void CudaIntegrateRPMDStepKernel::execute(ContextImpl& context, const RPMDIntegrator& integrator, bool forcesAreValid) {
cu.setAsCurrent();
CudaIntegrationUtilities& integration = cu.getIntegrationUtilities();
// Loop over copies and compute the force on each one.
if (!forcesAreValid)
computeForces(context);
// Apply the PILE-L thermostat.
bool useDoublePrecision = (cu.getUseDoublePrecision() || cu.getUseMixedPrecision());
double dt = integrator.getStepSize();
float dtFloat = (float) dt;
void* dtPtr = (useDoublePrecision ? (void*) &dt : (void*) &dtFloat);
double kT = integrator.getTemperature()*BOLTZ;
float kTFloat = (float) kT;
void* kTPtr = (useDoublePrecision ? (void*) &kT : (void*) &kTFloat);
double friction = integrator.getFriction();
float frictionFloat = (float) friction;
void* frictionPtr = (useDoublePrecision ? (void*) &friction : (void*) &frictionFloat);
int randomIndex = integration.prepareRandomNumbers(numParticles*numCopies);
void* pileArgs[] = {&velocities->getDevicePointer(), &integration.getRandom().getDevicePointer(), &randomIndex, dtPtr, kTPtr, frictionPtr};
cu.executeKernel(pileKernel, pileArgs, numParticles*numCopies, workgroupSize);
// Update positions and velocities.
void* stepArgs[] = {&positions->getDevicePointer(), &velocities->getDevicePointer(), &forces->getDevicePointer(), dtPtr, kTPtr};
cu.executeKernel(stepKernel, stepArgs, numParticles*numCopies, workgroupSize);
// Calculate forces based on the updated positions.
computeForces(context);
// Update velocities.
void* velocitiesArgs[] = {&velocities->getDevicePointer(), &forces->getDevicePointer(), dtPtr};
cu.executeKernel(velocitiesKernel, velocitiesArgs, numParticles*numCopies, workgroupSize);
// Apply the PILE-L thermostat again.
randomIndex = integration.prepareRandomNumbers(numParticles*numCopies);
cu.executeKernel(pileKernel, pileArgs, numParticles*numCopies, workgroupSize);
// Update the time and step count.
cu.setTime(cu.getTime()+dt);
cu.setStepCount(cu.getStepCount()+1);
cu.reorderAtoms();
if (cu.getAtomsWereReordered() && cu.getNonbondedUtilities().getUsePeriodic()) {
// Atoms may have been translated into a different periodic box, so apply
// the same translation to all the beads.
int i = numCopies-1;
void* args[] = {&positions->getDevicePointer(), &cu.getPosq().getDevicePointer(), &cu.getAtomIndexArray().getDevicePointer(), &i};
cu.executeKernel(translateKernel, args, cu.getNumAtoms());
}
}
void ReferenceIntegrateRPMDStepKernel::execute(ContextImpl& context, const RPMDIntegrator& integrator, bool forcesAreValid) {
const int numCopies = positions.size();
const int numParticles = positions[0].size();
const RealOpenMM dt = integrator.getStepSize();
const RealOpenMM halfdt = 0.5*dt;
const System& system = context.getSystem();
vector<RealVec>& pos = extractPositions(context);
vector<RealVec>& vel = extractVelocities(context);
vector<RealVec>& f = extractForces(context);
// Loop over copies and compute the force on each one.
if (!forcesAreValid)
computeForces(context, integrator);
// Apply the PILE-L thermostat.
vector<t_complex> v(numCopies);
vector<t_complex> q(numCopies);
const RealOpenMM hbar = 1.054571628e-34*AVOGADRO/(1000*1e-12);
const RealOpenMM scale = 1.0/sqrt((RealOpenMM) numCopies);
const RealOpenMM nkT = numCopies*BOLTZ*integrator.getTemperature();
const RealOpenMM twown = 2.0*nkT/hbar;
const RealOpenMM c1_0 = exp(-halfdt*integrator.getFriction());
const RealOpenMM c2_0 = sqrt(1.0-c1_0*c1_0);
if (integrator.getApplyThermostat()) {
for (int particle = 0; particle < numParticles; particle++) {
if (system.getParticleMass(particle) == 0.0)
continue;
const RealOpenMM c3_0 = c2_0*sqrt(nkT/system.getParticleMass(particle));
for (int component = 0; component < 3; component++) {
for (int k = 0; k < numCopies; k++)
v[k] = t_complex(scale*velocities[k][particle][component], 0.0);
fftpack_exec_1d(fft, FFTPACK_FORWARD, &v[0], &v[0]);
// Apply a local Langevin thermostat to the centroid mode.
v[0].re = v[0].re*c1_0 + c3_0*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber();
// Use critical damping white noise for the remaining modes.
