#include <klass.hh>

Inheritance diagram for Klass:

Collaboration diagram for Klass:

Public Member Functions
	Klass (const string &name, const string &super, int numInputs, int numOutputs, bool __vec=false)
virtual	~Klass ()
void	openLoop (const string &size)
	Open a non-recursive loop on top of the stack of open loops.
void	openLoop (Tree recsymbol, const string &size)
	Open a recursive loop on top of the stack of open loops.
void	closeLoop (Tree sig)
	Close the top loop and either keep it or absorb it within its enclosing loop.
void	setLoopProperty (Tree sig, Loop *l)
	Store the loop used to compute a signal.
bool	getLoopProperty (Tree sig, Loop *&l)
	Returns the loop used to compute a signal.
const string &	getClassName () const
	Returns the name of the class.
Loop *	topLoop ()
void	buildTasksList ()
void	addIncludeFile (const string &str)
void	addLibrary (const string &str)
void	rememberNeedPowerDef ()
void	collectIncludeFile (set< string > &S)
void	collectLibrary (set< string > &S)
void	addSubKlass (Klass *son)
void	addDeclCode (const string &str)
void	addInitCode (const string &str)
void	addStaticInitCode (const string &str)
void	addStaticFields (const string &str)
void	addUICode (const string &str)
void	addUIMacro (const string &str)
void	incUIActiveCount ()
void	incUIPassiveCount ()
void	addSharedDecl (const string &str)
void	addFirstPrivateDecl (const string &str)
void	addZone1 (const string &str)
void	addZone2 (const string &str)
void	addZone2b (const string &str)
void	addZone2c (const string &str)
void	addZone3 (const string &str)
void	addPreCode (const string &str)
void	addExecCode (const string &str)
void	addPostCode (const string &str)
virtual void	println (int n, ostream &fout)
	Print a full C++ class corresponding to a Faust dsp.
virtual void	printComputeMethod (int n, ostream &fout)
	Print Compute() method according to the various switch.
virtual void	printComputeMethodScalar (int n, ostream &fout)
virtual void	printComputeMethodVectorFaster (int n, ostream &fout)
	Uses loops of constant gVecSize boundary in order to provide the C compiler with more optimisation opportunities.
virtual void	printComputeMethodVectorSimple (int n, ostream &fout)
	Simple loop layout, generally less efficient than printComputeMethodVectorFaster.
virtual void	printComputeMethodOpenMP (int n, ostream &fout)
virtual void	printComputeMethodScheduler (int n, ostream &fout)
virtual void	printLoopGraphScalar (int n, ostream &fout)
	Print the loop graph (scalar mode)
virtual void	printLoopGraphVector (int n, ostream &fout)
	Print the loop graph (used for vector code)
virtual void	printLoopGraphOpenMP (int n, ostream &fout)
	Print the loop graph as a serie of parallel loops.
virtual void	printLoopGraphScheduler (int n, ostream &fout)
	Print the loop graph as a serie of parallel loops.
virtual void	printLoopGraphInternal (int n, ostream &fout)
	Print the loop graph (used for internals classes)
virtual void	printGraphDotFormat (ostream &fout)
	Print the loop graph in dot format.
virtual void	printLoopDeepFirst (int n, ostream &fout, Loop l, set< Loop > &visited)
	Print a loop graph deep first.
virtual void	printLastLoopLevelScheduler (int n, int lnum, const lset &L, ostream &fout)
	Print the 'level' of the loop graph as a set of parallel loops.
virtual void	printLoopLevelScheduler (int n, int lnum, const lset &L, ostream &fout)
	Print the 'level' of the loop graph as a set of parallel loops.
virtual void	printOneLoopScheduler (lset::const_iterator p, int n, ostream &fout)
virtual void	printLoopLevelOpenMP (int n, int lnum, const lset &L, ostream &fout)
	Print the 'level' of the loop graph as a set of parallel loops.
virtual void	printMetadata (int n, const map< Tree, set< Tree > > &S, ostream &fout)
	Print metadata declaration.
virtual void	printIncludeFile (ostream &fout)
	Print the required include files.
virtual void	printLibrary (ostream &fout)
	Print the required C++ libraries as comments in source code.
virtual void	printAdditionalCode (ostream &fout)
	Print additional functions required by the generated code.
int	inputs ()
int	outputs ()
Protected Attributes
string	fKlassName
string	fSuperKlassName
int	fNumInputs
int	fNumOutputs
int	fNumActives
	number of active controls in the UI (sliders, buttons, etc.)
int	fNumPassives
	number of passive widgets in the UI (bargraphs, etc.)
set< string >	fIncludeFileSet
set< string >	fLibrarySet
list< Klass * >	fSubClassList
list< string >	fDeclCode
list< string >	fStaticInitCode
	static init code for class constant tables
list< string >	fStaticFields
	static fields after class
list< string >	fInitCode
list< string >	fUICode
list< string >	fUIMacro
list< string >	fSharedDecl
	shared declarations
list< string >	fFirstPrivateDecl
	first private declarations
list< string >	fZone1Code
	shared vectors
list< string >	fZone2Code
	first private
list< string >	fZone2bCode
	single once per block
list< string >	fZone2cCode
	single once per block
list< string >	fZone3Code
	private every sub block
Loop *	fTopLoop
	active loops currently open
property< Loop * >	fLoopProperty
	loops used to compute some signals
bool	fVec
Static Protected Attributes
static bool	fNeedPowerDef = false
	true when faustpower definition is needed

