Skip to content

Commit

Permalink
dispatch with distributed needs to throw with half
Browse files Browse the repository at this point in the history
  • Loading branch information
@yhmtsai
yhmtsai committed Nov 5, 2024
1 parent ce1d7ed commit de5c2a7
Show file tree
Hide file tree
Showing 4 changed files with 350 additions and 247 deletions.
14 changes: 9 additions & 5 deletions core/distributed/helpers.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -122,11 +122,15 @@ void vector_dispatch(T* linop, F&& f, Args&&... args)
{
#if GINKGO_BUILD_MPI
if (is_distributed(linop)) {
using type = std::conditional_t<
std::is_const<T>::value,
const experimental::distributed::Vector<ValueType>,
experimental::distributed::Vector<ValueType>>;
f(dynamic_cast<type*>(linop), std::forward<Args>(args)...);
if constexpr (std::is_same_v<remove_complex<ValueType>, half>) {
GKO_NOT_SUPPORTED(linop);
} else {
using type = std::conditional_t<
std::is_const<T>::value,
const experimental::distributed::Vector<ValueType>,
experimental::distributed::Vector<ValueType>>;
f(dynamic_cast<type*>(linop), std::forward<Args>(args)...);
}
} else
#endif
{
Expand Down
266 changes: 138 additions & 128 deletions core/multigrid/pgm.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -389,137 +389,147 @@ void Pgm<ValueType, IndexType>::generate()
#if GINKGO_BUILD_MPI
if (std::dynamic_pointer_cast<
const experimental::distributed::DistributedBase>(system_matrix_)) {
auto convert_fine_op = [&](auto matrix) {
using global_index_type = typename std::decay_t<
decltype(*matrix)>::result_type::global_index_type;
auto exec = as<LinOp>(matrix)->get_executor();
auto comm = as<experimental::distributed::DistributedBase>(matrix)
->get_communicator();
auto fine = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
matrix::Csr<ValueType, IndexType>::create(exec),
matrix::Csr<ValueType, IndexType>::create(exec)));
matrix->convert_to(fine);
this->set_fine_op(fine);
};
auto setup_fine_op = [&](auto matrix) {
// Only support csr matrix currently.
auto local_csr = std::dynamic_pointer_cast<const csr_type>(
matrix->get_local_matrix());
auto non_local_csr = std::dynamic_pointer_cast<const csr_type>(
matrix->get_non_local_matrix());
// If system matrix is not csr or need sorting, generate the csr.
if (!parameters_.skip_sorting || !local_csr || !non_local_csr) {
if constexpr (std::is_same_v<remove_complex<ValueType>, half>) {
GKO_NOT_SUPPORTED(nullptr);
} else {
auto convert_fine_op = [&](auto matrix) {
using global_index_type = typename std::decay_t<
decltype(*matrix)>::result_type::global_index_type;
auto exec = as<LinOp>(matrix)->get_executor();
auto comm =
as<experimental::distributed::DistributedBase>(matrix)
->get_communicator();
auto fine = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
matrix::Csr<ValueType, IndexType>::create(exec),
matrix::Csr<ValueType, IndexType>::create(exec)));
matrix->convert_to(fine);
this->set_fine_op(fine);
};
auto setup_fine_op = [&](auto matrix) {
// Only support csr matrix currently.
auto local_csr = std::dynamic_pointer_cast<const csr_type>(
matrix->get_local_matrix());
auto non_local_csr = std::dynamic_pointer_cast<const csr_type>(
matrix->get_non_local_matrix());
// If system matrix is not csr or need sorting, generate the
// csr.
if (!parameters_.skip_sorting || !local_csr || !non_local_csr) {
using global_index_type = typename std::decay_t<
decltype(*matrix)>::global_index_type;
convert_fine_op(
as<ConvertibleTo<experimental::distributed::Matrix<
ValueType, IndexType, global_index_type>>>(matrix));
}
};

