Skip to content

Commit

Permalink
enable distributed multigrid
Browse files Browse the repository at this point in the history 
Co-authored-by: Pratik Nayak <pratikvn@protonmail.com>
  • Loading branch information
@yhmtsai @pratikvn
yhmtsai and pratikvn committed Jan 16, 2024
1 parent 4558fb2 commit 8add2c1
Show file tree
Hide file tree
Showing 2 changed files with 168 additions and 47 deletions.
195 changes: 149 additions & 46 deletions core/solver/multigrid.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,8 @@
#include <ginkgo/core/base/math.hpp>
#include <ginkgo/core/base/utils.hpp>
#include <ginkgo/core/base/utils_helper.hpp>
#include <ginkgo/core/distributed/matrix.hpp>
#include <ginkgo/core/distributed/vector.hpp>
#include <ginkgo/core/factorization/lu.hpp>
#include <ginkgo/core/matrix/dense.hpp>
#include <ginkgo/core/preconditioner/jacobi.hpp>
Expand All @@ -27,6 +29,7 @@

#include "core/base/dispatch_helper.hpp"
#include "core/components/fill_array_kernels.hpp"
#include "core/distributed/helpers.hpp"
#include "core/solver/ir_kernels.hpp"
#include "core/solver/multigrid_kernels.hpp"
#include "core/solver/solver_base.hpp"
Expand Down Expand Up @@ -80,26 +83,6 @@ casting(const T& x)
return static_cast<ValueType>(real(x));
}

/**
* as_vec gives a shortcut for casting pointer to dense.
*/
template <typename ValueType>
auto as_vec(std::shared_ptr<LinOp> x)
{
return std::static_pointer_cast<matrix::Dense<ValueType>>(x);
}


/**
* as_real_vec gives a shortcut for casting pointer to dense with real type.
*/
template <typename ValueType>
auto as_real_vec(std::shared_ptr<LinOp> x)
{
return std::static_pointer_cast<matrix::Dense<remove_complex<ValueType>>>(
x);
}


/**
* handle_list generate the smoother for each MultigridLevel
Expand Down Expand Up @@ -212,17 +195,43 @@ struct MultigridState {
/**
* allocate_memory is a helper function to allocate the memory of one level
*
* @tparam ValueType the value type of memory
* @tparam VectorType the vector type
*
* @param level the current level index
* @param cycle the multigrid cycle
* @param current_nrows the number of rows of current fine matrix
* @param next_nrows the number of rows of next coarse matrix
*/
template <typename ValueType>
template <typename VectorType>
void allocate_memory(int level, multigrid::cycle cycle,
size_type current_nrows, size_type next_nrows);

#if GINKGO_BUILD_MPI
/**
* allocate_memory is a helper function to allocate the memory of one level
*
* @tparam VectorType the vector type
*
* @param level the current level index
* @param cycle the multigrid cycle
* @param current_comm the communicator of the current fine matrix
* @param next_comm the communicator of the next coarse matrix
* @param current_nrows the number of rows of the current fine matrix
* @param next_nrows the number of rows of the next coarse matrix
* @param current_local_nrows the number of rows of the local operator of
* the current fine matrix
* @param next_local_nrows the number of rows of the local operator of the
* next coarse matrix
*/
template <typename VectorType>
void allocate_memory(int level, multigrid::cycle cycle,
experimental::mpi::communicator& current_comm,
experimental::mpi::communicator& next_comm,
size_type current_nrows, size_type next_nrows,
size_type current_local_nrows,
size_type next_local_nrows);
#endif

