git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@14265 f3b2605a-c512-4ea7-a41b-209d697bcdaa

This commit is contained in:
sjplimp
2015-11-18 18:23:45 +00:00
parent e3c4db746c
commit f6c76f4623
9 changed files with 568 additions and 582 deletions

View File

@ -27,15 +27,15 @@ static Tersoff<PRECISION,ACC_PRECISION> TSMF;
// ---------------------------------------------------------------------------
// Allocate memory on host and device and copy constants to device
// ---------------------------------------------------------------------------
int tersoff_gpu_init(const int ntypes, const int inum, const int nall, const int max_nbors,
int tersoff_gpu_init(const int ntypes, const int inum, const int nall, const int max_nbors,
const double cell_size, int &gpu_mode, FILE *screen,
int* host_map, const int nelements, int*** host_elem2param, const int nparams,
const double* ts_lam1, const double* ts_lam2, const double* ts_lam3,
const double* ts_powermint, const double* ts_biga, const double* ts_bigb,
const double* ts_bigr, const double* ts_bigd,
const double* ts_c1, const double* ts_c2, const double* ts_c3, const double* ts_c4,
const double* ts_c, const double* ts_d, const double* ts_h,
const double* ts_gamma, const double* ts_beta,
const double* ts_bigr, const double* ts_bigd,
const double* ts_c1, const double* ts_c2, const double* ts_c3, const double* ts_c4,
const double* ts_c, const double* ts_d, const double* ts_h,
const double* ts_gamma, const double* ts_beta,
const double* ts_powern, const double* ts_cutsq) {
TSMF.clear();
gpu_mode=TSMF.device->gpu_mode();
@ -47,13 +47,9 @@ int tersoff_gpu_init(const int ntypes, const int inum, const int nall, const int
int procs_per_gpu=TSMF.device->procs_per_gpu();
// disable host/device split for now
if (gpu_split != 1.0)
if (gpu_split != 1.0)
return -8;
// disable multiple threads per atom for now
if (TSMF.device->threads_per_atom() != 1)
return -10;
TSMF.device->init_message(screen,"tersoff/gpu",first_gpu,last_gpu);
bool message=false;
@ -69,9 +65,9 @@ int tersoff_gpu_init(const int ntypes, const int inum, const int nall, const int
if (world_me==0)
init_ok=TSMF.init(ntypes, inum, nall, 300, cell_size, gpu_split, screen,
host_map, nelements, host_elem2param, nparams,
ts_lam1, ts_lam2, ts_lam3, ts_powermint,
ts_biga, ts_bigb, ts_bigr, ts_bigd,
ts_c1, ts_c2, ts_c3, ts_c4, ts_c, ts_d, ts_h,
ts_lam1, ts_lam2, ts_lam3, ts_powermint,
ts_biga, ts_bigb, ts_bigr, ts_bigd,
ts_c1, ts_c2, ts_c3, ts_c4, ts_c, ts_d, ts_h,
ts_gamma, ts_beta, ts_powern, ts_cutsq);
TSMF.device->world_barrier();
@ -90,13 +86,13 @@ int tersoff_gpu_init(const int ntypes, const int inum, const int nall, const int
if (gpu_rank==i && world_me!=0)
init_ok=TSMF.init(ntypes, inum, nall, 300, cell_size, gpu_split, screen,
host_map, nelements, host_elem2param, nparams,
ts_lam1, ts_lam2, ts_lam3, ts_powermint,
ts_biga, ts_bigb, ts_bigr, ts_bigd,
ts_c1, ts_c2, ts_c3, ts_c4, ts_c, ts_d, ts_h,
ts_lam1, ts_lam2, ts_lam3, ts_powermint,
ts_biga, ts_bigb, ts_bigr, ts_bigd,
ts_c1, ts_c2, ts_c3, ts_c4, ts_c, ts_d, ts_h,
ts_gamma, ts_beta, ts_powern, ts_cutsq);
TSMF.device->gpu_barrier();
if (message)
if (message)
fprintf(screen,"Done.\n");
}
if (message)
@ -121,12 +117,12 @@ int ** tersoff_gpu_compute_n(const int ago, const int inum_full,
return TSMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
subhi, tag, nspecial, special, eflag, vflag, eatom,
vatom, host_start, ilist, jnum, cpu_time, success);
}
}
void tersoff_gpu_compute(const int ago, const int nlocal, const int nall,
const int nlist, double **host_x, int *host_type,
int *ilist, int *numj, int **firstneigh, const bool eflag,
const bool vflag, const bool eatom, const bool vatom,
void tersoff_gpu_compute(const int ago, const int nlocal, const int nall,
const int nlist, double **host_x, int *host_type,
int *ilist, int *numj, int **firstneigh, const bool eflag,
const bool vflag, const bool eatom, const bool vatom,
int &host_start, const double cpu_time, bool &success) {
TSMF.compute(ago,nlocal,nall,nlist,host_x,host_type,ilist,numj,
firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success);