diff --git a/algo/CMakeLists.txt b/algo/CMakeLists.txt
index 67f3fae56624c33dce195a12c791c017fc860af9..fa27d8b4c8e6ad3599121e88a77b4c63f5df6ea3 100644
--- a/algo/CMakeLists.txt
+++ b/algo/CMakeLists.txt
@@ -102,6 +102,11 @@ target_link_libraries(Algo
target_compile_definitions(Algo PUBLIC NO_ROOT)
xpu_attach(Algo ${DEVICE_SRCS})
+# Link against OpenMP if available
+if (OpenMP_CXX_FOUND)
+ target_link_libraries(Algo PUBLIC OpenMP::OpenMP_CXX)
+endif()
+
# Try to enable parallel execution in c++17 if TBB is available
if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
list(APPEND CMAKE_PREFIX_PATH "/opt/intel/oneapi/tbb/latest/")
diff --git a/algo/base/BuildInfo.h b/algo/base/BuildInfo.h
index bdacc7a166b19cf390075c6ec7e28415c32f9fe4..e96a9cb89c3d4ffa69058a13f983fca5c417276d 100644
--- a/algo/base/BuildInfo.h
+++ b/algo/base/BuildInfo.h
@@ -10,6 +10,10 @@
#define HAVE_PARALLEL_ALGORITHM
#endif
+#if __has_include(<omp.h>)
+#define HAVE_OMP
+#endif
+
namespace cbm::algo::BuildInfo
{
@@ -31,6 +35,13 @@ namespace cbm::algo::BuildInfo
false;
#endif
+ inline constexpr bool WITH_OMP =
+#ifdef HAVE_OMP
+ true;
+#else
+ false;
+#endif
+
} // namespace cbm::algo::BuildInfo
#endif // CBM_ALGO_BUILD_INFO_H
diff --git a/algo/base/Options.cxx b/algo/base/Options.cxx
index cc1be8932090bd96c5f80e2ea1c600d649ff1075..8a30d58307849426b188bf3c8a66b2f417759589 100644
--- a/algo/base/Options.cxx
+++ b/algo/base/Options.cxx
@@ -84,6 +84,8 @@ Options::Options(int argc, char** argv)
"Stop after <num> timeslices (-1 = all)")
("skip-ts", po::value(&fSkipTimeslices)->default_value(0)->value_name("<num>"),
"Skip first <num> timeslices")
+ ("omp", po::value(&fNumOMPThreads)->default_value(-1)->value_name("<num>"),
+ "Set number of OpenMP threads (-1 = use OMP_NUM_THREADS environment variable)")
("times,t", po::bool_switch(&fCollectKernelTimes)->default_value(false),
"print kernel times")
("help,h",
diff --git a/algo/base/Options.h b/algo/base/Options.h
index d3c387e467ec9425831a752b399e5da631dd6cb7..182a76e6d10e050928075f547540254d8ff8e540 100644
--- a/algo/base/Options.h
+++ b/algo/base/Options.h
@@ -35,6 +35,13 @@ namespace cbm::algo
bool CollectKernelTimes() const { return fCollectKernelTimes; }
int NumTimeslices() const { return fNumTimeslices; }
int SkipTimeslices() const { return fSkipTimeslices; }
+
+ std::optional<int> NumOMPThreads() const
+ {
+ // omp doesn't allow negative number of threads, so we use -1 to indicate that the user didn't specify a number
+ // and omp should use OMP_NUM_THREADS environment variable or the default instead
+ return fNumOMPThreads > 0 ? std::make_optional(fNumOMPThreads) : std::nullopt;
+ }
const std::string& ChildId() const { return fChildId; }
const std::vector<Step>& Steps() const { return fRecoSteps; }
@@ -64,6 +71,7 @@ namespace cbm::algo
bool fCollectKernelTimes = false;
int fNumTimeslices = -1;
int fSkipTimeslices = 0;
+ int fNumOMPThreads = -1;
std::vector<Step> fRecoSteps;
std::vector<RecoData> fOutputTypes;
std::vector<fles::Subsystem> fDetectors;
