ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Performance enhancements for USER-DPD

Browse Source 
1) For an ideal fluid benchmark: ~15% speedup and ~5% memory savings
2) Removes duCond and duMech, and replaces dvSSA with v_t0 (ghosts only)
3) Eliminates the need to transmit deltas and uCond/uMech in forward_comm
4) Reduces adds and loads for computing updated velocities from v + dvSSA
This commit is contained in:
Tim Mattox
2016-06-09 21:58:14 -04:00
parent 470cefbba1
commit cce37f71ff
6 changed files with 82 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/USER-DPD/atom_vec_dpd.cpp
View File

@ -76,8 +76,6 @@ void AtomVecDPD::grow(int n)
uChem = memory->grow(atom->uChem,nmax,"atom:uChem");
uCG = memory->grow(atom->uCG,nmax,"atom:uCG");
uCGnew = memory->grow(atom->uCGnew,nmax,"atom:uCGnew");
duCond = memory->grow(atom->duCond,nmax,"atom:duCond");
duMech = memory->grow(atom->duMech,nmax,"atom:duMech");
duChem = memory->grow(atom->duChem,nmax,"atom:duChem");
ssaAIR = memory->grow(atom->ssaAIR,nmax,"atom:ssaAIR");
@ -102,8 +100,6 @@ void AtomVecDPD::grow_reset()
uChem = atom->uChem;
uCG = atom->uCG;
uCGnew = atom->uCGnew;
duCond = atom->duCond;
duMech = atom->duMech;
duChem = atom->duChem;
ssaAIR = atom->ssaAIR;
}
@ -845,8 +841,6 @@ void AtomVecDPD::create_atom(int itype, double *coord)
uChem[nlocal] = 0.0;
uCG[nlocal] = 0.0;
uCGnew[nlocal] = 0.0;
duCond[nlocal] = 0.0;
duMech[nlocal] = 0.0;
duChem[nlocal] = 0.0;
ssaAIR[nlocal] = 1; /* coord2ssaAIR(x[nlocal]) */
@ -983,8 +977,6 @@ bigint AtomVecDPD::memory_usage()
if (atom->memcheck("uChem")) bytes += memory->usage(uChem,nmax);
if (atom->memcheck("uCG")) bytes += memory->usage(uCG,nmax);
if (atom->memcheck("uCGnew")) bytes += memory->usage(uCGnew,nmax);
if (atom->memcheck("duCond")) bytes += memory->usage(duCond,nmax);
if (atom->memcheck("duMech")) bytes += memory->usage(duMech,nmax);
if (atom->memcheck("duChem")) bytes += memory->usage(duChem,nmax);
if (atom->memcheck("ssaAIR")) bytes += memory->usage(ssaAIR,nmax);

2
src/USER-DPD/atom_vec_dpd.h
View File

@ -59,7 +59,7 @@ class AtomVecDPD : public AtomVec {
int write_data_hybrid(FILE *, double *);
bigint memory_usage();
double *uCond,*uMech,*uChem,*uCG,*uCGnew,*rho,*dpdTheta;
double *duCond,*duMech,*duChem;
double *duChem;
int *ssaAIR; // Shardlow Splitting Algorithm Active Interaction Region number
protected:

