Merge pull request #4645 from peastman/customcpp

Prevent deadlock using CustomCPPForceImpl with multiple GPUs

platforms/common/src/CommonKernels.cpp

@@ -8245,15 +8245,29 @@ void CommonCalcCustomCPPForceKernel::initialize(const System& system, CustomCPPF
     addForcesKernel->addArg(cc.getLongForceBuffer());
     addForcesKernel->addArg(cc.getAtomIndexArray());
     forceGroupFlag = (1<<force.getOwner().getForceGroup());
+    if (cc.getNumContexts() == 1) {
         cc.addPreComputation(new StartCalculationPreComputation(*this));
         cc.addPostComputation(new AddForcesPostComputation(*this));
+    }
 }
 
 double CommonCalcCustomCPPForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
+    if (cc.getNumContexts() == 1) {
         // This method does nothing.  The actual calculation is started by the pre-computation, continued on
         // the worker thread, and finished by the post-computation.
 
         return 0;
+    }
+
+    // When using multiple GPUs, this method is itself called from the worker thread.
+    // Submitting additional tasks and waiting for them to complete would lead to
+    // a deadlock.
+
+    if (cc.getContextIndex() != 0)
+        return 0.0;
+    contextImpl.getPositions(positionsVec);
+    executeOnWorkerThread(includeForces);
+    return addForces(includeForces, includeEnergy, -1);
 }
 
 void CommonCalcCustomCPPForceKernel::beginComputation(bool includeForces, bool includeEnergy, int groups) {
@@ -8290,6 +8304,7 @@ double CommonCalcCustomCPPForceKernel::addForces(bool includeForces, bool includ
 
     // Wait until executeOnWorkerThread() is finished.
     
+    if (cc.getNumContexts() == 1)
         cc.getWorkThread().flush();
 
     // Add in the forces.

platforms/cuda/tests/TestCudaCustomCPPForce.cpp

@@ -33,5 +33,7 @@
 #include "TestCustomCPPForce.h"
 
 void runPlatformTests() {
+    platform.setPropertyDefaultValue("DeviceIndex", "0,0");
+    testForce();
 }
 

platforms/hip/tests/TestHipCustomCPPForce.cpp

@@ -33,4 +33,6 @@
 #include "TestCustomCPPForce.h"
 
 void runPlatformTests() {
+    platform.setPropertyDefaultValue("DeviceIndex", "0,0");
+    testForce();
 }

platforms/opencl/tests/TestOpenCLCustomCPPForce.cpp

@@ -33,5 +33,7 @@
 #include "TestCustomCPPForce.h"
 
 void runPlatformTests() {
+    platform.setPropertyDefaultValue("DeviceIndex", "0,0");
+    testForce();
 }