Skip to content

fix(cc): fix message passing when nghost is 0 and send list is empty #4237

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
iProzd merged 9 commits into from
Oct 23, 2024
Merged

fix(cc): fix message passing when nghost is 0 and send list is empty #4237

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
8e9694d
fix error in nopbc (nghost = 0)
CaRoLZhangxy Oct 22, 2024
90d0648
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Oct 22, 2024
550f18d
fix dpa2 c++ when no lammps comm data pass in
CaRoLZhangxy Oct 22, 2024
d8541da
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Oct 22, 2024
50621bd
remove c++ lmp nlist ut due to no comm data
CaRoLZhangxy Oct 22, 2024
389b914
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Oct 22, 2024
8e82b9f
support old nlist api
CaRoLZhangxy Oct 23, 2024
410da82
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Oct 23, 2024
e3f9433
add tests
CaRoLZhangxy Oct 23, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 10 additions & 12 deletions source/api_cc/src/DeepPotPT.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,7 +168,7 @@ void DeepPotPT::compute(ENERGYVTYPE& ener,
nlist_data.copy_from_nlist(lmp_list);
nlist_data.shuffle_exclude_empty(fwd_map);
nlist_data.padding();
if (do_message_passing == 1 && nghost > 0) {
if (do_message_passing == 1) {
int nswap = lmp_list.nswap;
torch::Tensor sendproc_tensor =
torch::from_blob(lmp_list.sendproc, {nswap}, int32_option);
Expand All @@ -180,10 +180,14 @@ void DeepPotPT::compute(ENERGYVTYPE& ener,
torch::from_blob(lmp_list.recvnum, {nswap}, int32_option);
torch::Tensor sendnum_tensor =
torch::from_blob(lmp_list.sendnum, {nswap}, int32_option);
torch::Tensor communicator_tensor = torch::from_blob(
const_cast<void*>(lmp_list.world), {1}, torch::kInt64);
// torch::Tensor communicator_tensor =
// torch::tensor(lmp_list.world, int32_option);
torch::Tensor communicator_tensor;
if (lmp_list.world == 0) {
communicator_tensor = torch::empty({1}, torch::kInt64);
} else {
communicator_tensor = torch::from_blob(
const_cast<void*>(lmp_list.world), {1}, torch::kInt64);
}