for (int k = 1; k <= numCopies/2; k++) {
const bool isCenter = (numCopies%2 == 0 && k == numCopies/2);
const RealOpenMM wk = twown*sin(k*M_PI/numCopies);
const RealOpenMM c1 = exp(-2.0*wk*halfdt);
const RealOpenMM c2 = sqrt((1.0-c1*c1)/2) * (isCenter ? sqrt(2.0) : 1.0);
const RealOpenMM c3 = c2*sqrt(nkT/system.getParticleMass(particle));
RealOpenMM rand1 = c3*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber();
RealOpenMM rand2 = (isCenter ? 0.0 : c3*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber());
v[k] = v[k]*c1 + t_complex(rand1, rand2);
if (k < numCopies-k)
v[numCopies-k] = v[numCopies-k]*c1 + t_complex(rand1, -rand2);
}
fftpack_exec_1d(fft, FFTPACK_BACKWARD, &v[0], &v[0]);
for (int k = 0; k < numCopies; k++)
velocities[k][particle][component] = scale*v[k].re;
}
}
}
// Update velocities.
for (int i = 0; i < numCopies; i++)
for (int j = 0; j < numParticles; j++)
if (system.getParticleMass(j) != 0.0)
velocities[i][j] += forces[i][j]*(halfdt/system.getParticleMass(j));
// Evolve the free ring polymer by transforming to the frequency domain.
for (int particle = 0; particle < numParticles; particle++) {
if (system.getParticleMass(particle) == 0.0)
continue;
for (int component = 0; component < 3; component++) {
for (int k = 0; k < numCopies; k++) {
q[k] = t_complex(scale*positions[k][particle][component], 0.0);
v[k] = t_complex(scale*velocities[k][particle][component], 0.0);
}
fftpack_exec_1d(fft, FFTPACK_FORWARD, &q[0], &q[0]);
fftpack_exec_1d(fft, FFTPACK_FORWARD, &v[0], &v[0]);
q[0] += v[0]*dt;
for (int k = 1; k < numCopies; k++) {
const RealOpenMM wk = twown*sin(k*M_PI/numCopies);
const RealOpenMM wt = wk*dt;
const RealOpenMM coswt = cos(wt);
const RealOpenMM sinwt = sin(wt);
const RealOpenMM wm = wk*system.getParticleMass(particle);
const t_complex vprime = v[k]*coswt - q[k]*(wk*sinwt); // Advance velocity from t to t+dt
q[k] = v[k]*(sinwt/wk) + q[k]*coswt; // Advance position from t to t+dt
v[k] = vprime;
}
fftpack_exec_1d(fft, FFTPACK_BACKWARD, &q[0], &q[0]);
fftpack_exec_1d(fft, FFTPACK_BACKWARD, &v[0], &v[0]);
for (int k = 0; k < numCopies; k++) {
positions[k][particle][component] = scale*q[k].re;
velocities[k][particle][component] = scale*v[k].re;
}
}
}
// Calculate forces based on the updated positions.
computeForces(context, integrator);
// Update velocities.
for (int i = 0; i < numCopies; i++)
for (int j = 0; j < numParticles; j++)
if (system.getParticleMass(j) != 0.0)
velocities[i][j] += forces[i][j]*(halfdt/system.getParticleMass(j));
// Apply the PILE-L thermostat again.
if (integrator.getApplyThermostat()) {
for (int particle = 0; particle < numParticles; particle++) {
if (system.getParticleMass(particle) == 0.0)
continue;
const RealOpenMM c3_0 = c2_0*sqrt(nkT/system.getParticleMass(particle));
for (int component = 0; component < 3; component++) {
for (int k = 0; k < numCopies; k++)
v[k] = t_complex(scale*velocities[k][particle][component], 0.0);
fftpack_exec_1d(fft, FFTPACK_FORWARD, &v[0], &v[0]);
// Apply a local Langevin thermostat to the centroid mode.
v[0].re = v[0].re*c1_0 + c3_0*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber();
// Use critical damping white noise for the remaining modes.
for (int k = 1; k <= numCopies/2; k++) {
const bool isCenter = (numCopies%2 == 0 && k == numCopies/2);
const RealOpenMM wk = twown*sin(k*M_PI/numCopies);
const RealOpenMM c1 = exp(-2.0*wk*halfdt);
const RealOpenMM c2 = sqrt((1.0-c1*c1)/2) * (isCenter ? sqrt(2.0) : 1.0);
const RealOpenMM c3 = c2*sqrt(nkT/system.getParticleMass(particle));
RealOpenMM rand1 = c3*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber();
RealOpenMM rand2 = (isCenter ? 0.0 : c3*SimTKOpenMMUtilities::getNormallyDistributedRandomNumber());
v[k] = v[k]*c1 + t_complex(rand1, rand2);
if (k < numCopies-k)
v[numCopies-k] = v[numCopies-k]*c1 + t_complex(rand1, -rand2);
}
fftpack_exec_1d(fft, FFTPACK_BACKWARD, &v[0], &v[0]);
for (int k = 0; k < numCopies; k++)
velocities[k][particle][component] = scale*v[k].re;
}
}
}
// Update the time.
context.setTime(context.getTime()+dt);
}