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

Updating to master

Browse Source
This commit is contained in:
Stan Moore
2016-12-12 09:18:20 -07:00
parent a5e46e3e6a
commit a31b00965a
8 changed files with 114 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/KOKKOS/Install.sh
View File

@ -174,6 +174,8 @@ action pair_reax_c_kokkos.cpp pair_reax_c.cpp
action pair_reax_c_kokkos.h pair_reax_c.h
action pair_sw_kokkos.cpp pair_sw.cpp
action pair_sw_kokkos.h pair_sw.h
action pair_vashishta_kokkos.cpp pair_vashishta.cpp
action pair_vashishta_kokkos.h pair_vashishta.h
action pair_table_kokkos.cpp
action pair_table_kokkos.h
action pair_tersoff_kokkos.cpp pair_tersoff.cpp

14
src/KOKKOS/atom_vec_kokkos.h
View File

@ -83,8 +83,13 @@ class AtomVecKokkos : public AtomVec {
std::is_same<typename ViewType::execution_space,LMPDeviceType>::value,
Kokkos::CudaHostPinnedSpace,typename ViewType::memory_space>::type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > mirror_type;
if(buffer_size < src.capacity())
if (buffer_size == 0) {
buffer = Kokkos::kokkos_malloc<Kokkos::CudaHostPinnedSpace>(src.capacity());
buffer_size = src.capacity();
} else if (buffer_size < src.capacity()) {
buffer = Kokkos::kokkos_realloc<Kokkos::CudaHostPinnedSpace>(buffer,src.capacity());
buffer_size = src.capacity();
}
return mirror_type( buffer ,
src.dimension_0() ,
src.dimension_1() ,
@ -104,8 +109,13 @@ class AtomVecKokkos : public AtomVec {
std::is_same<typename ViewType::execution_space,LMPDeviceType>::value,
Kokkos::CudaHostPinnedSpace,typename ViewType::memory_space>::type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > mirror_type;
if(buffer_size < src.capacity())
if (buffer_size == 0) {
buffer = Kokkos::kokkos_malloc<Kokkos::CudaHostPinnedSpace>(src.capacity()*sizeof(typename ViewType::value_type));
buffer_size = src.capacity();
} else if (buffer_size < src.capacity()) {
buffer = Kokkos::kokkos_realloc<Kokkos::CudaHostPinnedSpace>(buffer,src.capacity()*sizeof(typename ViewType::value_type));
buffer_size = src.capacity();
}
mirror_type tmp_view( (typename ViewType::value_type*)buffer ,
src.dimension_0() ,
src.dimension_1() ,

39
src/KOKKOS/neighbor_kokkos.cpp
View File

@ -74,6 +74,19 @@ void NeighborKokkos::init()
atomKK = (AtomKokkos *) atom;
Neighbor::init();
// 1st time allocation of xhold
if (dist_check)
@ -358,8 +371,8 @@ void NeighborKokkos::init_topology() {
normally built with pair lists, but USER-CUDA separates them
------------------------------------------------------------------------- */
void NeighborKokkos::build_topology() {
if (device_flag) {
void NeighborKokkos::build_topology_kokkos() {
if (nlist_device) {
neighbond_device.build_topology_kk();
k_bondlist = neighbond_device.k_bondlist;
@ -376,22 +389,34 @@ void NeighborKokkos::build_topology() {
k_anglelist.modify<LMPDeviceType>();
k_dihedrallist.modify<LMPDeviceType>();
k_improperlist.modify<LMPDeviceType>();
} else {
// Transfer topology neighbor lists to Host for non-Kokkos styles
if (force->bond && force->bond->execution_space == Host)
k_bondlist.sync<LMPHostType>();
if (force->angle && force->angle->execution_space == Host)
k_anglelist.sync<LMPHostType>();
if (force->dihedral && force->dihedral->execution_space == Host)
k_dihedrallist.sync<LMPHostType>();
if (force->improper && force->improper->execution_space == Host)
k_improperlist.sync<LMPHostType>();
} else {
neighbond_host.build_topology_kk();
k_bondlist = neighbond_host.k_bondlist;
k_anglelist = neighbond_host.k_anglelist;
k_dihedrallist = neighbond_host.k_dihedrallist;
k_improperlist = neighbond_host.k_improperlist;
k_bondlist.sync<LMPHostType>();
k_anglelist.sync<LMPHostType>();
k_dihedrallist.sync<LMPHostType>();
k_improperlist.sync<LMPHostType>();
k_bondlist.modify<LMPHostType>();
k_anglelist.modify<LMPHostType>();
k_dihedrallist.modify<LMPHostType>();
k_improperlist.modify<LMPHostType>();
}
}
}

40
src/KOKKOS/pair_tersoff_kokkos.cpp
View File

@ -125,26 +125,26 @@ void PairTersoffKokkos<DeviceType>::setup_params()
for (i = 1; i <= n; i++)
for (j = 1; j <= n; j++)
for (k = 1; k <= n; k++) {
m = elem2param[i-1][j-1][k-1];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).gamma = params[m].gamma;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).c = params[m].c;
k_params.h_view(i,j,k).d = params[m].d;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
m = elem2param[map[i]][map[j]][map[k]];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).gamma = params[m].gamma;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).c = params[m].c;
k_params.h_view(i,j,k).d = params[m].d;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
}
k_params.template modify<LMPHostType>();

