Skip to content

Refactor: Optimize memory management in RT-TDDFT to enable large-scale GPU calculations #6995

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mohanchen merged 2 commits into from
Mar 6, 2026
Merged

Refactor: Optimize memory management in RT-TDDFT to enable large-scale GPU calculations #6995

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
487bc78
Optimize memory management of RT-TDDFT to enable larger scale calcula…
AsTonyshment Mar 5, 2026
1d79950
Fixed a bug where calculation results were incorrect due to not alloc…
AsTonyshment Mar 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 36 additions & 17 deletions source/source_esolver/esolver_ks_lcao_tddft.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
#include "esolver_ks_lcao_tddft.h"

//----------------IO-----------------
#include "source_io/module_dipole/dipole_io.h"
#include "source_io/module_ctrl/ctrl_output_td.h"
#include "source_io/module_current/td_current_io.h"
#include "source_io/module_dipole/dipole_io.h"
#include "source_io/module_output/output_log.h"
#include "source_io/module_wf/read_wfc_nao.h"
//------LCAO HSolver ElecState-------
Expand Down Expand Up @@ -425,32 +425,51 @@ void ESolver_KS_LCAO_TDDFT<TR, Device>::store_h_s_psi(UnitCell& ucell,
// Store H and S matrices to Hk_laststep and Sk_laststep
if (use_tensor && use_lapack)
{
// Gather H and S matrices to root process
#ifdef __MPI
int myid = 0;
int num_procs = 1;
MPI_Comm_rank(MPI_COMM_WORLD, &myid);
MPI_Comm_size(MPI_COMM_WORLD, &num_procs);

// Global matrix structure
std::complex<double>* h_ptr = nullptr;
std::complex<double>* s_ptr = nullptr;

// Define containers for gathered data (only needed for multi-process)
module_rt::Matrix_g<std::complex<double>> h_mat_g;
module_rt::Matrix_g<std::complex<double>> s_mat_g;

// Collect H matrix
module_rt::gatherMatrix(myid, 0, h_mat, h_mat_g);
BlasConnector::copy(len_HS_ik,
h_mat_g.p.get(),
1,
this->Hk_laststep.template data<std::complex<double>>() + ik * len_HS_ik,
1);
if (num_procs == 1)
{
// Single process: directly point to local data without gather
h_ptr = h_mat.p;
s_ptr = s_mat.p;
}
else
{
// Multiple processes: gather data to the root process (myid == 0) and point to the gathered data
module_rt::gatherMatrix(myid, 0, h_mat, h_mat_g);
module_rt::gatherMatrix(myid, 0, s_mat, s_mat_g);
if (myid == 0)
{
h_ptr = h_mat_g.p.get();
s_ptr = s_mat_g.p.get();
}
}

// Collect S matrix
module_rt::gatherMatrix(myid, 0, s_mat, s_mat_g);
BlasConnector::copy(len_HS_ik,
s_mat_g.p.get(),
1,
this->Sk_laststep.template data<std::complex<double>>() + ik * len_HS_ik,
1);
// Only the root process (myid == 0) performs the copy
if (myid == 0 && h_ptr != nullptr && s_ptr != nullptr)
{
BlasConnector::copy(len_HS_ik,
h_ptr,
1,
this->Hk_laststep.template data<std::complex<double>>() + ik * len_HS_ik,
1);
BlasConnector::copy(len_HS_ik,
s_ptr,
1,
this->Sk_laststep.template data<std::complex<double>>() + ik * len_HS_ik,
1);
}
#endif
}
else
Expand Down
Loading
Loading