in_list
objectsAllocated(SgStatement *statement, StopCond stopCond)
{
in_list objs;
SgStatement *parent;
do
{
#if 0
in_list::iterator startOfCurrent = objs.end();
#endif
parent = isSgStatement(statement->get_parent());
//cout << "parent = " << parent << endl;
ROSE_ASSERT(parent);
in_list blockObjs = blockObjectsAllocated(parent, statement);
objs.insert(objs.end(), blockObjs.begin(), blockObjs.end());
#if 0
switch (parent->variantT())
{
case V_SgBasicBlock:
{
SgBasicBlock *b = isSgBasicBlock(parent);
ROSE_ASSERT(b);
SgStatementPtrList &stmts = b->get_statements();
for (SgStatementPtrList::iterator i = stmts.begin(); i != stmts.end(); ++i)
{
if (*i == statement) break;
addStmtVarsIfAny(objs, startOfCurrent, *i);
}
}
break;
case V_SgForStatement:
{
SgForStatement *fs = isSgForStatement(parent);
SgStatementPtrList &stmts = fs->get_init_stmt();
bool done = false;
if (!omitInit)
{
for (SgStatementPtrList::iterator i = stmts.begin(); i != stmts.end(); ++i)
{
if (*i == statement)
{
done = true;
break;
}
addStmtVarsIfAny(objs, startOfCurrent, *i);
}
}
if (!done)
{
if (fs->get_test() != statement)
{
addStmtVarsIfAny(objs, startOfCurrent, fs->get_test());
}
}
}
break;
default:
break;
}
#endif
statement = parent;
} while (!stopCond(parent));
return objs;
}
void MintCudaMidend::lowerMinttoCuda(SgSourceFile* file)
{
//Searches mint pragmas, and performs necessary transformation
//We also check the mint pragmas syntactically
//At this point, we only care parallel regions and for loops
//But do not process forloops at this point (note that this is bottom-up)
//We process forloops when we see a parallel region pragma because they are always
//inside of a parallel region.
//TODO: Sometimes a forloop is merged with a parallel region. Need to handle these.
ROSE_ASSERT(file != NULL);
//replaces all the occurrences of mint parallel with omp parallel
replaceMintParallelWithOmpParallel(file);
replaceMintForWithOmpFor(file);
//adds the private and first private into private clause explicitly
patchUpPrivateVariables(file); //uses ROSE's
patchUpFirstprivateVariables(file); //uses ROSE's
//insert openmp specific headers
//insertRTLHeaders(file);
//check if mint pragma declarations are correct
mintPragmasFrontendProcessing(file);
//the map has the mapping from the host variables to device variables
//where we copy the data
Rose_STL_Container<SgNode*> nodeList = NodeQuery::querySubTree(file, V_SgStatement);
Rose_STL_Container<SgNode*>::reverse_iterator nodeListIterator = nodeList.rbegin();
for ( ;nodeListIterator !=nodeList.rend(); ++nodeListIterator)
{
SgStatement* node = isSgStatement(*nodeListIterator);
ROSE_ASSERT(node != NULL);
switch (node->variantT())
{
case V_SgOmpParallelStatement:
{
//first we handle data transfer pragmas
MintHostSymToDevInitMap_t hostToDevVars;
processDataTransferPragmas(node, hostToDevVars);
#ifdef VERBOSE_2
cout << " INFO:Mint: @ Line " << node->get_file_info()->get_line() << endl;
cout << " Processing Mint Parallel Statement" << endl << endl;
#endif
MintCudaMidend::transOmpParallel(node, hostToDevVars);
break;
}
case V_SgOmpTaskStatement:
{
//transOmpTask(node);
break;
}
case V_SgOmpForStatement:
{
//cout << "INFO-mint: Omp For Statement (skipped processing it)" << endl;
//LoweringToCuda::transOmpFor(node);
//OmpSupport::transOmpFor(node);
break;
}
case V_SgOmpBarrierStatement:
{
#ifdef VERBOSE_2
cout << " INFO:Mint: @ Line " << node->get_file_info()->get_line() << endl;
cout << " Processing Omp Barrier Statement" << endl;
#endif
transOmpBarrierToCudaBarrier(node);
break;
}
case V_SgOmpFlushStatement:
{
cout << " INFO:Mint: Processing Omp Flush Statement" << endl;
transOmpFlushToCudaBarrier(node);
break;
}
case V_SgOmpThreadprivateStatement:
{
//transOmpThreadprivate(node);
break;
}
case V_SgOmpTaskwaitStatement:
{
//transOmpTaskwait(node);
break;
}
case V_SgOmpSingleStatement:
{
//TODO: we need to check if the loop body becomes a cuda kernel or not.
