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

Only transfer data arrays that are needed in each kernel

Browse Source
This commit is contained in:
Trung Nguyen
2021-10-02 00:56:15 -05:00
parent f4d3d3a2b5
commit 5a6426bf96
2 changed files with 55 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
lib/gpu/lal_base_amoeba.cpp
View File

@ -350,8 +350,7 @@ int** BaseAmoebaT::precompute(const int ago, const int inum_full, const int nall
const bool eflag_in, const bool vflag_in, const bool eflag_in, const bool vflag_in,
const bool eatom, const bool vatom, int &host_start, const bool eatom, const bool vatom, int &host_start,
int **&ilist, int **&jnum, const double cpu_time, int **&ilist, int **&jnum, const double cpu_time,
bool &success, double *host_q, double *boxlo, bool &success, double *host_q, double *boxlo, double *prd) {
double *prd) {
acc_timers(); acc_timers();
if (eatom) _eflag=2; if (eatom) _eflag=2;
else if (eflag_in) _eflag=1; else if (eflag_in) _eflag=1;
@ -509,7 +508,7 @@ int** BaseAmoebaT::compute_udirect2b(const int ago, const int inum_full,
int** firstneigh = nullptr; int** firstneigh = nullptr;
cast_extra_data(host_amtype, host_amgroup, host_rpole, host_uind, host_uinp, host_pval); cast_extra_data(host_amtype, host_amgroup, host_rpole, host_uind, host_uinp, host_pval);
atom->add_extra_data(); atom->add_extra_data();
// ------------------- Resize _fieldp array ------------------------ // ------------------- Resize _fieldp array ------------------------
@ -647,30 +646,34 @@ void BaseAmoebaT::cast_extra_data(int* amtype, int* amgroup, double** rpole,
int n = 0; int n = 0;
int nstride = 4; int nstride = 4;
for (int i = 0; i < _nall; i++) { if (rpole) {
int idx = n+i*nstride; for (int i = 0; i < _nall; i++) {
pextra[idx] = rpole[i][0]; int idx = n+i*nstride;
pextra[idx+1] = rpole[i][1]; pextra[idx] = rpole[i][0];
pextra[idx+2] = rpole[i][2]; pextra[idx+1] = rpole[i][1];
pextra[idx+3] = rpole[i][3]; pextra[idx+2] = rpole[i][2];
} pextra[idx+3] = rpole[i][3];
}
n += nstride*_nall; n += nstride*_nall;
for (int i = 0; i < _nall; i++) { for (int i = 0; i < _nall; i++) {
int idx = n+i*nstride; int idx = n+i*nstride;
pextra[idx] = rpole[i][4]; pextra[idx] = rpole[i][4];
pextra[idx+1] = rpole[i][5]; pextra[idx+1] = rpole[i][5];
pextra[idx+2] = rpole[i][6]; pextra[idx+2] = rpole[i][6];
pextra[idx+3] = rpole[i][8]; pextra[idx+3] = rpole[i][8];
} }
n += nstride*_nall; n += nstride*_nall;
for (int i = 0; i < _nall; i++) { for (int i = 0; i < _nall; i++) {
int idx = n+i*nstride; int idx = n+i*nstride;
pextra[idx] = rpole[i][9]; pextra[idx] = rpole[i][9];
pextra[idx+1] = rpole[i][12]; pextra[idx+1] = rpole[i][12];
pextra[idx+2] = (numtyp)amtype[i]; pextra[idx+2] = (numtyp)amtype[i];
pextra[idx+3] = (numtyp)amgroup[i]; pextra[idx+3] = (numtyp)amgroup[i];
}
} else {
n += 2*nstride*_nall;
} }
n += nstride*_nall; n += nstride*_nall;

