Open-Harbor/osrm-backend
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 3 Wiki Activity

Use Intel TBB's parallel_sort even for nested parallelism.

Browse Source 
TBB has a global task scheduler (that's one of the reason TBB is not
linked statically but dyanmically instead). This allows control over all
running threads, enabling us to use nested parallelism and the scheduler
doing all the task allocation itself.

That is, nested parallel execution such as in

    parallel_for(seq, [](const auto& rng){
      parallel_sort(rng);
    });

is no problem at all, as the scheduler still claims control over the
global environment.

Therefore, use `parallel_sort` Range overload where possible.

References:

- https://www.threadingbuildingblocks.org/docs/help/hh_goto.htm#reference/algorithms.htm
- https://www.threadingbuildingblocks.org/docs/help/hh_goto.htm#reference/algorithms/parallel_sort_func.htm
- https://www.threadingbuildingblocks.org/docs/help/hh_goto.htm#reference/task_scheduler.htm
- https://www.threadingbuildingblocks.org/docs/help/hh_goto.htm#reference/task_scheduler/task_scheduler_init_cls.htm
- https://www.threadingbuildingblocks.org/docs/help/hh_goto.htm#tbb_userguide/Initializing_and_Terminating_the_Library.htm
This commit is contained in:
Daniel J. Hofmann
2015-09-09 17:22:51 +02:00
parent dfac34beac
commit 9231335eef
11 changed files with 93 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files
contractor/contractor.hpp
+21 -15
View File
@@ -50,6 +50,7 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include <algorithm>
#include <limits>
#include <memory>
#include <vector>
class Contractor
@@ -172,7 +173,8 @@ class Contractor
SimpleLogger().Write(logWARNING)
<< "Edge weight large -> "
<< static_cast<unsigned int>(std::max(diter->weight, 1)) << " : "
<< static_cast<unsigned int>(diter->source) << " -> " << static_cast<unsigned int>(diter->target);
<< static_cast<unsigned int>(diter->source) << " -> "
<< static_cast<unsigned int>(diter->target);
}
#endif
edges.emplace_back(diter->source, diter->target,
@@ -190,7 +192,7 @@ class Contractor
// FIXME not sure if we need this
edges.shrink_to_fit();
tbb::parallel_sort(edges.begin(), edges.end());
tbb::parallel_sort(edges);
NodeID edge = 0;
for (NodeID i = 0; i < edges.size();)
{
@@ -284,7 +286,7 @@ class Contractor
~Contractor() {}
void Run( double core_factor = 1.0 )
void Run(double core_factor = 1.0)
{
// for the preperation we can use a big grain size, which is much faster (probably cache)
constexpr size_t InitGrainSize = 100000;
@@ -330,13 +332,15 @@ class Contractor
this->EvaluateNodePriority(data, &node_data[x], x);
}
});
std::cout << "ok" << std::endl << "preprocessing " << number_of_nodes << " nodes ..."
<< std::flush;
std::cout << "ok" << std::endl
<< "preprocessing " << number_of_nodes << " nodes ..." << std::flush;
bool flushed_contractor = false;
while (number_of_nodes > 2 && number_of_contracted_nodes < static_cast<NodeID>(number_of_nodes * core_factor) )
while (number_of_nodes > 2 &&
number_of_contracted_nodes < static_cast<NodeID>(number_of_nodes * core_factor))
{
if (!flushed_contractor && (number_of_contracted_nodes > static_cast<NodeID>(number_of_nodes * 0.65 * core_factor)))
if (!flushed_contractor && (number_of_contracted_nodes >
static_cast<NodeID>(number_of_nodes * 0.65 * core_factor)))
{
DeallocatingVector<ContractorEdge> new_edge_set; // this one is not explicitely
// cleared since it goes out of
@@ -360,7 +364,8 @@ class Contractor
for (const auto new_node_id : osrm::irange<std::size_t>(0, remaining_nodes.size()))
{
// create renumbering maps in both directions
orig_node_id_from_new_node_id_map[new_node_id] = remaining_nodes[new_node_id].id;
orig_node_id_from_new_node_id_map[new_node_id] =
remaining_nodes[new_node_id].id;
new_node_id_from_orig_id_map[remaining_nodes[new_node_id].id] = new_node_id;
new_node_priority[new_node_id] =
node_priorities[remaining_nodes[new_node_id].id];
@@ -411,7 +416,7 @@ class Contractor
contractor_graph.reset();
// create new graph
std::sort(new_edge_set.begin(), new_edge_set.end());
tbb::parallel_sort(new_edge_set);
contractor_graph =
std::make_shared<ContractorGraph>(remaining_nodes.size(), new_edge_set);
@@ -463,7 +468,7 @@ class Contractor
[&](const ThreadDataContainer::EnumerableThreadData::range_type &range)
{
for (auto &data : range)
std::sort(data->inserted_edges.begin(), data->inserted_edges.end());
tbb::parallel_sort(data->inserted_edges);
});
tbb::parallel_for(
tbb::blocked_range<int>(first_independent_node, last, DeleteGrainSize),
@@ -550,7 +555,7 @@ class Contractor
if (remaining_nodes.size() > 2)
{
// TODO: for small cores a sorted array of core ids might also work good
for (const auto& node : remaining_nodes)
for (const auto &node : remaining_nodes)
{
auto orig_id = orig_node_id_from_new_node_id_map[node.id];
is_core_node[orig_id] = true;
@@ -563,7 +568,8 @@ class Contractor
is_core_node.clear();
}
SimpleLogger().Write() << "[core] " << remaining_nodes.size() << " nodes " << contractor_graph->GetNumberOfEdges() << " edges." << std::endl;
SimpleLogger().Write() << "[core] " << remaining_nodes.size() << " nodes "
<< contractor_graph->GetNumberOfEdges() << " edges." << std::endl;
thread_data_list.data.clear();
}
@@ -861,7 +867,7 @@ class Contractor
}
}
// eliminate duplicate entries ( forward + backward edges )
std::sort(neighbours.begin(), neighbours.end());
tbb::parallel_sort(neighbours);
neighbours.resize(std::unique(neighbours.begin(), neighbours.end()) - neighbours.begin());
for (const auto i : osrm::irange<std::size_t>(0, neighbours.size()))
@@ -890,7 +896,7 @@ class Contractor
node_data[u].depth = (std::max)(node_data[node].depth + 1, node_data[u].depth);
}
// eliminate duplicate entries ( forward + backward edges )
std::sort(neighbours.begin(), neighbours.end());
tbb::parallel_sort(neighbours);
neighbours.resize(std::unique(neighbours.begin(), neighbours.end()) - neighbours.begin());
// re-evaluate priorities of neighboring nodes
@@ -933,7 +939,7 @@ class Contractor
neighbours.push_back(target);
}
std::sort(neighbours.begin(), neighbours.end());
tbb::parallel_sort(neighbours);
neighbours.resize(std::unique(neighbours.begin(), neighbours.end()) - neighbours.begin());
// examine all neighbours that are at most 2 hops away
contractor/processing_chain.cpp
+1 -1
View File
@@ -183,7 +183,7 @@ void Prepare::FindComponents(unsigned max_edge_id,
}
}
tbb::parallel_sort(edges.begin(), edges.end());
tbb::parallel_sort(edges);
auto new_end = std::unique(edges.begin(), edges.end());
edges.resize(new_end - edges.begin());