using fst_mtx_type =
experimental::distributed::Matrix<ValueType, IndexType,
IndexType>;
using snd_mtx_type =
experimental::distributed::Matrix<ValueType, IndexType, int64>;
// setup the fine op using Csr with current ValueType
// we do not use dispatcher run in the first place because we have
// the fallback option for that.
if (auto obj = std::dynamic_pointer_cast<const fst_mtx_type>(
system_matrix_)) {
setup_fine_op(obj);
} else if (auto obj = std::dynamic_pointer_cast<const snd_mtx_type>(
system_matrix_)) {
setup_fine_op(obj);
} else {
// handle other ValueTypes.
run<ConvertibleTo, fst_mtx_type, snd_mtx_type>(obj,
convert_fine_op);
}

auto distributed_setup = [&](auto matrix) {
auto exec = gko::as<LinOp>(matrix)->get_executor();
auto comm =
gko::as<experimental::distributed::DistributedBase>(matrix)
->get_communicator();
auto num_rank = comm.size();
auto pgm_local_op =
gko::as<const csr_type>(matrix->get_local_matrix());
auto result = this->generate_local(pgm_local_op);

auto non_local_csr =
as<const csr_type>(matrix->get_non_local_matrix());
auto non_local_size = non_local_csr->get_size()[1];
array<IndexType> non_local_agg(exec, non_local_size);
// get agg information (prolong_row_gather row idx)
communicate(matrix, agg_, non_local_agg);
// generate non_local_col_map
non_local_agg.set_executor(exec->get_master());
array<IndexType> non_local_col_map(exec->get_master(),
non_local_size);
// add additional entry in tail such that the offset easily
// handle it.
array<IndexType> renumber(exec->get_master(),
non_local_size + 1);
auto recv_offsets = matrix->recv_offsets_;
generate_non_local_map(recv_offsets, non_local_agg,
non_local_col_map, renumber);

// get new recv_size and recv_offsets
std::vector<experimental::distributed::comm_index_type>
new_recv_size(num_rank);
std::vector<experimental::distributed::comm_index_type>
new_recv_offsets(num_rank + 1);
array<IndexType> new_recv_gather_idxs(exec->get_master());
compute_communication(recv_offsets, non_local_agg, renumber,
new_recv_size, new_recv_offsets,
new_recv_gather_idxs);

non_local_col_map.set_executor(exec);
IndexType non_local_num_agg = new_recv_gather_idxs.get_size();
// build csr from row and col map
// unlike non-distributed version, generate_coarse uses
// different row and col maps.
auto result_non_local_csr = generate_coarse(
exec, non_local_csr.get(),
static_cast<IndexType>(std::get<1>(result)->get_size()[0]),
agg_, non_local_num_agg, non_local_col_map);
// use local and non-local to build coarse matrix
// also restriction and prolongation (Local-only-global matrix)
auto coarse_size =
static_cast<int64>(std::get<1>(result)->get_size()[0]);
comm.all_reduce(exec->get_master(), &coarse_size, 1, MPI_SUM);
new_recv_gather_idxs.set_executor(exec);

// setup the generated linop.
using global_index_type =
typename std::decay_t<decltype(*matrix)>::global_index_type;
convert_fine_op(
as<ConvertibleTo<experimental::distributed::Matrix<
ValueType, IndexType, global_index_type>>>(matrix));
}
};

using fst_mtx_type =
experimental::distributed::Matrix<ValueType, IndexType, IndexType>;
using snd_mtx_type =
experimental::distributed::Matrix<ValueType, IndexType, int64>;
// setup the fine op using Csr with current ValueType
// we do not use dispatcher run in the first place because we have the
// fallback option for that.
if (auto obj =
std::dynamic_pointer_cast<const fst_mtx_type>(system_matrix_)) {
setup_fine_op(obj);
} else if (auto obj = std::dynamic_pointer_cast<const snd_mtx_type>(
system_matrix_)) {
setup_fine_op(obj);
} else {
// handle other ValueTypes.
run<ConvertibleTo, fst_mtx_type, snd_mtx_type>(obj,
convert_fine_op);
auto coarse = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm, gko::dim<2>(coarse_size, coarse_size),
std::get<1>(result), result_non_local_csr,
new_recv_size, new_recv_offsets,
new_recv_gather_idxs));
auto restrict_op = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
dim<2>(coarse_size,
gko::as<LinOp>(matrix)->get_size()[0]),
std::get<2>(result)));
auto prolong_op = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
dim<2>(gko::as<LinOp>(matrix)->get_size()[0],
coarse_size),
std::get<0>(result)));
this->set_multigrid_level(prolong_op, coarse, restrict_op);
};