204
src/USER-DPD/fix_shardlow.cpp
View File

@ -96,10 +96,10 @@ FixShardlow::FixShardlow(LAMMPS *lmp, int narg, char **arg) :
pairDPDE = (PairDPDfdtEnergy *) force->pair_match("dpd/fdt/energy",1);
if(pairDPDE){
comm_forward = 10;
comm_forward = 3;
comm_reverse = 5;
} else {
comm_forward = 6;
comm_forward = 3;
comm_reverse = 3;
}
@ -175,8 +175,6 @@ void FixShardlow::initial_integrate(int vflag)
int *ssaAIR = atom->ssaAIR;
double *uCond = atom->uCond;
double *uMech = atom->uMech;
double *duCond = atom->duCond;
double *duMech = atom->duMech;
double *dpdTheta = atom->dpdTheta;
double kappa_ij, alpha_ij, theta_ij, gamma_ij, sigma_ij;
double vxi, vyi, vzi, vxj, vyj, vzj;
@ -204,24 +202,10 @@ void FixShardlow::initial_integrate(int vflag)
if(rcut >= bbx || rcut >= bby || rcut>= bbz )
error->all(FLERR,"Shardlow algorithm requires sub-domain length > 2*(rcut+skin). Either reduce the number of processors requested, or change the cutoff/skin\n");
// Allocate memory for the dvSSA arrays
dvSSA = new double*[nall];
for (ii = 0; ii < nall; ii++) {
dvSSA[ii] = new double[3];
}
// Allocate memory for v_t0 to hold the initial velocities for the ghosts
v_t0 = (double (*)[3]) memory->smalloc(sizeof(double)*3*nghost, "FixShardlow:v_t0");
// Zero the momenta
for (ii = 0; ii < nlocal; ii++) {
dvSSA[ii][0] = double(0.0);
dvSSA[ii][1] = double(0.0);
dvSSA[ii][2] = double(0.0);
if(pairDPDE){
duCond[ii] = double(0.0);
duMech[ii] = double(0.0);
}
}
// Communicate the updated momenta and velocities to all nodes
// Communicate the current velocities to all nodes
comm->forward_comm_fix(this);
// Define pointers to access the neighbor list
@ -240,6 +224,10 @@ void FixShardlow::initial_integrate(int vflag)
//Loop over all 14 directions (8 stages)
for (int idir = 1; idir <=8; idir++){
// Zero out the ghosts' uCond & uMech to be used as delta accumulators
memset(&(uCond[nlocal]), 0, sizeof(double)*nghost);
memset(&(uMech[nlocal]), 0, sizeof(double)*nghost);
// Loop over neighbors of my atoms
for (ii = 0; ii < inum; ii++) {
i = ilist[ii];
@ -248,6 +236,11 @@ void FixShardlow::initial_integrate(int vflag)
ytmp = x[i][1];
ztmp = x[i][2];
// load velocity for i from memory
vxi = v[i][0];
vyi = v[i][1];
vzi = v[i][2];
itype = type[i];
jlist = firstneigh[i];
jnum = numneigh[i];
@ -272,20 +265,20 @@ void FixShardlow::initial_integrate(int vflag)
cut = pairDPD->cut[itype][jtype];
}
// if (rsq < pairDPD->cutsq[itype][jtype]) {
// if (rsq < pairDPD->cutsq[itype][jtype])
if (rsq < cut2) {
r = sqrt(rsq);
if (r < EPSILON) continue; // r can be 0.0 in DPD systems
rinv = double(1.0)/r;
// Store the velocities from previous Shardlow step
vx0i = v[i][0] + dvSSA[i][0];
vy0i = v[i][1] + dvSSA[i][1];
vz0i = v[i][2] + dvSSA[i][2];
// Keep a copy of the velocities from previous Shardlow step
vx0i = vxi;
vy0i = vyi;
vz0i = vzi;
vx0j = v[j][0] + dvSSA[j][0];
vy0j = v[j][1] + dvSSA[j][1];
vz0j = v[j][2] + dvSSA[j][2];
vx0j = vxj = v[j][0];
vy0j = vyj = v[j][1];
vz0j = vzj = v[j][2];
// Compute the velocity difference between atom i and atom j
delvx = vx0i - vx0j;
@ -331,22 +324,22 @@ void FixShardlow::initial_integrate(int vflag)
massinv_i = double(1.0) / mass_i;
massinv_j = double(1.0) / mass_j;
// Update the delta velocity on i
dvSSA[i][0] += dpx*force->ftm2v*massinv_i;
dvSSA[i][1] += dpy*force->ftm2v*massinv_i;
dvSSA[i][2] += dpz*force->ftm2v*massinv_i;
// Update the velocity on i
vxi += dpx*force->ftm2v*massinv_i;
vyi += dpy*force->ftm2v*massinv_i;
vzi += dpz*force->ftm2v*massinv_i;
if (newton_pair || j < nlocal) {
// Update the delta velocity on j
dvSSA[j][0] -= dpx*force->ftm2v*massinv_j;
dvSSA[j][1] -= dpy*force->ftm2v*massinv_j;
dvSSA[j][2] -= dpz*force->ftm2v*massinv_j;
// Update the velocity on j
vxj -= dpx*force->ftm2v*massinv_j;
vyj -= dpy*force->ftm2v*massinv_j;
vzj -= dpz*force->ftm2v*massinv_j;
}
//ii. Compute the velocity diff
delvx = v[i][0] + dvSSA[i][0] - v[j][0] - dvSSA[j][0];
delvy = v[i][1] + dvSSA[i][1] - v[j][1] - dvSSA[j][1];
delvz = v[i][2] + dvSSA[i][2] - v[j][2] - dvSSA[j][2];
delvx = vxi - vxj;
delvy = vyi - vyj;
delvz = vzi - vzj;
dot = delx*delvx + dely*delvy + delz*delvz;
@ -363,16 +356,20 @@ void FixShardlow::initial_integrate(int vflag)
dpy = (factor_dpd*dot*rinv/force->ftm2v + factor_dpd1)*dely*rinv;
dpz = (factor_dpd*dot*rinv/force->ftm2v + factor_dpd1)*delz*rinv;
//Update the velocity change on i
dvSSA[i][0] += dpx*force->ftm2v*massinv_i;
dvSSA[i][1] += dpy*force->ftm2v*massinv_i;
dvSSA[i][2] += dpz*force->ftm2v*massinv_i;
// Update the velocity on i
vxi += dpx*force->ftm2v*massinv_i;
vyi += dpy*force->ftm2v*massinv_i;
vzi += dpz*force->ftm2v*massinv_i;
if (newton_pair || j < nlocal) {
//Update the velocity change on j
dvSSA[j][0] -= dpx*force->ftm2v*massinv_j;
dvSSA[j][1] -= dpy*force->ftm2v*massinv_j;
dvSSA[j][2] -= dpz*force->ftm2v*massinv_j;
// Update the velocity on j
vxj -= dpx*force->ftm2v*massinv_j;
