Skip to content

Performance Optimizations #28

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

Open
John-194 wants to merge 10 commits into
base: master
Choose a base branch
from
Open

Performance Optimizations #28

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
39805c2
Replace timing-based granularity with simple heuristic in parfor
John-194 Apr 3, 2026
8da5e33
Fix idle CPU usage using progressive backoff and CV based sleep in wo…
John-194 Apr 3, 2026
96c9cff
Pre-compute cell keys for grid sort avoiding floor()
John-194 Apr 3, 2026
4a4025f
Defer grid cell/mutex init to used cells only, reduce hash table size
John-194 Apr 3, 2026
607417e
O(n) cluster renumbering via prefix sum, replace O(n log n) sort + ha…
John-194 Apr 3, 2026
847e798
Eliminate sqrt from distance comparisons, use ternary for branchless …
John-194 Apr 3, 2026
909a621
Dynamic thread count via getWorkers(), fix hardcoded P=36*8
John-194 Apr 3, 2026
e83de83
Fix undefined behaviour in CAS operations
John-194 Mar 28, 2026
f0fb24f
Fix DCLP race on nbrCache
John-194 Mar 28, 2026
2a83e39
Add static_assert for deque age_t size
John-194 Mar 28, 2026
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
75 changes: 19 additions & 56 deletions include/dbscan/algo.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ int DBSCAN(intT n, floatT* PF, double epsilon, intT minPts, bool* coreFlagOut, i
intT count = 0;
auto isCore = [&] (pointT *p) {
if(count >= minPts) return true;
if(p->distSqr(P[i]) <= epsSqr) {//todo sqrt opt
if(p->distSqr(P[i]) <= epsSqr) {
count ++;}
return false;};
G->nghPointMap(P[i].coordinate(), isCore);
Expand Down Expand Up @@ -165,67 +165,30 @@ int DBSCAN(intT n, floatT* PF, double epsilon, intT minPts, bool* coreFlagOut, i
delete G;

//improving cluster representation
auto cluster2 = newA(intT, n);
auto flag = newA(intT, n+1);
parallel_for(0, n, [&](intT i){cluster2[i] = cluster[i];});
sampleSort(cluster, n, std::less<intT>());

flag[0] = 1;
parallel_for(1, n, [&](intT i){
if (cluster[i] != cluster[i-1])
flag[i] = 1;
else
flag[i] = 0;
});
flag[n] = sequence::prefixSum(flag, 0, n);

// typedef pair<intT,intT> eType;
struct myPair {
intT first;
intT second;
myPair(intT _first, intT _second): first(_first), second(_second) {}
myPair(): first(-1), second(-1) {}
inline bool operator==(myPair a) {
if(a.first==first && a.second== second)
return true;
else
return false;
}
};

typedef Table<hashSimplePair<myPair>,intT> tableT;
auto T = new tableT(n, hashSimplePair<myPair>());
parallel_for(0, n, [&] (intT i) {
if (flag[i] != flag[i+1]) {
// T->insert(make_pair(cluster[i], flag[i]));
T->insert(myPair(cluster[i], flag[i]));
}
});
// O(n) renumbering: cluster IDs are point indices ∈ [0,n), use prefix sum
// instead of O(n log n) sort + hash table.
auto idMap = newA(intT, n);
parallel_for(0, n, [&](intT i) { idMap[i] = 0; });
parallel_for(0, n, [&](intT i) {
if (cluster[i] >= 0) idMap[cluster[i]] = 1;
});
sequence::prefixSum(idMap, 0, n); // exclusive: idMap[cid] = sequential ID for cid

if(T->find(-1).second < 0) {
parallel_for(0, n, [&](intT i){
cluster2[i] = T->find(cluster2[i]).second;
});
} else {
parallel_for(0, n, [&](intT i){
if (cluster2[i] > 0)
cluster2[i] = T->find(cluster2[i]).second-1;
});
}
// Remap to sequential IDs (separate buffer to avoid read/write conflict)
auto cluster2 = newA(intT, n);
parallel_for(0, n, [&](intT i) {
cluster2[i] = (cluster[i] >= 0) ? idMap[cluster[i]] : cluster[i];
});

