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

git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@9581 f3b2605a-c512-4ea7-a41b-209d697bcdaa

Browse Source
This commit is contained in:
sjplimp
2013-03-06 16:44:43 +00:00
parent f4b458219d
commit adb129fe99
4 changed files with 41 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
src/USER-OMP/fix_omp.cpp
View File

@ -68,7 +68,7 @@ static int get_tid()
FixOMP::FixOMP(LAMMPS *lmp, int narg, char **arg) FixOMP::FixOMP(LAMMPS *lmp, int narg, char **arg)
: Fix(lmp, narg, arg), : Fix(lmp, narg, arg),
thr(NULL), last_omp_style(NULL), last_pair_hybrid(NULL), thr(NULL), last_omp_style(NULL), last_pair_hybrid(NULL),
_nthr(-1), _neighbor(true), _mixed(false) _nthr(-1), _neighbor(true), _mixed(false), _reduced(true)
{ {
if ((narg < 4) || (narg > 7)) error->all(FLERR,"Illegal package omp command"); if ((narg < 4) || (narg > 7)) error->all(FLERR,"Illegal package omp command");
if (strcmp(arg[1],"all") != 0) error->all(FLERR,"fix OMP has to operate on group 'all'"); if (strcmp(arg[1],"all") != 0) error->all(FLERR,"fix OMP has to operate on group 'all'");
@ -87,8 +87,10 @@ FixOMP::FixOMP(LAMMPS *lmp, int narg, char **arg)
if (nthreads < 1) if (nthreads < 1)
error->all(FLERR,"Illegal number of OpenMP threads requested"); error->all(FLERR,"Illegal number of OpenMP threads requested");
int reset_thr = 0;
if (nthreads != comm->nthreads) { if (nthreads != comm->nthreads) {
#if defined(_OPENMP) #if defined(_OPENMP)
reset_thr = 1;
omp_set_num_threads(nthreads); omp_set_num_threads(nthreads);
#endif #endif
comm->nthreads = nthreads; comm->nthreads = nthreads;
@ -110,25 +112,21 @@ FixOMP::FixOMP(LAMMPS *lmp, int narg, char **arg)
// print summary of settings // print summary of settings
if (comm->me == 0) { if (comm->me == 0) {
const char * const nmode = _neighbor ? "OpenMP capable" : "serial"; const char * const nmode = _neighbor ? "multi-threaded" : "serial";
const char * const kmode = _mixed ? "mixed" : "double"; const char * const kmode = _mixed ? "mixed" : "double";
if (screen) { if (screen) {
fprintf(screen," reset %d OpenMP thread(s) per MPI task\n", nthreads); if (reset_thr)
fprintf(screen," using %s neighbor list subroutines\n", nmode); fprintf(screen,"set %d OpenMP thread(s) per MPI task\n", nthreads);
if (_mixed) fprintf(screen,"using %s neighbor list subroutines\n", nmode);
fputs(" using mixed precision OpenMP force kernels where available\n", screen); fprintf(screen,"prefer %s precision OpenMP force kernels\n", kmode);
else
fputs(" using double precision OpenMP force kernels\n", screen);
} }
if (logfile) { if (logfile) {
fprintf(logfile," reset %d OpenMP thread(s) per MPI task\n", nthreads); if (reset_thr)
fprintf(logfile," using %s neighbor list subroutines\n", nmode); fprintf(logfile,"set %d OpenMP thread(s) per MPI task\n", nthreads);
if (_mixed) fprintf(logfile,"using %s neighbor list subroutines\n", nmode);
fputs(" using mixed precision OpenMP force kernels where available\n", logfile); fprintf(logfile,"prefer %s precision OpenMP force kernels\n", kmode);
else
fputs(" using double precision OpenMP force kernels\n", logfile);
} }
} }
@ -137,7 +135,6 @@ FixOMP::FixOMP(LAMMPS *lmp, int narg, char **arg)
// encourage the OS to use storage that is "close" to each thread's CPU. // encourage the OS to use storage that is "close" to each thread's CPU.
thr = new ThrData *[nthreads]; thr = new ThrData *[nthreads];
_nthr = nthreads; _nthr = nthreads;
_clearforce = true;
#if defined(_OPENMP) #if defined(_OPENMP)
#pragma omp parallel default(none) #pragma omp parallel default(none)
#endif #endif
@ -188,8 +185,9 @@ int FixOMP::setmask()
void FixOMP::init() void FixOMP::init()
{ {
if (strstr(update->integrate_style,"respa") != NULL) if ((strstr(update->integrate_style,"respa") != NULL)
error->all(FLERR,"Cannot use r-RESPA with /omp styles"); && (strstr(update->integrate_style,"respa/omp") == NULL))
error->all(FLERR,"Need to use respa/omp for r-RESPA with /omp styles");
int check_hybrid, kspace_split; int check_hybrid, kspace_split;
last_pair_hybrid = NULL; last_pair_hybrid = NULL;
@ -282,7 +280,6 @@ void FixOMP::init()
fprintf(logfile,"Last active /omp style is %s_style %s\n", fprintf(logfile,"Last active /omp style is %s_style %s\n",
last_force_name, last_omp_name); last_force_name, last_omp_name);
} else { } else {
_clearforce = false;
if (screen) if (screen)
fprintf(screen,"No /omp style for force computation currently active\n"); fprintf(screen,"No /omp style for force computation currently active\n");
if (logfile) if (logfile)
@ -322,18 +319,16 @@ void FixOMP::pre_force(int)
double *de = atom->de; double *de = atom->de;
double *drho = atom->drho; double *drho = atom->drho;
if (_clearforce) {
#if defined(_OPENMP) #if defined(_OPENMP)
#pragma omp parallel default(none) shared(f,torque,erforce,de,drho) #pragma omp parallel default(none) shared(f,torque,erforce,de,drho)
#endif #endif
{ {
const int tid = get_tid(); const int tid = get_tid();
thr[tid]->check_tid(tid); thr[tid]->check_tid(tid);
thr[tid]->init_force(nall,f,torque,erforce,de,drho); thr[tid]->init_force(nall,f,torque,erforce,de,drho);
} } // end of omp parallel region
} else {
thr[0]->init_force(nall,f,torque,erforce,de,drho); _reduced = false;
}
} }
/* ---------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- */