vyj -= dpy*force->ftm2v*massinv_j;
vzj -= dpz*force->ftm2v*massinv_j;
// Store updated velocity for j
v[j][0] = vxj;
v[j][1] = vyj;
v[j][2] = vzj;
}
if(pairDPDE){
@ -385,20 +382,12 @@ void FixShardlow::initial_integrate(int vflag)
factor_dpd = kappa_ij*(double(1.0)/dpdTheta[i] - double(1.0)/dpdTheta[j])*wd*dt;
factor_dpd += randPair;
duCond[i] += factor_dpd;
uCond[i] += factor_dpd;
if (newton_pair || j < nlocal) {
duCond[j] -= factor_dpd;
uCond[j] -= factor_dpd;
}
// Compute uMech
vxi = v[i][0] + dvSSA[i][0];
vyi = v[i][1] + dvSSA[i][1];
vzi = v[i][2] + dvSSA[i][2];
vxj = v[j][0] + dvSSA[j][0];
vyj = v[j][1] + dvSSA[j][1];
vzj = v[j][2] + dvSSA[j][2];
dot1 = vxi*vxi + vyi*vyi + vzi*vzi;
dot2 = vxj*vxj + vyj*vyj + vzj*vzj;
dot3 = vx0i*vx0i + vy0i*vy0i + vz0i*vz0i;
@ -410,44 +399,29 @@ void FixShardlow::initial_integrate(int vflag)
dot4 = dot4*mass_j;
factor_dpd = double(0.25)*(dot1+dot2-dot3-dot4)/force->ftm2v;
duMech[i] -= factor_dpd;
uMech[i] -= factor_dpd;
if (newton_pair || j < nlocal) {
duMech[j] -= factor_dpd;
uMech[j] -= factor_dpd;
}
}
}
}
// store updated velocity for i
v[i][0] = vxi;
v[i][1] = vyi;
v[i][2] = vzi;
}
// Communicate the ghost delta velocities to the locally owned atoms
// Communicate the ghost deltas to the atom owners
comm->reverse_comm_fix(this);
// Zero the dv
for (ii = 0; ii < nlocal; ii++) {
// Shardlow update
v[ii][0] += dvSSA[ii][0];
v[ii][1] += dvSSA[ii][1];
v[ii][2] += dvSSA[ii][2];
dvSSA[ii][0] = double(0.0);
dvSSA[ii][1] = double(0.0);
dvSSA[ii][2] = double(0.0);
if(pairDPDE){
uCond[ii] += duCond[ii];
uMech[ii] += duMech[ii];
duCond[ii] = double(0.0);
duMech[ii] = double(0.0);
}
}
// Communicate the updated momenta and velocities to all nodes
// Communicate the updated velocities to all nodes
comm->forward_comm_fix(this);
} //End Loop over all directions For idir = Top, Top-Right, Right, Bottom-Right, Back
for (ii = 0; ii < nall; ii++) {
delete dvSSA[ii];
}
delete [] dvSSA;
memory->sfree(v_t0);
v_t0 = NULL;
}
/* ---------------------------------------------------------------------- */
@ -456,26 +430,13 @@ int FixShardlow::pack_forward_comm(int n, int *list, double *buf, int pbc_flag,
{
int ii,jj,m;
double **v = atom->v;
double *duCond = atom->duCond;
double *duMech = atom->duMech;
double *uCond = atom->uCond;
double *uMech = atom->uMech;
m = 0;
for (ii = 0; ii < n; ii++) {
jj = list[ii];
buf[m++] = dvSSA[jj][0];
buf[m++] = dvSSA[jj][1];
buf[m++] = dvSSA[jj][2];
buf[m++] = v[jj][0];
buf[m++] = v[jj][1];
buf[m++] = v[jj][2];
if(pairDPDE){
buf[m++] = duCond[jj];
buf[m++] = duMech[jj];
buf[m++] = uCond[jj];
buf[m++] = uMech[jj];
}
}
return m;
}
@ -485,27 +446,15 @@ int FixShardlow::pack_forward_comm(int n, int *list, double *buf, int pbc_flag,
void FixShardlow::unpack_forward_comm(int n, int first, double *buf)
{
int ii,m,last;
int nlocal = atom->nlocal;
double **v = atom->v;
double *duCond = atom->duCond;
double *duMech = atom->duMech;
double *uCond = atom->uCond;
double *uMech = atom->uMech;
m = 0;
last = first + n ;
for (ii = first; ii < last; ii++) {
dvSSA[ii][0] = buf[m++];
dvSSA[ii][1] = buf[m++];
dvSSA[ii][2] = buf[m++];
v[ii][0] = buf[m++];
v[ii][1] = buf[m++];
v[ii][2] = buf[m++];
if(pairDPDE){
duCond[ii] = buf[m++];
duMech[ii] = buf[m++];
uCond[ii] = buf[m++];
uMech[ii] = buf[m++];
}
v_t0[ii - nlocal][0] = v[ii][0] = buf[m++];
v_t0[ii - nlocal][1] = v[ii][1] = buf[m++];
v_t0[ii - nlocal][2] = v[ii][2] = buf[m++];
}
}
@ -514,18 +463,20 @@ void FixShardlow::unpack_forward_comm(int n, int first, double *buf)
int FixShardlow::pack_reverse_comm(int n, int first, double *buf)
{
int i,m,last;
double *duCond = atom->duCond;
double *duMech = atom->duMech;
int nlocal = atom->nlocal;
double **v = atom->v;
double *uCond = atom->uCond;
double *uMech = atom->uMech;
m = 0;
last = first + n;
for (i = first; i < last; i++) {
buf[m++] = dvSSA[i][0];
buf[m++] = dvSSA[i][1];
buf[m++] = dvSSA[i][2];
buf[m++] = v[i][0] - v_t0[i - nlocal][0];
buf[m++] = v[i][1] - v_t0[i - nlocal][1];
buf[m++] = v[i][2] - v_t0[i - nlocal][2];
if(pairDPDE){
buf[m++] = duCond[i];
buf[m++] = duMech[i];
buf[m++] = uCond[i]; // for ghosts, this is an accumulated delta
buf[m++] = uMech[i]; // for ghosts, this is an accumulated delta
}
}
return m;
@ -536,19 +487,20 @@ int FixShardlow::pack_reverse_comm(int n, int first, double *buf)
void FixShardlow::unpack_reverse_comm(int n, int *list, double *buf)
{
int i,j,m;
double *duCond = atom->duCond;
double *duMech = atom->duMech;
double **v = atom->v;
double *uCond = atom->uCond;
double *uMech = atom->uMech;
m = 0;
for (i = 0; i < n; i++) {
j = list[i];
dvSSA[j][0] += buf[m++];
dvSSA[j][1] += buf[m++];
dvSSA[j][2] += buf[m++];
v[j][0] += buf[m++];
v[j][1] += buf[m++];
v[j][2] += buf[m++];
if(pairDPDE){
duCond[j] += buf[m++];
duMech[j] += buf[m++];
uCond[j] += buf[m++]; // add in the accumulated delta
uMech[j] += buf[m++]; // add in the accumulated delta
}
}
}