MintCudaMidend::transOmpSingle(node);
break;
}
case V_SgOmpMasterStatement:
{
//TODO: we need to check if the loop body becomes a cuda kernel or not.
MintCudaMidend::transOmpMaster(node);
break;
}
case V_SgOmpAtomicStatement:
{
//transOmpAtomic(node);
break;
}
case V_SgOmpOrderedStatement:
{
//transOmpOrdered(node);
break;
}
case V_SgOmpCriticalStatement:
{
//transOmpCritical(node);
break;
}
default:
{
//This is any other statement in the source code which is not omp pragma
//cout<< node->unparseToString()<<" at line:"<< (node->get_file_info())->get_line()<<endl;
// do nothing here
}
}// switch
}
#if 0
//3. Special handling for files with main()
// rename main() to user_main()
SgFunctionDeclaration * mainFunc = findMain(cur_file);
if (mainFunc)
{
renameMainToUserMain(mainFunc);
}
#endif
}
void MPI_Code_Generator::lower_xomp (SgSourceFile* file)
{
ROSE_ASSERT(file != NULL);
Rose_STL_Container<SgNode*> nodeList = NodeQuery::querySubTree(file, V_SgStatement);
Rose_STL_Container<SgNode*>::reverse_iterator nodeListIterator = nodeList.rbegin();
for ( ;nodeListIterator !=nodeList.rend(); ++nodeListIterator)
{
SgStatement* node = isSgStatement(*nodeListIterator);
ROSE_ASSERT(node != NULL);
//debug the order of the statements
// cout<<"Debug lower_omp(). stmt:"<<node<<" "<<node->class_name() <<" "<< node->get_file_info()->get_line()<<endl;
switch (node->variantT())
{
#if 0
case V_SgOmpParallelStatement:
{
// check if this parallel region is under "omp target"
SgNode* parent = node->get_parent();
ROSE_ASSERT (parent != NULL);
if (isSgBasicBlock(parent)) // skip the padding block in between.
parent= parent->get_parent();
if (isSgOmpTargetStatement(parent))
transOmpTargetParallel(node);
else
transOmpParallel(node);
break;
}
case V_SgOmpForStatement:
case V_SgOmpDoStatement:
{
// check if the loop is part of the combined "omp parallel for" under the "omp target" directive
// TODO: more robust handling of this logic, not just fixed AST form
bool is_target_loop = false;
SgNode* parent = node->get_parent();
ROSE_ASSERT (parent != NULL);
// skip a possible BB between omp parallel and omp for, especially when the omp parallel has multiple omp for loops
if (isSgBasicBlock(parent))
parent = parent->get_parent();
SgNode* grand_parent = parent->get_parent();
ROSE_ASSERT (grand_parent != NULL);
if (isSgOmpParallelStatement (parent) && isSgOmpTargetStatement(grand_parent) )
is_target_loop = true;
if (is_target_loop)
{
// transOmpTargetLoop (node);
// use round-robin scheduler for larger iteration space and better performance
transOmpTargetLoop_RoundRobin(node);
}
else
{
transOmpLoop(node);
}
break;
}
#endif
// transform combined "omp target parallel for", represented as separated three directives: omp target, omp parallel, and omp for
case V_SgOmpForStatement:
{
SgOmpTargetStatement * omp_target;
SgOmpParallelStatement* omp_parallel;
if (isCombinedTargetParallelFor (isSgOmpForStatement(node),&omp_target, &omp_parallel ))
{
transOmpTargetParallelLoop (isSgOmpForStatement(node));
}
break;
}
case V_SgOmpTargetStatement:
{
SgOmpTargetStatement* t_stmt = isSgOmpTargetStatement(node);
ROSE_ASSERT (t_stmt != NULL);
SgStatement* body_stmt = t_stmt->get_body();
SgBasicBlock * body_block = isSgBasicBlock (body_stmt);
// transOmpTarget(node);
if (isMPIAllBegin (t_stmt))
{
// move all body statements to be after omp target
if (body_block != NULL)
{
stripOffBasicBlock (body_block, t_stmt);
