Include Likwid Markers. Allow to switch between SP and DP floats.
This commit is contained in:
69
src/main.c
69
src/main.c
@@ -28,6 +28,8 @@
|
||||
#include <math.h>
|
||||
#include <float.h>
|
||||
|
||||
#include <likwid-marker.h>
|
||||
|
||||
#include <timing.h>
|
||||
#include <allocate.h>
|
||||
#include <neighbor.h>
|
||||
@@ -100,11 +102,13 @@ double reneighbour(
|
||||
double S, E;
|
||||
|
||||
S = getTimeStamp();
|
||||
LIKWID_MARKER_START("reneighbour");
|
||||
updateAtomsPbc(atom, param);
|
||||
setupPbc(atom, param);
|
||||
updatePbc(atom, param);
|
||||
/* sortAtom(); */
|
||||
buildNeighbor(atom, neighbor);
|
||||
LIKWID_MARKER_STOP("reneighbour");
|
||||
E = getTimeStamp();
|
||||
|
||||
return E-S;
|
||||
@@ -112,9 +116,9 @@ double reneighbour(
|
||||
|
||||
void initialIntegrate(Parameter *param, Atom *atom)
|
||||
{
|
||||
double* x = atom->x; double* y = atom->y; double* z = atom->z;
|
||||
double* fx = atom->fx; double* fy = atom->fy; double* fz = atom->fz;
|
||||
double* vx = atom->vx; double* vy = atom->vy; double* vz = atom->vz;
|
||||
MD_FLOAT* x = atom->x; MD_FLOAT* y = atom->y; MD_FLOAT* z = atom->z;
|
||||
MD_FLOAT* fx = atom->fx; MD_FLOAT* fy = atom->fy; MD_FLOAT* fz = atom->fz;
|
||||
MD_FLOAT* vx = atom->vx; MD_FLOAT* vy = atom->vy; MD_FLOAT* vz = atom->vz;
|
||||
|
||||
for(int i = 0; i < atom->Nlocal; i++) {
|
||||
vx[i] += param->dtforce * fx[i];
|
||||
@@ -128,8 +132,8 @@ void initialIntegrate(Parameter *param, Atom *atom)
|
||||
|
||||
void finalIntegrate(Parameter *param, Atom *atom)
|
||||
{
|
||||
double* fx = atom->fx; double* fy = atom->fy; double* fz = atom->fz;
|
||||
double* vx = atom->vx; double* vy = atom->vy; double* vz = atom->vz;
|
||||
MD_FLOAT* fx = atom->fx; MD_FLOAT* fy = atom->fy; MD_FLOAT* fz = atom->fz;
|
||||
MD_FLOAT* vx = atom->vx; MD_FLOAT* vy = atom->vy; MD_FLOAT* vz = atom->vz;
|
||||
|
||||
for(int i = 0; i < atom->Nlocal; i++) {
|
||||
vx[i] += param->dtforce * fx[i];
|
||||
@@ -142,12 +146,12 @@ double computeForce(Parameter *param, Atom *atom, Neighbor *neighbor)
|
||||
{
|
||||
int Nlocal = atom->Nlocal;
|
||||
int* neighs;
|
||||
double cutforcesq = param->cutforce * param->cutforce;
|
||||
double sigma6 = param->sigma6;
|
||||
double epsilon = param->epsilon;
|
||||
double* x = atom->x; double* y = atom->y; double* z = atom->z;
|
||||
double* fx = atom->fx; double* fy = atom->fy; double* fz = atom->fz;
|
||||
double S, E;
|
||||
MD_FLOAT cutforcesq = param->cutforce * param->cutforce;
|
||||
MD_FLOAT sigma6 = param->sigma6;
|
||||
MD_FLOAT epsilon = param->epsilon;
|
||||
MD_FLOAT* x = atom->x; MD_FLOAT* y = atom->y; MD_FLOAT* z = atom->z;
|
||||
MD_FLOAT* fx = atom->fx; MD_FLOAT* fy = atom->fy; MD_FLOAT* fz = atom->fz;
|
||||
MD_FLOAT S, E;
|
||||
|
||||
S = getTimeStamp();
|
||||
for(int i = 0; i < Nlocal; i++) {
|
||||
@@ -156,29 +160,31 @@ double computeForce(Parameter *param, Atom *atom, Neighbor *neighbor)
