- redo changes of rev22212, but this time without circular dependency

git-svn-id: https://openmodelica.org/svn/OpenModelica/trunk@22234 f25d12d1-65f4-0310-ae8a-bbce733d8d8e

Compiler/BackEnd/HpcOmScheduler.mo

@@ -1097,10 +1097,11 @@ algorithm
   end match;
 end convertScheduleStrucToInfo1;
 
-protected function convertScheduleStrucToInfoLevel
+protected function convertScheduleStrucToInfoLevel "author: marcusw
+  Convert the given task list, representing a level of the level-scheduler, to a scheduler info."
   input list<HpcOmSimCode.TaskList> taskLst;
   input Integer sectionsNumber;
-  input array<tuple<Integer,Integer,Real>> iScheduleInfo;
+  input array<tuple<Integer,Integer,Real>> iScheduleInfo; //maps each Task to <threadId, orderId, startCalcTime>
   output array<tuple<Integer,Integer,Real>> oScheduleInfo;
 algorithm
   oScheduleInfo := matchcontinue(taskLst,sectionsNumber,iScheduleInfo)
@@ -1129,22 +1130,24 @@ protected function convertScheduleStrucToInfoLevel1
   input list<HpcOmSimCode.Task> tasks;
   input Integer sectionsNumber;
   input Integer sectionIdx;
-  input array<tuple<Integer,Integer,Real>> iScheduleInfo;
+  input array<tuple<Integer,Integer,Real>> iScheduleInfo; //maps each Task to <threadId, orderId, startCalcTime>
   output array<tuple<Integer,Integer,Real>> oScheduleInfo;
 algorithm
   oScheduleInfo := matchcontinue(tasks,sectionsNumber,sectionIdx,iScheduleInfo)
     local
-      Integer numNodes;
+      Integer numNodes, threadIdx;
       list<Integer> nodeIdc;
       array<tuple<Integer,Integer,Real>> scheduleInfo;
       list<tuple<Integer,Integer,Real>> tuplLst;
       list<HpcOmSimCode.Task> rest;
+      Option<Integer> threadIdxOpt;
     case({},_,_,_)
       then iScheduleInfo;
-    case(HpcOmSimCode.CALCTASK_LEVEL(nodeIdc=nodeIdc)::rest,_,_,_)
+    case(HpcOmSimCode.CALCTASK_LEVEL(nodeIdc=nodeIdc,threadIdx=threadIdxOpt)::rest,_,_,_)
       equation
         numNodes = listLength(nodeIdc);
-        tuplLst = List.threadMap1(List.fill(sectionsNumber,numNodes),List.fill(sectionIdx,numNodes),Util.make3Tuple,0.0);
+        threadIdx = Util.getOptionOrDefault(threadIdxOpt,-1);
+        tuplLst = List.threadMap1(List.fill(threadIdx,numNodes),List.fill(-1,numNodes),Util.make3Tuple,0.0);
         List.threadMap1_0(nodeIdc,tuplLst,Util.arrayUpdateIndexFirst,iScheduleInfo);
       then convertScheduleStrucToInfoLevel1(rest,sectionsNumber,sectionIdx+1,iScheduleInfo);
   end matchcontinue;
@@ -1680,22 +1683,25 @@ public function createFixedLevelSchedule
   output HpcOmTaskGraph.TaskGraphMeta oMeta;
 protected
   list<list<Integer>> levelTasks;
+  //for each node (arrayIdx): list of threads that handle predecessor tasks. If a node has multiple  
+  //predecessors handled by the same thread, the thread-index occurs multiple times in the list.
   array<list<Integer>> adviceLists;
   HpcOmSimCode.Schedule tmpSchedule;
   list<HpcOmSimCode.TaskList> levelTaskLists;
 algorithm
   // 1. Create a task list for each thread and a advice list for each task which is empty at beginning
   // 2. Iterate over all levels
-  //  2.1. Create an ready-list and set all values to 0
+  //  2.1. Create an thread-ready-list and set all values to 0
   //  2.2. Iterate over all tasks of the current level
   //    2.2.1. Find the thread that should calulcate the task
-  //        2.2.1. (1) This could be the thread with the lowest value in the ready list if no task is in the advice list
+  //        2.2.1. (1) This could be the thread with the lowest value in the ready list if no thread is in the advice list
   //        2.2.1. (2) This could be the first thread in the advice list, if the thread has not already an execution time > (sum(exec(levelTasks)) / numThreads)
   //        2.2.1. (3) Otherwise case (1)
   //    2.2.2. Append the task to the thread and add the execution cost the the ready list
   //    2.2.3. Add the thread to the advice-list of all successor tasks
   levelTasks := HpcOmTaskGraph.getLevelNodes(iGraph);
   adviceLists := arrayCreate(arrayLength(iGraph), {});
+  HpcOmTaskGraph.printTaskGraph(iGraph);
   levelTaskLists := List.fold5(levelTasks, createFixedLevelScheduleForLevel, adviceLists, iGraph, iMeta, iNumberOfThreads, iSccSimEqMapping, {});
   levelTaskLists := listReverse(levelTaskLists);
   oSchedule := HpcOmSimCode.LEVELSCHEDULE(levelTaskLists);
@@ -1722,6 +1728,7 @@ protected
   list<HpcOmSimCode.Task> tasksOfLevel;
   array<list<Integer>> inComps;
 algorithm
+  print("createFixedLevelScheduleForLevel: entering next level\n");
   HpcOmTaskGraph.TASKGRAPHMETA(exeCosts=exeCosts,inComps=inComps) := iMeta;
   levelExecCosts := HpcOmTaskGraph.getCostsForContractedNodes(iTasksOfLevel, exeCosts);
   threadReadyList := arrayCreate(iNumberOfThreads, 0.0);
@@ -1731,6 +1738,7 @@ algorithm
   ((_,tasksOfLevel)) := Util.arrayFold2(threadTaskList, createFixedLevelScheduleForLevel0, inComps, iSccSimEqMapping, (1,{}));
   taskList := HpcOmSimCode.PARALLELTASKLIST(tasksOfLevel);
   oLevelTaskLists := taskList :: iLevelTaskLists;
+  print("createFixedLevelScheduleForLevel: leaving level\n");
 end createFixedLevelScheduleForLevel;
 
