/*
* ======================================================
* This function appends all additional subroutines
* called inside the user subroutine. It is specialised
* for CUDA in the related subclass
* ======================================================
*/
void FortranCUDAUserSubroutine::appendAdditionalSubroutines ( SgScopeStatement * moduleScope,
FortranParallelLoop * parallelLoop, FortranProgramDeclarationsAndDefinitions * declarations,
FortranConstantDeclarations * CUDAconstants, std::vector < SgProcedureHeaderStatement * > * allCalledRoutines)
{
using std::vector;
using boost::iequals;
/*
* ======================================================
* First removes duplicates in calledRoutines itself
* ======================================================
*/
sort ( calledRoutines.begin(), calledRoutines.end() );
calledRoutines.erase ( unique ( calledRoutines.begin(), calledRoutines.end() ), calledRoutines.end() );
Debug::getInstance ()->debugMessage ("Before removing, the list of routine calls found in the user kernels is: ",
Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
vector < SgProcedureHeaderStatement * > :: iterator routinesIt2;
for ( routinesIt2 = calledRoutines.begin (); routinesIt2 != calledRoutines.end (); routinesIt2++ )
{
string appendingSubroutine = (*routinesIt2)->get_name ().getString ();
Debug::getInstance ()->debugMessage (appendingSubroutine,
Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
}
/*
* ======================================================
* The removes routines already appended by other user
* kernels, using the list in allCalledRoutines
* ======================================================
*/
Debug::getInstance ()->debugMessage ("Removing global duplicates, the number of routines in the list is: '"
+ boost::lexical_cast<string> ((int) calledRoutines.size()) + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
vector < SgProcedureHeaderStatement * > :: iterator routinesIt;
for ( routinesIt = calledRoutines.begin (); routinesIt != calledRoutines.end (); ) //routinesIt++ )
{
string appendingSubroutine = (*routinesIt)->get_name ().getString ();
Debug::getInstance ()->debugMessage ("Checking routine for deletion: '"
+ appendingSubroutine + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
bool foundAndErased = false;
vector < SgProcedureHeaderStatement * > :: iterator finder;
for ( finder = allCalledRoutines->begin (); finder != allCalledRoutines->end (); finder++ )
{
Debug::getInstance ()->debugMessage ("Checking against: '"
+ (*finder)->get_name ().getString () + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
if ( iequals ((*finder)->get_name ().getString (), appendingSubroutine) )
{
/*
* ======================================================
* Routine already appended by another user kernel:
* delete it from list of routines to be appended for
* this user kernel, and exit this loop
* ======================================================
*/
Debug::getInstance ()->debugMessage ("Deleting: '"
+ appendingSubroutine + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
calledRoutines.erase (routinesIt++);
if ( calledRoutines.empty () ) return;
foundAndErased = true;
routinesIt--;
break;
}
}
if ( foundAndErased == false )
{
/*
* ======================================================
* New routine: it must be added to the list of
* routines called by all previous user kernels because
* recursively called routines need to discard those
* already appended by this routine
* ======================================================
*/
Debug::getInstance ()->debugMessage ("Not found, appending: '"
+ appendingSubroutine + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
allCalledRoutines->push_back ( *routinesIt );
routinesIt++;
}
}
vector < SgProcedureHeaderStatement * > :: iterator it;
for ( it = calledRoutines.begin(); it != calledRoutines.end(); it++ )
{
string calledSubroutineName = (*it)->get_name ().getString ();
Debug::getInstance ()->debugMessage ("Appending new subroutine '"
+ calledSubroutineName + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
FortranCUDAUserSubroutine * newRoutine = new FortranCUDAUserSubroutine ( moduleScope,
parallelLoop, declarations, calledSubroutineName );
newRoutine->createFormalParameterDeclarations ();
newRoutine->createStatements ();
additionalSubroutines.push_back (newRoutine);
}
vector < FortranUserSubroutine * > :: iterator itRecursive;
for ( itRecursive = additionalSubroutines.begin(); itRecursive != additionalSubroutines.end(); itRecursive++ )
{
FortranCUDAUserSubroutine * cudaSubroutineCasting = (FortranCUDAUserSubroutine *) *itRecursive;
CUDAconstants->patchReferencesToConstants (
(cudaSubroutineCasting )->getSubroutineHeaderStatement ());
RoseHelper::forceOutputOfCodeToFile (
