Skip to content

Commit

Permalink
implement other gain types for deterministic label prop refiner
Browse files Browse the repository at this point in the history
  • Loading branch information
@larsgottesbueren
larsgottesbueren committed Sep 29, 2023
1 parent c7536ff commit fa53892
Show file tree
Hide file tree
Showing 4 changed files with 48 additions and 56 deletions.
2 changes: 1 addition & 1 deletion mt-kahypar/partition/factories.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -93,7 +93,7 @@ using LabelPropagationDispatcher = kahypar::meta::StaticMultiDispatchFactory<
using DeterministicLabelPropagationDispatcher = kahypar::meta::StaticMultiDispatchFactory<
DeterministicLabelPropagationRefiner,
IRefiner,
kahypar::meta::Typelist<TypeTraitsList>>;
kahypar::meta::Typelist<TypeTraitsList, GainTypes>>;

using FMFactory = kahypar::meta::Factory<FMAlgorithm,
IRefiner* (*)(HypernodeID, HyperedgeID, const Context&, gain_cache_t, IRebalancer&)>;
Expand Down
72 changes: 42 additions & 30 deletions mt-kahypar/partition/refinement/deterministic/deterministic_label_propagation.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -32,14 +32,16 @@
#include "mt-kahypar/parallel/chunking.h"
#include "mt-kahypar/parallel/parallel_counting_sort.h"
#include "mt-kahypar/utils/cast.h"
#include "mt-kahypar/partition/refinement/gains/gain_definitions.h"

#include <tbb/parallel_sort.h>
#include <tbb/parallel_reduce.h>

namespace mt_kahypar {

template<typename TypeTraits>
bool DeterministicLabelPropagationRefiner<TypeTraits>::refineImpl(

template<typename TypeTraits, typename GainTypes>
bool DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::refineImpl(
mt_kahypar_partitioned_hypergraph_t& hypergraph,
const vec<HypernodeID>&,
Metrics& best_metrics,
Expand All @@ -49,6 +51,9 @@ namespace mt_kahypar {
constexpr size_t num_buckets = utils::ParallelPermutation<HypernodeID>::num_buckets;
size_t num_sub_rounds = context.refinement.deterministic_refinement.num_sub_rounds_sync_lp;

using GainComputation = typename GainTypes::GainComputation;
GainComputation gain_computation(context);

for (size_t iter = 0; iter < context.refinement.label_propagation.maximum_iterations; ++iter) {
if (context.refinement.deterministic_refinement.use_active_node_set && ++round == 0) {
std::fill(last_moved_in_round.begin(), last_moved_in_round.end(), CAtomic<uint32_t>(0));
Expand All @@ -70,22 +75,23 @@ namespace mt_kahypar {
bool increase_sub_rounds = false;
for (size_t sub_round = 0; sub_round < num_sub_rounds; ++sub_round) {
auto[first_bucket, last_bucket] = parallel::chunking::bounds(sub_round, num_buckets, num_buckets_per_sub_round);
assert(first_bucket < last_bucket && last_bucket < permutation.bucket_bounds.size());
ASSERT(first_bucket < last_bucket && last_bucket < permutation.bucket_bounds.size());
size_t first = permutation.bucket_bounds[first_bucket], last = permutation.bucket_bounds[last_bucket];
moves.clear();

// calculate moves
if (phg.k() == 2) {
tbb::parallel_for(HypernodeID(first), HypernodeID(last), [&](const HypernodeID position) {
assert(position < permutation.permutation.size());
calculateAndSaveBestMoveTwoWay(phg, permutation.at(position));
});
} else {
tbb::parallel_for(HypernodeID(first), HypernodeID(last), [&](const HypernodeID position) {
assert(position < permutation.permutation.size());
calculateAndSaveBestMove(phg, permutation.at(position));
});
}
tbb::parallel_for(HypernodeID(first), HypernodeID(last), [&](const HypernodeID position) {
ASSERT(position < permutation.permutation.size());
const HypernodeID u = permutation.at(position);
ASSERT(u < phg.initialNumNodes());
if (!phg.nodeIsEnabled(u) || !phg.isBorderNode(u)) return;
Move move = gain_computation.computeMaxGainMove(phg, u, /*rebalance=*/false, /*consider_non_adjacent_blocks=*/false, /*allow_imbalance=*/true);
move.gain = -move.gain;
if (move.gain > 0 && move.to != phg.partID(u)) {
moves.push_back_buffered(move);
}
});

moves.finalize();

