Merge branch 'master' of gitlab.vision.in.tum.de:basalt/basalt

CMakeLists.txt

@@ -129,7 +129,7 @@ endif()
 # Set platform / compiler specific compile flags and checks
 if(APPLE)
   # Need to investigate how to reliably detect and use OpenMP on macOS...
-  set(USE_OPENMP_DEFAULT OFF)
+#  set(USE_OPENMP_DEFAULT OFF)
 
   # Among others, setting CMAKE_FIND_FRAMEWORK to LAST fixed issues
   # with installed Mono that contains old headers (libpng, ...).
@@ -163,7 +163,7 @@ if(APPLE)
   endif()
 
 elseif(UNIX)
-  set(USE_OPENMP_DEFAULT ON)
+#  set(USE_OPENMP_DEFAULT ON)
 
   # assume libstdc++
   set(STD_CXX_FS stdc++fs)
@@ -186,13 +186,37 @@ endif()
 
 
 # OpenMP option and compile flags
-option(USE_OPENMP "Use OpenMP (e.g. for parallel computation in Eigen)" ${USE_OPENMP_DEFAULT})
+#
-if(USE_OPENMP)
+# Note: OpenMP and TBB don't mix well, so we disable Eigen's parallelization.
-  message(STATUS "OpenMP Enabled")
+# It's trying to paralellize matrix products during SC, which we run in a parallel_reduce using TBB.
-  set(BASALT_CXX_FLAGS "${BASALT_CXX_FLAGS} -fopenmp")
+# Turns out using OpenMP can slow down the computby factor 10-100x! So for now we discable it completely.
-else()
+# One place where Eigen's parallelization could still have been useful is the CG solver in the mapper.
-  message(STATUS "OpenMP Disabled")
+# We could in the future investiagte other implementations (paralellized with TBB) or selectively enabling
-endif()
+# Eigen's parallelization just for CG, setting number of threads to 1 everywhere else.
+# Another way to ensure Eigen doesn't use OpenMP regardless of how it was built is setting the environment
+# variable OMP_NUM_THREADS=1 beofre running the application.
+#
+# See: https://eigen.tuxfamily.org/dox/TopicMultiThreading.html
+#
+# If we enable BLAS / LAPACK either directly or via thirdparty libs like ceres,
+# make sure to disable OpenMP for the linked BLAS library. In particular on Ubuntu it seems OpenBLAS is often installed,
+# and it can have similar issues in multithreaded applications if it's own parallelization with OpenMP is enabled.
+# You can set the environment varaibles OPENBLAS_NUM_THREADS=1 or OMP_NUM_THREADS=1. This is also mentioned in the ceres
+# installation documentation.
+#
+# See also: https://github.com/xianyi/OpenBLAS/wiki/faq#multi-threaded
+#
+# Set EIGEN_DONT_PARALLELIZE to be sure it doesn't use openmp,
+# just in case some dependency enables openmp without us noticing.
+set(BASALT_CXX_FLAGS "${BASALT_CXX_FLAGS} -DEIGEN_DONT_PARALLELIZE")
+
+#option(USE_OPENMP "Use OpenMP (e.g. for parallel computation in Eigen)" ${USE_OPENMP_DEFAULT})
+#if(USE_OPENMP)
+#  message(STATUS "OpenMP Enabled")
+#  set(BASALT_CXX_FLAGS "${BASALT_CXX_FLAGS} -fopenmp")
+#else()
+#  message(STATUS "OpenMP Disabled")
+#endif()
 
 
 # setup combined compiler flags

include/basalt/optimization/accumulator.h

@@ -201,6 +201,10 @@ class SparseHashAccumulator {
     VectorX res;
 
     if (iterative_solver) {
+      // NOTE: since we have to disable Eigen's parallelization with OpenMP
+      // (interferes with TBB), the current CG is single-threaded, and we
+      // can expect a substantial speedup by switching to a parallel
+      // implementation of CG.
       Eigen::ConjugateGradient<Eigen::SparseMatrix<double>,
                                Eigen::Lower | Eigen::Upper>
           cg;

src/rs_t265_vio.cpp

@@ -137,6 +137,12 @@ int main(int argc, char** argv) {
   app.add_option("--num-threads", num_threads, "Number of threads.");
   app.add_option("--step-by-step", step_by_step, "Path to config file.");
 
+  try {
+    app.parse(argc, argv);
+  } catch (const CLI::ParseError& e) {
+    return app.exit(e);
+  }
+
   // global thread limit is in effect until global_control object is destroyed
   std::unique_ptr<tbb::global_control> tbb_global_control;
   if (num_threads > 0) {
@@ -144,12 +150,6 @@ int main(int argc, char** argv) {
         tbb::global_control::max_allowed_parallelism, num_threads);
   }
 
-  try {
-    app.parse(argc, argv);
-  } catch (const CLI::ParseError& e) {
-    return app.exit(e);
-  }
-
   if (!config_path.empty()) {
     vio_config.load(config_path);
   } else {

src/vio.cpp

@@ -222,6 +222,12 @@ int main(int argc, char** argv) {
                  "Save trajectory. Supported formats <tum, euroc, kitti>");
   app.add_option("--use-imu", use_imu, "Use IMU.");
 
+  try {
+    app.parse(argc, argv);
+  } catch (const CLI::ParseError& e) {
+    return app.exit(e);
+  }
+
   // global thread limit is in effect until global_control object is destroyed
   std::unique_ptr<tbb::global_control> tbb_global_control;
   if (num_threads > 0) {
@@ -229,12 +235,6 @@ int main(int argc, char** argv) {
         tbb::global_control::max_allowed_parallelism, num_threads);
   }
 
-  try {
-    app.parse(argc, argv);
-  } catch (const CLI::ParseError& e) {
-    return app.exit(e);
-  }
-
   if (!config_path.empty()) {
     vio_config.load(config_path);
 

thirdparty/basalt-headers

@@ -1 +1 @@
-Subproject commit b44542afd6c1aaccfc7e029771d025d066501030
+Subproject commit 4b62bcbdb343cc9464ba1de800ddcfd9942a9a6f