ktyl
/
basalt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Disable Eigen's parallelization with OpenMP (interferes with TBB)

This commit is contained in:
Nikolaus Demmel 2021-01-29 15:14:58 +01:00
parent 86d77a6744
commit bf94c9ee61
2 changed files with 37 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
CMakeLists.txt
View File

@ -129,7 +129,7 @@ endif()
# Set platform / compiler specific compile flags and checks # Set platform / compiler specific compile flags and checks
if(APPLE) if(APPLE)
# Need to investigate how to reliably detect and use OpenMP on macOS... # 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 # Among others, setting CMAKE_FIND_FRAMEWORK to LAST fixed issues
# with installed Mono that contains old headers (libpng, ...). # with installed Mono that contains old headers (libpng, ...).
@ -163,7 +163,7 @@ if(APPLE)
endif() endif()
elseif(UNIX) elseif(UNIX)
set(USE_OPENMP_DEFAULT ON) # set(USE_OPENMP_DEFAULT ON)
# assume libstdc++ # assume libstdc++
set(STD_CXX_FS stdc++fs) set(STD_CXX_FS stdc++fs)
@ -186,13 +186,37 @@ endif()
# OpenMP option and compile flags # 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 # setup combined compiler flags

4
include/basalt/optimization/accumulator.h
View File

@ -201,6 +201,10 @@ class SparseHashAccumulator {
VectorX res; VectorX res;
if (iterative_solver) { 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::ConjugateGradient<Eigen::SparseMatrix<double>,
Eigen::Lower | Eigen::Upper> Eigen::Lower | Eigen::Upper>
cg; cg;