Add configure-time CELERITAS_OPENMP switch to change threading (cel…

…eritas-project#1222)

* Comment cmake config

* Add cmake for openmp

* Switch openmp pragmas based on configure flag

* Fix ccache key for ultralite builds

* Update warning message in accel

* Overwrite tags

* Address review feedback

* Escape macros for launcher

.github/workflows/build-ultralite.yml

@@ -28,7 +28,7 @@ jobs:
         uses: actions/cache@v4
         with:
           path: ${{env.CCACHE_DIR}}
-          key: ccache-ultralite-${{github.run_id}}
+          key: ccache-ultralite-ubuntu-${{github.run_id}}
           restore-keys: |
             ccache-ultralite
       - name: Zero ccache stats
@@ -80,7 +80,7 @@ jobs:
         uses: actions/cache@v4
         with:
           path: ${{env.CCACHE_DIR}}
-          key: ccache-ultralite-${{github.run_id}}
+          key: ccache-ultralite-windows-${{github.run_id}}
           restore-keys: |
             ccache-ultralite
       - name: Zero ccache stats

CMakeLists.txt

@@ -80,16 +80,31 @@ cmake_dependent_option(CELERITAS_TEST_VERBOSE
 # CELERITAS CORE IMPLEMENTATION OPTIONS
 #----------------------------------------------------------------------------#
 
-# CELERITAS_CORE_RNG: random number generator selection
+if(CELERITAS_USE_CUDA OR CELERITAS_USE_HIP)
+  set(CELERITAS_MAX_BLOCK_SIZE 256
+    CACHE STRING "Threads-per-block launch bound for Celeritas action kernels"
+  )
+else()
+  set(CELERITAS_MAX_BLOCK_SIZE 0)
+endif()
+
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# CELERITAS_CORE_RNG
+# Random number generator selection
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 celeritas_setup_option(CELERITAS_CORE_RNG xorwow)
 celeritas_setup_option(CELERITAS_CORE_RNG cuRAND CELERITAS_USE_CUDA)
 celeritas_setup_option(CELERITAS_CORE_RNG hipRAND CELERITAS_USE_HIP)
-# TODO: allow wrapper to standard library RNG when not building for device?
+# TODO: add wrapper to standard library RNG when not building for device?
+# TODO: add ranluxpp?
 # TODO: maybe even add wrapper to Geant4 RNG??
 celeritas_define_options(CELERITAS_CORE_RNG
   "Celeritas runtime random number generator")
 
-# CELERITAS_CORE_GEO: runtime geometry selection
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# CELERITAS_CORE_GEO
+# Runtime geometry selection
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 if(CELERITAS_USE_VecGeom AND NOT CELERITAS_USE_HIP)
   set(_allow_vecgeom TRUE)
 else()
@@ -98,7 +113,7 @@ else()
     message(SEND_ERROR "VecGeom core geometry is incompatible with HIP")
   endif()
 endif()
-if(CELERITAS_USE_Geant4 AND NOT (CELERITAS_USE_HIP OR CELERITAS_USE_CUDA OR CELERITAS_USE_OpenMP))
+if(CELERITAS_USE_Geant4 AND NOT (CELERITAS_USE_HIP OR CELERITAS_USE_CUDA OR CELERITAS_OPENMP))
   set(_allow_g4 TRUE)
 else()
   if(CELERITAS_CORE_GEO STREQUAL "Geant4")
@@ -111,14 +126,26 @@ celeritas_setup_option(CELERITAS_CORE_GEO ORANGE)
 celeritas_setup_option(CELERITAS_CORE_GEO Geant4 _allow_g4)
 celeritas_define_options(CELERITAS_CORE_GEO "Celeritas runtime geometry")
 
-if(CELERITAS_USE_CUDA OR CELERITAS_USE_HIP)
-  set(CELERITAS_MAX_BLOCK_SIZE 256
-    CACHE STRING "Threads-per-block launch bound for Celeritas action kernels"
-  )
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# CELERITAS_OPENMP
+# Thread parallelism
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+if(CELERITAS_USE_OpenMP)
+  set(_disable_openmp FALSE)
 else()
-  set(CELERITAS_MAX_BLOCK_SIZE 0)
+  set(_disable_openmp TRUE)
 endif()
 
+celeritas_setup_option(CELERITAS_OPENMP disabled _disable_openmp)
+celeritas_setup_option(CELERITAS_OPENMP event CELERITAS_USE_OpenMP)
+celeritas_setup_option(CELERITAS_OPENMP track CELERITAS_USE_OpenMP)
+celeritas_define_options(CELERITAS_OPENMP "Celeritas OpenMP parallelism")
+
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# CELERITAS_UNITS
+# Unit system for celeritas runtime
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 if(CELERITAS_CORE_GEO STREQUAL Geant4)
   set(_allow_single_prec FALSE)
 else()
@@ -131,6 +158,10 @@ celeritas_setup_option(CELERITAS_UNITS CLHEP)
 celeritas_define_options(CELERITAS_UNITS
   "Native unit system for Celeritas")
 