Gain sub_round_improvement = 0;
Expand Down Expand Up @@ -127,12 +133,13 @@ namespace mt_kahypar {
/*
* for configs where we don't know exact gains --> have to trace the overall improvement with attributed gains
*/
template<typename TypeTraits>
Gain DeterministicLabelPropagationRefiner<TypeTraits>::performMoveWithAttributedGain(
template<typename TypeTraits, typename GainTypes>
Gain DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::performMoveWithAttributedGain(
PartitionedHypergraph& phg, const Move& m, bool activate_neighbors) {
Gain attributed_gain = 0;
auto objective_delta = [&](const SynchronizedEdgeUpdate& sync_update) {
attributed_gain -= Km1AttributedGains::gain(sync_update);
using AttributedGains = typename GainTypes::AttributedGains;
attributed_gain -= AttributedGains::gain(sync_update);
};
const bool was_moved = phg.changeNodePart(m.node, m.from, m.to, objective_delta);
if (context.refinement.deterministic_refinement.use_active_node_set && activate_neighbors && was_moved) {
Expand All @@ -156,9 +163,9 @@ namespace mt_kahypar {
return attributed_gain;
}

template<typename TypeTraits>
template<typename TypeTraits, typename GainTypes>
template<typename Predicate>
Gain DeterministicLabelPropagationRefiner<TypeTraits>::applyMovesIf(
Gain DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::applyMovesIf(
PartitionedHypergraph& phg, const vec<Move>& my_moves, size_t end, Predicate&& predicate) {
auto range = tbb::blocked_range<size_t>(UL(0), end);
auto accum = [&](const tbb::blocked_range<size_t>& r, const Gain& init) -> Gain {
Expand Down Expand Up @@ -196,8 +203,8 @@ namespace mt_kahypar {
return res;
}

template<typename TypeTraits>
Gain DeterministicLabelPropagationRefiner<TypeTraits>::applyMovesSortedByGainAndRevertUnbalanced(PartitionedHypergraph& phg) {
template<typename TypeTraits, typename GainTypes>
Gain DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::applyMovesSortedByGainAndRevertUnbalanced(PartitionedHypergraph& phg) {
const size_t num_moves = moves.size();
tbb::parallel_sort(moves.begin(), moves.begin() + num_moves, [](const Move& m1, const Move& m2) {
return m1.gain > m2.gain || (m1.gain == m2.gain && m1.node < m2.node);
Expand Down Expand Up @@ -287,8 +294,8 @@ namespace mt_kahypar {
return gain;
}

template<typename TypeTraits>
std::pair<Gain, bool> DeterministicLabelPropagationRefiner<TypeTraits>::applyMovesByMaximalPrefixesInBlockPairs(PartitionedHypergraph& phg) {
template<typename TypeTraits, typename GainTypes>
std::pair<Gain, bool> DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::applyMovesByMaximalPrefixesInBlockPairs(PartitionedHypergraph& phg) {
PartitionID k = phg.k();
PartitionID max_key = k * k;
auto index = [&](PartitionID b1, PartitionID b2) { return b1 * k + b2; };
Expand Down Expand Up @@ -397,8 +404,8 @@ namespace mt_kahypar {
return std::make_pair(actual_gain, revert_all);
}

template<typename TypeTraits>
std::pair<size_t, size_t> DeterministicLabelPropagationRefiner<TypeTraits>::findBestPrefixesRecursive(
template<typename TypeTraits, typename GainTypes>
std::pair<size_t, size_t> DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::findBestPrefixesRecursive(
size_t p1_begin, size_t p1_end, size_t p2_begin, size_t p2_end,
size_t p1_invalid, size_t p2_invalid,
HypernodeWeight lb_p1, HypernodeWeight ub_p2)
Expand Down Expand Up @@ -473,8 +480,8 @@ namespace mt_kahypar {
}
}

template<typename TypeTraits>
std::pair<size_t, size_t> DeterministicLabelPropagationRefiner<TypeTraits>::findBestPrefixesSequentially(
template<typename TypeTraits, typename GainTypes>
std::pair<size_t, size_t> DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::findBestPrefixesSequentially(
size_t p1_begin, size_t p1_end, size_t p2_begin, size_t p2_end, size_t p1_inv, size_t p2_inv,
HypernodeWeight lb_p1, HypernodeWeight ub_p2)
{
Expand Down Expand Up @@ -502,8 +509,8 @@ namespace mt_kahypar {
return std::make_pair(invalid_pos, invalid_pos);
}

template<typename TypeTraits>
Gain DeterministicLabelPropagationRefiner<TypeTraits>::applyMovesSortedByGainWithRecalculation(PartitionedHypergraph& phg) {
template<typename TypeTraits, typename GainTypes>
Gain DeterministicLabelPropagationRefiner<TypeTraits, GainTypes>::applyMovesSortedByGainWithRecalculation(PartitionedHypergraph& phg) {
if (last_recalc_round.empty() || ++recalc_round == std::numeric_limits<uint32_t>::max()) {
last_recalc_round.assign(max_num_edges, CAtomic<uint32_t>(0));
}
Expand Down Expand Up @@ -635,5 +642,10 @@ namespace mt_kahypar {
return best_gain;
}

