void CudaIntegrateRPMDStepKernel::computeForces(ContextImpl& context) {
// Compute forces from all groups that didn't have a specified contraction.
for (int i = 0; i < numCopies; i++) {
void* copyToContextArgs[] = {&velocities->getDevicePointer(), &cu.getVelm().getDevicePointer(), &positions->getDevicePointer(),
&cu.getPosq().getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(copyToContextKernel, copyToContextArgs, cu.getNumAtoms());
context.computeVirtualSites();
Vec3 initialBox[3];
context.getPeriodicBoxVectors(initialBox[0], initialBox[1], initialBox[2]);
context.updateContextState();
Vec3 finalBox[3];
context.getPeriodicBoxVectors(finalBox[0], finalBox[1], finalBox[2]);
if (initialBox[0] != finalBox[0] || initialBox[1] != finalBox[1] || initialBox[2] != finalBox[2])
throw OpenMMException("Standard barostats cannot be used with RPMDIntegrator. Use RPMDMonteCarloBarostat instead.");
context.calcForcesAndEnergy(true, false, groupsNotContracted);
void* copyFromContextArgs[] = {&cu.getForce().getDevicePointer(), &forces->getDevicePointer(), &cu.getVelm().getDevicePointer(),
&velocities->getDevicePointer(), &cu.getPosq().getDevicePointer(), &positions->getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(copyFromContextKernel, copyFromContextArgs, cu.getNumAtoms());
}
// Now loop over contractions and compute forces from them.
for (map<int, int>::const_iterator iter = groupsByCopies.begin(); iter != groupsByCopies.end(); ++iter) {
int copies = iter->first;
int groupFlags = iter->second;
// Find the contracted positions.
void* contractPosArgs[] = {&positions->getDevicePointer(), &contractedPositions->getDevicePointer()};
cu.executeKernel(positionContractionKernels[copies], contractPosArgs, numParticles*numCopies, workgroupSize);
// Compute forces.
for (int i = 0; i < copies; i++) {
void* copyToContextArgs[] = {&velocities->getDevicePointer(), &cu.getVelm().getDevicePointer(), &contractedPositions->getDevicePointer(),
&cu.getPosq().getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(copyToContextKernel, copyToContextArgs, cu.getNumAtoms());
context.computeVirtualSites();
context.calcForcesAndEnergy(true, false, groupFlags);
void* copyFromContextArgs[] = {&cu.getForce().getDevicePointer(), &contractedForces->getDevicePointer(), &cu.getVelm().getDevicePointer(),
&velocities->getDevicePointer(), &cu.getPosq().getDevicePointer(), &contractedPositions->getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(copyFromContextKernel, copyFromContextArgs, cu.getNumAtoms());
}
// Apply the forces to the original copies.
void* contractForceArgs[] = {&forces->getDevicePointer(), &contractedForces->getDevicePointer()};
cu.executeKernel(forceContractionKernels[copies], contractForceArgs, numParticles*numCopies, workgroupSize);
}
if (groupsByCopies.size() > 0) {
// Ensure the Context contains the positions from the last copy, since we'll assume that later.
int i = numCopies-1;
void* copyToContextArgs[] = {&velocities->getDevicePointer(), &cu.getVelm().getDevicePointer(), &positions->getDevicePointer(),
&cu.getPosq().getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(copyToContextKernel, copyToContextArgs, cu.getNumAtoms());
}
}
void ReferenceIntegrateRPMDStepKernel::computeForces(ContextImpl& context, const RPMDIntegrator& integrator) {
const int totalCopies = positions.size();
const int numParticles = positions[0].size();
vector<RealVec>& pos = extractPositions(context);
vector<RealVec>& vel = extractVelocities(context);
vector<RealVec>& f = extractForces(context);
// Compute forces from all groups that didn't have a specified contraction.
for (int i = 0; i < totalCopies; i++) {
pos = positions[i];
vel = velocities[i];
context.computeVirtualSites();
Vec3 initialBox[3];
context.getPeriodicBoxVectors(initialBox[0], initialBox[1], initialBox[2]);
context.updateContextState();
Vec3 finalBox[3];
context.getPeriodicBoxVectors(finalBox[0], finalBox[1], finalBox[2]);
if (initialBox[0] != finalBox[0] || initialBox[1] != finalBox[1] || initialBox[2] != finalBox[2]) {
// A barostat was applied during updateContextState(). Adjust the particle positions in all the
// other copies to match this one.
for (int j = 0; j < numParticles; j++) {
Vec3 delta = pos[j]-positions[i][j];
for (int k = 0; k < totalCopies; k++)
if (k != i)
positions[k][j] += delta;
}
}
positions[i] = pos;
velocities[i] = vel;
context.calcForcesAndEnergy(true, false, groupsNotContracted);
forces[i] = f;
}
// Now loop over contractions and compute forces from them.
for (map<int, int>::const_iterator iter = groupsByCopies.begin(); iter != groupsByCopies.end(); ++iter) {
int copies = iter->first;
int groupFlags = iter->second;
fftpack* shortFFT = contractionFFT[copies];
// Find the contracted positions.
vector<t_complex> q(totalCopies);
const RealOpenMM scale1 = 1.0/totalCopies;
for (int particle = 0; particle < numParticles; particle++) {
for (int component = 0; component < 3; component++) {
// Transform to the frequency domain, set high frequency components to zero, and transform back.
for (int k = 0; k < totalCopies; k++)
q[k] = t_complex(positions[k][particle][component], 0.0);
fftpack_exec_1d(fft, FFTPACK_FORWARD, &q[0], &q[0]);
if (copies > 1) {
int start = (copies+1)/2;
int end = totalCopies-copies+start;
for (int k = end; k < totalCopies; k++)
q[k-(totalCopies-copies)] = q[k];
fftpack_exec_1d(shortFFT, FFTPACK_BACKWARD, &q[0], &q[0]);
}
for (int k = 0; k < copies; k++)
contractedPositions[k][particle][component] = scale1*q[k].re;
}
}
// Compute forces.
for (int i = 0; i < copies; i++) {
pos = contractedPositions[i];
context.computeVirtualSites();
context.calcForcesAndEnergy(true, false, groupFlags);
contractedForces[i] = f;
}
// Apply the forces to the original copies.
const RealOpenMM scale2 = 1.0/copies;
for (int particle = 0; particle < numParticles; particle++) {
for (int component = 0; component < 3; component++) {
// Transform to the frequency domain, pad with zeros, and transform back.
for (int k = 0; k < copies; k++)
q[k] = t_complex(contractedForces[k][particle][component], 0.0);
if (copies > 1)
fftpack_exec_1d(shortFFT, FFTPACK_FORWARD, &q[0], &q[0]);
int start = (copies+1)/2;
int end = totalCopies-copies+start;
for (int k = end; k < totalCopies; k++)
q[k] = q[k-(totalCopies-copies)];
for (int k = start; k < end; k++)
q[k] = t_complex(0, 0);
fftpack_exec_1d(fft, FFTPACK_BACKWARD, &q[0], &q[0]);
for (int k = 0; k < totalCopies; k++)
forces[k][particle][component] += scale2*q[k].re;
}
}
}
}