Detailed Description

Definition at line 55 of file klass.hh.

Constructor & Destructor Documentation

Klass::Klass	(	const string &	name,
		const string &	super,
		int	numInputs,
		int	numOutputs,
		bool	__vec = `false`
	)		`[inline]`

Definition at line 108 of file klass.hh.

      :     fKlassName(name), fSuperKlassName(super), fNumInputs(numInputs), fNumOutputs(numOutputs),
            fNumActives(0), fNumPassives(0),
            fTopLoop(new Loop(0, "count")), fVec(__vec)
    {}

virtual Klass::~Klass ( ) [inline, virtual]

Definition at line 114 of file klass.hh.

{}

Member Function Documentation

void Klass::addDeclCode ( const string & str ) [inline]

Definition at line 140 of file klass.hh.

Referenced by buildTasksList(), SchedulerCompiler::dlineLoop(), VectorCompiler::dlineLoop(), and SchedulerCompiler::vectorLoop().

{ fDeclCode.push_back(str); }

Here is the caller graph for this function:

void Klass::addExecCode ( const string & str ) [inline]

Definition at line 166 of file klass.hh.

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), SchedulerCompiler::dlineLoop(), VectorCompiler::dlineLoop(), SchedulerCompiler::vectorLoop(), and VectorCompiler::vectorLoop().

{ fTopLoop->addExecCode(str); }

Here is the caller graph for this function:

void Klass::addFirstPrivateDecl ( const string & str ) [inline]

Definition at line 157 of file klass.hh.

Referenced by SchedulerCompiler::dlineLoop(), and VectorCompiler::dlineLoop().

{ fFirstPrivateDecl.push_back(str); }

Here is the caller graph for this function:

void Klass::addIncludeFile ( const string & str ) [inline]

Definition at line 128 of file klass.hh.

Referenced by Compiler::addIncludeFile().

{ fIncludeFileSet.insert(str); }

Here is the caller graph for this function:

void Klass::addInitCode ( const string & str ) [inline]

Definition at line 142 of file klass.hh.

Referenced by buildTasksList(), SchedulerCompiler::dlineLoop(), and VectorCompiler::dlineLoop().

{ fInitCode.push_back(str); }

Here is the caller graph for this function:

void Klass::addLibrary ( const string & str ) [inline]

Definition at line 130 of file klass.hh.

Referenced by Compiler::addLibrary().

{ fLibrarySet.insert(str); }

Here is the caller graph for this function:

void Klass::addPostCode ( const string & str ) [inline]

Definition at line 167 of file klass.hh.

Referenced by SchedulerCompiler::dlineLoop(), and VectorCompiler::dlineLoop().

{ fTopLoop->addPostCode(str); }

Here is the caller graph for this function:

void Klass::addPreCode ( const string & str ) [inline]

Definition at line 165 of file klass.hh.