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+# CELERITAS_REAL_TYPE
+# Precision for real numbers
+#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 celeritas_setup_option(CELERITAS_REAL_TYPE double)
 celeritas_setup_option(CELERITAS_REAL_TYPE float)
 celeritas_define_options(CELERITAS_REAL_TYPE
@@ -269,6 +300,11 @@ endif()
 
 if(CELERITAS_USE_Geant4 AND NOT Geant4_FOUND)
   find_package(Geant4 REQUIRED)
+  if(CELERITAS_OPENMP STREQUAL "track" AND Geant4_multithreaded_FOUND)
+    message(WARNING "Track-level OpenMP will conflict with MT runs of geant4:"
+      "consider setting CELERITAS_OPENMP to \"event\" or disabling OpenMP"
+    )
+  endif()
 endif()
 
 if(CELERITAS_USE_HepMC3)

app/celer-sim/Runner.cc

@@ -89,7 +89,7 @@ namespace
 size_type calc_num_streams(RunnerInput const& inp, size_type num_events)
 {
     size_type num_threads = 1;
-#if CELERITAS_USE_OPENMP
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_EVENT
     if (!inp.merge_events)
     {
 #    pragma omp parallel

app/celer-sim/celer-sim.cc

@@ -125,7 +125,7 @@ void run(std::istream* is, std::shared_ptr<OutputRegistry> output)
         CELER_LOG(status) << "Transporting " << run_stream.num_events()
                           << " on " << num_streams << " threads";
         MultiExceptionHandler capture_exception;
-#ifdef _OPENMP
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_EVENT
 #    pragma omp parallel for
 #endif
         for (size_type event = 0; event < run_stream.num_events(); ++event)

scripts/ci/run-ci.sh

@@ -20,7 +20,7 @@ if [ -f "${_ENV_SCRIPT}" ]; then
 fi
 
 # Fetch tags for version provenance
-git fetch --tags
+git fetch -f --tags
 
 # Clean older builds from jenkins *BEFORE* setting up presets
 git clean -fxd

src/CMakeLists.txt

@@ -25,10 +25,11 @@ set(CELERITAS_USE_VECGEOM ${CELERITAS_USE_VecGeom})
 # Start counter from 1 because undefined config have the implicit value of 0 in
 # the C preprocessor, so any unavailable options (e.g. CELERITAS_USE_CURAND
 # when HIP is in use) will implicitly be zero.
-celeritas_generate_option_config(CELERITAS_CORE_RNG)
 celeritas_generate_option_config(CELERITAS_CORE_GEO)
-celeritas_generate_option_config(CELERITAS_UNITS)
+celeritas_generate_option_config(CELERITAS_CORE_RNG)
+celeritas_generate_option_config(CELERITAS_OPENMP)
 celeritas_generate_option_config(CELERITAS_REAL_TYPE)
+celeritas_generate_option_config(CELERITAS_UNITS)
 celeritas_configure_file("celeritas_config.h.in" "celeritas_config.h" @ONLY)
 
 #----------------------------------------------------------------------------#
@@ -47,7 +48,7 @@ endif()
 
 set(CELERITAS_CMAKE_STRINGS)
 set(CELERITAS_BUILD_TYPE ${CMAKE_BUILD_TYPE})
-foreach(_var BUILD_TYPE HOSTNAME REAL_TYPE CORE_GEO CORE_RNG UNITS)
+foreach(_var BUILD_TYPE HOSTNAME REAL_TYPE UNITS OPENMP CORE_GEO CORE_RNG)
   set(_var "CELERITAS_${_var}")
   string(TOLOWER "${_var}" _lower)
   string(APPEND CELERITAS_CMAKE_STRINGS

src/accel/LocalTransporter.cc

@@ -73,11 +73,12 @@ LocalTransporter::LocalTransporter(SetupOptions const& options,
         << params.Params()->max_streams() << ")");
 
     // Check that OpenMP and Geant4 threading models don't collide
-    if (CELERITAS_USE_OPENMP && !celeritas::device()
+    if (CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK && !celeritas::device()
         && G4Threading::IsMultithreadedApplication())
     {
         auto msg = CELER_LOG_LOCAL(warning);
-        msg << "Using multithreaded Geant4 with Celeritas OpenMP";
+        msg << "Using multithreaded Geant4 with Celeritas track-level OpenMP "
+               "parallelism";
         if (std::string const& nt_str = celeritas::getenv("OMP_NUM_THREADS");
             !nt_str.empty())
         {

src/celeritas/global/ActionLauncher.hh

@@ -7,6 +7,7 @@
 //---------------------------------------------------------------------------//
 #pragma once
 