 protected function createFixedLevelScheduleForLevel0
@@ -1768,10 +1776,16 @@ protected
   Integer threadIdx;
   Real threadReadyTime, exeCost;
 algorithm
+  print("\tcreateFixedLevelScheduleForTask: handling task: " +& intString(iTaskIdx) +& "\n");
   adviceElem := arrayGet(iAdviceList, iTaskIdx);
+  print("\t\tAdvice-list: " +& stringDelimitList(List.map(adviceElem, intString), ",") +& "\n");
+  adviceElem := flattenAdviceList(adviceElem, arrayLength(iThreadReadyList));
+  print("\t\tAdvice-list-flattened: " +& stringDelimitList(List.map(adviceElem, intString), ",") +& "\n");
   threadIdx := getBestFittingThread(adviceElem, iLevelExecCosts, iThreadReadyList);
+  print("\t\tBest-thread: " +& intString(threadIdx) +& "\n");
   threadTasks := arrayGet(iThreadTasks, threadIdx);
-  successorList := arrayGet(iGraph, threadIdx);
+  successorList := arrayGet(iGraph, iTaskIdx);
+  print("\t\tSuccessors: " +& stringDelimitList(List.map(successorList, intString), ",") +& "\n");
   //update the advice list
   _ := List.fold1(successorList, createFixedLevelScheduleForTask0, threadIdx, iAdviceList);
   threadReadyTime := arrayGet(iThreadReadyList, threadIdx);
@@ -1785,6 +1799,8 @@ algorithm
 end createFixedLevelScheduleForTask;
 
 protected function createFixedLevelScheduleForTask0
+  "author: marcusw
+   Update the given advice list, by adding the iThreadAdvice to the successor-task-entry."
   input Integer iSuccessor;
   input Integer iThreadAdvice;
   input array<list<Integer>> iAdviceList;
@@ -1797,8 +1813,81 @@ algorithm
   oAdviceList := arrayUpdate(iAdviceList, iSuccessor, adviceElem);
 end createFixedLevelScheduleForTask0;
 