Referenced by SchedulerCompiler::dlineLoop(), and VectorCompiler::dlineLoop().

{ fTopLoop->addPreCode(str); }

Here is the caller graph for this function:

void Klass::addSharedDecl ( const string & str ) [inline]

Definition at line 156 of file klass.hh.

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), SchedulerCompiler::dlineLoop(), VectorCompiler::dlineLoop(), SchedulerCompiler::vectorLoop(), and VectorCompiler::vectorLoop().

{ fSharedDecl.push_back(str); }

Here is the caller graph for this function:

void Klass::addStaticFields ( const string & str ) [inline]

Definition at line 146 of file klass.hh.

{ fStaticFields.push_back(str); }

void Klass::addStaticInitCode ( const string & str ) [inline]

Definition at line 144 of file klass.hh.

{ fStaticInitCode.push_back(str); }

void Klass::addSubKlass ( Klass * son ) [inline]

Definition at line 138 of file klass.hh.

{ fSubClassList.push_back(son); }

void Klass::addUICode ( const string & str ) [inline]

Definition at line 148 of file klass.hh.

Referenced by Compiler::generateUserInterfaceTree(), and Compiler::generateWidgetCode().

{ fUICode.push_back(str); }

Here is the caller graph for this function:

void Klass::addUIMacro ( const string & str ) [inline]

Definition at line 150 of file klass.hh.

Referenced by Compiler::generateWidgetMacro().

{ fUIMacro.push_back(str); }

Here is the caller graph for this function:

void Klass::addZone1 ( const string & str ) [inline]

Definition at line 159 of file klass.hh.

Referenced by VectorCompiler::dlineLoop(), and VectorCompiler::vectorLoop().

{ fZone1Code.push_back(str); }

Here is the caller graph for this function:

void Klass::addZone2 ( const string & str ) [inline]

Definition at line 160 of file klass.hh.

Referenced by SchedulerCompiler::dlineLoop(), and VectorCompiler::dlineLoop().

{ fZone2Code.push_back(str); }

Here is the caller graph for this function:

void Klass::addZone2b ( const string & str ) [inline]

Definition at line 161 of file klass.hh.

{ fZone2bCode.push_back(str); }

void Klass::addZone2c ( const string & str ) [inline]

Definition at line 162 of file klass.hh.

Referenced by buildTasksList().

{ fZone2cCode.push_back(str); }

Here is the caller graph for this function:

void Klass::addZone3 ( const string & str ) [inline]

Definition at line 163 of file klass.hh.

Referenced by buildTasksList(), SchedulerCompiler::compileMultiSignal(), and VectorCompiler::compileMultiSignal().

{ fZone3Code.push_back(str); }

Here is the caller graph for this function:

void Klass::buildTasksList ( )

Definition at line 323 of file klass.cpp.

References addDeclCode(), addInitCode(), addZone2c(), addZone3(), computeUseCount(), Loop::fIndex, fTopLoop, gGroupTaskSwitch, groupSeqLoops(), gTaskCount, LAST_TASK_INDEX, sortGraph(), START_TASK_INDEX, START_TASK_MAX, subst(), and T().

Referenced by SchedulerCompiler::compileMultiSignal().

{
    lgraph G;

    if (gGroupTaskSwitch) {
        computeUseCount(fTopLoop);
        groupSeqLoops(fTopLoop);
    }

    sortGraph(fTopLoop, G);
    int index_task = START_TASK_INDEX;

    addDeclCode("TaskGraph fGraph;");
    addDeclCode("FAUSTFLOAT** input;");
    addDeclCode("FAUSTFLOAT** output;");
    addDeclCode("volatile bool fIsFinished;");
    addDeclCode("int fFullCount;");
    addDeclCode("int fIndex;");
    addDeclCode("DSPThreadPool* fThreadPool;");
    addDeclCode("int fStaticNumThreads;");
    addDeclCode("int fDynamicNumThreads;");

    // Compute forward dependencies
    for (int l=G.size()-1; l>=0; l--) {
        for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
            for (lset::const_iterator p1 = (*p)->fBackwardLoopDependencies.begin(); p1!=(*p)->fBackwardLoopDependencies.end(); p1++) {
                (*p1)->fForwardLoopDependencies.insert((*p));
            }
            (*p)->fIndex = index_task;
            index_task++;
        }
    }

