Merge pull request #439 from peastman/master

Fixed bug using multiple devices with PME

platforms/cuda/src/CudaKernels.cpp

@@ -1520,7 +1520,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
     else
         dispersionCoefficient = 0.0;
     alpha = 0;
-    if (force.getNonbondedMethod() == NonbondedForce::Ewald && cu.getContextIndex() == 0) {
+    if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
         // Compute the Ewald parameters.
 
         int kmaxx, kmaxy, kmaxz;
@@ -1528,6 +1528,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
         defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
+        if (cu.getContextIndex() == 0) {
             ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
 
             // Create the reciprocal space kernels.
@@ -1547,7 +1548,8 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
             int elementSize = (cu.getUseDoublePrecision() ? sizeof(double2) : sizeof(float2));
             cosSinSums = new CudaArray(cu, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), elementSize, "cosSinSums");
         }
-    else if (force.getNonbondedMethod() == NonbondedForce::PME && cu.getContextIndex() == 0) {
+    }
+    else if (force.getNonbondedMethod() == NonbondedForce::PME) {
         // Compute the PME parameters.
 
         int gridSizeX, gridSizeY, gridSizeZ;
@@ -1560,6 +1562,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
         defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
+        if (cu.getContextIndex() == 0) {
             ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
             pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
             pmeDefines["NUM_ATOMS"] = cu.intToString(numParticles);
@@ -1698,6 +1701,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
                 }
             }
         }
+    }
     else
         ewaldSelfEnergy = 0.0;
 

platforms/cuda/tests/TestCudaNonbondedForce.cpp

@@ -748,7 +748,7 @@ void testChangingParameters() {
     ASSERT_EQUAL_TOL(cuState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
 }
 
-void testParallelComputation(bool useCutoff) {
+void testParallelComputation(NonbondedForce::NonbondedMethod method) {
     System system;
     const int numParticles = 200;
     for (int i = 0; i < numParticles; i++)
@@ -756,9 +756,9 @@ void testParallelComputation(bool useCutoff) {
     NonbondedForce* force = new NonbondedForce();
     for (int i = 0; i < numParticles; i++)
         force->addParticle(i%2-0.5, 0.5, 1.0);
-    if (useCutoff)
-        force->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
+    force->setNonbondedMethod(method);
     system.addForce(force);
+    system.setDefaultPeriodicBoxVectors(Vec3(5,0,0), Vec3(0,5,0), Vec3(0,0,5));
     OpenMM_SFMT::SFMT sfmt;
     init_gen_rand(0, sfmt);
     vector<Vec3> positions(numParticles);
@@ -877,8 +877,9 @@ int main(int argc, char* argv[]) {
         //testBlockInteractions(true);
         testDispersionCorrection();
         testChangingParameters();
-        testParallelComputation(false);
-        testParallelComputation(true);
+        testParallelComputation(NonbondedForce::NoCutoff);
+        testParallelComputation(NonbondedForce::Ewald);
+        testParallelComputation(NonbondedForce::PME);
         testSwitchingFunction(NonbondedForce::CutoffNonPeriodic);
         testSwitchingFunction(NonbondedForce::PME);
     }

platforms/opencl/src/OpenCLKernels.cpp

@@ -1492,7 +1492,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
     else
         dispersionCoefficient = 0.0;
     alpha = 0;
-    if (force.getNonbondedMethod() == NonbondedForce::Ewald && cl.getContextIndex() == 0) {
+    if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
         // Compute the Ewald parameters.
 
         int kmaxx, kmaxy, kmaxz;
@@ -1500,6 +1500,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["EWALD_ALPHA"] = cl.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cl.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
+        if (cl.getContextIndex() == 0) {
             ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
 
             // Create the reciprocal space kernels.
@@ -1516,7 +1517,8 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
             int elementSize = (cl.getUseDoublePrecision() ? sizeof(mm_double2) : sizeof(mm_float2));
             cosSinSums = new OpenCLArray(cl, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), elementSize, "cosSinSums");
         }
-    else if (force.getNonbondedMethod() == NonbondedForce::PME && cl.getContextIndex() == 0) {
+    }
+    else if (force.getNonbondedMethod() == NonbondedForce::PME) {
         // Compute the PME parameters.
 
         int gridSizeX, gridSizeY, gridSizeZ;
@@ -1527,6 +1529,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["EWALD_ALPHA"] = cl.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cl.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
+        if (cl.getContextIndex() == 0) {
             ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
             pmeDefines["PME_ORDER"] = cl.intToString(PmeOrder);
             pmeDefines["NUM_ATOMS"] = cl.intToString(numParticles);
@@ -1644,6 +1647,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                 }
             }
         }
+    }
     else
         ewaldSelfEnergy = 0.0;
 

platforms/opencl/tests/TestOpenCLNonbondedForce.cpp

@@ -751,7 +751,7 @@ void testChangingParameters() {
     ASSERT_EQUAL_TOL(clState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
 }
 
-void testParallelComputation(bool useCutoff) {
+void testParallelComputation(NonbondedForce::NonbondedMethod method) {
     System system;
     const int numParticles = 200;
     for (int i = 0; i < numParticles; i++)
@@ -759,9 +759,9 @@ void testParallelComputation(bool useCutoff) {
     NonbondedForce* force = new NonbondedForce();
     for (int i = 0; i < numParticles; i++)
         force->addParticle(i%2-0.5, 0.5, 1.0);
-    if (useCutoff)
-        force->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
+    force->setNonbondedMethod(method);
     system.addForce(force);
+    system.setDefaultPeriodicBoxVectors(Vec3(5,0,0), Vec3(0,5,0), Vec3(0,0,5));
     OpenMM_SFMT::SFMT sfmt;
     init_gen_rand(0, sfmt);
     vector<Vec3> positions(numParticles);
@@ -880,8 +880,9 @@ int main(int argc, char* argv[]) {
 //        testBlockInteractions(true);
         testDispersionCorrection();
         testChangingParameters();
-        testParallelComputation(false);
-        testParallelComputation(true);
+        testParallelComputation(NonbondedForce::NoCutoff);
+        testParallelComputation(NonbondedForce::Ewald);
+        testParallelComputation(NonbondedForce::PME);
         testSwitchingFunction(NonbondedForce::CutoffNonPeriodic);
         testSwitchingFunction(NonbondedForce::PME);
     }