-protected function getBestFittingThread
+protected function flattenAdviceList
+  "author: marcusw
+   Flatten the given advice list and order the entries regarding their occurency count.
+   For example: {2,3,1,1,2,2} -> {2,1,3}"
   input list<Integer> iAdviceList;
+  input Integer iNumOfThreads;
+  output list<Integer> oAdviceList;
+protected
+  array<Integer> counterArray;
+  list<tuple<Integer,Integer>> tupleList;
+algorithm
+  counterArray := arrayCreate(iNumOfThreads,0);
+  counterArray := List.fold(iAdviceList, flattenAdviceListElem, counterArray);
+  tupleList := arrayToTupleListZeroRemoved(counterArray, 1, {});
+  oAdviceList := List.map(List.sort(tupleList, intTpl22Gt),Util.tuple21);
+end flattenAdviceList;
+
+protected function flattenAdviceListElem "author: marcusw
+   Increment the value in the counter array of the given thread (iAdviceElem)."
+  input Integer iAdviceElem;
+  input array<Integer> iCounterArray;
+  output array<Integer> oCounterArray;
+protected
+  Integer counter;
+algorithm
+  counter := arrayGet(iCounterArray, iAdviceElem);
+  counter := counter + 1;
+  oCounterArray := arrayUpdate(iCounterArray, iAdviceElem, counter);
+end flattenAdviceListElem;
+
+protected function arrayToTupleListZeroRemoved "author: marcusw
+   Convert a integer array, to list of tuples <arrayIndex, value> if the value is != 0.
+   For example: [1,4,2] -> {<1,1>,<2,4>,<3,2>}"
+  input array<Integer> iArray;
+  input Integer iCurrentIdx;
+  input list<tuple<Integer,Integer>> iTupleList;
+  output list<tuple<Integer,Integer>> oTupleList;
+protected
+  list<tuple<Integer,Integer>> tmpTupleList;
+  Integer currentValue;
+algorithm
+  oTupleList := matchcontinue(iArray, iCurrentIdx, iTupleList)
+    case(_,_,_)
+      equation
+        true = intLe(iCurrentIdx, arrayLength(iArray));
+        currentValue = arrayGet(iArray, iCurrentIdx);
+        true = intNe(currentValue, 0);
+        tmpTupleList = (iCurrentIdx, currentValue)::iTupleList;
+        tmpTupleList = arrayToTupleListZeroRemoved(iArray, iCurrentIdx+1, tmpTupleList);
+      then tmpTupleList;
+    case(_,_,_)
+      equation
+        true = intLe(iCurrentIdx, arrayLength(iArray));
+        tmpTupleList = arrayToTupleListZeroRemoved(iArray, iCurrentIdx+1, iTupleList);
+      then tmpTupleList;
+    else then iTupleList;
+  end matchcontinue;
+end arrayToTupleListZeroRemoved;
+
+protected function intTpl22Gt
+  input tuple<Integer,Integer> iTpl1;
+  input tuple<Integer,Integer> iTpl2;
+  output Boolean oRes;
+protected
+  Integer val1, val2;
+algorithm
+  (_,val1) := iTpl1;
+  (_,val2) := iTpl2;
+  oRes := intGt(val1, val2);
+end intTpl22Gt;
+
+protected function getBestFittingThread
+  "author: marcusw
+  Get the optimal thread for the task, regarding the given advice list."
+  input list<Integer> iAdviceList; //advice list of the task - the list is traversed from front to back, until a suitable thread is found
   input Real iLevelExecCosts; //sum of all execosts
   input array<Real> iThreadReadyList;
   output Integer oThreadIdx;

Compiler/BackEnd/HpcOmSimCode.mo

@@ -32,6 +32,7 @@
 encapsulated package HpcOmSimCode
 
   public import HashTableCrILst;
+  //public import SimCode;
 
   public uniontype MemoryMap //stores information to organize the memory for the parallel code in an efficient way
     record MEMORYMAP_ARRAY

Compiler/BackEnd/HpcOmSimCodeMain.mo

@@ -161,7 +161,7 @@ algorithm
 
       Boolean taskGraphMetaValid, numFixed;
       String taskGraphMetaMessage, criticalPathInfo;
-      array<tuple<Integer,Integer,Real>> schedulerInfo;
+      array<tuple<Integer,Integer,Real>> schedulerInfo; //maps each Task to <threadId, orderId, startCalcTime>
       HpcOmSimCode.Schedule schedule;
       array<tuple<Integer,Integer,Integer>> eqCompMapping, varCompMapping;
       Real graphCosts;
@@ -331,7 +331,7 @@ algorithm
       SimCodeUtil.execStat("hpcom create schedule");
 
       fileName = ("taskGraph"+&filenamePrefix+&"ODE_schedule.graphml");
-      HpcOmTaskGraph.dumpAsGraphMLSccLevel(taskGraphScheduled, taskGraphDataScheduled, inBackendDAE, fileName, criticalPathInfo, HpcOmTaskGraph.convertNodeListToEdgeTuples(List.first(criticalPaths)), HpcOmTaskGraph.convertNodeListToEdgeTuples(List.first(criticalPathsWoC)), sccSimEqMapping, schedulerInfo, HpcOmTaskGraph.GRAPHDUMPOPTIONS(true,true,true,true));
+      HpcOmTaskGraph.dumpAsGraphMLSccLevel(taskGraphScheduled, taskGraphDataScheduled, inBackendDAE, fileName, criticalPathInfo, HpcOmTaskGraph.convertNodeListToEdgeTuples(List.first(criticalPaths)), HpcOmTaskGraph.convertNodeListToEdgeTuples(List.first(criticalPathsWoC)), sccSimEqMapping, schedulerInfo, HpcOmTaskGraph.GRAPHDUMPOPTIONS(true,false,true,true));
       //HpcOmScheduler.printSchedule(schedule);
 