    // Compute ready tasks list
    vector<int> task_num;
    for (int l=G.size()-1; l>=0; l--) {
        lset::const_iterator next;
        for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
            if ((*p)->fBackwardLoopDependencies.size() == 0) {
                task_num.push_back((*p)->fIndex);
            }
        }
    }

    if (task_num.size() < START_TASK_MAX) {

        // Push ready tasks thread 0, execute one task directly

        addZone3("if (cur_thread == 0) {");

        Loop* keep = NULL;
        for (int l=G.size()-1; l>=0; l--) {
            lset::const_iterator next;
            for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
                if ((*p)->fBackwardLoopDependencies.size() == 0) {
                    if (keep == NULL) {
                        keep = *p;
                    } else {
                        addZone3(subst("    taskqueue.PushHead($0);", T((*p)->fIndex)));
                    }
                }
            }
        }

        if (keep != NULL) {
            addZone3(subst("    tasknum = $0;", T(keep->fIndex)));
        }

        addZone3("} else {");
        addZone3("    tasknum = TaskQueue::GetNextTask(cur_thread, fDynamicNumThreads);");
        addZone3("}");

    } else {

        // Cut ready tasks list and have each thread (dynamically) use a subpart
        addZone3(subst("int task_list_size = $0;", T((int)task_num.size())));
        stringstream buf;
        buf << "int task_list[" << task_num.size() << "] = {";
        for(size_t i = 0; i < task_num.size(); i++) {
            buf << task_num[i];
            if (i != (task_num.size() - 1))
                buf << ",";
        }
        buf << "};";

        addZone3(buf.str());
        addZone3("taskqueue.InitTaskList(task_list_size, task_list, fDynamicNumThreads, cur_thread, tasknum);");
    }

    // Last stage connected to end task
    if (G[0].size() > 1) {
        addZone2c("// Initialize end task, if more than one input");
        addZone2c(subst("fGraph.InitTask($0,$1);", T(LAST_TASK_INDEX), T((int)G[0].size())));
    } else {
        addZone2c("// End task has only one input, so will be directly activated");
    }

    // Compute init section
    addZone2c("// Only initialize taks with more than one input");
    for (int l=G.size()-1; l>=0; l--) {
        for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
            if ((*p)->fBackwardLoopDependencies.size() > 1)  { // Only initialize taks with more than 1 input, since taks with one input are "directly" activated.
                addZone2c(subst("fGraph.InitTask($0,$1);", T(START_TASK_INDEX + gTaskCount++), T((int)(*p)->fBackwardLoopDependencies.size())));
            } else {
                gTaskCount++;
            }
        }
    }

    addInitCode("fStaticNumThreads = get_max_cpu();");
    addInitCode("fDynamicNumThreads = getenv(\"OMP_NUM_THREADS\") ? atoi(getenv(\"OMP_NUM_THREADS\")) : fStaticNumThreads;");
    addInitCode("fThreadPool = DSPThreadPool::Init();");
    addInitCode("fThreadPool->StartAll(fStaticNumThreads - 1, false);");

    gTaskCount = 0;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::closeLoop ( Tree sig )

Close the top loop and either keep it or absorb it within its enclosing loop.

Definition at line 116 of file klass.cpp.

References Loop::absorb(), Loop::fBackwardLoopDependencies, Loop::fEnclosingLoop, Loop::fRecSymbolSet, fTopLoop, Loop::hasRecDependencyIn(), hd(), Loop::isEmpty(), isNil(), setLoopProperty(), symlist(), and tl().

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), VectorCompiler::generateCodeRecursions(), and VectorCompiler::generateLoopCode().