/**
* run the cycle of the level
*
Expand All @@ -240,11 +249,11 @@ struct MultigridState {
/**
* @copydoc run_cycle
*
* @tparam ValueType the value type
* @tparam VectorType the vector type
*
* @note it is the version with known ValueType
*/
template <typename ValueType>
template <typename VectorType>
void run_cycle(multigrid::cycle cycle, size_type level,
const std::shared_ptr<const LinOp>& matrix, const LinOp* b,
LinOp* x, cycle_mode mode);
Expand Down Expand Up @@ -291,13 +300,44 @@ void MultigridState::generate(const LinOp* system_matrix_in,
mg_level,
[&, this](auto mg_level, auto i, auto cycle, auto current_nrows,
auto next_nrows) {
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
using vec = matrix::Dense<value_type>;
this->allocate_memory<value_type>(i, cycle, current_nrows,
next_nrows);
auto exec = as<LinOp>(multigrid->get_mg_level_list().at(i))
->get_executor();
#if GINKGO_BUILD_MPI
if (gko::detail::is_distributed(system_matrix_in)) {
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
using VectorType =
experimental::distributed::Vector<value_type>;
auto fine = mg_level->get_fine_op().get();
auto coarse = mg_level->get_coarse_op().get();
auto current_comm =
dynamic_cast<
const experimental::distributed::DistributedBase*>(
fine)
->get_communicator();
auto next_comm =
dynamic_cast<
const experimental::distributed::DistributedBase*>(
coarse)
->get_communicator();
auto current_local_nrows =
dynamic_cast<const experimental::distributed::
DistributedLocalSize*>(fine)
->get_local_size()[0];
auto next_local_nrows =
dynamic_cast<const experimental::distributed::
DistributedLocalSize*>(coarse)
->get_local_size()[0];
this->allocate_memory<VectorType>(
i, cycle, current_comm, next_comm, current_nrows,
next_nrows, current_local_nrows, next_local_nrows);
} else
#endif
{
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
using VectorType = matrix::Dense<value_type>;
this->allocate_memory<VectorType>(i, cycle, current_nrows,
next_nrows);
}
},
i, cycle, current_nrows, next_nrows);

Expand All @@ -306,13 +346,13 @@ void MultigridState::generate(const LinOp* system_matrix_in,
}


template <typename ValueType>
template <class VectorType>
void MultigridState::allocate_memory(int level, multigrid::cycle cycle,
size_type current_nrows,
size_type next_nrows)
{
using vec = matrix::Dense<ValueType>;
using norm_vec = matrix::Dense<remove_complex<ValueType>>;
using value_type = typename VectorType::value_type;
using vec = matrix::Dense<value_type>;

auto exec =
as<LinOp>(multigrid->get_mg_level_list().at(level))->get_executor();
Expand All @@ -321,19 +361,70 @@ void MultigridState::allocate_memory(int level, multigrid::cycle cycle,
// allocate the previous level
g_list.emplace_back(vec::create(exec, dim<2>{current_nrows, nrhs}));
e_list.emplace_back(vec::create(exec, dim<2>{current_nrows, nrhs}));
next_one_list.emplace_back(initialize<vec>({one<ValueType>()}, exec));
next_one_list.emplace_back(initialize<vec>({one<value_type>()}, exec));
}
if (level + 1 == multigrid->get_mg_level_list().size()) {
// the last level allocate the g, e for coarsest solver
g_list.emplace_back(vec::create(exec, dim<2>{next_nrows, nrhs}));
e_list.emplace_back(vec::create(exec, dim<2>{next_nrows, nrhs}));
next_one_list.emplace_back(initialize<vec>({one<ValueType>()}, exec));
next_one_list.emplace_back(initialize<vec>({one<value_type>()}, exec));
}
one_list.emplace_back(initialize<vec>({one<ValueType>()}, exec));
neg_one_list.emplace_back(initialize<vec>({-one<ValueType>()}, exec));
one_list.emplace_back(initialize<vec>({one<value_type>()}, exec));
neg_one_list.emplace_back(initialize<vec>({-one<value_type>()}, exec));
}


#if GINKGO_BUILD_MPI


template <typename VectorType>
void MultigridState::allocate_memory(
int level, multigrid::cycle cycle,
experimental::mpi::communicator& current_comm,
experimental::mpi::communicator& next_comm, size_type current_nrows,
size_type next_nrows, size_type current_local_nrows,
size_type next_local_nrows)
{
using value_type = typename VectorType::value_type;
using vec = VectorType;
using dense_vec = matrix::Dense<value_type>;

auto exec =
as<LinOp>(multigrid->get_mg_level_list().at(level))->get_executor();
r_list.emplace_back(vec::create(exec, current_comm,
dim<2>{current_nrows, nrhs},
dim<2>{current_local_nrows, nrhs}));
if (level != 0) {
// allocate the previous level
g_list.emplace_back(vec::create(exec, current_comm,
dim<2>{current_nrows, nrhs},
dim<2>{current_local_nrows, nrhs}));
e_list.emplace_back(vec::create(exec, current_comm,
dim<2>{current_nrows, nrhs},
dim<2>{current_local_nrows, nrhs}));
next_one_list.emplace_back(
initialize<dense_vec>({one<value_type>()}, exec));
}
if (level + 1 == multigrid->get_mg_level_list().size()) {
// the last level allocate the g, e for coarsest solver
g_list.emplace_back(vec::create(exec, next_comm,
dim<2>{next_nrows, nrhs},
dim<2>{next_local_nrows, nrhs}));
e_list.emplace_back(vec::create(exec, next_comm,
dim<2>{next_nrows, nrhs},
dim<2>{next_local_nrows, nrhs}));
next_one_list.emplace_back(
initialize<dense_vec>({one<value_type>()}, exec));
}
one_list.emplace_back(initialize<dense_vec>({one<value_type>()}, exec));
neg_one_list.emplace_back(
initialize<dense_vec>({-one<value_type>()}, exec));
}