       SimCodeUtil.execStat("hpcom dump schedule TaskGraph");

Compiler/Template/CodegenCpp.tpl

@@ -1269,28 +1269,28 @@ case SIMCODE(modelInfo=MODELINFO(__), makefileParams=MAKEFILE_PARAMS(__)) then
   #include <Core/ModelicaDefine.h>
   #include <SimCoreFactory/Policies/FactoryConfig.h>
   #include <SimController/ISimController.h>
-  <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then '#include "Core/Utils/extension/measure_time.hpp"' %>
+  <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then '#include "Core/Utils/extension/measure_time_rdtsc.hpp"' %>
   #if defined(_MSC_VER) || defined(__MINGW32__)
   #include <tchar.h>
   int _tmain(int argc, const _TCHAR* argv[])
   #else
   int main(int argc, const char* argv[])
   #endif
   {
-  <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then 'RDTSC_MeasureTime::initialize();' %>
+      <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then 'MeasureTimeRDTSC::initialize();' %>
       try
       {
-      boost::shared_ptr<OMCFactory>  _factory =  boost::shared_ptr<OMCFactory>(new OMCFactory());
+            boost::shared_ptr<OMCFactory>  _factory =  boost::shared_ptr<OMCFactory>(new OMCFactory());
             //SimController to start simulation
 
             std::pair<boost::shared_ptr<ISimController>,SimSettings> simulation =  _factory->createSimulation(argc,argv);
 
-        //create Modelica system
+            //create Modelica system
             std::pair<boost::shared_ptr<IMixedSystem>,boost::shared_ptr<ISimData> > system = simulation.first->LoadSystem("OMCpp<%fileNamePrefix%><%makefileParams.dllext%>","<%lastIdentOfPath(modelInfo.name)%>");
 
             simulation.first->Start(system.first,simulation.second,"<%lastIdentOfPath(modelInfo.name)%>");
-      <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then '//RDTSC_MeasureTime::deinitialize();' %>
-      return 0;
+            <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then '//MeasureTimeRDTSC::deinitialize();' %>
+            return 0;
 
       }
       catch(std::exception& ex)
@@ -1651,12 +1651,14 @@ case SIMCODE(modelInfo = MODELINFO(__)) then
         >>
         %>
         <%if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then
+        let numOfEqs = SimCodeUtil.getMaxSimEqSystemIndex(simCode)
         <<
-        MeasureTime::data ref;
-        ref.max_time = 0;
-        ref.num_calcs = 0;
-        ref.sum_time = 0;
-        measureTimeArray = std::vector<MeasureTime::data>(<%intAdd(listLength(simCode.allEquations), listLength(simCode.initialEquations))%>,ref);
+        measureTimeArray = std::vector<MeasureTimeData>(<%numOfEqs%>);
+        measuredStartValues = MeasureTime::getZeroValues();
+        measuredEndValues = MeasureTime::getZeroValues();
+        
+        for(int i = 0; i < <%numOfEqs%>; i++)
+            measureTimeArray[i] = MeasureTimeData();
         >>
         %>
         //DAE's are not supported yet, Index reduction is enabled
@@ -1681,9 +1683,9 @@ case SIMCODE(modelInfo = MODELINFO(__)) then
 
         <%if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then
         <<
-            MeasureTime::getInstance()->writeTimeToJason("<%dotPath(modelInfo.name)%>",measureTimeArray);
+        MeasureTime::getInstance()->writeTimeToJason("<%dotPath(modelInfo.name)%>",measureTimeArray);
         >>
-      %>
+        %>
     }
 
 
@@ -4284,7 +4286,11 @@ match modelInfo
 
      boost::shared_ptr<ISimData> _simData;
 
-  <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then 'std::vector<MeasureTime::data> measureTimeArray;' %>
+     <% if boolNot(stringEq(getConfigString(PROFILING_LEVEL),"none")) then
+     <<
+     std::vector<MeasureTimeData> measureTimeArray;
+     MeasureTimeValues *measuredStartValues, *measuredEndValues, *measuredOverheadValues;
+     >>%>
 