2
src/USER-DPD/fix_shardlow.h
View File

@ -41,7 +41,7 @@ class FixShardlow : public Fix {
class PairDPDfdt *pairDPD;
class PairDPDfdtEnergy *pairDPDE;
double **dvSSA;
double (*v_t0)[3];
private:
class NeighList *list;

4
src/atom.cpp
View File

@ -94,7 +94,7 @@ Atom::Atom(LAMMPS *lmp) : Pointers(lmp)
// USER-DPD
uCond = uMech = uChem = uCG = uCGnew = NULL;
duCond = duMech = duChem = NULL;
duChem = NULL;
dpdTheta = NULL;
ssaAIR = NULL;
@ -285,8 +285,6 @@ Atom::~Atom()
memory->destroy(uChem);
memory->destroy(uCG);
memory->destroy(uCGnew);
memory->destroy(duCond);
memory->destroy(duMech);
memory->destroy(duChem);
memory->destroy(ssaAIR);

2
src/atom.h
View File

@ -88,7 +88,7 @@ class Atom : protected Pointers {
// USER-DPD package
double *uCond,*uMech,*uChem,*uCGnew,*uCG;
double *duCond,*duMech,*duChem;
double *duChem;
double *dpdTheta;
int nspecies_dpd;
int *ssaAIR; // Shardlow Splitting Algorithm Active Interaction Region number