42
src/KOKKOS/pair_tersoff_mod_kokkos.cpp
View File

@ -124,27 +124,27 @@ void PairTersoffMODKokkos<DeviceType>::setup_params()
for (i = 1; i <= n; i++)
for (j = 1; j <= n; j++)
for (k = 1; k <= n; k++) {
m = elem2param[i-1][j-1][k-1];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
k_params.h_view(i,j,k).c5 = params[m].c5;
k_params.h_view(i,j,k).ca1 = params[m].ca1;
k_params.h_view(i,j,k).ca4 = params[m].ca4;
k_params.h_view(i,j,k).powern_del = params[m].powern_del;
m = elem2param[map[i]][map[j]][map[k]];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
k_params.h_view(i,j,k).c5 = params[m].c5;
k_params.h_view(i,j,k).ca1 = params[m].ca1;
k_params.h_view(i,j,k).ca4 = params[m].ca4;
k_params.h_view(i,j,k).powern_del = params[m].powern_del;
}
k_params.template modify<LMPHostType>();

48
src/KOKKOS/pair_tersoff_zbl_kokkos.cpp
View File

@ -135,30 +135,30 @@ void PairTersoffZBLKokkos<DeviceType>::setup_params()
for (i = 1; i <= n; i++)
for (j = 1; j <= n; j++)
for (k = 1; k <= n; k++) {
m = elem2param[i-1][j-1][k-1];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).gamma = params[m].gamma;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).c = params[m].c;
k_params.h_view(i,j,k).d = params[m].d;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
k_params.h_view(i,j,k).Z_i = params[m].Z_i;
k_params.h_view(i,j,k).Z_j = params[m].Z_j;
k_params.h_view(i,j,k).ZBLcut = params[m].ZBLcut;
k_params.h_view(i,j,k).ZBLexpscale = params[m].ZBLexpscale;
m = elem2param[map[i]][map[j]][map[k]];
k_params.h_view(i,j,k).powerm = params[m].powerm;
k_params.h_view(i,j,k).gamma = params[m].gamma;
k_params.h_view(i,j,k).lam3 = params[m].lam3;
k_params.h_view(i,j,k).c = params[m].c;
k_params.h_view(i,j,k).d = params[m].d;
k_params.h_view(i,j,k).h = params[m].h;
k_params.h_view(i,j,k).powern = params[m].powern;
k_params.h_view(i,j,k).beta = params[m].beta;
k_params.h_view(i,j,k).lam2 = params[m].lam2;
k_params.h_view(i,j,k).bigb = params[m].bigb;
k_params.h_view(i,j,k).bigr = params[m].bigr;
k_params.h_view(i,j,k).bigd = params[m].bigd;
k_params.h_view(i,j,k).lam1 = params[m].lam1;
k_params.h_view(i,j,k).biga = params[m].biga;
k_params.h_view(i,j,k).cutsq = params[m].cutsq;
k_params.h_view(i,j,k).c1 = params[m].c1;
k_params.h_view(i,j,k).c2 = params[m].c2;
k_params.h_view(i,j,k).c3 = params[m].c3;
k_params.h_view(i,j,k).c4 = params[m].c4;
k_params.h_view(i,j,k).Z_i = params[m].Z_i;
k_params.h_view(i,j,k).Z_j = params[m].Z_j;
k_params.h_view(i,j,k).ZBLcut = params[m].ZBLcut;
k_params.h_view(i,j,k).ZBLexpscale = params[m].ZBLexpscale;
}
k_params.template modify<LMPHostType>();

2
src/KOKKOS/region_block_kokkos.h
View File

@ -33,10 +33,10 @@ template<class DeviceType>
class RegBlockKokkos : public RegBlock {
friend class FixPour;
public:
typedef DeviceType device_type;
typedef ArrayTypes<DeviceType> AT;
public:
RegBlockKokkos(class LAMMPS *, int, char **);
~RegBlockKokkos();
void match_all_kokkos(int, DAT::t_int_1d);

4
src/finish.cpp
View File

@ -856,7 +856,7 @@ void mpi_timings(const char *label, Timer *t, enum Timer::ttype tt,
time_cpu = tmp/nprocs*100.0;
// % variance from the average as measure of load imbalance
if (time > 1.0e-10)
if ((time_sq/time - time) > 1.0e-10)
time_sq = sqrt(time_sq/time - time)*100.0;
else
time_sq = 0.0;
@ -908,7 +908,7 @@ void omp_times(FixOMP *fix, const char *label, enum Timer::ttype which,
time_std /= nthreads;
time_total /= nthreads;
if (time_avg > 1.0e-10)
if ((time_std/time_avg -time_avg) > 1.0e-10)
time_std = sqrt(time_std/time_avg - time_avg)*100.0;
else
time_std = 0.0;