      <%memberfuncs%>
 
@@ -7278,20 +7284,23 @@ end equation_function_call;
 template measureTimeStart(String eq_number)
 ::=
   <<
-    long long unsigned startMeasure_t, endMeasure_t,measure_t;
-    startMeasure_t = MeasureTime::getTime();
+  MeasureTime::getTimeValues(measuredStartValues);
+  for(int i = 0; i < 10; i++) {
   >>
 end measureTimeStart;
 
 template measureTimeStop(String eq_number)
 ::=
   let eqNumber = intSub(stringInt(eq_number),1)
   <<
-    endMeasure_t = MeasureTime::getTime();
-    measure_t = endMeasure_t - startMeasure_t;
-    if(measure_t > measureTimeArray[<%eqNumber%>].max_time) measureTimeArray[<%eqNumber%>].max_time = measure_t;
-    measureTimeArray[<%eqNumber%>].sum_time += measure_t;
-    ++(measureTimeArray[<%eqNumber%>].num_calcs);
+  }
+  MeasureTime::getTimeValues(measuredEndValues);
+  measuredEndValues->sub(measuredStartValues);
+  measuredEndValues->div(10);
+  //measuredEndValues->sub(MeasureTime::getOverhead());
+  //if(measure_t > measureTimeArray[<%eqNumber%>].max_time) measureTimeArray[<%eqNumber%>].max_time = measuredEndValues;
+  (measureTimeArray[<%eqNumber%>].sumMeasuredValues)->add(measuredEndValues);
+  ++(measureTimeArray[<%eqNumber%>].numCalcs);
   >>
 end measureTimeStop;
 
@@ -7339,10 +7348,10 @@ template equation_function_create_single_func(SimEqSystem eq, Context context, S
     */
     void <%lastIdentOfPathFromSimCode(simCode)%><%classnameext%>::<%method%>_<%ix_str%>()
     {
-    <%measureTimeStartVar%>
       <%varDeclsLocal%>
-      <%body%>
-    <%measureTimeEndVar%>
+      <%measureTimeStartVar%>
+        <%body%>
+      <%measureTimeEndVar%>
     }
   >>
 

Compiler/Template/SimCodeTV.mo

@@ -835,6 +835,10 @@ package SimCodeUtil
     output Integer outVariableIndex;
   end getVariableIndex;
 
+  function getMaxSimEqSystemIndex
+    input SimCode.SimCode simCode;
+    output Integer idxOut;
+  end getMaxSimEqSystemIndex;
 end SimCodeUtil;
 
 
@@ -1129,9 +1133,6 @@ package DAE
     record BCONST
       Boolean bool;
     end BCONST;
-    record CLKCONST
-      Absyn.ClockKind clk;
-    end CLKCONST;
     record ENUM_LITERAL
       Absyn.Path name;
       Integer index;
@@ -3110,10 +3111,6 @@ package Types
     input list<DAE.Exp> inDimensionLst;
     output DAE.Type outType;
   end liftArrayListExp;
- function unliftArray
-    input DAE.Type inType;
-    output DAE.Type outType;
-  end unliftArray;
 end Types;
 
 package FMI

SimulationRuntime/cpp/Core/Utils/extension/CMakeLists.txt

@@ -2,7 +2,7 @@ cmake_minimum_required (VERSION 2.6)
 
 project(${ExtensionUtilitiesName})
 
-add_library(${ExtensionUtilitiesName} SHARED measure_time.cpp)
+add_library(${ExtensionUtilitiesName} SHARED measure_time.cpp measure_time_rdtsc.cpp)
 target_link_libraries (${ExtensionUtilitiesName} ${Boost_LIBRARIES})
 
 if (UNIX)
@@ -11,5 +11,6 @@ endif(UNIX)
 
 install (TARGETS ${ExtensionUtilitiesName} DESTINATION lib/omc/${LIBINSTALLEXT})
 
-install (FILES ${CMAKE_SOURCE_DIR}/Include/Core/Utils/extension/measure_time.hpp
-          DESTINATION include/omc/cpp/Core/Utils/extension)
+install (FILES  ${CMAKE_SOURCE_DIR}/Include/Core/Utils/extension/measure_time.hpp
+                ${CMAKE_SOURCE_DIR}/Include/Core/Utils/extension/measure_time_rdtsc.hpp
+         DESTINATION include/omc/cpp/Core/Utils/extension)