Skip to content

Improved worksharing across threads for find_overlaps_groups. #65

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
jkanche merged 2 commits into from
Sep 16, 2025
Merged

Improved worksharing across threads for find_overlaps_groups. #65

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f8e7a11
Improved worksharing across threads for find_overlaps_groups.
LTLA Sep 12, 2025
e580aab
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 12, 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
110 changes: 69 additions & 41 deletions lib/src/nclsoverlaps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,38 +183,72 @@ pybind11::tuple perform_find_overlaps_groups(
query_group_info[i] = {static_cast<const Index*>(q_req.ptr), static_cast<std::size_t>(q_req.shape[0])};
}

// Building the indices in parallel with a simple worker pool.
std::vector<nclist::Nclist<Index, Position> > built(n_groups);
{
std::vector<std::thread> workers;
workers.reserve(num_threads);
std::mutex mut;
std::size_t group = 0;

for (int i = 0; i < num_threads; ++i) {
workers.emplace_back([&]() -> void {
while (1) {
std::unique_lock lck(mut);
if (group == n_groups) {
break;
}
const auto curgroup = group;
++group;
lck.unlock();

const auto& s_info = self_group_info[curgroup];
if (s_info.size != 0) {
built[curgroup] = nclist::build<Index, Position>(s_info.size, s_info.ptr, s_starts_ptr, s_ends_ptr);
}
}
});
}

for (auto& worker : workers) {
worker.join();
}
}

// Now running through all groups to find overlaps in parallel.
std::vector<std::vector<std::vector<Index> > > all_group_results(n_groups);
std::vector<std::thread> workers;
workers.reserve(num_threads);
std::vector<std::size_t> total_hits_per_thread(num_threads);

int num_jobs = n_groups / num_threads;
int num_remaining = n_groups % num_threads;
int jobs_so_far = 0;

for (int i = 0; i < num_threads; ++i) {
int current_jobs = num_jobs + (i < num_remaining);
if (current_jobs == 0) {
break;
for (std::size_t g = 0; g < n_groups; ++g) {
const auto& s_info = self_group_info[g];
const auto& q_info = query_group_info[g];
if (s_info.size == 0 || q_info.size == 0) {
continue;
}

workers.emplace_back([&](int first, int length) -> void {
std::vector<Index> single_query_matches;

for (Index j = first, last = first + length; j < last; ++j) {
const std::size_t num_jobs = q_info.size / num_threads;
const std::size_t num_remaining = q_info.size % num_threads;
std::size_t jobs_so_far = 0;

const auto& s_info = self_group_info[j];
const auto& q_info = query_group_info[j];
const auto& nclist_obj = built[g];
auto& current_group_results = all_group_results[g];
current_group_results.resize(q_info.size);

if (s_info.size == 0 || q_info.size == 0) {
continue;
}
workers.clear();
for (int i = 0; i < num_threads; ++i) {
const std::size_t current_jobs = num_jobs + (i < num_remaining);
if (current_jobs == 0) {
break;
}

auto nclist_obj = nclist::build<Index, Position>(s_info.size, s_info.ptr, s_starts_ptr, s_ends_ptr);
std::vector<std::vector<Index> > local_group_results;
local_group_results.resize(q_info.size);
workers.emplace_back([&](int thread, std::size_t first, std::size_t length) -> void {
std::size_t current_total_hits = 0;

for (std::size_t k = 0; k < q_info.size; ++k) {
Index original_query_idx = q_info.ptr[k];
for (std::size_t k = first, last = first + length; k < last; ++k) {
const Index original_query_idx = q_info.ptr[k];
auto& single_query_matches = current_group_results[k];

if (query_type == "any") {
nclist::OverlapsAnyWorkspace<Index> ws_any;
Expand Down Expand Up @@ -248,35 +282,29 @@ pybind11::tuple perform_find_overlaps_groups(
nclist::overlaps_within(nclist_obj, q_starts_ptr[original_query_idx], q_ends_ptr[original_query_idx], params, ws_within, single_query_matches);
}

if (!single_query_matches.empty()) {
if (!single_query_matches.empty()) { // WTF is this
if (select == "last" && !quit_on_first) {
local_group_results[k] = {single_query_matches.back()};
} else {
local_group_results[k] = single_query_matches;
}
single_query_matches.front() = single_query_matches.back();
single_query_matches.resize(1);
}
}
}

if (!local_group_results.empty()){
all_group_results[j] = std::move(local_group_results);
current_total_hits += single_query_matches.size();
}
}
}, jobs_so_far, current_jobs);

jobs_so_far += current_jobs;
}
// Don't add directly to this value in the inner loop, to reduce the risk of false sharing.
total_hits_per_thread[thread] += current_total_hits;
}, i, jobs_so_far, current_jobs);

for (auto& worker : workers) {
worker.join();
}
jobs_so_far += current_jobs;
}

std::size_t total_hits = 0;
for(const auto& group_res : all_group_results) {
for (const auto& query_res : group_res) {
total_hits += query_res.size();
for (auto& worker : workers) {
worker.join();
}
}

const std::size_t total_hits = std::accumulate(total_hits_per_thread.begin(), total_hits_per_thread.end(), static_cast<std::size_t>(0));
py::array_t<Index> query_hits(total_hits);
py::array_t<Index> self_hits(total_hits);
auto q_res_ptr = static_cast<Index*>(query_hits.request().ptr);
Expand Down