Merge branch 'feature/passiveParallelRegions' into develop

OpDiLib handles passive parallel regions and assigns the corresponding
tapes.

Author: jblueh

include/opdi/logic/omp/implicitTaskOmpLogic.cpp

@@ -24,6 +24,7 @@
  */
 
 #include "../../config.hpp"
+#include "../../helpers/exceptions.hpp"
 #include "../../tool/toolInterface.hpp"
 
 #include "instrument/ompLogicInstrumentInterface.hpp"
@@ -55,7 +56,10 @@ void* opdi::ImplicitTaskOmpLogic::onImplicitTaskBegin(int actualParallelism, int
     data->parallelData = parallelData;
 
     void* newTape = this->tapePool.getTape(parallelData->masterTape, index);
-    tool->setActive(newTape, true);
+
+    if (parallelData->activeParallelRegion) {
+      tool->setActive(newTape, true);
+    }
 
     data->parallelData->tapes[index] = newTape;
 
@@ -94,6 +98,14 @@ void opdi::ImplicitTaskOmpLogic::onImplicitTaskEnd(void* dataPtr) {
     tool->getTapePosition(data->parallelData->tapes[data->index],
                           data->parallelData->positions[data->index].back());
 
+    if (!data->parallelData->activeParallelRegion) {
+      if (tool->comparePosition(data->parallelData->positions[data->index].front(),
+                                data->parallelData->positions[data->index].back()) != 0) {
+        OPDI_WARNING("Something became active during a passive parallel region. This is not supported and will not be ",
+                     "differentiated correctly.");
+      }
+    }
+
     #if OPDI_OMP_LOGIC_INSTRUMENT
       for (auto& instrument : ompLogicInstruments) {
         instrument->onImplicitTaskEnd(data);
@@ -102,7 +114,7 @@ void opdi::ImplicitTaskOmpLogic::onImplicitTaskEnd(void* dataPtr) {
 
     tool->setActive(data->parallelData->tapes[data->index], false);
 
-    if (data->oldTape == data->parallelData->masterTape) {
+    if (data->oldTape == data->parallelData->masterTape && data->parallelData->activeParallelRegion) {
       tool->setActive(data->oldTape, true);
     }
 

include/opdi/logic/omp/parallelOmpLogic.cpp

@@ -32,6 +32,8 @@
 
 #include "parallelOmpLogic.hpp"
 
+int opdi::ParallelOmpLogic::skipParallelHandling = 0;
+
 void opdi::ParallelOmpLogic::reverseFunc(void* dataPtr) {
 
   Data* data = (Data*) dataPtr;
@@ -42,6 +44,8 @@ void opdi::ParallelOmpLogic::reverseFunc(void* dataPtr) {
     }
   #endif
 
+  ++ParallelOmpLogic::skipParallelHandling;
+
   #pragma omp parallel num_threads(data->actualThreads)
   {
     int threadNum = omp_get_thread_num();
@@ -70,6 +74,8 @@ void opdi::ParallelOmpLogic::reverseFunc(void* dataPtr) {
                      data->adjointAccessModes[threadNum][j - 1] == AdjointAccessMode::Atomic);
     }
 
+    --ParallelOmpLogic::skipParallelHandling;
+
     tool->setThreadLocalTape(oldTape);
 
     #if OPDI_OMP_LOGIC_INSTRUMENT
@@ -90,6 +96,8 @@ void opdi::ParallelOmpLogic::deleteFunc(void* dataPtr) {
 
   Data* data = (Data*) dataPtr;
 
+  ++ParallelOmpLogic::skipParallelHandling;
+
   // this triggers possibly pending implicit task end events
   #pragma omp parallel num_threads(data->actualThreads)
   {
@@ -108,17 +116,20 @@ void opdi::ParallelOmpLogic::deleteFunc(void* dataPtr) {
     }
   }
 
+  --ParallelOmpLogic::skipParallelHandling;
+
   delete [] data->tapes;
   delete data;
 }
 
 void* opdi::ParallelOmpLogic::onParallelBegin(int maxThreads) {
 
-  if (tool->getThreadLocalTape() != nullptr && tool->isActive(tool->getThreadLocalTape())) {
+  if (tool->getThreadLocalTape() != nullptr && ParallelOmpLogic::skipParallelHandling == 0) {
 
     Data* data = new Data;
 
     data->maxThreads = maxThreads;
+    data->activeParallelRegion = tool->isActive(tool->getThreadLocalTape());
     data->masterTape = tool->getThreadLocalTape();
     data->tapes = new void*[maxThreads]();
     data->positions.resize(maxThreads);
@@ -155,14 +166,20 @@ void opdi::ParallelOmpLogic::onParallelEnd(void* dataPtr) {
       }
     #endif
 
-    Handle* handle = new Handle;
-    handle->data = (void*) data;
-    handle->reverseFunc = ParallelOmpLogic::reverseFunc;
-    handle->deleteFunc = ParallelOmpLogic::deleteFunc;
+    if (data->activeParallelRegion) {
+
+      Handle* handle = new Handle;
+      handle->data = (void*) data;
+      handle->reverseFunc = ParallelOmpLogic::reverseFunc;
+      handle->deleteFunc = ParallelOmpLogic::deleteFunc;
 
-    tool->pushExternalFunction(data->masterTape, handle);
+      tool->pushExternalFunction(data->masterTape, handle);
 
-    // do not delete data, it is deleted with the handle
+      // do not delete data, it is deleted with the handle
+
+    } else {
+      deleteFunc(data);
+    }
   }
   #if OPDI_OMP_LOGIC_INSTRUMENT
   else {

include/opdi/logic/omp/parallelOmpLogic.hpp

@@ -47,6 +47,7 @@ namespace opdi {
         public:
           int maxThreads;
           int actualThreads;
+          bool activeParallelRegion;
           void* masterTape;
           void** tapes;
           std::vector<std::deque<void*>> positions;
@@ -56,6 +57,9 @@ namespace opdi {
 
     private:
 
+      static int skipParallelHandling;
+      #pragma omp threadprivate(skipParallelHandling)
+
       static void reverseFunc(void* dataPtr);
       static void deleteFunc(void* dataPtr);