(cudaSubroutineCasting )->getSubroutineHeaderStatement ());
}
for ( itRecursive = additionalSubroutines.begin(); itRecursive != additionalSubroutines.end(); itRecursive++ )
{
FortranCUDAUserSubroutine * cudaSubroutineCasting = (FortranCUDAUserSubroutine *) *itRecursive;
cudaSubroutineCasting->appendAdditionalSubroutines (moduleScope, parallelLoop, declarations, CUDAconstants, allCalledRoutines);
}
}
void
FortranCUDASubroutinesGeneration::createSubroutines ()
{
using std::string;
using std::map;
CUDAconstants->appendConstantInitialisationToModule ( moduleScope, declarations, /* isCuda = */ true );
/*
* ======================================================
* This vector contains all subroutines called by user
* kernels, to avoid their duplication in the output
* file
* ======================================================
*/
vector < SgProcedureHeaderStatement * > allCalledRoutines;
for (map <string, ParallelLoop *>::const_iterator it =
declarations->firstParallelLoop (); it
!= declarations->lastParallelLoop (); ++it)
{
string const userSubroutineName = it->first;
Debug::getInstance ()->debugMessage ("Analysing user subroutine '"
+ userSubroutineName + "' with subroutines already defined in previous kernels (number = "
+ boost::lexical_cast<string> (allCalledRoutines.size()) + ": ", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
vector < SgProcedureHeaderStatement * > :: iterator finder;
for ( finder = allCalledRoutines.begin (); finder != allCalledRoutines.end (); finder++ )
{
Debug::getInstance ()->debugMessage ("Routine: '" + (*finder)->get_name ().getString () + "'", Debug::FUNCTION_LEVEL, __FILE__, __LINE__);
}
FortranParallelLoop * parallelLoop =
static_cast <FortranParallelLoop *> (it->second);
FortranCUDAUserSubroutine * userDeviceSubroutine =
new FortranCUDAUserSubroutine (moduleScope, parallelLoop, declarations);
userDeviceSubroutine->createStatements ();
CUDAconstants->patchReferencesToConstants (
userDeviceSubroutine->getSubroutineHeaderStatement ());
/*
* ======================================================
* We have to set each node in the AST representation of
* this subroutine as compiler generated, otherwise chunks
* of the user kernel are missing in the output
* ======================================================
*/
RoseHelper::forceOutputOfCodeToFile (
userDeviceSubroutine->getSubroutineHeaderStatement ());
/*
* ======================================================
* When the user subroutine has calls to other user
* subroutines we need to add them to the generated file
* This call also eliminates automatically the duplicates
* of routines already contained in allCalledRoutines
* ======================================================
*/
userDeviceSubroutine->appendAdditionalSubroutines (moduleScope, parallelLoop, declarations, CUDAconstants, &allCalledRoutines);
// vector < FortranUserSubroutine * > additionalSubroutines = userDeviceSubroutine->getAdditionalSubroutines ();
/*
* ======================================================
* Appending new routines to the list of all routines
* appended
* ======================================================
*/
//allCalledRoutines.insert(allCalledRoutines.end(), additionalSubroutines.begin(), additionalSubroutines.end());
FortranCUDAKernelSubroutine * kernelSubroutine;
if (parallelLoop->isDirectLoop ())
{
kernelSubroutine
= new FortranCUDAKernelSubroutineDirectLoop (
moduleScope,
userDeviceSubroutine,
parallelLoop,
reductionSubroutines,
cardinalitiesDeclarations[userSubroutineName],
dimensionsDeclarations[userSubroutineName],
static_cast <FortranCUDAModuleDeclarations *> (moduleDeclarations[userSubroutineName]));
hostSubroutines[userSubroutineName]
= new FortranCUDAHostSubroutineDirectLoop (
moduleScope,
kernelSubroutine,
parallelLoop,
cardinalitiesDeclarations[userSubroutineName],
dimensionsDeclarations[userSubroutineName],
static_cast <FortranCUDAModuleDeclarations *> (moduleDeclarations[userSubroutineName]));
}
else
{
kernelSubroutine
= new FortranCUDAKernelSubroutineIndirectLoop (
moduleScope,
userDeviceSubroutine,
parallelLoop,
reductionSubroutines,
static_cast <FortranCUDAOpDatCardinalitiesDeclarationIndirectLoop *> (cardinalitiesDeclarations[userSubroutineName]),
dimensionsDeclarations[userSubroutineName],
static_cast <FortranCUDAModuleDeclarations *> (moduleDeclarations[userSubroutineName]));
hostSubroutines[userSubroutineName]
= new FortranCUDAHostSubroutineIndirectLoop (
moduleScope,
kernelSubroutine,
parallelLoop,
static_cast <FortranCUDAOpDatCardinalitiesDeclarationIndirectLoop *> (cardinalitiesDeclarations[userSubroutineName]),
dimensionsDeclarations[userSubroutineName],
static_cast <FortranCUDAModuleDeclarations *> (moduleDeclarations[userSubroutineName]));
}
}
}
