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

Add direct CUDA resident-state sync path and profiling hooks

Browse Source
This commit is contained in:
ianchb
2026-04-13 00:57:05 +08:00
parent 7f2a391dd2
commit 636e35bfd8
5 changed files with 1188 additions and 527 deletions
Download Patch File Download Diff File Expand all files Collapse all files

175
AMSS_NCKU_source/bssn_class.C
View File

@@ -76,6 +76,48 @@ bool fill_bssn_cuda_views(Block *cg, MyList<var> *vars,
return idx == BSSN_CUDA_STATE_COUNT && vars == 0;
}
bool bssn_cuda_use_resident_sync(int lev)
{
#ifdef WithShell
(void)lev;
return false;
#else
return lev == 0;
#endif
}
void bssn_cuda_download_level_state(MyList<Patch> *PatL, MyList<var> *vars, int myrank)
{
MyList<Patch> *Pp = PatL;
while (Pp)
{
MyList<Block> *BP = Pp->data->blb;
while (BP)
{
Block *cg = BP->data;
if (myrank == cg->rank)
{
double *state_out[BSSN_CUDA_STATE_COUNT];
if (!fill_bssn_cuda_views(cg, vars, state_out))
{
cout << "CUDA BSSN state list mismatch on resident state download" << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
if (bssn_cuda_download_resident_state(cg, cg->shape, state_out))
{
cout << "CUDA resident state download failed" << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
bssn_cuda_release_step_ctx(cg);
}
if (BP == Pp->data->ble)
break;
BP = BP->next;
}
Pp = Pp->next;
}
}
} // namespace
#endif
@@ -3058,13 +3100,18 @@ void bssn_class::RecursiveStep(int lev, int num) // in all 2^(lev+1)-1 steps
#if (PSTR == 0)
#if 1
void bssn_class::Step(int lev, int YN)
{
setpbh(BH_num, Porg0, Mass, BH_num_input);
double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
// new code 2013-2-15, zjcao
void bssn_class::Step(int lev, int YN)
{
setpbh(BH_num, Porg0, Mass, BH_num_input);
double dT_lev = dT * pow(0.5, Mymax(lev, trfls));
#if USE_CUDA_BSSN
const bool use_cuda_resident_sync = bssn_cuda_use_resident_sync(lev);
#else
const bool use_cuda_resident_sync = false;
#endif
// new code 2013-2-15, zjcao
#if (MAPBH == 1)
// for black hole position
if (BH_num > 0 && lev == GH->levels - 1)
@@ -3128,15 +3175,17 @@ void bssn_class::Step(int lev, int YN)
Block *cg = BP->data;
if (myrank == cg->rank)
{
#if (AGM == 0)
f_enforce_ga(cg->shape,
cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn],
cg->fgfs[gyy0->sgfn], cg->fgfs[gyz0->sgfn], cg->fgfs[gzz0->sgfn],
cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn],
cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn]);
#endif
#if (AGM == 0)
if (!use_cuda_resident_sync)
f_enforce_ga(cg->shape,
cg->fgfs[gxx0->sgfn], cg->fgfs[gxy0->sgfn], cg->fgfs[gxz0->sgfn],
cg->fgfs[gyy0->sgfn], cg->fgfs[gyz0->sgfn], cg->fgfs[gzz0->sgfn],
cg->fgfs[Axx0->sgfn], cg->fgfs[Axy0->sgfn], cg->fgfs[Axz0->sgfn],
cg->fgfs[Ayy0->sgfn], cg->fgfs[Ayz0->sgfn], cg->fgfs[Azz0->sgfn]);
#endif
bool used_gpu_substep = false;
bool used_gpu_resident_state = false;
#if USE_CUDA_BSSN
{
double *state_in[BSSN_CUDA_STATE_COUNT];
@@ -3154,6 +3203,11 @@ void bssn_class::Step(int lev, int YN)
MPI_Abort(MPI_COMM_WORLD, 1);
}
int apply_bam_bc = 0;
int keep_resident_state = use_cuda_resident_sync ? 1 : 0;
int apply_enforce_ga = 0;
#if (AGM == 0)
apply_enforce_ga = 1;
#endif
#if (SommerType == 0)
#ifndef WithShell
apply_bam_bc = (lev == 0) ? 1 : 0;
@@ -3164,7 +3218,8 @@ void bssn_class::Step(int lev, int YN)
state_in, state_out, matter,
propspeed, soa_flat, Pp->data->bbox,
dT_lev, TRK4, iter_count, apply_bam_bc,
Symmetry, lev, ndeps, pre))
Symmetry, lev, ndeps, pre,
keep_resident_state, apply_enforce_ga, chitiny))
{
cout << "CUDA predictor substep failed in domain: ("
<< cg->bbox[0] << ":" << cg->bbox[3] << ","
@@ -3173,6 +3228,7 @@ void bssn_class::Step(int lev, int YN)
ERROR = 1;
}
used_gpu_substep = true;
used_gpu_resident_state = (keep_resident_state != 0);
}
#endif
if (!used_gpu_substep)
@@ -3277,8 +3333,9 @@ void bssn_class::Step(int lev, int YN)
varlrhs = varlrhs->next;
}
}
f_lowerboundset(cg->shape, cg->fgfs[phi->sgfn], chitiny);
}
if (!used_gpu_resident_state)
f_lowerboundset(cg->shape, cg->fgfs[phi->sgfn], chitiny);