102
lib/gpu/lal_hippo.cpp
View File

@ -314,23 +314,12 @@ int** HippoT::compute_dispersion_real(const int ago, const int inum_full,
const double cpu_time, bool &success, const double cpu_time, bool &success,
const double aewald, const double off2_disp, const double aewald, const double off2_disp,
double *host_q, double *boxlo, double *prd) { double *host_q, double *boxlo, double *prd) {
// reallocate per-atom arrays, transfer data from the host
// and build the neighbor lists if needed
// NOTE:
// For now we invoke precompute() again here,
// to be able to turn on/off the udirect2b kernel (which comes before this)
// We only need to cast necesary data arrays from host to device here
// because the neighbor lists are rebuilt and other per-atom arrays
// (x, type) are ready on the device.
int** firstneigh = nullptr; // cast necessary data arrays from host to device
firstneigh = this->precompute(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, this->cast_extra_data(host_amtype, host_amgroup, host_rpole,
nullptr, nullptr, nullptr, sublo, subhi, tag, nullptr, nullptr, nullptr);
nspecial, special, nspecial15, special15, this->atom->add_extra_data();
eflag_in, vflag_in, eatom, vatom,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
this->_off2_disp = off2_disp; this->_off2_disp = off2_disp;
this->_aewald = aewald; this->_aewald = aewald;
@ -344,7 +333,7 @@ int** HippoT::compute_dispersion_real(const int ago, const int inum_full,
this->hd_balancer.stop_timer(); this->hd_balancer.stop_timer();
return firstneigh; // nbor->host_jlist.begin()-host_start; return nullptr; // nbor->host_jlist.begin()-host_start;
} }
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
@ -407,25 +396,11 @@ int** HippoT::compute_multipole_real(const int ago, const int inum_full,
const double aewald, const double felec, const double aewald, const double felec,
const double off2_mpole, double *host_q, const double off2_mpole, double *host_q,
double *boxlo, double *prd, void **tep_ptr) { double *boxlo, double *prd, void **tep_ptr) {
// reallocate per-atom arrays, transfer data from the host
// and build the neighbor lists if needed
// NOTE:
// For now we invoke precompute() again here,
// to be able to turn on/off the udirect2b kernel (which comes before this)
// Once all the kernels are ready, precompute() is needed only once
// in the first kernel in a time step.
// We only need to cast uind and uinp from host to device here
// if the neighbor lists are rebuilt and other per-atom arrays
// (x, type, amtype, amgroup, rpole) are ready on the device.
int** firstneigh = nullptr; // cast necessary data arrays from host to device
firstneigh = this->precompute(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, this->cast_extra_data(nullptr, nullptr, nullptr, nullptr, nullptr, host_pval);
nullptr, nullptr, host_pval, sublo, subhi, tag, this->atom->add_extra_data();
nspecial, special, nspecial15, special15,
eflag_in, vflag_in, eatom, vatom,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
// ------------------- Resize _tep array ------------------------ // ------------------- Resize _tep array ------------------------
@ -451,7 +426,7 @@ int** HippoT::compute_multipole_real(const int ago, const int inum_full,
this->_tep.update_host(this->_max_tep_size*4,false); this->_tep.update_host(this->_max_tep_size*4,false);
return firstneigh; // nbor->host_jlist.begin()-host_start; return nullptr; // nbor->host_jlist.begin()-host_start;
} }
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
@ -516,17 +491,11 @@ int** HippoT::compute_udirect2b(const int ago, const int inum_full,
const double aewald, const double off2_polar, const double aewald, const double off2_polar,
double *host_q, double *boxlo, double *prd, double *host_q, double *boxlo, double *prd,
void** fieldp_ptr) { void** fieldp_ptr) {
// reallocate per-atom arrays, transfer data from the host
// and build the neighbor lists if needed
int** firstneigh = nullptr; // all the necessary data arrays are already copied from host to device
