuPme() {
+        static const std::string key = "UseCpuPme";
+        return key;
+    }
     /**
      * This is the name of the parameter for specifying the path to the CUDA compiler.
      */
@@ -100,14 +107,14 @@ public:
 class OPENMM_EXPORT_CUDA CudaPlatform::PlatformData {
 public:
     PlatformData(ContextImpl* context, const System& system, const std::string& deviceIndexProperty, const std::string& blockingProperty, const std::string& precisionProperty,
-            const std::string& compilerProperty, const std::string& tempProperty);
+            const std::string& cpuPmeProperty, const std::string& compilerProperty, const std::string& tempProperty);
     ~PlatformData();
     void initializeContexts(const System& system);
     void syncContexts();
     ContextImpl* context;
     std::vector<CudaContext*> contexts;
     std::vector<double> contextEnergy;
-    bool removeCM, peerAccessSupported;
+    bool removeCM, peerAccessSupported, useCpuPme;
     int cmMotionFrequency;
     int stepCount, computeForceCount;
     double time;

platforms/cuda/src/CudaKernels.cpp

@@ -1569,8 +1569,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
         if (cu.getUseDoublePrecision())
             pmeDefines["USE_DOUBLE_PRECISION"] = "1";
         CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaKernelSources::pme, pmeDefines);
-        bool useCpuPme = true;
-        if (useCpuPme) {
+        if (cu.getPlatformData().useCpuPme) {
             try {
                 cpuPme = getPlatform().createKernel(CalcPmeReciprocalForceKernel::Name(), *cu.getPlatformData().context);
                 cpuPme.getAs<CalcPmeReciprocalForceKernel>().initialize(gridSizeX, gridSizeY, gridSizeZ, numParticles, alpha);

platforms/cuda/src/CudaPlatform.cpp

@@ -86,6 +86,7 @@ CudaPlatform::CudaPlatform() {
     platformProperties.push_back(CudaDeviceIndex());
     platformProperties.push_back(CudaDeviceName());
     platformProperties.push_back(CudaUseBlockingSync());
+    platformProperties.push_back(UseCpuPme());
     platformProperties.push_back(CudaPrecision());
     platformProperties.push_back(CudaCompiler());
     platformProperties.push_back(CudaTempDirectory());
@@ -93,6 +94,7 @@ CudaPlatform::CudaPlatform() {
     setPropertyDefaultValue(CudaDeviceName(), "");
     setPropertyDefaultValue(CudaUseBlockingSync(), "true");
     setPropertyDefaultValue(CudaPrecision(), "single");
+    setPropertyDefaultValue(UseCpuPme(), "false");
 #ifdef _MSC_VER
     char* bindir = getenv("CUDA_BIN_PATH");
     string nvcc = (bindir == NULL ? "nvcc.exe" : string(bindir)+"\\nvcc.exe");
@@ -141,13 +143,20 @@ void CudaPlatform::contextCreated(ContextImpl& context, const map<string, string
             getPropertyDefaultValue(CudaUseBlockingSync()) : properties.find(CudaUseBlockingSync())->second);
     string precisionPropValue = (properties.find(CudaPrecision()) == properties.end() ?
             getPropertyDefaultValue(CudaPrecision()) : properties.find(CudaPrecision())->second);
+    string cpuPmePropValue = (properties.find(UseCpuPme()) == properties.end() ?
+            getPropertyDefaultValue(UseCpuPme()) : properties.find(UseCpuPme())->second);
     const string& compilerPropValue = (properties.find(CudaCompiler()) == properties.end() ?
             getPropertyDefaultValue(CudaCompiler()) : properties.find(CudaCompiler())->second);
     const string& tempPropValue = (properties.find(CudaTempDirectory()) == properties.end() ?
             getPropertyDefaultValue(CudaTempDirectory()) : properties.find(CudaTempDirectory())->second);
     transform(blockingPropValue.begin(), blockingPropValue.end(), blockingPropValue.begin(), ::tolower);
     transform(precisionPropValue.begin(), precisionPropValue.end(), precisionPropValue.begin(), ::tolower);
-    context.setPlatformData(new PlatformData(&context, context.getSystem(), devicePropValue, blockingPropValue, precisionPropValue, compilerPropValue, tempPropValue));
+    transform(cpuPmePropValue.begin(), cpuPmePropValue.end(), cpuPmePropValue.begin(), ::tolower);
+    vector<string> pmeKernelName;
+    pmeKernelName.push_back(CalcPmeReciprocalForceKernel::Name());
+    if (!supportsKernels(pmeKernelName))
+        cpuPmePropValue = "false";
+    context.setPlatformData(new PlatformData(&context, context.getSystem(), devicePropValue, blockingPropValue, precisionPropValue, cpuPmePropValue, compilerPropValue, tempPropValue));
 }
 
 void CudaPlatform::contextDestroyed(ContextImpl& context) const {
@@ -156,8 +165,9 @@ void CudaPlatform::contextDestroyed(ContextImpl& context) const {
 }
 
 CudaPlatform::PlatformData::PlatformData(ContextImpl* context, const System& system, const string& deviceIndexProperty, const string& blockingProperty, const string& precisionProperty,
-            const string& compilerProperty, const string& tempProperty) : context(context), removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
+            const string& cpuPmeProperty, const string& compilerProperty, const string& tempProperty) : context(context), removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
     bool blocking = (blockingProperty == "true");
+    useCpuPme = (cpuPmeProperty == "true");
     vector<string> devices;
     size_t searchPos = 0, nextPos;
     while ((nextPos = deviceIndexProperty.find_first_of(", ", searchPos)) != string::npos) {
@@ -189,6 +199,7 @@ CudaPlatform::PlatformData::PlatformData(ContextImpl* context, const System& sys
     propertyValues[CudaPlatform::CudaDeviceName()] = deviceName.str();
     propertyValues[CudaPlatform::CudaUseBlockingSync()] = blocking ? "true" : "false";
     propertyValues[CudaPlatform::CudaPrecision()] = precisionProperty;
+    propertyValues[CudaPlatform::UseCpuPme()] = useCpuPme ? "true" : "false";
     propertyValues[CudaPlatform::CudaCompiler()] = compilerProperty;
     propertyValues[CudaPlatform::CudaTempDirectory()] = tempProperty;
     contextEnergy.resize(contexts.size());

platforms/cuda/tests/TestCudaRandom.cpp

@@ -54,7 +54,7 @@ void testGaussian() {
     System system;
     for (int i = 0; i < numAtoms; i++)
         system.addParticle(1.0);
-    CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
+    CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"), "false",
             platform.getPropertyDefaultValue(CudaPlatform::CudaCompiler()), platform.getPropertyDefaultValue(CudaPlatform::CudaTempDirectory()));
     CudaContext& context = *platformData.contexts[0];
     context.initialize();

platforms/cuda/tests/TestCudaSort.cpp

@@ -64,7 +64,7 @@ void verifySorting(vector<float> array) {
 
     System system;
     system.addParticle(0.0);
-    CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
+    CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"), "false",
             platform.getPropertyDefaultValue(CudaPlatform::CudaCompiler()), platform.getPropertyDefaultValue(CudaPlatform::CudaTempDirectory()));
     CudaContext& context = *platformData.contexts[0];
     context.initialize();