+#include "celeritas_config.h"
 #include "corecel/Assert.hh"
 #include "corecel/Types.hh"
 #include "corecel/sys/MultiExceptionHandler.hh"
@@ -34,7 +35,7 @@ void launch_action(ExplicitActionInterface const& action,
                    F&& execute_thread)
 {
     MultiExceptionHandler capture_exception;
-#ifdef _OPENMP
+#if defined(_OPENMP) && CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #    pragma omp parallel for
 #endif
     for (size_type i = 0; i < num_threads; ++i)

src/celeritas/track/detail/TrackSortUtils.cc

@@ -114,7 +114,9 @@ void count_tracks_per_action(
     std::fill(offsets.begin(), offsets.end(), ThreadId{});
     auto const size = states.size();
     // if get_action(0) != get_action(1), get_action(0) never gets initialized
-#pragma omp parallel for
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
+#    pragma omp parallel for
+#endif
     for (size_type i = 1; i < size; ++i)
     {
         ActionId current_action = get_action(ThreadId{i});

src/celeritas/user/detail/SimpleCaloImpl.cc

@@ -7,6 +7,7 @@
 //---------------------------------------------------------------------------//
 #include "SimpleCaloImpl.hh"
 
+#include "celeritas_config.h"
 #include "corecel/Types.hh"
 #include "corecel/sys/MultiExceptionHandler.hh"
 #include "corecel/sys/ThreadId.hh"
@@ -27,7 +28,9 @@ void simple_calo_accum(HostRef<StepStateData> const& step,
     CELER_EXPECT(step && calo);
     MultiExceptionHandler capture_exception;
     SimpleCaloExecutor execute{step, calo};
-#pragma omp parallel for
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
+#    pragma omp parallel for
+#endif
     for (ThreadId::size_type i = 0; i < step.size(); ++i)
     {
         CELER_TRY_HANDLE(execute(ThreadId{i}), capture_exception);

src/celeritas_config.h.in

@@ -31,6 +31,8 @@
 
 @CELERITAS_UNITS_CONFIG@
 
+@CELERITAS_OPENMP_CONFIG@
+
 @CELERITAS_CORE_GEO_CONFIG@
 
 @CELERITAS_CORE_RNG_CONFIG@

src/corecel/math/Atomics.hh

@@ -19,6 +19,13 @@ namespace celeritas
 //---------------------------------------------------------------------------//
 /*!
  * Add to a value, returning the original value.
+ *
+ * Note that on CPU, this assumes the atomic add is being done in with \em
+ * track-level parallelism rather than \em event-level because these utilities
+ * are meant for "kernel" code.
+ *
+ * \warning Multiple events must not use this function to simultaneously modify
+ * shared data.
  */
 template<class T>
 CELER_FORCEINLINE_FUNCTION T atomic_add(T* address, T value)
@@ -28,7 +35,7 @@ CELER_FORCEINLINE_FUNCTION T atomic_add(T* address, T value)
 #else
     CELER_EXPECT(address);
     T initial;
-#    ifdef _OPENMP
+#    if defined(_OPENMP) && CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #        pragma omp atomic capture
 #    endif
     {
@@ -78,7 +85,7 @@ CELER_FORCEINLINE_FUNCTION T atomic_min(T* address, T value)
 #else
     CELER_EXPECT(address);
     T initial;
-#    ifdef _OPENMP
+#    if defined(_OPENMP) && CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #        pragma omp atomic capture
 #    endif
     {
@@ -101,7 +108,7 @@ CELER_FORCEINLINE_FUNCTION T atomic_max(T* address, T value)
 #else
     CELER_EXPECT(address);
     T initial;
-#    ifdef _OPENMP
+#    if defined(_OPENMP) && CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #        pragma omp atomic capture
 #    endif
     {

src/corecel/sys/MultiExceptionHandler.cc

@@ -106,7 +106,7 @@ MultiExceptionHandler::~MultiExceptionHandler()
  */
 void MultiExceptionHandler::operator()(std::exception_ptr p)
 {
-#ifdef _OPENMP
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #    pragma omp critical(MultiExceptionHandler)
 #endif
     {

src/geocel/rasterize/RaytraceImager.t.hh

@@ -137,7 +137,7 @@ void RaytraceImager<G>::launch_raytrace_kernel(
     size_type const num_threads = geo_states.size();
 
     MultiExceptionHandler capture_exception;
-#ifdef _OPENMP
+#if CELERITAS_OPENMP == CELERITAS_OPENMP_TRACK
 #    pragma omp parallel for
 #endif
     for (size_type i = 0; i < num_threads; ++i)