Skip to content

Utility functions to create an edge_src|dst_property view objects. #5078

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
merged 2 commits into from
May 20, 2025
Merged

Utility functions to create an edge_src|dst_property view objects. #5078

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4bbdf82
add a utility function to create an edge_endpoint_property_view object
seunghwak May 17, 2025
7737228
silence a build warning
seunghwak May 19, 2025
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
176 changes: 170 additions & 6 deletions cpp/include/cugraph/edge_src_dst_property.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -294,9 +294,9 @@ class edge_major_property_t {

void clear()
{
edge_partition_keys_ = std::nullopt;
edge_partition_key_chunk_start_offsets_ = std::nullopt;
key_chunk_size_ = std::nullopt;
edge_partition_keys_.reset();
edge_partition_key_chunk_start_offsets_.reset();
key_chunk_size_.reset();

buffers_.clear();
buffers_.shrink_to_fit();
Expand Down Expand Up @@ -424,9 +424,9 @@ class edge_minor_property_t {

void clear()
{
keys_ = std::nullopt;
key_chunk_start_offsets_ = std::nullopt;
key_chunk_size_ = std::nullopt;
keys_.reset();
key_chunk_start_offsets_.reset();
key_chunk_size_.reset();

rmm::cuda_stream_view stream{};
if constexpr (std::is_arithmetic_v<T>) {
Expand Down Expand Up @@ -546,6 +546,12 @@ class edge_src_property_t {
public:
static_assert(is_arithmetic_or_thrust_tuple_of_arithmetic<T>::value);

using value_type = T;
using value_iterator =
dataframe_buffer_iterator_type_t<std::conditional_t<std::is_same_v<T, bool>, uint32_t, T>>;
using const_value_iterator = dataframe_buffer_const_iterator_type_t<
std::conditional_t<std::is_same_v<T, bool>, uint32_t, T>>;

edge_src_property_t(raft::handle_t const& handle) {}

template <typename GraphViewType>
Expand Down Expand Up @@ -599,6 +605,10 @@ class edge_dst_property_t {
static_assert(is_arithmetic_or_thrust_tuple_of_arithmetic<T>::value);

using value_type = T;
using value_iterator =
dataframe_buffer_iterator_type_t<std::conditional_t<std::is_same_v<T, bool>, uint32_t, T>>;
using const_value_iterator = dataframe_buffer_const_iterator_type_t<
std::conditional_t<std::is_same_v<T, bool>, uint32_t, T>>;

edge_dst_property_t(raft::handle_t const& handle) {}

Expand Down Expand Up @@ -661,6 +671,160 @@ class edge_dst_dummy_property_t {
auto view() const { return detail::edge_endpoint_dummy_property_view_t{}; }
};

// SG-only (use a vertex property buffer instead of creating a new edge_src_property_t object to
// save memory)
template <typename vertex_t, typename T, typename GraphViewType>
detail::edge_endpoint_property_view_t<vertex_t,
typename edge_src_property_t<vertex_t,
T>::const_value_iterator,
T>
make_edge_src_property_view(
GraphViewType const& graph_view,
typename edge_src_property_t<vertex_t, T>::const_value_iterator value_first,
size_t num_values)
{
using const_value_iterator = typename edge_src_property_t<vertex_t, T>::const_value_iterator;

CUGRAPH_EXPECTS(!GraphViewType::is_multi_gpu,
"Invalid input argument: this function is only for single-GPU.");

vertex_t range_size{};
if constexpr (GraphViewType::is_storage_transposed) { // minor
range_size = graph_view.local_edge_partition_src_range_size();
} else { // major
range_size = graph_view.local_edge_partition_src_range_size(size_t{0});
}
auto expected_num_values = std::is_same_v<T, bool>
? cugraph::packed_bool_size(static_cast<size_t>(range_size))
: static_cast<size_t>(range_size);
CUGRAPH_EXPECTS(
num_values == expected_num_values,
"Invalid input argument: num_values does not match the expected number of values.");

if constexpr (GraphViewType::is_storage_transposed) { // minor
return detail::edge_endpoint_property_view_t<vertex_t, const_value_iterator, T>(value_first,
vertex_t{0});
} else { // major
return detail::edge_endpoint_property_view_t<vertex_t, const_value_iterator, T>(
std::vector<const_value_iterator>{value_first}, std::vector<vertex_t>{vertex_t{0}});
}
}

