Created a CudaPlatform property for specifying whether CUDA should sync or...

Created a CudaPlatform property for specifying whether CUDA should sync or spin loop while waiting for results

platforms/cuda/include/CudaPlatform.h

@@ -64,6 +64,13 @@ public:
         static const std::string key = "CudaDevice";
         return key;
     }
+    /**
+     * This is the name of the parameter for selecting whether CUDA should sync or spin loop while waiting for results.
+     */
+    static const std::string& CudaUseBlockingSync() {
+        static const std::string key = "CudaUseBlockingSync";
+        return key;
+    }
 private:
     CudaStreamFactory defaultStreamFactory;
 };

platforms/cuda/src/CudaPlatform.cpp

@@ -63,7 +63,9 @@ CudaPlatform::CudaPlatform() {
     registerKernelFactory(CalcKineticEnergyKernel::Name(), factory);
     registerKernelFactory(RemoveCMMotionKernel::Name(), factory);
     platformProperties.push_back(CudaDevice());
+    platformProperties.push_back(CudaUseBlockingSync());
     setPropertyDefaultValue(CudaDevice(), "0");
+    setPropertyDefaultValue(CudaUseBlockingSync(), "false");
 }
 
 bool CudaPlatform::supportsDoublePrecision() const {
@@ -92,7 +94,7 @@ void CudaPlatform::contextCreated(ContextImpl& context) const {
     if (devicePropValue.length() > 0)
         stringstream(devicePropValue) >> device;
     int numParticles = context.getSystem().getNumParticles();
-    _gpuContext* gpu = (_gpuContext*) gpuInit(numParticles, device);
+    _gpuContext* gpu = (_gpuContext*) gpuInit(numParticles, device, getPropertyDefaultValue(CudaUseBlockingSync()) == "true");
     context.setPlatformData(new PlatformData(gpu));
 }
 
@@ -107,4 +109,5 @@ CudaPlatform::PlatformData::PlatformData(_gpuContext* gpu) : gpu(gpu), removeCM(
     stringstream device;
     device << gpu->device;
     propertyValues[CudaPlatform::CudaDevice()] = device.str();
+    propertyValues[CudaPlatform::CudaUseBlockingSync()] = (gpu->useBlockingSync ? "true" : "false");
 }

platforms/cuda/src/kernels/gpu.cpp

@@ -1118,7 +1118,7 @@ bool gpuIsAvailable()
 }
 
 extern "C"
-void* gpuInit(int numAtoms, unsigned int device)
+void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync)
 {
     gpuContext gpu = new _gpuContext;
     int LRFSize = 0;
@@ -1134,6 +1134,9 @@ void* gpuInit(int numAtoms, unsigned int device)
         cudaSetDevice(device); // Ignore errors
     status = cudaGetDevice(&gpu->device);
     RTERROR(status, "Error getting CUDA device")
+    status = cudaSetDeviceFlags(useBlockingSync ? cudaDeviceBlockingSync : cudaDeviceScheduleAuto);
+    RTERROR(status, "Error setting device flags")
+    gpu->useBlockingSync = useBlockingSync;
 
     // Determine kernel call configuration
     cudaDeviceProp deviceProp;

platforms/cuda/src/kernels/gputypes.h

@@ -59,6 +59,7 @@ struct _gpuContext {
     //have to be repeatedly passed around
     int natoms;
     int device;
+    bool useBlockingSync;
     gpuAtomType* gpAtomTable;
     int gAtomTypes;
     cudaGmxSimulation sim;
@@ -206,7 +207,7 @@ extern "C"
 void gpuInitializeRandoms(gpuContext gpu);
 
 extern "C"
-void* gpuInit(int numAtoms, unsigned int device);
+void* gpuInit(int numAtoms, unsigned int device = 0, bool useBlockingSync = false);
 
 extern "C"
 void gpuSetLangevinIntegrationParameters(gpuContext gpu, float tau, float deltaT, float temperature, float errorTol);