//restoring order
parallel_for(0, n, [&](intT i){
cluster[I[i]] = cluster2[i];
});
parallel_for(0, n, [&](intT i){
coreFlagOut[I[i]] = coreFlag[i];
});
parallel_for(0, n, [&](intT i) {
cluster[I[i]] = cluster2[i];
coreFlagOut[I[i]] = coreFlag[i];
});

free(I);
free(cluster2);
free(flag);
T->del(); // Required to clean-up T's internals
delete T;
free(idMap);
free(P);
#ifdef VERBOSE
cout << "output-time = " << tt.stop() << endl;
Expand Down
4 changes: 4 additions & 0 deletions include/dbscan/cell.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,7 +172,11 @@ struct cellHash {
bool replaceQ(eType c1, eType c2) {return 0;}

bool cas(eType* p, eType o, eType n) {
#ifdef _MSC_VER
return std::atomic_compare_exchange_strong_explicit(
reinterpret_cast<std::atomic<eType>*>(p), &o, n, std::memory_order_acq_rel, std::memory_order_acquire);
#else
return __atomic_compare_exchange(p, &o, &n, false, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE);
#endif
}
};
32 changes: 17 additions & 15 deletions include/dbscan/coreBccp.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,17 +29,18 @@
#include "pbbs/parallel.h"
#include "pbbs/utils.h"

// r holds squared distance; using distSqr and nodeDistanceSqr avoids sqrt in hot path
template<class nodeT, class objT>
inline void compBcpCoreHSerial(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, objT* P) {
if (n1->nodeDistance(n2) > *r) return;
if (n1->nodeDistanceSqr(n2) > *r) return;

if (n1->isLeaf() && n2->isLeaf()) {//basecase
for (intT i=0; i<n1->size(); ++i) {
for (intT j=0; j<n2->size(); ++j) {
auto pi = n1->getItem(i);
auto pj = n2->getItem(j);
if (coreFlag[pi - P] && coreFlag[pj - P]) {
floatT dist = pi->dist(*pj);
floatT dist = pi->distSqr(*pj);
r[0] = min(r[0], dist);
}
}
Expand Down Expand Up @@ -78,30 +79,31 @@ inline void compBcpCoreHSerial(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag,

template<class nodeT, class objT>
inline void compBcpCoreHBase(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, objT* P) {
if (n1->nodeDistance(n2) > *r) return;
if (n1->nodeDistanceSqr(n2) > *r) return;

if (n1->isLeaf() && n2->isLeaf()) {//basecase
for (intT i=0; i<n1->size(); ++i) {
for (intT j=0; j<n2->size(); ++j) {
auto pi = n1->getItem(i);
auto pj = n2->getItem(j);
if (coreFlag[pi - P] && coreFlag[pj - P]) {
floatT dist = pi->dist(*pj);
floatT dist = pi->distSqr(*pj);
utils::writeMin(r, dist);
}
}
}
} else {//recursive, todo consider call order, might help
} else {//recursive
if (n1->isLeaf()) {
if (n1->nodeDistance(n2->L()) < n1->nodeDistance(n2->R())) {
// nodeDistanceSqr avoids sqrt; monotonicity preserves ordering
if (n1->nodeDistanceSqr(n2->L()) < n1->nodeDistanceSqr(n2->R())) {
compBcpCoreH(n1, n2->L(), r, coreFlag, P);
compBcpCoreH(n1, n2->R(), r, coreFlag, P);
} else {
compBcpCoreH(n1, n2->R(), r, coreFlag, P);
compBcpCoreH(n1, n2->L(), r, coreFlag, P);
}
} else if (n2->isLeaf()) {
if (n2->nodeDistance(n1->L()) < n2->nodeDistance(n1->R())) {
if (n2->nodeDistanceSqr(n1->L()) < n2->nodeDistanceSqr(n1->R())) {
compBcpCoreH(n2, n1->L(), r, coreFlag, P);
compBcpCoreH(n2, n1->R(), r, coreFlag, P);
} else {
Expand All @@ -115,7 +117,7 @@ inline void compBcpCoreHBase(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, ob
ordering[2] = make_pair(n2->L(), n1->R());
ordering[3] = make_pair(n2->R(), n1->R());
auto bbd = [&](pair<nodeT*,nodeT*> p1, pair<nodeT*,nodeT*> p2) {
return p1.first->nodeDistance(p1.second) < p2.first->nodeDistance(p2.second);};
return p1.first->nodeDistanceSqr(p1.second) < p2.first->nodeDistanceSqr(p2.second);};
quickSortSerial(ordering, 4, bbd);
for (intT o=0; o<4; ++o) {
compBcpCoreH(ordering[o].first, ordering[o].second, r, coreFlag, P);}
Expand All @@ -125,21 +127,21 @@ inline void compBcpCoreHBase(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, ob

template<class nodeT, class objT>
inline void compBcpCoreH(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, objT* P) {
if (n1->nodeDistance(n2) > *r) return;
if (n1->nodeDistanceSqr(n2) > *r) return;

