core: Add semantic ROOT::EnableImplicitMT

Instead of specificy the number of core, the new overload can
be specified to select a configuration type.  Currently the following
2 configuration are supported:

ROOT::EIMTConfig::kWholeMachine
ROOT::EIMTConfig::kExistingTBBArena

calling
```
ROOT::EnableImplicitMT(ROOT::EIMTConfig::kExistingTBBArena);
```
will attach to an existing TBB Arena instead of creating a new one.

This fixes #18301

Author: pcanal

core/base/inc/TROOT.h

@@ -80,12 +80,21 @@ namespace Internal {
 } } // End ROOT::Internal
 
 namespace ROOT {
+   enum class EIMTConfig {
+      kWholeMachine = 0,     ///< Default configuration
+      kExistingTBBArena = 1, ///< Use the existing TBB arena
+      kNumConfigs = 2        ///< Number of support IMT semantic configurations
+
+   };
    /// \brief Enable support for multi-threading within the ROOT code
    /// in particular, enables the global mutex to make ROOT thread safe/aware.
    void EnableThreadSafety();
    /// \brief Enable ROOT's implicit multi-threading for all objects and methods that provide an internal
    /// parallelisation mechanism.
    void EnableImplicitMT(UInt_t numthreads = 0);
+   /// \brief Enable ROOT's implicit multi-threading for all objects and methods that provide an internal
+   /// parallelisation mechanism.
+   void EnableImplicitMT(ROOT::EIMTConfig config);
    void DisableImplicitMT();
    Bool_t IsImplicitMTEnabled();
    UInt_t GetThreadPoolSize();

core/base/src/TROOT.cxx

@@ -551,6 +551,29 @@ namespace Internal {
 #endif
    }
 
+   ////////////////////////////////////////////////////////////////////////////////
+   /// @param[in] config Configuration to use. The default is kWholeMachine, which
+   ///                   will create a thread pool that spans the whole machine.
+   ///
+   /// EnableImplicitMT calls in turn EnableThreadSafety.
+   /// If ImplicitMT is already enabled, this function does nothing.
+
+
+   void EnableImplicitMT(ROOT::EIMTConfig config)
+   {
+#ifdef R__USE_IMT
+      if (ROOT::Internal::IsImplicitMTEnabledImpl())
+         return;
+      EnableThreadSafety();
+      static void (*sym)(ROOT::EIMTConfig) = (void(*)(ROOT::EIMTConfig))Internal::GetSymInLibImt("ROOT_TImplicitMT_EnableImplicitMT_Config");
+      if (sym)
+         sym(config);
+      ROOT::Internal::IsImplicitMTEnabledImpl() = true;
+#else
+      ::Warning("EnableImplicitMT", "Cannot enable implicit multi-threading with %d threads, please build ROOT with -Dimt=ON", numthreads);
+#endif
+   }
+
    ////////////////////////////////////////////////////////////////////////////////
    /// Disables the implicit multi-threading in ROOT (see EnableImplicitMT).
    void DisableImplicitMT()

core/imt/inc/ROOT/RTaskArena.hxx

@@ -24,6 +24,7 @@
 #define ROOT_RTaskArena
 
 #include "RConfigure.h"
+#include "TROOT.h" // For ROOT::EIMTConfig
 #include <memory>
 
 // exclude in case ROOT does not have IMT support
@@ -65,9 +66,13 @@ public:
    ~RTaskArenaWrapper(); // necessary to set size back to zero
    static unsigned TaskArenaSize(); // A static getter lets us check for RTaskArenaWrapper's existence
    ROOT::ROpaqueTaskArena &Access();
-private:
+   struct Attach {}; ///< Marker for attaching to an existing tbb::task_arena
+
    RTaskArenaWrapper(unsigned maxConcurrency = 0);
-   friend std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(unsigned maxConcurrency);
+   RTaskArenaWrapper(Attach);
+private:
+
+   friend std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(bool, unsigned, ROOT::EIMTConfig);
    std::unique_ptr<ROOT::ROpaqueTaskArena> fTBBArena;
    static unsigned fNWorkers;
 };
@@ -81,6 +86,7 @@ private:
 /// references to the previous one are gone and the object destroyed.
 ////////////////////////////////////////////////////////////////////////////////
 std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(unsigned maxConcurrency = 0);
+std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(ROOT::EIMTConfig config);
 
 } // namespace Internal
 } // namespace ROOT

core/imt/src/ROpaqueTaskArena.hxx

@@ -1,5 +1,8 @@
 #include "tbb/task_arena.h"
 
 namespace ROOT {
-class ROpaqueTaskArena: public tbb::task_arena {};
+class ROpaqueTaskArena: public tbb::task_arena {
+   public:
+      using tbb::task_arena::task_arena;
+};
 }

core/imt/src/RTaskArena.cxx

@@ -108,6 +108,21 @@ RTaskArenaWrapper::RTaskArenaWrapper(unsigned maxConcurrency) : fTBBArena(new RO
    ROOT::EnableThreadSafety();
 }
 
