64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reduce staged GPU host-device copies

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 16:44:08 +08:00
parent 42e851d19a
commit 5b00d49070
5 changed files with 349 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
AMSS_NCKU_source/bssn_cuda_step.C
View File

@@ -83,7 +83,7 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
cg->fgfs[varlr->data->sgfn],
varl0->data->propspeed,
varl0->data->SoA,
Symmetry, lev, rk_stage))
Symmetry, lev, rk_stage, false))
{
cerr << "GPU rk4/boundary failure: lev=" << lev
<< " rk_stage=" << rk_stage
@@ -101,6 +101,43 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
}
};
auto stage_download_var_list =
[&](Block *cg, MyList<var> *var_list) {
while (var_list)
{
if (bssn_cuda_download_buffer(cg->shape, cg->fgfs[var_list->data->sgfn]))
{
cerr << "GPU stage download failure: lev=" << lev
<< " var=" << var_list->data->name
<< " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
<< cg->bbox[1] << ":" << cg->bbox[4] << ","
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
break;
}
var_list = var_list->next;
}
};
auto stage_upload_var_list =
[&](Block *cg, MyList<var> *var_list) {
const int n = cg->shape[0] * cg->shape[1] * cg->shape[2];
while (var_list)
{
if (bssn_gpu_stage_upload_buffer(cg->fgfs[var_list->data->sgfn], n))
{
cerr << "GPU state upload failure: lev=" << lev
<< " var=" << var_list->data->name
<< " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
<< cg->bbox[1] << ":" << cg->bbox[4] << ","
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
break;
}
var_list = var_list->next;
}
};
MyList<Patch> *Pp = GH->PatL[lev];
while (Pp)
{
@@ -110,12 +147,13 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
Block *cg = BP->data;
if (myrank == cg->rank)
{
stage_upload_var_list(cg, StateList);
if (gpu_rhs(CALLED_BY_STEP, myrank, RHS_PARA_CALLED_FIRST_TIME))
ERROR = 1;
run_stage_on_block(cg, Pp->data, StateList, StateList, SynchList_pre, RHSList, iter_count);
if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi->sgfn], chitiny))
if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi->sgfn], chitiny, false))
{
cerr << "GPU lowerbound failure: lev=" << lev
<< " rk_stage=" << iter_count
@@ -125,6 +163,8 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
}
if (!ERROR)
stage_download_var_list(cg, SynchList_pre);
}
if (BP == Pp->data->ble)
break;
@@ -194,12 +234,13 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
Block *cg = BP->data;
if (myrank == cg->rank)
{
stage_upload_var_list(cg, SynchList_pre);
if (gpu_rhs(CALLED_BY_STEP, myrank, RHS_PARA_CALLED_THEN))
ERROR = 1;
run_stage_on_block(cg, Pp->data, StateList, SynchList_pre, SynchList_cor, RHSList, iter_count);
if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi1->sgfn], chitiny))
if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi1->sgfn], chitiny, false))
{
cerr << "GPU lowerbound failure: lev=" << lev
<< " rk_stage=" << iter_count
@@ -209,6 +250,8 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
}
if (!ERROR)
stage_download_var_list(cg, SynchList_cor);
}
if (BP == Pp->data->ble)