Skip to content

Commit

Permalink
small optimization to parallel component paths
Browse files Browse the repository at this point in the history
  • Loading branch information
@jeizenga
jeizenga committed Nov 8, 2021
1 parent 2a569b8 commit d141f31
Showing 1 changed file with 18 additions and 18 deletions.
36 changes: 18 additions & 18 deletions src/algorithms/component_paths.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -207,9 +207,9 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
// keep the memory use down
// note: this has to be done on the heap because the deleted copy assignment
// and copy construction for atomic ints means vectors can never be resized
vector<atomic<int8_t>>* id_vec_8 = new vector<atomic<int8_t>>(id_range);
vector<atomic<int16_t>>* id_vec_16 = nullptr;
vector<atomic<int32_t>>* id_vec_32 = nullptr;
vector<atomic<uint8_t>>* id_vec_8 = new vector<atomic<uint8_t>>(id_range);
vector<atomic<uint16_t>>* id_vec_16 = nullptr;
vector<atomic<uint32_t>>* id_vec_32 = nullptr;
// in parallel initialize with sentinels, which will be replaced by search IDs
static const size_t block_size = 4096;
atomic<size_t> block_idx(0);
Expand All @@ -222,7 +222,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
break;
}
for (size_t j = begin, end = min<size_t>(begin + block_size, id_vec_8->size()); j < end; ++j) {
(*id_vec_8)[j].store(-1);
(*id_vec_8)[j].store(0);
}
}
});
Expand All @@ -238,10 +238,10 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
// the values 0, 1, or 2
uint8_t which_vec = 0;
// the last search ID we can accommodate for each vector
const int32_t max_search_id[3] = {
numeric_limits<int8_t>::max(),
numeric_limits<int16_t>::max(),
numeric_limits<int32_t>::max()
const uint32_t max_search_id[3] = {
numeric_limits<uint8_t>::max(),
numeric_limits<uint16_t>::max(),
numeric_limits<uint32_t>::max()
};


Expand All @@ -254,7 +254,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr

// perform atomic load on whichever vector we're currently using
auto load = [&](int64_t i) {
int32_t loaded;
uint32_t loaded;
switch (which_vec) {
case 0:
loaded = (*id_vec_8)[i].load();
Expand All @@ -269,13 +269,13 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
return loaded;
};
// perform atomic compare-exchange on whichever vector we're currently using
auto compare_exchange = [&](int64_t i, int32_t& expected, int32_t desired) {
auto compare_exchange = [&](int64_t i, uint32_t& expected, uint32_t desired) {
bool exchanged;
switch (which_vec) {
case 0:
{
int8_t expected_8 = expected;
int8_t desired_8 = desired;
uint8_t expected_8 = expected;
uint8_t desired_8 = desired;
exchanged = (*id_vec_8)[i].compare_exchange_strong(expected_8, desired_8);
if (!exchanged) {
expected = expected_8;
Expand All @@ -284,8 +284,8 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
}
case 1:
{
int16_t expected_16 = expected;
int16_t desired_16 = desired;
uint16_t expected_16 = expected;
uint16_t desired_16 = desired;
exchanged = (*id_vec_16)[i].compare_exchange_strong(expected_16, desired_16);
if (!exchanged) {
expected = expected_16;
Expand All @@ -305,7 +305,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
// to keep track of the index of the path we will use to seed a BFS next
atomic<int64_t> next_path(0);
// to keep track of the ID of the next seeded search
atomic<int32_t> next_search_id(0);
atomic<uint32_t> next_search_id(1);
// initialize the swarm of workers
vector<thread> workers;
for (int i = 0; i < thread_count; ++i) {
Expand Down Expand Up @@ -333,7 +333,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
// seed a BFS search off of the first node in this path
handle_t seed = graph.get_handle_of_step(graph.path_begin(path));

if (load(graph.get_id(seed) - min_id) != -1) {
if (load(graph.get_id(seed) - min_id) != 0) {
// another thread has already traversed over this node, no need
// to start a search here
#ifdef debug_parallel_component_paths
Expand All @@ -346,7 +346,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr


// we're going to initiate a BFS from the seed, assign a new search ID
int32_t search_id = next_search_id++;
uint32_t search_id = next_search_id++;

// TODO: add finer-grain reallocations so that neighbors and
// search_path_sets don't need to get so large to guarantee that
Expand Down Expand Up @@ -388,7 +388,7 @@ vector<unordered_set<path_handle_t>> component_paths_parallel(const PathHandleGr
// function to call on each subsequent handle we navigate to
function<bool(const handle_t&)> record_paths_and_enqueue = [&](const handle_t& here) {
int64_t idx = graph.get_id(here) - min_id;
int32_t visit_id = -1;
uint32_t visit_id = 0;
bool exchanged = compare_exchange(idx, visit_id, search_id);
if (exchanged) {
// we found the unvisited sentinel and replaced it with our search ID
Expand Down

1 comment on commit d141f31

@adamnovak
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

vg CI tests complete for branch parallel-components-paths. View the full report here.

16 tests passed, 0 tests failed and 0 tests skipped in 11743 seconds

Please sign in to comment.