// SG-only (use a vertex property buffer instead of creating a new edge_src_property_t object to
// save memory)
template <typename vertex_t, typename T, typename GraphViewType>
detail::edge_endpoint_property_view_t<vertex_t,
typename edge_src_property_t<vertex_t, T>::value_iterator,
T>
make_edge_src_property_mutable_view(
GraphViewType const& graph_view,
typename edge_src_property_t<vertex_t, T>::value_iterator value_first,
size_t num_values)
{
using value_iterator = typename edge_src_property_t<vertex_t, T>::value_iterator;

CUGRAPH_EXPECTS(!GraphViewType::is_multi_gpu,
"Invalid input argument: this function is only for single-GPU.");

vertex_t range_size{};
if constexpr (GraphViewType::is_storage_transposed) { // minor
range_size = graph_view.local_edge_partition_src_range_size();
} else { // major
range_size = graph_view.local_edge_partition_src_range_size(size_t{0});
}
auto expected_num_values = std::is_same_v<T, bool>
? cugraph::packed_bool_size(static_cast<size_t>(range_size))
: static_cast<size_t>(range_size);
CUGRAPH_EXPECTS(
num_values == expected_num_values,
"Invalid input argument: num_values does not match the expected number of values.");

if constexpr (GraphViewType::is_storage_transposed) { // minor
return detail::edge_endpoint_property_view_t<vertex_t, value_iterator, T>(value_first,
vertex_t{0});
} else { // major
return detail::edge_endpoint_property_view_t<vertex_t, value_iterator, T>(
std::vector<value_iterator>{value_first}, std::vector<vertex_t>{vertex_t{0}});
}
}

// SG-only (use a vertex property buffer instead of creating a new edge_dst_property_t object to
// save memory)
template <typename vertex_t, typename T, typename GraphViewType>
detail::edge_endpoint_property_view_t<vertex_t,
typename edge_dst_property_t<vertex_t,
T>::const_value_iterator,
T>
make_edge_dst_property_view(
GraphViewType const& graph_view,
typename edge_dst_property_t<vertex_t, T>::const_value_iterator value_first,
size_t num_values)
{
using const_value_iterator = typename edge_dst_property_t<vertex_t, T>::const_value_iterator;

CUGRAPH_EXPECTS(!GraphViewType::is_multi_gpu,
"Invalid input argument: this function is only for single-GPU.");

vertex_t range_size{};
if constexpr (GraphViewType::is_storage_transposed) { // major
range_size = graph_view.local_edge_partition_dst_range_size(size_t{0});
} else { // minor
range_size = graph_view.local_edge_partition_dst_range_size();
}
auto expected_num_values = std::is_same_v<T, bool>
? cugraph::packed_bool_size(static_cast<size_t>(range_size))
: static_cast<size_t>(range_size);
CUGRAPH_EXPECTS(
num_values == expected_num_values,
"Invalid input argument: num_values does not match the expected number of values.");

if constexpr (GraphViewType::is_storage_transposed) { // major
return detail::edge_endpoint_property_view_t<vertex_t, const_value_iterator, T>(
std::vector<const_value_iterator>{value_first}, std::vector<vertex_t>{vertex_t{0}});
} else { // minor
return detail::edge_endpoint_property_view_t<vertex_t, const_value_iterator, T>(value_first,
vertex_t{0});
}
}

// SG-only (use a vertex property buffer instead of creating a new edge_dst_property_t object to
// save memory)
template <typename vertex_t, typename T, typename GraphViewType>
detail::edge_endpoint_property_view_t<vertex_t,
typename edge_dst_property_t<vertex_t, T>::value_iterator,
T>
make_edge_dst_property_mutable_view(
GraphViewType const& graph_view,
typename edge_dst_property_t<vertex_t, T>::value_iterator value_first,
size_t num_values)
{
using value_iterator = typename edge_dst_property_t<vertex_t, T>::value_iterator;