if ((n1->isLeaf() && n2->isLeaf()) || (n1->size()+n2->size() < 2000)) {
return compBcpCoreHBase(n1, n2, r, coreFlag, P);
} else {//recursive, todo consider call order, might help
} else {//recursive
if (n1->isLeaf()) {
if (n1->nodeDistance(n2->L()) < n1->nodeDistance(n2->R())) {
if (n1->nodeDistanceSqr(n2->L()) < n1->nodeDistanceSqr(n2->R())) {
par_do([&](){compBcpCoreH(n1, n2->L(), r, coreFlag, P);},
[&](){compBcpCoreH(n1, n2->R(), r, coreFlag, P);});
} else {
par_do([&](){compBcpCoreH(n1, n2->R(), r, coreFlag, P);},
[&](){compBcpCoreH(n1, n2->L(), r, coreFlag, P);});
}
} else if (n2->isLeaf()) {
if (n2->nodeDistance(n1->L()) < n2->nodeDistance(n1->R())) {
if (n2->nodeDistanceSqr(n1->L()) < n2->nodeDistanceSqr(n1->R())) {
par_do([&](){compBcpCoreH(n2, n1->L(), r, coreFlag, P);},
[&](){compBcpCoreH(n2, n1->R(), r, coreFlag, P);});
} else {
Expand All @@ -153,7 +155,7 @@ inline void compBcpCoreH(nodeT* n1, nodeT* n2, floatT* r, intT* coreFlag, objT*
ordering[2] = make_pair(n2->L(), n1->R());
ordering[3] = make_pair(n2->R(), n1->R());
auto bbd = [&](pair<nodeT*,nodeT*> p1, pair<nodeT*,nodeT*> p2) {
return p1.first->nodeDistance(p1.second) < p2.first->nodeDistance(p2.second);};
return p1.first->nodeDistanceSqr(p1.second) < p2.first->nodeDistanceSqr(p2.second);};
quickSortSerial(ordering, 4, bbd);
parallel_for (0, 4, [&](intT o) {
compBcpCoreH(ordering[o].first, ordering[o].second, r, coreFlag, P);}, 1);
Expand All @@ -179,11 +181,11 @@ inline bool hasEdge(intT n1, intT n2, intT* coreFlag, objT* P, floatT epsilon, c

if (!trees[n1])
trees[n1] = new treeT(cells[n1].getItem(), cells[n1].size(), false);//todo allocation, parallel
if (!trees[n2])
if (!trees[n2])
trees[n2] = new treeT(cells[n2].getItem(), cells[n2].size(), false);//todo allocation, parallel
floatT r = floatMax();
compBcpCoreH(trees[n1]->rootNode(), trees[n2]->rootNode(), &r, coreFlag, P);
return r <= epsilon;
return r <= epsilon * epsilon; // r holds squared distance now
}

#endif
90 changes: 62 additions & 28 deletions include/dbscan/grid.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@

#pragma once

#include <atomic>
#include <mutex>
#include "cell.h"
#include "point.h"
Expand Down Expand Up @@ -77,7 +78,7 @@ struct grid {
tableT* table=NULL;
treeT* tree=NULL;
intT totalPoints;
cellBuf **nbrCache;
std::atomic<cellBuf*>* nbrCache;
std::mutex* cacheLocks;

/**
Expand All @@ -90,26 +91,29 @@ struct grid {
r(rr), pMin(pMinn), cellCapacity(cellMax), totalPoints(0) {

cells = newA(cellT, cellCapacity);
nbrCache = newA(cellBuf*, cellCapacity);
nbrCache = new std::atomic<cellBuf*>[cellCapacity];
cacheLocks = (std::mutex*) malloc(cellCapacity * sizeof(std::mutex));
parallel_for(0, cellCapacity, [&](intT i) {
new (&cacheLocks[i]) std::mutex();
nbrCache[i] = NULL;
cells[i].init();
});
// cacheLocks/nbrCache/cells are initialized lazily in insertParallel over
// numCells only, so nothing to initialize here.
numCells = 0;

myHash = new cellHashT(pMinn, r);
table = new tableT(cellMax*2, cellHash<dim, objT>(myHash));//todo load
// Hash table sized for expected number of cells, with safety rebuild
// in insertParallel if numCells exceeds the estimate.
intT tableHint = std::max((intT)2048, cellMax / 4);
table = new tableT(tableHint, cellHash<dim, objT>(myHash));
}