// the fine op is using csr with the current ValueType
run<fst_mtx_type, snd_mtx_type>(this->get_fine_op(),
distributed_setup);
}

auto distributed_setup = [&](auto matrix) {
auto exec = gko::as<LinOp>(matrix)->get_executor();
auto comm =
gko::as<experimental::distributed::DistributedBase>(matrix)
->get_communicator();
auto num_rank = comm.size();
auto pgm_local_op =
gko::as<const csr_type>(matrix->get_local_matrix());
auto result = this->generate_local(pgm_local_op);

auto non_local_csr =
as<const csr_type>(matrix->get_non_local_matrix());
auto non_local_size = non_local_csr->get_size()[1];
array<IndexType> non_local_agg(exec, non_local_size);
// get agg information (prolong_row_gather row idx)
communicate(matrix, agg_, non_local_agg);
// generate non_local_col_map
non_local_agg.set_executor(exec->get_master());
array<IndexType> non_local_col_map(exec->get_master(),
non_local_size);
// add additional entry in tail such that the offset easily handle
// it.
array<IndexType> renumber(exec->get_master(), non_local_size + 1);
auto recv_offsets = matrix->recv_offsets_;
generate_non_local_map(recv_offsets, non_local_agg,
non_local_col_map, renumber);

// get new recv_size and recv_offsets
std::vector<experimental::distributed::comm_index_type>
new_recv_size(num_rank);
std::vector<experimental::distributed::comm_index_type>
new_recv_offsets(num_rank + 1);
array<IndexType> new_recv_gather_idxs(exec->get_master());
compute_communication(recv_offsets, non_local_agg, renumber,
new_recv_size, new_recv_offsets,
new_recv_gather_idxs);

non_local_col_map.set_executor(exec);
IndexType non_local_num_agg = new_recv_gather_idxs.get_size();
// build csr from row and col map
// unlike non-distributed version, generate_coarse uses different
// row and col maps.
auto result_non_local_csr = generate_coarse(
exec, non_local_csr.get(),
static_cast<IndexType>(std::get<1>(result)->get_size()[0]),
agg_, non_local_num_agg, non_local_col_map);
// use local and non-local to build coarse matrix
// also restriction and prolongation (Local-only-global matrix)
auto coarse_size =
static_cast<int64>(std::get<1>(result)->get_size()[0]);
comm.all_reduce(exec->get_master(), &coarse_size, 1, MPI_SUM);
new_recv_gather_idxs.set_executor(exec);

// setup the generated linop.
using global_index_type =
typename std::decay_t<decltype(*matrix)>::global_index_type;
auto coarse = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm, gko::dim<2>(coarse_size, coarse_size),
std::get<1>(result), result_non_local_csr,
new_recv_size, new_recv_offsets, new_recv_gather_idxs));
auto restrict_op = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
dim<2>(coarse_size,
gko::as<LinOp>(matrix)->get_size()[0]),
std::get<2>(result)));
auto prolong_op = share(
experimental::distributed::
Matrix<ValueType, IndexType, global_index_type>::create(
exec, comm,
dim<2>(gko::as<LinOp>(matrix)->get_size()[0],
coarse_size),
std::get<0>(result)));
this->set_multigrid_level(prolong_op, coarse, restrict_op);
};

// the fine op is using csr with the current ValueType
run<fst_mtx_type, snd_mtx_type>(this->get_fine_op(), distributed_setup);
} else
#endif // GINKGO_BUILD_MPI
{
Expand Down
Loading

0 comments on commit de5c2a7

Please sign in to comment.