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

Reverted the block size tuning, which caused bugs for low atom counts (will revisit later)

Browse Source
This commit is contained in:
Trung Nguyen
2022-11-04 13:45:59 -05:00
parent 2f1f7ee0fa
commit a3cc0e8432
3 changed files with 26 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
lib/gpu/lal_amoeba.cpp
View File

@ -278,14 +278,16 @@ int AmoebaT::polar_real(const int eflag, const int vflag) {
int nbor_pitch=this->nbor->nbor_pitch();
// Compute the block size and grid size to keep all cores busy
const int BX=this->block_size();
const int GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
/*
const int cus = this->device->gpu->cus();
int BX=this->block_size();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
while (GX < cus) {
while (GX < cus && GX > 1) {
BX /= 2;
GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
}
*/
this->time_pair.start();
// Build the short neighbor list if not done yet

22
lib/gpu/lal_base_amoeba.cpp
View File

@ -727,14 +727,16 @@ int BaseAmoebaT::fphi_uind() {
return 0;
// Compute the block size and grid size to keep all cores busy
const int cus = device->gpu->cus();
int BX=block_size();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
while (GX < cus) {
const int BX=block_size();
const int GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
/*
const int cus = this->device->gpu->cus();
while (GX < cus && GX > 1) {
BX /= 2;
GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
}
*/
time_pair.start();
int ngridxy = _ngridx * _ngridy;
k_fphi_uind.set_size(GX,BX);
@ -793,14 +795,16 @@ int BaseAmoebaT::fphi_mpole() {
int nbor_pitch=nbor->nbor_pitch();
// Compute the block size and grid size to keep all cores busy
const int BX=block_size();
const int GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
/*
const int cus = device->gpu->cus();
int BX=block_size();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
while (GX < cus) {
while (GX < cus && GX > 1) {
BX /= 2;
GX=static_cast<int>(ceil(static_cast<double>(ainum)/BX));
}
*/
time_pair.start();
int ngridxy = _ngridx * _ngridy;
k_fphi_mpole.set_size(GX,BX);

12
lib/gpu/lal_hippo.cpp
View File

@ -619,14 +619,16 @@ int HippoT::polar_real(const int eflag, const int vflag) {
int nbor_pitch=this->nbor->nbor_pitch();
// Compute the block size and grid size to keep all cores busy
const int cus = this->device->gpu->cus();
int BX=this->block_size();
int GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
while (GX < cus) {
const int BX=this->block_size();
const int GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
/*
const int cus = this->device->gpu->cus();
while (GX < cus && GX > 1) {
BX /= 2;
GX=static_cast<int>(ceil(static_cast<double>(ainum)/(BX/this->_threads_per_atom)));
}
*/
this->time_pair.start();
// Build the short neighbor list if not done yet