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

Download staged GPU sync regions incrementally

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 18:23:05 +08:00
parent 3b16795e78
commit 5bc67ded06
3 changed files with 147 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
AMSS_NCKU_source/bssn_cuda_step.C
View File

@@ -243,6 +243,69 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
}
};
auto refresh_stage_host_before_sync =
[&](MyList<var> *var_list, Parallel::SyncCache &cache) -> bool {
if (!cache.valid || !cache.combined_src || myrank < 0 || myrank >= cache.cpusize)
return false;
MyList<Patch> *patch_it = GH->PatL[lev];
while (patch_it)
{
MyList<Block> *block_it = patch_it->data->blb;
while (block_it)
{
Block *cg = block_it->data;
if (myrank == cg->rank)
{
std::vector<Parallel::gridseg *> local_segments;
MyList<Parallel::gridseg> *seg = cache.combined_src[myrank];
while (seg)
{
if (seg->data && seg->data->Bg == cg)
local_segments.push_back(seg->data);
seg = seg->next;
}
if (!local_segments.empty())
{
MyList<var> *var_it = var_list;
while (var_it)
{
double *host_ptr = cg->fgfs[var_it->data->sgfn];
for (size_t i = 0; i < local_segments.size(); ++i)
{
Parallel::gridseg *src_seg = local_segments[i];
if (bssn_gpu_stage_download_region(host_ptr,
cg->shape,
cg->bbox,
cg->bbox + dim,
src_seg->shape,
src_seg->llb))
{
cerr << "GPU sync region download failure: lev=" << lev
<< " var=" << var_it->data->name
<< " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
<< cg->bbox[1] << ":" << cg->bbox[4] << ","
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
return true;
}
}
var_it = var_it->next;
}
}
}
if (block_it == patch_it->data->ble)
break;
block_it = block_it->next;
}
patch_it = patch_it->next;
}
return true;
};
MyList<Patch> *Pp = GH->PatL[lev];
while (Pp)
{
@@ -268,7 +331,7 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
}
if (!ERROR)
if (!ERROR && !sync_cache_pre[lev].valid)
stage_download_var_list(cg, SynchList_pre);
}
if (BP == Pp->data->ble)
@@ -278,6 +341,9 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
Pp = Pp->next;
}
if (!ERROR && sync_cache_pre[lev].valid)
refresh_stage_host_before_sync(SynchList_pre, sync_cache_pre[lev]);
MPI_Request err_req_pre;
{
int erh = ERROR;
@@ -348,15 +414,15 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
if (bssn_cuda_lowerbound(cg->shape, cg->fgfs[phi1->sgfn], chitiny, false))
{
cerr << "GPU lowerbound failure: lev=" << lev
<< " rk_stage=" << iter_count
<< " var=" << phi1->name
cerr << "GPU lowerbound failure: lev=" << lev
<< " rk_stage=" << iter_count
<< " var=" << phi1->name
<< " bbox=(" << cg->bbox[0] << ":" << cg->bbox[3] << ","
<< cg->bbox[1] << ":" << cg->bbox[4] << ","
<< cg->bbox[2] << ":" << cg->bbox[5] << ")" << endl;
ERROR = 1;
}
if (!ERROR)
if (!ERROR && (!sync_cache_cor[lev].valid || iter_count == 3))
stage_download_var_list(cg, SynchList_cor);
}
@@ -367,6 +433,9 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
Pp = Pp->next;
}
if (!ERROR && sync_cache_cor[lev].valid && iter_count < 3)
refresh_stage_host_before_sync(SynchList_cor, sync_cache_cor[lev]);
MPI_Request err_req_cor;
{
int erh = ERROR;