}
else
{
//TODO: ideally, the body_stmt should be normalized to be a BB even it is only a single statement
removeStatement (body_stmt);
insertStatementAfter (t_stmt, body_stmt, false);
}
// remove the pragma stmt after the translation
removeStatement (t_stmt);
}
else if (isMPIMasterBegin (t_stmt))
{
transMPIDeviceMaster (t_stmt);
}
else
{
// other target directive with followed omp parallel for will be handled when parallel for is translated
// cerr<<"Error. Unhandled target directive:" <<t_stmt->unparseToString()<<endl;
//ROSE_ASSERT (false);
}
break;
}
default:
{
// do nothing here
}
}// switch
} // end for
}
void MintCudaMidend::processLoopsInParallelRegion(SgNode* parallelRegionNode,
MintHostSymToDevInitMap_t hostToDevVars,
ASTtools::VarSymSet_t& deviceSyms,
MintSymSizesMap_t& trfSizes,
std::set<SgInitializedName*>& readOnlyVars,
const SgVariableSymbol* dev_struct)
{
Rose_STL_Container<SgNode*> nodeList = NodeQuery::querySubTree(parallelRegionNode, V_SgStatement);
Rose_STL_Container<SgNode*>::reverse_iterator nodeListIterator = nodeList.rbegin();
for ( ;nodeListIterator !=nodeList.rend(); ++nodeListIterator)
{
SgStatement* node = isSgStatement(*nodeListIterator);
ROSE_ASSERT(node != NULL);
switch (node->variantT())
{
case V_SgOmpForStatement:
{
#ifdef VERBOSE_2
cout << " INFO:Mint: @ Line " << node->get_file_info()->get_line() << endl;
cout << " Processing Omp For Statement" << endl << endl;
#endif
//DataTransferSizes::findTransferSizes(node, trfSizes);
bool isBoundaryCond = LoweringToCuda::isBoundaryConditionLoop(node);
SgFunctionDeclaration* kernel;
MintForClauses_t clauseList;
//kernel= LoweringToCuda::transOmpFor(node, hostToDevVars, deviceSyms, readOnlyVars,clauseList, dev_struct) ;
kernel= LoweringToCuda::transOmpFor(node, hostToDevVars, deviceSyms, clauseList, dev_struct) ;
//swap anyways // x swapping is buggy, need to fix that before allowing x as well
if(clauseList.chunksize.x == 1 && ( clauseList.chunksize.z != 1 || clauseList.chunksize.y != 1 ))
{
//if(MintOptions::GetInstance()->isSwapOpt())
CudaOptimizer::swapLoopAndIf(kernel, clauseList);
}
if (!isBoundaryCond && MintOptions::GetInstance()->optimize())
//if (MintOptions::GetInstance()->optimize())
{
cout << "\n\n INFO:Mint: Optimization is ON. Optimizing ...\n\n" ;
CudaOptimizer::optimize(kernel, clauseList);
}
// MintTools::printAllStatements(isSgNode(kernel));
//MintArrayInterface::linearizeArrays(kernel);
break;
}
default:
{
//cout << " INFO:Mint: @ Line " << node->get_file_info()->get_line() << endl;
//cout << " Currently we only handle for loops" << endl << endl;
//do nothing
//currently we only handle for loops
break;
}
}
}
for (ASTtools::VarSymSet_t::const_iterator i = deviceSyms.begin (); i!= deviceSyms.end (); ++i)
{
SgVariableSymbol* sym= const_cast<SgVariableSymbol*> (*i);
SgInitializedName* name = sym->get_declaration();
SgType* type = name->get_type();
if(isSgArrayType(type) || isSgPointerType(type)){
//Check if is of the fields of the struct
if(hostToDevVars.find(sym) == hostToDevVars.end())
{
string name_str = name->get_name().str();
cerr << " ERR:Mint: Ooops! Did you forget to insert a copy pragma for the variable ("<< name_str << ") ?"<< endl;
cerr << " ERR:Mint: Please insert the copy pragma and compile again "<< endl;
cerr << " INFO:Mint: Note that copy pragmas should appear right before and after a parallel region" << endl;
ROSE_ABORT();
}
}
}
}