{
    assert(fTopLoop);
    Loop* l = fTopLoop;
    fTopLoop = l->fEnclosingLoop;
    assert(fTopLoop);

    //l->println(4, cerr);
    //cerr << endl;

    Tree S = symlist(sig);
    //cerr << "CLOSE LOOP :" << l << " with symbols " << *S  << endl;
    if (l->isEmpty() || fTopLoop->hasRecDependencyIn(S)) {
        //cout << " will absorb" << endl;
        // empty or dependent loop -> absorbed by enclosing one
        //cerr << "absorbed by : " << fTopLoop << endl;
        fTopLoop->absorb(l);
        //delete l; HACK !!!
    } else {
        // cout << " will NOT absorb" << endl;
        // we have an independent loop
        setLoopProperty(sig,l);     // associate the signal
        fTopLoop->fBackwardLoopDependencies.insert(l);
        // we need to indicate that all recursive symbols defined
        // in this loop are defined in this loop
        for (Tree lsym=l->fRecSymbolSet; !isNil(lsym); lsym=tl(lsym)) {
            this->setLoopProperty(hd(lsym), l);
            //cerr << "loop " << l << " defines " << *hd(lsym) << endl;
        }
    }
    //cerr << "\n" << endl;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::collectIncludeFile ( set< string > & S )

Definition at line 1201 of file klass.cpp.

References fIncludeFileSet, fSubClassList, and merge().

Referenced by printIncludeFile().

{
    list<Klass* >::iterator     k;

    for (k = fSubClassList.begin(); k != fSubClassList.end(); k++)  (*k)->collectIncludeFile(S);
    merge(S, fIncludeFileSet);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::collectLibrary ( set< string > & S )

Definition at line 1209 of file klass.cpp.

References fLibrarySet, fSubClassList, and merge().

Referenced by printLibrary().

{
    list<Klass* >::iterator     k;

    for (k = fSubClassList.begin(); k != fSubClassList.end(); k++)  (*k)->collectLibrary(S);
    merge(S, fLibrarySet);
}

Here is the call graph for this function:

Here is the caller graph for this function:

const string& Klass::getClassName ( ) const [inline]

Returns the name of the class.

Definition at line 122 of file klass.hh.

bool Klass::getLoopProperty	(	Tree	sig,
		Loop *&	l
	)

Returns the loop used to compute a signal.

Definition at line 86 of file klass.cpp.

References fLoopProperty, and property< P >::get().

Referenced by VectorCompiler::CS().

{
    return  fLoopProperty.get(sig, l);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::incUIActiveCount ( ) [inline]

Definition at line 152 of file klass.hh.

Referenced by Compiler::generateWidgetCode().

{ fNumActives++; }

Here is the caller graph for this function:

void Klass::incUIPassiveCount ( ) [inline]

Definition at line 153 of file klass.hh.

Referenced by Compiler::generateWidgetCode().

{ fNumPassives++; }

Here is the caller graph for this function:

int Klass::inputs ( ) [inline]

Definition at line 200 of file klass.hh.

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), and main().

{ return fNumInputs; }

Here is the caller graph for this function:

void Klass::openLoop ( const string & size )

Open a non-recursive loop on top of the stack of open loops.

Parameters:

size	the number of iterations of the loop

Definition at line 95 of file klass.cpp.

References fTopLoop.

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), VectorCompiler::generateCodeRecursions(), and VectorCompiler::generateLoopCode().

{
    fTopLoop = new Loop(fTopLoop, size);
    //cerr << "\nOPEN SHARED LOOP(" << size << ") ----> " << fTopLoop << endl;
}

Here is the caller graph for this function:

void Klass::openLoop	(	Tree	recsymbol,
		const string &	size
	)

Open a recursive loop on top of the stack of open loops.

Parameters:

recsymbol	the recursive symbol defined in this loop
size	the number of iterations of the loop

Definition at line 106 of file klass.cpp.

References fTopLoop.

{
    fTopLoop = new Loop(recsymbol, fTopLoop, size);
    //cerr << "\nOPEN REC LOOP(" << *recsymbol << ", " << size << ") ----> " << fTopLoop << endl;
}

int Klass::outputs ( ) [inline]

Definition at line 201 of file klass.hh.

Referenced by SchedulerCompiler::compileMultiSignal(), VectorCompiler::compileMultiSignal(), and main().

{ return fNumOutputs; }

Here is the caller graph for this function:

void Klass::printAdditionalCode ( ostream & fout ) [virtual]

Print additional functions required by the generated code.

Definition at line 216 of file klass.cpp.

References fNeedPowerDef.