#endif


void MultigridState::run_mg_cycle(multigrid::cycle cycle, size_type level,
const std::shared_ptr<const LinOp>& matrix,
const LinOp* b, LinOp* x, cycle_mode mode)
Expand All @@ -346,23 +437,36 @@ void MultigridState::run_mg_cycle(multigrid::cycle cycle, size_type level,
run<gko::multigrid::EnableMultigridLevel, float, double,
std::complex<float>, std::complex<double>>(
mg_level, [&, this](auto mg_level) {
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
this->run_cycle<value_type>(cycle, level, matrix, b, x, mode);
#if GINKGO_BUILD_MPI
if (gko::detail::is_distributed(matrix.get())) {
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
this->run_cycle<
typename experimental::distributed::Vector<value_type>>(
cycle, level, matrix, b, x, mode);
} else
#endif
{
using value_type =
typename std::decay_t<decltype(*mg_level)>::value_type;
this->run_cycle<typename matrix::Dense<value_type>>(
cycle, level, matrix, b, x, mode);
}
});
}


template <typename ValueType>
template <typename VectorType>
void MultigridState::run_cycle(multigrid::cycle cycle, size_type level,
const std::shared_ptr<const LinOp>& matrix,
const LinOp* b, LinOp* x, cycle_mode mode)
{
using value_type = typename VectorType::value_type;
auto total_level = multigrid->get_mg_level_list().size();

auto r = r_list.at(level);
auto g = g_list.at(level);
auto e = e_list.at(level);
auto e = as<VectorType>(e_list.at(level));
// get mg_level
auto mg_level = multigrid->get_mg_level_list().at(level);
// get the pre_smoother
Expand Down Expand Up @@ -392,8 +496,7 @@ void MultigridState::run_cycle(multigrid::cycle cycle, size_type level,
} else {
// x in first level is already filled by zero outside.
if (level != 0) {
dynamic_cast<matrix::Dense<ValueType>*>(x)->fill(
zero<ValueType>());
dynamic_cast<VectorType*>(x)->fill(zero<value_type>());
}
pre_smoother->apply(b, x);
}
Expand All @@ -414,7 +517,7 @@ void MultigridState::run_cycle(multigrid::cycle cycle, size_type level,
// next level
if (level + 1 == total_level) {
// the coarsest solver use the last level valuetype
as_vec<ValueType>(e)->fill(zero<ValueType>());
e->fill(zero<value_type>());
}
auto next_level_matrix =
(level + 1 < total_level)
Expand Down
20 changes: 19 additions & 1 deletion include/ginkgo/core/distributed/matrix.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -127,6 +127,21 @@ template <typename ValueType>
class Vector;


/**
* DistributedLocalSize is a feature class providing `get_local_size` which
* return the size of local operator.
*/
class DistributedLocalSize {
public:
/**
* get the size of local operator
*
* @return the size of local operator
*/
virtual dim<2> get_local_size() const = 0;
};


/**
* The Matrix class defines a (MPI-)distributed matrix.
*
Expand Down Expand Up @@ -240,7 +255,8 @@ class Matrix
Matrix<ValueType, LocalIndexType, GlobalIndexType>>,
public ConvertibleTo<
Matrix<next_precision<ValueType>, LocalIndexType, GlobalIndexType>>,
public DistributedBase {
public DistributedBase,
public DistributedLocalSize {
friend class EnableCreateMethod<Matrix>;
friend class EnableDistributedPolymorphicObject<Matrix, LinOp>;
friend class Matrix<next_precision<ValueType>, LocalIndexType,
Expand Down Expand Up @@ -339,6 +355,8 @@ class Matrix
ptr_param<const Partition<local_index_type, global_index_type>>
col_partition);

dim<2> get_local_size() const override { return local_mtx_->get_size(); }

/**
* Get read access to the stored local matrix.
*
Expand Down

0 comments on commit 8add2c1

Please sign in to comment.