i/rtabmap
1
0
Fork 0
mirror of https://github.com/introlab/rtabmap.git synced 2026-10-07 02:27:47 +08:00
Code Issues Packages Projects Releases Wiki Activity

VWDictionary: use multi-core FLANN kNN search (#1760)

Browse Source 
* VWDictionary: use multi-core FLANN kNN search

* Add parameter with default 1 thread

* Kp/FlannTreads plumbing  to UI. Also added to performance tests for comparison.

* fixing ci error

* dump debug data for windows ci

* Adding  more dll debugging report windows ci

* install vc2012 runtime explicitly

* updated comment

---------

Co-authored-by: matlabbe <[email protected]>
This commit is contained in:
Torjus Iveland
2026-09-10 23:22:00 -07:00
committed by GitHub
co-authored by matlabbe
parent fb457255b7
commit 2fbbe19d70
10 changed files with 283 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
corelib/test/perf_flann_index.cpp
+41 -6
View File
@@ -11,6 +11,10 @@
// version can be compared to what it replaces.
#include "FlannIndexBackends.h"
#ifdef _OPENMP
#include <omp.h>
#endif
// The times are reported rather than asserted on: which backend is the fastest
// depends on the machine. They are here so that a change of backend, of
// parameters or of nanoflann version can be compared to what it replaces.
@@ -517,16 +521,31 @@ TEST(FlannIndexPerfTest, RegistrationGuessMatching)
// A factor of 1 for the rtflann rows keeps their per-point bookkeeping out
// of the measurement, and picks the nanoflann tree that is built once.
const Backend backends[] = {
{"cv BFMatcher ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true},
std::vector<Backend> backends = {
{"cv BFMatcher ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true, 1},
{"cv BFMatcher threaded ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true, 0},
{"rtflann kd-tree (4 randomized) ", FlannIndex::FLANN_INDEX_KDTREE, 1.0f},
{"rtflann kd-tree single ", FlannIndex::FLANN_INDEX_KDTREE_SINGLE, 1.0f},
{"nanoflann kd-tree single ", FlannIndex::NANOFLANN_INDEX_KDTREE_SINGLE, 1.0f},
{"nanoflann kd-tree single incremental", FlannIndex::NANOFLANN_INDEX_KDTREE_SINGLE, 2.0f},
};
#ifdef _OPENMP
// rtflann threads a radius search over its batch of queries the same way it
// threads a kNN one, so the two trees come back with one thread per core.
// The tree is still built on one core, and here it is rebuilt every frame,
// which caps what threading can take off the total: the times say how much
// of a frame is the search rather than the build.
backends.push_back({"rtflann kd-tree (4 rand.) threaded", FlannIndex::FLANN_INDEX_KDTREE, 1.0f, false, 0});
backends.push_back({"rtflann kd-tree single threaded ", FlannIndex::FLANN_INDEX_KDTREE_SINGLE, 1.0f, false, 0});
#endif
std::cout << "[ ] " << keypoints << " keypoints indexed and as many looked up in a "
<< radius << " px radius, per frame" << std::endl;
#ifdef _OPENMP
std::cout << "[ ] the threaded rows search with " << omp_get_max_threads()
<< " threads, the others with one" << std::endl;
#endif
for(const Backend & backend: backends)
{
@@ -538,7 +557,7 @@ TEST(FlannIndexPerfTest, RegistrationGuessMatching)
{
FlannIndex index;
index.buildIndex(backend.algorithm, points, false, backend.rebalancingFactor);
index.radiusSearch(projected, indices, dists, radius, 0, 32, 0.0f, false);
index.radiusSearch(projected, indices, dists, radius, 0, 32, 0.0f, false, backend.cores);
}
const double perFrame = timer.ticks()/double(frames);
@@ -571,15 +590,31 @@ void compareDictionaryMatching(int indexedCount, int queriedCount)
// built once, so it is neither kept ready to be added to nor rebuilt. The
// incremental nanoflann tree is kept in the comparison to show what asking
// for one costs here.
const Backend backends[] = {
std::vector<Backend> backends = {
{"linear exhaustive ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f},
{"cv BFMatcher ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true},
{"cv BFMatcher ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true, 1},
{"cv BFMatcher threaded ", FlannIndex::FLANN_INDEX_LINEAR, 1.0f, true, 0},
{"rtflann kd-tree (4 randomized) ", FlannIndex::FLANN_INDEX_KDTREE, 1.0f},
{"rtflann kd-tree single ", FlannIndex::FLANN_INDEX_KDTREE_SINGLE, 1.0f},
{"nanoflann kd-tree single ", FlannIndex::NANOFLANN_INDEX_KDTREE_SINGLE, 1.0f},
{"nanoflann kd-tree single incremental", FlannIndex::NANOFLANN_INDEX_KDTREE_SINGLE, 2.0f},
};
#ifdef _OPENMP
// The same two rtflann trees searched with Kp/FlannThreads=0, one thread per
// core: rtflann is the only backend here threading a batch of queries, over
// an OpenMP loop. Only the search half of the times can improve, the trees
// are still built on one core. Queries are independent of each other, so
// threading doesn't change what is found: the exact tree holds its recall to
// the digit. The randomized one moves by a tenth of a percent from one run to
// the next whether threaded or not, it randomizes its splits on every build.
backends.push_back({"rtflann kd-tree (4 rand.) threaded", FlannIndex::FLANN_INDEX_KDTREE, 1.0f, false, 0});
backends.push_back({"rtflann kd-tree single threaded ", FlannIndex::FLANN_INDEX_KDTREE_SINGLE, 1.0f, false, 0});
std::cout << "[ ] the threaded rows search with " << omp_get_max_threads()
<< " threads, the others with one" << std::endl;
#endif
for(int dim: {32, 64, 128, 256})
{
const cv::Mat from = makeDescriptors(indexedCount, dim, clusterCount(indexedCount), 150);
@@ -599,7 +634,7 @@ void compareDictionaryMatching(int indexedCount, int queriedCount)
{
FlannIndex index;
index.buildIndex(backend.algorithm, from, false, backend.rebalancingFactor);
index.knnSearch(to, indices, dists, KNN);
index.knnSearch(to, indices, dists, KNN, 32, 0.0f, true, backend.cores);
}
const double perFrame = timer.ticks()/double(frames);