CUGRAPH_EXPECTS(!GraphViewType::is_multi_gpu,
"Invalid input argument: this function is only for single-GPU.");

vertex_t range_size{};
if constexpr (GraphViewType::is_storage_transposed) { // major
range_size = graph_view.local_edge_partition_dst_range_size(size_t{0});
} else { // minor
range_size = graph_view.local_edge_partition_dst_range_size();
}
auto expected_num_values = std::is_same_v<T, bool>
? cugraph::packed_bool_size(static_cast<size_t>(range_size))
: static_cast<size_t>(range_size);
CUGRAPH_EXPECTS(
num_values == expected_num_values,
"Invalid input argument: num_values does not match the expected number of values.");

if constexpr (GraphViewType::is_storage_transposed) { // major
return detail::edge_endpoint_property_view_t<vertex_t, value_iterator, T>(
std::vector<value_iterator>{value_first}, std::vector<vertex_t>{vertex_t{0}});
} else { // minor
return detail::edge_endpoint_property_view_t<vertex_t, value_iterator, T>(value_first,
vertex_t{0});
}
}

template <typename vertex_t, typename... Iters, typename... Types>
auto view_concat(detail::edge_endpoint_property_view_t<vertex_t, Iters, Types> const&... views)
{
Expand Down
61 changes: 32 additions & 29 deletions cpp/src/community/approx_weighted_matching_impl.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,9 +45,7 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
"Invalid input arguments: input graph for approximate_weighted_matching must "
"need to be symmetric");

using graph_view_t = cugraph::graph_view_t<vertex_t, edge_t, false, multi_gpu>;

graph_view_t current_graph_view(graph_view);
auto current_graph_view = graph_view;
if (current_graph_view.has_edge_mask()) { current_graph_view.clear_edge_mask(); }

cugraph::edge_property_t<edge_t, bool> edge_masks_even(handle, current_graph_view);
Expand Down Expand Up @@ -100,7 +98,7 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
cugraph::edge_src_property_t<vertex_t, bool> src_match_flags(handle);
cugraph::edge_dst_property_t<vertex_t, bool> dst_match_flags(handle);

if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
src_key_cache = cugraph::edge_src_property_t<vertex_t, vertex_t>(handle, current_graph_view);

update_edge_src_property(
Expand Down Expand Up @@ -129,11 +127,9 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
cugraph::edge_src_dummy_property_t{}.view(),
cugraph::edge_dst_dummy_property_t{}.view(),
edge_weight_view,
graph_view_t::is_multi_gpu
? src_key_cache.view()
: detail::edge_endpoint_property_view_t<vertex_t, vertex_t const*>(
std::vector<vertex_t const*>{local_vertices.begin()},
std::vector<vertex_t>{vertex_t{0}}),
multi_gpu ? src_key_cache.view()
: make_edge_src_property_view<vertex_t, vertex_t>(
current_graph_view, local_vertices.begin(), local_vertices.size()),
[] __device__(auto, auto dst, cuda::std::nullopt_t, cuda::std::nullopt_t, auto wt) {
return thrust::make_tuple(wt, dst);
},
Expand All @@ -145,7 +141,7 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
// For each target, find the best offer
//

if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
auto vertex_partition_range_lasts = current_graph_view.vertex_partition_range_lasts();

rmm::device_uvector<vertex_t> d_vertex_partition_range_lasts(
Expand Down Expand Up @@ -225,7 +221,7 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin

rmm::device_uvector<vertex_t> candidates_of_candidates(0, handle.get_stream());

if (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
auto& comm = handle.get_comms();
auto& major_comm = handle.get_subcomm(cugraph::partition_manager::major_comm_name());
auto const major_comm_size = major_comm.get_size();
Expand Down Expand Up @@ -302,15 +298,15 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin

