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

Reduce redundant GPU host buffer preparation

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 21:20:45 +08:00
parent cf3c6d6218
commit e1a0bff43c
2 changed files with 11 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
AMSS_NCKU_source/bssn_cuda_ops.cu
View File

@@ -787,12 +787,19 @@ int bssn_cuda_rk4_boundary_var(int *ex, double dT,
const bool refresh_rhs =
(rk_stage == 0) || !cache.rhs_resident || cache.host_rhs != rhs_accum;
double *stage_ptr = nullptr;
const double *mapped_state0_ptr = refresh_state0 ? bssn_gpu_find_device_buffer(state0) : cache.state0.ptr;
const double *mapped_boundary_ptr = need_boundary_input ? bssn_gpu_find_device_buffer(boundary_src) : nullptr;
const double *mapped_stage_ptr = need_stage_input ? bssn_gpu_find_device_buffer(stage_data) : nullptr;
const double *mapped_rhs_ptr = refresh_rhs ? bssn_gpu_find_device_buffer(rhs_accum) : cache.rhs.ptr;
bssn_gpu_prepare_host_buffer(state0, n);
if (need_boundary_input) bssn_gpu_prepare_host_buffer(boundary_src, n);
if (need_stage_input) bssn_gpu_prepare_host_buffer(stage_data, n);
bssn_gpu_prepare_host_buffer(rhs_accum, n);
if (refresh_state0 && !mapped_state0_ptr)
bssn_gpu_prepare_host_buffer(state0, n);
if (need_boundary_input && !mapped_boundary_ptr)
bssn_gpu_prepare_host_buffer(boundary_src, n);
if (need_stage_input && !mapped_stage_ptr)
bssn_gpu_prepare_host_buffer(stage_data, n);
if (refresh_rhs && !mapped_rhs_ptr)
bssn_gpu_prepare_host_buffer(rhs_accum, n);
ok = ok &&
(!refresh_state0 || copy_to_device_preferring_device(cache.state0, state0, bytes)) &&

20
AMSS_NCKU_source/bssn_cuda_step.C
View File

@@ -124,25 +124,6 @@ void bssn_class::Step_MainPath_GPU(int lev, int YN)
}
};
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;
}
};
auto ensure_stage_device_var_list =
[&](Block *cg, MyList<var> *var_list) {
const int n = cg->shape[0] * cg->shape[1] * cg->shape[2];
@@ -340,7 +321,6 @@ 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;