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

Stabilize GPU buffer lifecycle around regrid

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 20:48:06 +08:00
parent 46e94d1248
commit cf3c6d6218
8 changed files with 186 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
AMSS_NCKU_source/cgh.C
View File

@@ -23,10 +23,13 @@ using namespace std;
#include <mpi.h>
#include "macrodef.h"
#include "misc.h"
#include "cgh.h"
#include "Parallel.h"
#include "parameters.h"
#include "misc.h"
#include "cgh.h"
#include "Parallel.h"
#include "parameters.h"
#ifdef USE_GPU
#include "bssn_gpu.h"
#endif
//================================================================================================
@@ -881,13 +884,17 @@ void cgh::recompose_cgh(int nprocs, bool *lev_flag,
tmPat = construct_patchlist(lev, Symmetry);
// tmPat construction completes
Parallel::distribute(tmPat, nprocs, ingfs, fngfs, false);
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
#ifdef USE_GPU
bssn_gpu_clear_cached_device_buffers();
bssn_gpu_release_pinned_host_buffers();
#endif
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
#if (RPB == 1)
Parallel::destroypsuList_bam(bdsul[lev]);
Parallel::destroypsuList_bam(rsul[lev]);
@@ -910,13 +917,17 @@ void cgh::recompose_cgh(int nprocs, bool *lev_flag,
tmPat = construct_patchlist(lev, Symmetry);
// tmPat construction completes
Parallel::distribute(tmPat, end_rank[lev] - start_rank[lev] + 1, ingfs, fngfs, false, start_rank[lev], end_rank[lev]);
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
#ifdef USE_GPU
bssn_gpu_clear_cached_device_buffers();
bssn_gpu_release_pinned_host_buffers();
#endif
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
#if (RPB == 1)
#error "not support yet"
#endif
@@ -1518,13 +1529,17 @@ void cgh::recompose_cgh_Onelevel(int nprocs, int lev,
tmPat = construct_patchlist(lev, Symmetry);
// tmPat construction completes
Parallel::distribute(tmPat, nprocs, ingfs, fngfs, false);
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
#ifdef USE_GPU
bssn_gpu_clear_cached_device_buffers();
bssn_gpu_release_pinned_host_buffers();
#endif
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
}
#elif (PSTR == 1 || PSTR == 2 || PSTR == 3)
#warning "recompose_cgh_Onelevel is not implimented yet"
@@ -1540,14 +1555,18 @@ void cgh::recompose_cgh_Onelevel(int nprocs, int lev,
// tmPat construction completes
Parallel::distribute(tmPat, end_rank[lev] - start_rank[lev] + 1, ingfs, fngfs, false, start_rank[lev], end_rank[lev]);
misc::tillherecheck(Commlev[lev], start_rank[lev], "after distribute");
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
misc::tillherecheck(Commlev[lev], start_rank[lev], "after fill_level_data");
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
// checkPatchList(tmPat,true);
bool CC = (lev > trfls);
Parallel::fill_level_data(tmPat, PatL[lev], PatL[lev - 1], OldList, StateList, FutureList, tmList, Symmetry, BB, CC);
misc::tillherecheck(Commlev[lev], start_rank[lev], "after fill_level_data");
#ifdef USE_GPU
bssn_gpu_clear_cached_device_buffers();
bssn_gpu_release_pinned_host_buffers();
#endif
Parallel::KillBlocks(PatL[lev]);
PatL[lev]->destroyList();
PatL[lev] = tmPat;
}