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]));
}
}
}