+
+////////////////////////////////////////////////////////////////////////////////
+/// Initializes the tbb::task_arena within RTaskArenaWrapper by attaching to an
+/// existing arena.
+///
+/// * Can't be reinitialized
+////////////////////////////////////////////////////////////////////////////////
+RTaskArenaWrapper::RTaskArenaWrapper(RTaskArenaWrapper::Attach) :
+  fTBBArena(new ROpaqueTaskArena{tbb::task_arena::attach{}})
+{
+   fTBBArena->initialize(tbb::task_arena::attach{});
+   fNWorkers = fTBBArena->max_concurrency();
+   ROOT::EnableThreadSafety();
+}
+
 RTaskArenaWrapper::~RTaskArenaWrapper()
 {
    fNWorkers = 0u;
@@ -127,7 +142,7 @@ ROOT::ROpaqueTaskArena &RTaskArenaWrapper::Access()
    return *fTBBArena;
 }
 
-std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(unsigned maxConcurrency)
+std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(bool semantic, unsigned maxConcurrency, ROOT::EIMTConfig config)
 {
    static std::weak_ptr<ROOT::Internal::RTaskArenaWrapper> weak_GTAWrapper;
 
@@ -140,10 +155,28 @@ std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(unsigned m
       }
       return sp;
    }
-   std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> sp(new ROOT::Internal::RTaskArenaWrapper(maxConcurrency));
+   std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> sp;
+   if (semantic && config == ROOT::EIMTConfig::kExistingTBBArena) {
+      sp = std::make_shared<ROOT::Internal::RTaskArenaWrapper>(ROOT::Internal::RTaskArenaWrapper::Attach{});
+   } else {
+      if (semantic && config != ROOT::EIMTConfig::kWholeMachine) {
+         maxConcurrency = 0;
+      }
+      sp = std::make_shared<ROOT::Internal::RTaskArenaWrapper>(maxConcurrency);
+   }
    weak_GTAWrapper = sp;
    return sp;
 }
 
+std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(ROOT::EIMTConfig config)
+{
+   return GetGlobalTaskArena(true, 0, config);
+}
+
+std::shared_ptr<ROOT::Internal::RTaskArenaWrapper> GetGlobalTaskArena(unsigned maxConcurrency)
+{
+   return GetGlobalTaskArena(false, maxConcurrency, ROOT::EIMTConfig::kNumConfigs);
+}
+
 } // namespace Internal
 } // namespace ROOT

core/imt/src/TImplicitMT.cxx

@@ -18,6 +18,7 @@
 //                                                                      //
 //////////////////////////////////////////////////////////////////////////
 
+#include "TROOT.h"
 #include "TError.h"
 #include "ROOT/RTaskArena.hxx"
 #include <atomic>
@@ -55,6 +56,21 @@ extern "C" void ROOT_TImplicitMT_EnableImplicitMT(UInt_t numthreads)
    }
 };
 
+extern "C" void ROOT_TImplicitMT_EnableImplicitMT_Config(ROOT::EIMTConfig config)
+{
+   if (!GetImplicitMTFlag()) {
+      if (config < ROOT::EIMTConfig::kNumConfigs) {
+         R__GetTaskArena4IMT() = ROOT::Internal::GetGlobalTaskArena(config);
+      } else {
+         ::Warning("ROOT_TImplicitMT_EnableImplicitMT_Config", "Unknown enum value %d defaulting to EIMTCconfig::kWholeMachine", (int)config);
+         R__GetTaskArena4IMT() = ROOT::Internal::GetGlobalTaskArena(0);
+      }
+      GetImplicitMTFlag() = true;
+   } else {
+      ::Warning("ROOT_TImplicitMT_EnableImplicitMT_Config", "Implicit multi-threading is already enabled");
+   }
+};
+
 extern "C" void ROOT_TImplicitMT_DisableImplicitMT()
 {
    if (GetImplicitMTFlag()) {

core/imt/test/CMakeLists.txt

@@ -5,4 +5,5 @@
 # For the list of contributors see $ROOTSYS/README/CREDITS.
 
 ROOT_ADD_GTEST(testTaskArena testRTaskArena.cxx LIBRARIES Imt ${TBB_LIBRARIES} FAILREGEX "")
-ROOT_ADD_GTEST(testTBBGlobalControl testTBBGlobalControl.cxx LIBRARIES Imt ${TBB_LIBRARIES})
+ROOT_ADD_GTEST(testTBBGlobalControl testTBBGlobalControl.cxx testEnableImt.cxx LIBRARIES Imt ${TBB_LIBRARIES})
+ROOT_ADD_GTEST(testEnable testEnableImt.cxx LIBRARIES Imt ${TBB_LIBRARIES})

core/imt/test/testEnableImt.cxx

@@ -0,0 +1,26 @@
+#include "TROOT.h"
+#include "ROOT/RTaskArena.hxx"
+
+#include "tbb/task_arena.h"
+
+#include "ROOT/TestSupport.hxx"
+#include "gtest/gtest.h"
+
+#ifdef R__USE_IMT
+
+static const unsigned gMaxConcurrency = ROOT::Internal::LogicalCPUBandwidthControl();
+
+TEST(EnableImt, TBBAttach)
+{
+   tbb::task_arena main_arena;
+   main_arena.initialize(2);
+
+   ROOT::EnableImplicitMT(ROOT::EIMTConfig::kExistingTBBArena);
+
+   auto psize = ROOT::GetThreadPoolSize();
+
+   EXPECT_TRUE( psize > 1);
+   EXPECT_EQ( psize, 2); // gMaxConcurrency);
+}
+
+#endif