}
if (BP == Pp->data->ble)
break;
BP = BP->next;
@@ -3436,9 +3493,9 @@ void bssn_class::Step(int lev, int YN)
MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req);
}
#endif
Parallel::AsyncSyncState async_pre;
Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
Parallel::AsyncSyncState async_pre;
Parallel::Sync_start(GH->PatL[lev], SynchList_pre, Symmetry, sync_cache_pre[lev], async_pre);
#ifdef WithShell
if (lev == 0)
@@ -3455,9 +3512,9 @@ void bssn_class::Step(int lev, int YN)
<< (double)(curr_clock - prev_clock) / ((double)CLOCKS_PER_SEC)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
}
#endif
Parallel::Sync_finish(sync_cache_pre[lev], async_pre, SynchList_pre, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
@@ -3530,22 +3587,24 @@ void bssn_class::Step(int lev, int YN)
Block *cg = BP->data;
if (myrank == cg->rank)
{
#if (AGM == 0)
f_enforce_ga(cg->shape,
cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
#elif (AGM == 1)
if (iter_count == 3)
f_enforce_ga(cg->shape,
cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
#endif
#if (AGM == 0)
if (!use_cuda_resident_sync)
f_enforce_ga(cg->shape,
cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
#elif (AGM == 1)
if (iter_count == 3 && !use_cuda_resident_sync)
f_enforce_ga(cg->shape,
cg->fgfs[gxx->sgfn], cg->fgfs[gxy->sgfn], cg->fgfs[gxz->sgfn],
cg->fgfs[gyy->sgfn], cg->fgfs[gyz->sgfn], cg->fgfs[gzz->sgfn],
cg->fgfs[Axx->sgfn], cg->fgfs[Axy->sgfn], cg->fgfs[Axz->sgfn],
cg->fgfs[Ayy->sgfn], cg->fgfs[Ayz->sgfn], cg->fgfs[Azz->sgfn]);
#endif
bool used_gpu_substep = false;
bool used_gpu_resident_state = false;
#if USE_CUDA_BSSN
{
double *state_in[BSSN_CUDA_STATE_COUNT];
@@ -3563,6 +3622,13 @@ void bssn_class::Step(int lev, int YN)
MPI_Abort(MPI_COMM_WORLD, 1);
}
int apply_bam_bc = 0;
int keep_resident_state = use_cuda_resident_sync ? 1 : 0;
int apply_enforce_ga = 0;
#if (AGM == 0)
apply_enforce_ga = 1;
#elif (AGM == 1)
apply_enforce_ga = (iter_count == 3) ? 1 : 0;
#endif
#if (SommerType == 0)
#ifndef WithShell
apply_bam_bc = (lev == 0) ? 1 : 0;
@@ -3573,7 +3639,8 @@ void bssn_class::Step(int lev, int YN)
state_in, state_out, matter,
propspeed, soa_flat, Pp->data->bbox,
dT_lev, TRK4, iter_count, apply_bam_bc,
Symmetry, lev, ndeps, cor))
Symmetry, lev, ndeps, cor,
keep_resident_state, apply_enforce_ga, chitiny))
{
cout << "CUDA corrector substep failed in domain: ("
<< cg->bbox[0] << ":" << cg->bbox[3] << ","
@@ -3582,6 +3649,7 @@ void bssn_class::Step(int lev, int YN)
ERROR = 1;
}
used_gpu_substep = true;
used_gpu_resident_state = (keep_resident_state != 0);
}
#endif
if (!used_gpu_substep)
@@ -3686,8 +3754,9 @@ void bssn_class::Step(int lev, int YN)
varlrhs = varlrhs->next;
}
}
f_lowerboundset(cg->shape, cg->fgfs[phi1->sgfn], chitiny);
}
if (!used_gpu_resident_state)
f_lowerboundset(cg->shape, cg->fgfs[phi1->sgfn], chitiny);
}
if (BP == Pp->data->ble)
break;
BP = BP->next;
@@ -3842,9 +3911,9 @@ void bssn_class::Step(int lev, int YN)
MPI_Iallreduce(&erh, &ERROR, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD, &err_req_cor);
}
#endif
Parallel::AsyncSyncState async_cor;
Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
Parallel::AsyncSyncState async_cor;
Parallel::Sync_start(GH->PatL[lev], SynchList_cor, Symmetry, sync_cache_cor[lev], async_cor);
#ifdef WithShell
if (lev == 0)
@@ -3862,8 +3931,8 @@ void bssn_class::Step(int lev, int YN)
<< " seconds! " << endl;
}
}
#endif
Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
#endif
Parallel::Sync_finish(sync_cache_cor[lev], async_cor, SynchList_cor, Symmetry);
#ifdef WithShell
// Complete non-blocking error reduction and check
@@ -3968,10 +4037,14 @@ void bssn_class::Step(int lev, int YN)
}
#endif
}
}
#if (RPS == 0)
// mesh refinement boundary part
RestrictProlong(lev, YN, BB);
}
#if USE_CUDA_BSSN
if (use_cuda_resident_sync)
bssn_cuda_download_level_state(GH->PatL[lev], SynchList_cor, myrank);
#endif
#if (RPS == 0)
// mesh refinement boundary part
RestrictProlong(lev, YN, BB);
#ifdef WithShell
if (lev == 0)