if (current_graph_view.compute_number_of_edges(handle) == 0) { break; }

if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
cugraph::update_edge_src_property(
handle, current_graph_view, is_vertex_matched.begin(), src_match_flags.mutable_view());
cugraph::update_edge_dst_property(
handle, current_graph_view, is_vertex_matched.begin(), dst_match_flags.mutable_view());
}

if (loop_counter % 2 == 0) {
if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
cugraph::transform_e(
handle,
current_graph_view,
Expand All @@ -326,14 +322,17 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
cugraph::transform_e(
handle,
current_graph_view,
detail::edge_endpoint_property_view_t<vertex_t, bool const*>(
std::vector<bool const*>{is_vertex_matched.begin()},
std::vector<vertex_t>{vertex_t{0}}),
detail::edge_endpoint_property_view_t<vertex_t, bool const*>(is_vertex_matched.begin(),
vertex_t{0}),
cugraph::edge_src_dummy_property_t{}.view(),
cugraph::edge_dst_dummy_property_t{}.view(),
cugraph::edge_dummy_property_t{}.view(),
[] __device__(
auto src, auto dst, auto is_src_matched, auto is_dst_matched, cuda::std::nullopt_t) {
[is_vertex_matched = raft::device_span<bool const>(
is_vertex_matched.data(), is_vertex_matched.size())] __device__(auto src,
auto dst,
cuda::std::nullopt_t,
cuda::std::nullopt_t,
cuda::std::nullopt_t) {
auto is_src_matched = is_vertex_matched[src];
auto is_dst_matched = is_vertex_matched[dst];
return !((is_src_matched == true) || (is_dst_matched == true));
},
edge_masks_odd.mutable_view());
Expand All @@ -344,7 +343,7 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
handle, current_graph_view, edge_masks_even.mutable_view(), bool{false});
current_graph_view.attach_edge_mask(edge_masks_odd.view());
} else {
if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
cugraph::transform_e(
handle,
current_graph_view,
Expand All @@ -360,14 +359,17 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
cugraph::transform_e(
handle,
current_graph_view,
detail::edge_endpoint_property_view_t<vertex_t, bool const*>(
std::vector<bool const*>{is_vertex_matched.begin()},
std::vector<vertex_t>{vertex_t{0}}),
detail::edge_endpoint_property_view_t<vertex_t, bool const*>(is_vertex_matched.begin(),
vertex_t{0}),
cugraph::edge_src_dummy_property_t{}.view(),
cugraph::edge_dst_dummy_property_t{}.view(),
cugraph::edge_dummy_property_t{}.view(),
[] __device__(
auto src, auto dst, auto is_src_matched, auto is_dst_matched, cuda::std::nullopt_t) {
[is_vertex_matched = raft::device_span<bool const>(
is_vertex_matched.data(), is_vertex_matched.size())] __device__(auto src,
auto dst,
cuda::std::nullopt_t,
cuda::std::nullopt_t,
cuda::std::nullopt_t) {
auto is_src_matched = is_vertex_matched[src];
auto is_dst_matched = is_vertex_matched[dst];
return !((is_src_matched == true) || (is_dst_matched == true));
},
edge_masks_even.mutable_view());
Expand All @@ -385,13 +387,14 @@ std::tuple<rmm::device_uvector<vertex_t>, weight_t> approximate_weighted_matchin
weight_t sum_matched_edge_weights = thrust::reduce(
handle.get_thrust_policy(), offers_from_partners.begin(), offers_from_partners.end());

if constexpr (graph_view_t::is_multi_gpu) {
if constexpr (multi_gpu) {
sum_matched_edge_weights = host_scalar_allreduce(
handle.get_comms(), sum_matched_edge_weights, raft::comms::op_t::SUM, handle.get_stream());
}

return std::make_tuple(std::move(partners), sum_matched_edge_weights / 2.0);
}

} // namespace detail

template <typename vertex_t, typename edge_t, typename weight_t, bool multi_gpu>
Expand Down
Loading