torch::Tensor nswap_tensor = torch::tensor(nswap, int32_option);
int total_send =
std::accumulate(lmp_list.sendnum, lmp_list.sendnum + nswap, 0);
Expand All @@ -196,12 +200,6 @@ void DeepPotPT::compute(ENERGYVTYPE& ener,
comm_dict.insert("recv_num", recvnum_tensor);
comm_dict.insert("communicator", communicator_tensor);
}
if (do_message_passing == 1 && nghost == 0) {
// for the situation that no ghost atoms (e.g. serial nopbc)
// set the mapping arange(nloc) is enough
auto option = torch::TensorOptions().device(device).dtype(torch::kInt64);
mapping_tensor = at::arange(nloc_real, option).unsqueeze(0);
}
}
at::Tensor firstneigh = createNlistTensor(nlist_data.jlist);
firstneigh_tensor = firstneigh.to(torch::kInt64).to(device);
Expand All @@ -224,7 +222,7 @@ void DeepPotPT::compute(ENERGYVTYPE& ener,
.to(device);
}
c10::Dict<c10::IValue, c10::IValue> outputs =
(do_message_passing == 1 && nghost > 0)
(do_message_passing == 1)
? module
.run_method("forward_lower", coord_wrapped_Tensor, atype_Tensor,
firstneigh_tensor, mapping_tensor, fparam_tensor,
Expand Down
8 changes: 0 additions & 8 deletions source/lmp/tests/test_lammps_dpa_pt_nopbc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -681,14 +681,6 @@ def test_pair_deepmd_si(lammps_si):
[(["--balance"],), ([],)],
)
def test_pair_deepmd_mpi(balance_args: list):
if balance_args == []:
# python:5331 terminated with signal 11 at PC=7f3e940e3806 SP=7ffd5787edc0. Backtrace:
# /home/runner/work/deepmd-kit/deepmd-kit/dp_test/lib/libdeepmd_op_pt.so(+0x95806)[0x7f3e940e3806]
# /home/runner/work/deepmd-kit/deepmd-kit/dp_test/lib/libdeepmd_op_pt.so(+0x8f76e)[0x7f3e940dd76e]
# /home/runner/work/deepmd-kit/deepmd-kit/dp_test/lib/libdeepmd_op_pt.so(+0x9a38a)[0x7f3e940e838a]
# /home/runner/work/deepmd-kit/deepmd-kit/dp_test/lib/libdeepmd_op_pt.so(_Z9border_opRKN2at6TensorES2_S2_S2_S2_S2_S2_S2_S2_+0x8e)[0x7f3e940dda63]
# /home/runner/work/deepmd-kit/deepmd-kit/dp_test/lib/libdeepmd_op_pt.so(+0xaeac3)[0x7f3e940fcac3]
pytest.skip(reason="Known segfault, see comments for details")
with tempfile.NamedTemporaryFile() as f:
sp.check_call(
[
Expand Down
63 changes: 37 additions & 26 deletions source/op/pt/comm.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,16 +87,18 @@ class Border : public torch::autograd::Function<Border> {
int mpi_init = 0;
MPI_Initialized(&mpi_init);
int cuda_aware = 1;
int me;
int me = 0;
MPI_Comm world;
int world_size = 0;
unpack_communicator(communicator_tensor, world);
MPI_Comm_rank(world, &me);
MPI_Comm_size(world, &world_size);
if (mpi_init) {
unpack_communicator(communicator_tensor, world);
MPI_Comm_rank(world, &me);
MPI_Comm_size(world, &world_size);
}
MPI_Datatype mpi_type = get_mpi_type<FPTYPE>();
MPI_Request request;
#if defined(GOOGLE_CUDA) || defined(TENSORFLOW_USE_ROCM)
if (world_size != 1) {
if (world_size >= 1) {
int version, subversion;
MPI_Get_version(&version, &subversion);
if (version >= 4) {
Expand All @@ -120,11 +122,15 @@ class Border : public torch::autograd::Function<Border> {
for (int iswap = 0; iswap < nswap; ++iswap) {
int nrecv = recvnum[iswap];
int nsend = sendnum[iswap];
torch::Tensor isendlist =
torch::from_blob(sendlist[iswap], {nsend}, int32_options)
.to(recv_g1_tensor.device());
torch::Tensor send_g1_tensor = recv_g1_tensor.index_select(0, isendlist);
FPTYPE* send_g1 = send_g1_tensor.data_ptr<FPTYPE>();
torch::Tensor isendlist;
torch::Tensor send_g1_tensor;
FPTYPE* send_g1;
if (nsend != 0) {
isendlist = torch::from_blob(sendlist[iswap], {nsend}, int32_options)
.to(recv_g1_tensor.device());
send_g1_tensor = recv_g1_tensor.index_select(0, isendlist);
send_g1 = send_g1_tensor.data_ptr<FPTYPE>();
}
#ifdef USE_MPI
if (sendproc[iswap] != me) {
if (nrecv) {
Expand Down Expand Up @@ -207,15 +213,17 @@ class Border : public torch::autograd::Function<Border> {
MPI_Initialized(&mpi_init);
int world_size = 0;
int cuda_aware = 1;
int me = 0;
MPI_Comm world;
unpack_communicator(communicator_tensor, world);
int me;
MPI_Comm_rank(world, &me);
MPI_Comm_size(world, &world_size);
if (mpi_init) {
unpack_communicator(communicator_tensor, world);
MPI_Comm_rank(world, &me);
MPI_Comm_size(world, &world_size);
}
MPI_Datatype mpi_type = get_mpi_type<FPTYPE>();
MPI_Request request;
#if defined(GOOGLE_CUDA) || defined(TENSORFLOW_USE_ROCM)
if (world_size != 1) {
if (world_size >= 1) {
int version, subversion;
MPI_Get_version(&version, &subversion);
if (version >= 4) {
Expand Down Expand Up @@ -248,17 +256,20 @@ class Border : public torch::autograd::Function<Border> {
int nlocal = nlocal_tensor.item<int>();
int nghost = nghost_tensor.item<int>();
int ntotal = nlocal + nghost;

torch::Tensor send_g1_tensor = d_local_g1_tensor;

int max_recvnum = sendnum_tensor.max().item<int>();
auto options = torch::TensorOptions()
.dtype(d_local_g1_tensor.dtype())
.device(d_local_g1_tensor.device());
torch::Tensor recv_g1_tensor =
torch::empty({max_recvnum, tensor_size}, options);
FPTYPE* recv_g1 = recv_g1_tensor.data_ptr<FPTYPE>();
FPTYPE* send_g1 = send_g1_tensor.data_ptr<FPTYPE>() + ntotal * tensor_size;
torch::Tensor send_g1_tensor;
torch::Tensor recv_g1_tensor;
FPTYPE* recv_g1;
FPTYPE* send_g1;
if (nswap != 0) {
send_g1_tensor = d_local_g1_tensor;
int max_recvnum = sendnum_tensor.max().item<int>();
auto options = torch::TensorOptions()
.dtype(d_local_g1_tensor.dtype())
.device(d_local_g1_tensor.device());
recv_g1_tensor = torch::empty({max_recvnum, tensor_size}, options);
recv_g1 = recv_g1_tensor.data_ptr<FPTYPE>();
send_g1 = send_g1_tensor.data_ptr<FPTYPE>() + ntotal * tensor_size;
}

int end = ntotal;
auto int32_options = torch::TensorOptions().dtype(torch::kInt32);
Expand Down