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

Batch GA/BH subset sync with indexed GPU pack/unpack buffers

Browse Source
This commit is contained in:
ianchb
2026-04-13 20:27:30 +08:00
parent c5d1268dd1
commit e952ee8e91
4 changed files with 495 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
AMSS_NCKU_source/Parallel.C
View File

@@ -145,6 +145,18 @@ int cuda_seg_begin(const Parallel::gridseg *seg, Block *bg, int dir)
return (int)floor((seg->llb[dir] - bg->bbox[dir]) / dx + 0.5);
}
int cuda_state_var_count(MyList<var> *src_vars, MyList<var> *dst_vars)
{
int count = 0;
while (src_vars && dst_vars)
{
++count;
src_vars = src_vars->next;
dst_vars = dst_vars->next;
}
return (src_vars || dst_vars) ? -1 : count;
}
#if USE_CUDA_BSSN
bool cuda_can_direct_pack(const Parallel::gridseg *src, const Parallel::gridseg *dst, int type)
{
@@ -163,15 +175,17 @@ bool cuda_can_direct_unpack(const Parallel::gridseg *dst, int type)
bool cuda_direct_pack_segment(double *buffer,
const Parallel::gridseg *src,
const Parallel::gridseg *dst,
int state_index)
int state_count)
{
if (state_count <= 0 || state_count > BSSN_CUDA_STATE_COUNT)
return false;
const double t0 = sync_profile_enabled() ? MPI_Wtime() : 0.0;
const int i0 = cuda_seg_begin(dst, src->Bg, 0);
const int j0 = cuda_seg_begin(dst, src->Bg, 1);
const int k0 = cuda_seg_begin(dst, src->Bg, 2);
const bool ok = bssn_cuda_pack_state_region_to_host_buffer(src->Bg, state_index, buffer, src->Bg->shape,
i0, j0, k0,
dst->shape[0], dst->shape[1], dst->shape[2]) == 0;
const bool ok = bssn_cuda_pack_state_batch_to_host_buffer(src->Bg, state_count, buffer, src->Bg->shape,
i0, j0, k0,
dst->shape[0], dst->shape[1], dst->shape[2]) == 0;
if (sync_profile_enabled())
sync_profile_stats().direct_pack_sec += MPI_Wtime() - t0;
return ok;
@@ -179,15 +193,17 @@ bool cuda_direct_pack_segment(double *buffer,
bool cuda_direct_unpack_segment(double *buffer,
const Parallel::gridseg *dst,
int state_index)
int state_count)
{
if (state_count <= 0 || state_count > BSSN_CUDA_STATE_COUNT)
return false;
const double t0 = sync_profile_enabled() ? MPI_Wtime() : 0.0;
const int i0 = cuda_seg_begin(dst, dst->Bg, 0);
const int j0 = cuda_seg_begin(dst, dst->Bg, 1);
const int k0 = cuda_seg_begin(dst, dst->Bg, 2);
const bool ok = bssn_cuda_unpack_state_region_from_host_buffer(dst->Bg, state_index, buffer, dst->Bg->shape,
i0, j0, k0,
dst->shape[0], dst->shape[1], dst->shape[2]) == 0;
const bool ok = bssn_cuda_unpack_state_batch_from_host_buffer(dst->Bg, state_count, buffer, dst->Bg->shape,
i0, j0, k0,
dst->shape[0], dst->shape[1], dst->shape[2]) == 0;
if (sync_profile_enabled())
sync_profile_stats().direct_unpack_sec += MPI_Wtime() - t0;
return ok;
@@ -3921,21 +3937,14 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
if (!src || !dst)
return size_out;
MyList<var> *varls, *varld;
varls = VarLists;
varld = VarListd;
while (varls && varld)
{
varls = varls->next;
varld = varld->next;
}
if (varls || varld)
{
cout << "error in short data packer, var lists does not match." << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
MyList<var> *varls, *varld;
const int state_count = cuda_state_var_count(VarLists, VarListd);
if (state_count < 0)
{
cout << "error in short data packer, var lists does not match." << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
int type; /* 1 copy, 2 restrict, 3 prolong */
if (src->data->Bg->lev == dst->data->Bg->lev)
@@ -3961,7 +3970,7 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
bool handled_by_cuda = false;
if (dir == PACK && cuda_can_direct_pack(src->data, dst->data, type))
{
handled_by_cuda = cuda_direct_pack_segment(data + size_out, src->data, dst->data, state_idx);
handled_by_cuda = cuda_direct_pack_segment(data + size_out, src->data, dst->data, state_count);
if (!handled_by_cuda)
{
cout << "Parallel::data_packer: CUDA direct pack failed." << endl;
@@ -3970,7 +3979,7 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
}
else if (dir == UNPACK && cuda_can_direct_unpack(dst->data, type))
{
handled_by_cuda = cuda_direct_unpack_segment(data + size_out, dst->data, state_idx);
handled_by_cuda = cuda_direct_unpack_segment(data + size_out, dst->data, state_count);
if (!handled_by_cuda)
{
cout << "Parallel::data_packer: CUDA direct unpack failed." << endl;
@@ -4005,6 +4014,16 @@ int Parallel::data_packer(double *data, MyList<Parallel::gridseg> *src, MyList<P
dst->data->llb, dst->data->uub);
#if USE_CUDA_BSSN
}
else
{
size_out += (state_count - 1) * dst->data->shape[0] * dst->data->shape[1] * dst->data->shape[2];
while (varls->next && varld->next)
{
varls = varls->next;
varld = varld->next;
++state_idx;
}
}
#endif
}
size_out += dst->data->shape[0] * dst->data->shape[1] * dst->data->shape[2];