|
||||
fz[i] = 0.0;
|
||||
}
|
||||
|
||||
LIKWID_MARKER_START("force");
|
||||
|
||||
#pragma omp parallel for
|
||||
for(int i = 0; i < Nlocal; i++) {
|
||||
neighs = &neighbor->neighbors[i * neighbor->maxneighs];
|
||||
int numneighs = neighbor->numneigh[i];
|
||||
double xtmp = x[i];
|
||||
double ytmp = y[i];
|
||||
double ztmp = z[i];
|
||||
MD_FLOAT xtmp = x[i];
|
||||
MD_FLOAT ytmp = y[i];
|
||||
MD_FLOAT ztmp = z[i];
|
||||
|
||||
double fix = 0;
|
||||
double fiy = 0;
|
||||
double fiz = 0;
|
||||
MD_FLOAT fix = 0;
|
||||
MD_FLOAT fiy = 0;
|
||||
MD_FLOAT fiz = 0;
|
||||
|
||||
for(int k = 0; k < numneighs; k++) {
|
||||
int j = neighs[k];
|
||||
double delx = xtmp - x[j];
|
||||
double dely = ytmp - y[j];
|
||||
double delz = ztmp - z[j];
|
||||
double rsq = delx * delx + dely * dely + delz * delz;
|
||||
MD_FLOAT delx = xtmp - x[j];
|
||||
MD_FLOAT dely = ytmp - y[j];
|
||||
MD_FLOAT delz = ztmp - z[j];
|
||||
MD_FLOAT rsq = delx * delx + dely * dely + delz * delz;
|
||||
|
||||
if(rsq < cutforcesq) {
|
||||
double sr2 = 1.0 / rsq;
|
||||
double sr6 = sr2 * sr2 * sr2 * sigma6;
|
||||
double force = 48.0 * sr6 * (sr6 - 0.5) * sr2 * epsilon;
|
||||
MD_FLOAT sr2 = 1.0 / rsq;
|
||||
MD_FLOAT sr6 = sr2 * sr2 * sr2 * sigma6;
|
||||
MD_FLOAT force = 48.0 * sr6 * (sr6 - 0.5) * sr2 * epsilon;
|
||||
fix += delx * force;
|
||||
fiy += dely * force;
|
||||
fiz += delz * force;
|
||||
@@ -189,6 +195,7 @@ double computeForce(Parameter *param, Atom *atom, Neighbor *neighbor)
|
||||
fy[i] += fiy;
|
||||
fz[i] += fiz;
|
||||
}
|
||||
LIKWID_MARKER_STOP("force");
|
||||
E = getTimeStamp();
|
||||
|
||||
return E-S;
|
||||
@@ -214,6 +221,13 @@ int main (int argc, char** argv)
|
||||
Neighbor neighbor;
|
||||
Parameter param;
|
||||
|
||||
LIKWID_MARKER_INIT;
|
||||
#pragma omp parallel
|
||||
{
|
||||
LIKWID_MARKER_REGISTER("force");
|
||||
LIKWID_MARKER_REGISTER("reneighbour");
|
||||
LIKWID_MARKER_REGISTER("pbc");
|
||||
}
|
||||
init(¶m);
|
||||
|
||||
for(int i = 0; i < argc; i++)
|
||||
@@ -279,6 +293,12 @@ int main (int argc, char** argv)
|
||||
computeThermo(-1, ¶m, &atom);
|
||||
|
||||
printf(HLINE);
|
||||
#if PRECISION == 1
|
||||
printf("Using single precision floating point.\n");
|
||||
#else
|
||||
printf("Using double precision floating point.\n");
|
||||
#endif
|
||||
printf(HLINE);
|
||||
printf("System: %d atoms %d ghost atoms, Steps: %d\n", atom.Natoms, atom.Nghost, param.ntimes);
|
||||
printf("TOTAL %.2fs FORCE %.2fs NEIGH %.2fs REST %.2fs\n",
|
||||
timer[TOTAL], timer[FORCE], timer[NEIGH], timer[TOTAL]-timer[FORCE]-timer[NEIGH]);
|
||||
@@ -286,5 +306,6 @@ int main (int argc, char** argv)
|
||||
printf("Performance: %.2f million atom updates per second\n",
|
||||
1e-6 * (double) atom.Natoms * param.ntimes / timer[TOTAL]);
|
||||
|
||||
LIKWID_MARKER_CLOSE;
|
||||
return EXIT_SUCCESS;
|
||||
}
|
||||
|
Reference in New Issue
Block a user