diff --git a/algo/base/compat/OpenMP.h b/algo/base/compat/OpenMP.h
new file mode 100644
index 0000000000000000000000000000000000000000..9129cdde38b7e09184ad95943abc0ea1570ba305
--- /dev/null
+++ b/algo/base/compat/OpenMP.h
@@ -0,0 +1,28 @@
+/* Copyright (C) 2023 FIAS Frankfurt Institute for Advanced Studies, Frankfurt / Main
+ SPDX-License-Identifier: GPL-3.0-only
+ Authors: Felix Weiglhofer [committer] */
+#ifndef CBM_ALGO_BASE_COMPAT_OPENMP_H
+#define CBM_ALGO_BASE_COMPAT_OPENMP_H
+
+#include "BuildInfo.h"
+
+#ifdef HAVE_OMP
+#include <omp.h>
+#endif
+
+namespace cbm::algo::openmp
+{
+
+#ifndef HAVE_OMP
+ inline int GetMaxThreads() { return 1; }
+ inline int GetThreadNum() { return 0; }
+ inline void SetNumThreads(int) {}
+#else
+ inline int GetMaxThreads() { return omp_get_max_threads(); }
+ inline int GetThreadNum() { return omp_get_thread_num(); }
+ inline void SetNumThreads(int n) { omp_set_num_threads(n); }
+#endif
+
+} // namespace cbm::algo::openmp
+
+#endif // CBM_ALGO_BASE_COMPAT_OPENMP_H
diff --git a/algo/global/Reco.cxx b/algo/global/Reco.cxx
index b1a225fec74ff2d0f2a6ad6c8405dbca8118e48e..f47691b519bd87a2e3282300560e43ffd4659882 100644
--- a/algo/global/Reco.cxx
+++ b/algo/global/Reco.cxx
@@ -10,6 +10,7 @@
#include <xpu/host.h>
+#include "compat/OpenMP.h"
#include "config/Yaml.h"
#include "evbuild/Config.h"
#include "log.hpp"
@@ -48,7 +49,8 @@ void Reco::Init(const Options& opts)
fStsHitFinder.SetContext(&fContext);
xpu::device_prop props {xpu::device::active()};
- L_(info) << "Running CBM Reco on Device " << props.name();
+ L_(info) << "Running CBM Reco on Device '" << props.name() << "' (Using " << openmp::GetMaxThreads()
+ << " OpenMP threads)";
if (!opts.MonitorUri().empty()) {
fContext.monitor = std::make_unique<cbm::Monitor>(opts.MonitorUri());
diff --git a/reco/app/cbmreco/main.cxx b/reco/app/cbmreco/main.cxx
index 7edd4a4cb684c8ca67a2be6c9e01db2a4dcfe2d0..18ea5e3d0b2111e86330a32288c91239afc849fa 100644
--- a/reco/app/cbmreco/main.cxx
+++ b/reco/app/cbmreco/main.cxx
@@ -14,6 +14,7 @@
#include "BuildInfo.h"
#include "Options.h"
#include "Reco.h"
+#include "compat/OpenMP.h"
using namespace cbm::algo;
@@ -36,8 +37,15 @@ int main(int argc, char** argv)
xpu::initialize(settings);
xpu::preload<GPUReco>();
+ auto ompThreads = opts.NumOMPThreads();
+ if (ompThreads) {
+ L_(debug) << *ompThreads << " OpenMP threads requested";
+ openmp::SetNumThreads(*ompThreads);
+ }
+
L_(info) << "CBMRECO buildType=" << BuildInfo::BUILD_TYPE << " gpuDebug=" << BuildInfo::GPU_DEBUG
- << " parallelSTL=" << BuildInfo::WITH_PARALLEL_ALGORITHM << " commit=" << BuildInfo::GIT_HASH;
+ << " parallelSTL=" << BuildInfo::WITH_PARALLEL_ALGORITHM << " OMP=" << BuildInfo::WITH_OMP
+ << " commit=" << BuildInfo::GIT_HASH;
std::stringstream ss;
for (int i = 0; i < argc; i++) {
ss << argv[i] << " ";
@@ -47,7 +55,6 @@ int main(int argc, char** argv)
Reco reco;
reco.Init(opts);
-
Archive archive(ArchiveDescriptor {}); // TODO: use opts.Detector() once detector flag is merged
fles::TimesliceAutoSource source {opts.InputLocator()};
int tsIdx = 0;