12
src/USER-OMP/fix_omp.h
View File

@ -29,6 +29,7 @@ class ThrData;
class FixOMP : public Fix { class FixOMP : public Fix {
friend class ThrOMP; friend class ThrOMP;
friend class RespaOMP;
public: public:
FixOMP(class LAMMPS *, int, char **); FixOMP(class LAMMPS *, int, char **);
@ -51,6 +52,8 @@ class FixOMP : public Fix {
// to do the general force reduction // to do the general force reduction
void *last_pair_hybrid; // pointer to the pair style that needs void *last_pair_hybrid; // pointer to the pair style that needs
// to call virial_fdot_compute() // to call virial_fdot_compute()
// signal that an /omp style did the force reduction. needed by respa/omp
void did_reduce() { _reduced = true; }
public: public:
ThrData *get_thr(int tid) { return thr[tid]; } ThrData *get_thr(int tid) { return thr[tid]; }
@ -58,12 +61,13 @@ class FixOMP : public Fix {
bool get_neighbor() const { return _neighbor; } bool get_neighbor() const { return _neighbor; }
bool get_mixed() const { return _mixed; } bool get_mixed() const { return _mixed; }
bool get_reduced() const { return _reduced; }
private: private:
int _nthr; // number of currently active ThrData object int _nthr; // number of currently active ThrData objects
bool _neighbor; // en/disable threads for neighbor list construction bool _neighbor; // en/disable threads for neighbor list construction
bool _clearforce; // whether to clear per thread data objects bool _mixed; // whether to prefer mixed precision compute kernels
bool _mixed; // whether to use a mixed precision compute kernel bool _reduced; // whether forces have been reduced for this step
void set_neighbor_omp(); void set_neighbor_omp();
}; };

11
src/USER-OMP/thr_omp.cpp
View File

@ -183,6 +183,7 @@ void ThrOMP::reduce_thr(void *style, const int eflag, const int vflag,
// pair_style hybrid will compute fdotr for us // pair_style hybrid will compute fdotr for us
// but we first need to reduce the forces // but we first need to reduce the forces
data_reduce_thr(&(f[0][0]), nall, nthreads, 3, tid); data_reduce_thr(&(f[0][0]), nall, nthreads, 3, tid);
fix->did_reduce();
need_force_reduce = 0; need_force_reduce = 0;
} }
} }
@ -380,7 +381,11 @@ void ThrOMP::reduce_thr(void *style, const int eflag, const int vflag,
break; break;
case THR_KSPACE: case THR_KSPACE:
// nothing to do. XXX need to add support for per-atom info // nothing to do. XXX may need to add support for per-atom info
break;
case THR_INTGR:
// nothing to do
break; break;
default: default:
@ -389,8 +394,10 @@ void ThrOMP::reduce_thr(void *style, const int eflag, const int vflag,
} }
if (style == fix->last_omp_style) { if (style == fix->last_omp_style) {
if (need_force_reduce) if (need_force_reduce) {
data_reduce_thr(&(f[0][0]), nall, nthreads, 3, tid); data_reduce_thr(&(f[0][0]), nall, nthreads, 3, tid);
fix->did_reduce();
}
if (lmp->atom->torque) if (lmp->atom->torque)
data_reduce_thr(&(lmp->atom->torque[0][0]), nall, nthreads, 3, tid); data_reduce_thr(&(lmp->atom->torque[0][0]), nall, nthreads, 3, tid);

4
src/USER-OMP/thr_omp.h
View File

@ -58,8 +58,8 @@ class ThrOMP {
enum {THR_NONE=0,THR_PAIR=1,THR_BOND=1<<1,THR_ANGLE=1<<2, enum {THR_NONE=0,THR_PAIR=1,THR_BOND=1<<1,THR_ANGLE=1<<2,
THR_DIHEDRAL=1<<3,THR_IMPROPER=1<<4,THR_KSPACE=1<<5, THR_DIHEDRAL=1<<3,THR_IMPROPER=1<<4,THR_KSPACE=1<<5,
THR_CHARMM=1<<6, /*THR_PROXY=1<<7,*/ THR_HYBRID=1<<8, THR_CHARMM=1<<6, /*THR_PROXY=1<<7,THR_HYBRID=1<<8, */
THR_FIX=1<<9}; THR_FIX=1<<9,THR_INTGR=1<<10};
protected: protected:
// extra ev_tally setup work for threaded styles // extra ev_tally setup work for threaded styles