INSTANTIATE_CLASS_WITH_TYPE_TRAITS(DeterministicLabelPropagationRefiner)
namespace {
#define DETERMINISTIC_LABEL_PROPAGATION_REFINER(X, Y) DeterministicLabelPropagationRefiner<X, Y>
}


INSTANTIATE_CLASS_WITH_TYPE_TRAITS_AND_GAIN_TYPES(DETERMINISTIC_LABEL_PROPAGATION_REFINER)
} // namespace mt_kahypar
26 changes: 2 additions & 24 deletions mt-kahypar/partition/refinement/deterministic/deterministic_label_propagation.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@

namespace mt_kahypar {

template<typename TypeTraits>
template<typename TypeTraits, typename GainTypes>
class DeterministicLabelPropagationRefiner final : public IRefiner {

using PartitionedHypergraph = typename TypeTraits::PartitionedHypergraph;
Expand All @@ -55,7 +55,6 @@ class DeterministicLabelPropagationRefiner final : public IRefiner {
const HyperedgeID num_hyperedges,
const Context& context) :
context(context),
compute_gains(context),
cumulative_node_weights(num_hypernodes),
moves(num_hypernodes),
sorted_moves(num_hypernodes),
Expand Down Expand Up @@ -98,26 +97,6 @@ class DeterministicLabelPropagationRefiner final : public IRefiner {
size_t p1_begin, size_t p1_end, size_t p2_begin, size_t p2_end, size_t p1_inv, size_t p2_inv,
HypernodeWeight lb_p1, HypernodeWeight ub_p2);

MT_KAHYPAR_ATTRIBUTE_ALWAYS_INLINE
void calculateAndSaveBestMove(PartitionedHypergraph& phg, HypernodeID u) {
assert(u < phg.initialNumNodes());
if (!phg.nodeIsEnabled(u) || !phg.isBorderNode(u)) return;
//auto [to, gain] = compute_gains.local().computeBestTargetBlock(phg, u, context.partition.max_part_weights);
auto [to, gain] = compute_gains.local().computeBestTargetBlockIgnoringBalance(phg, u);
if (gain > 0 && to != kInvalidPartition) { // depending on apply moves function we might do gain >= 0
moves.push_back_buffered( { phg.partID(u), to, u, gain } );
}
}

MT_KAHYPAR_ATTRIBUTE_ALWAYS_INLINE
void calculateAndSaveBestMoveTwoWay(PartitionedHypergraph& phg, HypernodeID u) {
if (!phg.nodeIsEnabled(u) || !phg.isBorderNode(u)) return;
const Gain gain = TwoWayGainComputer::gainToOtherBlock(phg, u);
if (gain > 0) {
moves.push_back_buffered({ phg.partID(u), 1 - phg.partID(u), u, gain });
}
}

struct RecalculationData {
MoveID first_in, last_out;
HypernodeID remaining_pins;
Expand All @@ -129,13 +108,12 @@ class DeterministicLabelPropagationRefiner final : public IRefiner {
};

const Context& context;
tbb::enumerable_thread_specific<Km1GainComputer> compute_gains;
vec<HypernodeWeight> cumulative_node_weights;
ds::BufferedVector<Move> moves;
vec<Move> sorted_moves;

std::mt19937 prng;
utils::ParallelPermutation<HypernodeID> permutation; // gets memory only once used
utils::ParallelPermutation<HypernodeID> permutation;
ds::BufferedVector<HypernodeID> active_nodes;
vec<CAtomic<uint32_t>> last_moved_in_round;
uint32_t round = 0;
Expand Down
4 changes: 3 additions & 1 deletion mt-kahypar/partition/registries/register_refinement_algorithms.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -143,7 +143,9 @@ REGISTER_DISPATCHED_LP_REFINER(LabelPropagationAlgorithm::label_propagation,
REGISTER_DISPATCHED_LP_REFINER(LabelPropagationAlgorithm::deterministic,
DeterministicLabelPropagationDispatcher,
kahypar::meta::PolicyRegistry<mt_kahypar_partition_type_t>::getInstance().getPolicy(
context.partition.partition_type));
context.partition.partition_type),
kahypar::meta::PolicyRegistry<GainPolicy>::getInstance().getPolicy(
context.partition.gain_policy));
REGISTER_LP_REFINER(LabelPropagationAlgorithm::do_nothing, DoNothingRefiner, 1);

REGISTER_DISPATCHED_FM_REFINER(FMAlgorithm::kway_fm,
Expand Down

0 comments on commit fa53892

Please sign in to comment.