firstneigh = this->precompute(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, this->cast_extra_data(nullptr, nullptr, nullptr, host_uind, host_uinp, host_pval);
host_uind, host_uinp, host_pval, sublo, subhi, tag, this->atom->add_extra_data();
nspecial, special, nspecial15, special15,
eflag_in, vflag_in, eatom, vatom,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
// ------------------- Resize _fieldp array ------------------------ // ------------------- Resize _fieldp array ------------------------
@ -544,7 +513,7 @@ int** HippoT::compute_udirect2b(const int ago, const int inum_full,
this->_fieldp.update_host(this->_max_fieldp_size*8,false); this->_fieldp.update_host(this->_max_fieldp_size*8,false);
return firstneigh; //nbor->host_jlist.begin()-host_start; return nullptr; //nbor->host_jlist.begin()-host_start;
} }
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
@ -608,17 +577,11 @@ int** HippoT::compute_umutual2b(const int ago, const int inum_full,
const double aewald, const double off2_polar, const double aewald, const double off2_polar,
double *host_q, double *boxlo, double *prd, double *host_q, double *boxlo, double *prd,
void** fieldp_ptr) { void** fieldp_ptr) {
// reallocate per-atom arrays, transfer extra data from the host
// and build the neighbor lists if needed
int** firstneigh = nullptr; // cast necessary data arrays from host to device
firstneigh = this->precompute(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, this->cast_extra_data(nullptr, nullptr, nullptr, host_uind, host_uinp, nullptr);
host_uind, host_uinp, host_pval, sublo, subhi, tag, this->atom->add_extra_data();
nspecial, special, nspecial15, special15,
eflag_in, vflag_in, eatom, vatom,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
// ------------------- Resize _fieldp array ------------------------ // ------------------- Resize _fieldp array ------------------------
@ -636,7 +599,7 @@ int** HippoT::compute_umutual2b(const int ago, const int inum_full,
this->_fieldp.update_host(this->_max_fieldp_size*8,false); this->_fieldp.update_host(this->_max_fieldp_size*8,false);
return firstneigh; //nbor->host_jlist.begin()-host_start; return nullptr; //nbor->host_jlist.begin()-host_start;
} }
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
@ -698,23 +661,12 @@ int** HippoT::compute_polar_real(const int ago, const int inum_full,
const double aewald, const double felec, const double aewald, const double felec,
const double off2_polar, double *host_q, const double off2_polar, double *host_q,
double *boxlo, double *prd, void **tep_ptr) { double *boxlo, double *prd, void **tep_ptr) {
// reallocate per-atom arrays, transfer data from the host
// and build the neighbor lists if needed
// NOTE:
// For now we invoke precompute() again here,
// to be able to turn on/off the udirect2b kernel (which comes before this)
// We only need to cast uind and uinp from host to device here
// if the neighbor lists are rebuilt and other per-atom arrays
// (x, type, amtype, amgroup, rpole) are ready on the device.
int** firstneigh = nullptr; // cast necessary data arrays from host to device
firstneigh = this->precompute(ago, inum_full, nall, host_x, host_type,
host_amtype, host_amgroup, host_rpole, //this->cast_extra_data(host_amtype, host_amgroup, host_rpole, host_uind, host_uinp, host_pval);
host_uind, host_uinp, host_pval, sublo, subhi, tag, this->cast_extra_data(nullptr, nullptr, nullptr, host_uind, host_uinp, nullptr);
nspecial, special, nspecial15, special15, this->atom->add_extra_data();
eflag_in, vflag_in, eatom, vatom,
host_start, ilist, jnum, cpu_time,
success, host_q, boxlo, prd);
// ------------------- Resize _tep array ------------------------ // ------------------- Resize _tep array ------------------------
@ -740,7 +692,7 @@ int** HippoT::compute_polar_real(const int ago, const int inum_full,
this->_tep.update_host(this->_max_tep_size*4,false); this->_tep.update_host(this->_max_tep_size*4,false);
return firstneigh; // nbor->host_jlist.begin()-host_start; return nullptr;
} }
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------