~grid() {
free(cells);
free(cacheLocks);
parallel_for(0, cellCapacity, [&](intT i) {
if(nbrCache[i]) delete nbrCache[i];
// Only [0, numCells) were initialized in insertParallel; reading beyond
// that would touch uninitialized atomics.
parallel_for(0, numCells, [&](intT i) {
auto cached = nbrCache[i].load(std::memory_order_relaxed);
if(cached) delete cached;
});
free(nbrCache);
delete[] nbrCache;
if(myHash) delete myHash;
if(table) {
table->del();
Expand Down Expand Up @@ -147,22 +151,24 @@ struct grid {
}
return false;};//todo, optimize
int idx = bait - cells;
if (nbrCache[idx]) {
auto accum = nbrCache[idx];
for (auto accum_i : *accum) {
// Acquire ensures vector contents are visible if pointer is non-null
auto cached = nbrCache[idx].load(std::memory_order_acquire);
if (cached) {
for (auto accum_i : *cached) {
if(fWrap(accum_i)) break;
}
} else {
// wait for other threads to do their thing then try again
std::lock_guard<std::mutex> lock(cacheLocks[idx]);
if (nbrCache[idx]) {
auto accum = nbrCache[idx];
for (auto accum_i : *accum) {
cached = nbrCache[idx].load(std::memory_order_relaxed);
if (cached) {
for (auto accum_i : *cached) {
if (fWrap(accum_i)) break;
}
} else {
floatT hop = sqrt(dim + 3) * 1.0000001;
nbrCache[idx] = tree->rangeNeighbor(bait, r * hop, fStop, fWrap, true, nbrCache[idx]);
auto result = tree->rangeNeighbor(bait, r * hop, fStop, fWrap, true, (cellBuf*)nullptr);
// Release ensures vector contents are fully written before pointer is visible
nbrCache[idx].store(result, std::memory_order_release);
}
}
}
Expand All @@ -176,22 +182,22 @@ struct grid {
return false;
};
int idx = bait - cells;
if (nbrCache[idx]) {
auto accum = nbrCache[idx];
for (auto accum_i : *accum) {
auto cached = nbrCache[idx].load(std::memory_order_acquire);
if (cached) {
for (auto accum_i : *cached) {
if (fWrap(accum_i)) break;
}
} else {
// wait for other threads to do their thing then try again
std::lock_guard<std::mutex> lock(cacheLocks[idx]);
if (nbrCache[idx]) {
auto accum = nbrCache[idx];
for (auto accum_i : *accum) {
cached = nbrCache[idx].load(std::memory_order_relaxed);
if (cached) {
for (auto accum_i : *cached) {
if (fWrap(accum_i)) break;
}
} else {
floatT hop = sqrt(dim + 3) * 1.0000001;
nbrCache[bait-cells] = tree->rangeNeighbor(bait, r * hop, fStop, fWrap, true, nbrCache[idx]);
auto result = tree->rangeNeighbor(bait, r * hop, fStop, fWrap, true, (cellBuf*)nullptr);
nbrCache[idx].store(result, std::memory_order_release);
}
}
}
Expand All @@ -205,9 +211,24 @@ struct grid {
freeFlag=true;}

parallel_for(0, nn, [&](intT i){I[i] = i;});

// Pre-compute integer cell coordinates to avoid floor() in every sort comparison
auto cellKeys = newA(intT, nn * dim);
floatT invR = 1.0 / r;
parallel_for(0, nn, [&](intT i) {
for (int d = 0; d < dim; d++) {
cellKeys[i * dim + d] = (intT)floor((P[i][d] - pMin[d]) * invR);
}
});
auto ipLess = [&] (intT a, intT b) {
return pointGridCmp<dim, objT, geoPointT>(P[a], P[b], pMin, r);};
for (int d = 0; d < dim; d++) {
intT ca = cellKeys[a * dim + d];
intT cb = cellKeys[b * dim + d];
if (ca != cb) return ca < cb;
}
return false;};
sampleSort(I, nn, ipLess);
free(cellKeys);
parallel_for(0, nn, [&](intT i){PP[i] = P[I[i]];});

flag[0] = 1;
Expand All @@ -224,6 +245,19 @@ struct grid {
if (numCells > cellCapacity) {
cout << "error, grid insert exceeded cell capacity, abort()" << endl;abort();}

// Rebuild hash table if initial estimate was too small
if (numCells > cellCapacity / 8) {
table->del(); delete table;
table = new tableT(numCells * 2, cellHash<dim, objT>(myHash));
}

// Initialize only the cells that will actually be used
parallel_for(0, numCells, [&](intT i) {
new (&cacheLocks[i]) std::mutex();
nbrCache[i].store(nullptr, std::memory_order_relaxed);
cells[i].init();
});

parallel_for(0, nn, [&](intT i) {
if (flag[i] != flag[i+1]) {
auto c = &cells[flag[i]];
Expand Down
Loading
Loading