Referenced by main().

{
    if (fNeedPowerDef) {
        // Add faustpower definition to C++ code
        fout << "#include <cmath>" << endl;
        fout << "template <int N> inline float faustpower(float x)      { return powf(x,N); } " << endl;
        fout << "template <int N> inline double faustpower(double x)    { return pow(x,N); }"  << endl;
        fout << "template <int N> inline int faustpower(int x)          { return faustpower<N/2>(x) * faustpower<N-N/2>(x); } " << endl;
        fout << "template <>     inline int faustpower<0>(int x)        { return 1; }" << endl;
        fout << "template <>     inline int faustpower<1>(int x)        { return x; }" << endl;
    }

}

Here is the caller graph for this function:

void Klass::printComputeMethod	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print Compute() method according to the various switch.

Definition at line 813 of file klass.cpp.

References gOpenMPSwitch, gSchedulerSwitch, gVectorLoopVariant, gVectorSwitch, printComputeMethodOpenMP(), printComputeMethodScalar(), printComputeMethodScheduler(), printComputeMethodVectorFaster(), and printComputeMethodVectorSimple().

Referenced by println().

{
    if (gSchedulerSwitch) {
        printComputeMethodScheduler (n, fout);
    } else if (gOpenMPSwitch) {
        printComputeMethodOpenMP (n, fout);
    } else if (gVectorSwitch) {
        switch (gVectorLoopVariant) {
            case 0 : printComputeMethodVectorFaster(n, fout); break;
            case 1 : printComputeMethodVectorSimple(n, fout); break;
            default : cerr << "unknown loop variant " << gVectorLoopVariant << endl; exit(1);
        }
   } else {
        printComputeMethodScalar(n, fout);
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printComputeMethodOpenMP	(	int	n,
		ostream &	fout
	)		`[virtual]`

Definition at line 946 of file klass.cpp.

References fFirstPrivateDecl, fZone1Code, fZone2bCode, fZone2Code, fZone3Code, gVecSize, printdecllist(), printlines(), printLoopGraphOpenMP(), subst(), tab(), and xfloat().

Referenced by printComputeMethod().

{
    // in openMP mode we need to split loops in smaller pieces not larger
    // than gVecSize and add OpenMP pragmas
    tab(n+1,fout); fout << subst("virtual void compute (int fullcount, $0** input, $0** output) {", xfloat());
        printlines(n+2, fZone1Code, fout);
        printlines(n+2, fZone2Code, fout);
        tab(n+2,fout); fout << "#pragma omp parallel";
        printdecllist(n+3, "firstprivate", fFirstPrivateDecl, fout);

        tab(n+2,fout); fout << "{";
            if (!fZone2bCode.empty()) {
                tab(n+3,fout); fout << "#pragma omp single";
                tab(n+3,fout); fout << "{";
                    printlines(n+4, fZone2bCode, fout);
                tab(n+3,fout); fout << "}";
            }

            tab(n+3,fout); fout << "for (int index = 0; index < fullcount; index += " << gVecSize << ") {";
            tab(n+4,fout); fout << "int count = min ("<< gVecSize << ", fullcount-index);";

            printlines (n+4, fZone3Code, fout);
            printLoopGraphOpenMP (n+4,fout);

            tab(n+3,fout); fout << "}";

        tab(n+2,fout); fout << "}";
    tab(n+1,fout); fout << "}";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printComputeMethodScalar	(	int	n,
		ostream &	fout
	)		`[virtual]`

Definition at line 830 of file klass.cpp.

References fZone1Code, fZone2bCode, fZone2Code, fZone3Code, printlines(), printLoopGraphScalar(), subst(), tab(), and xfloat().

Referenced by printComputeMethod().

{
    tab(n+1,fout); fout << subst("virtual void compute (int count, $0** input, $0** output) {", xfloat());
        printlines (n+2, fZone1Code, fout);
        printlines (n+2, fZone2Code, fout);
        printlines (n+2, fZone2bCode, fout);
        printlines (n+2, fZone3Code, fout);
        printLoopGraphScalar (n+2,fout);
    tab(n+1,fout); fout << "}";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printComputeMethodScheduler	(	int	n,
		ostream &	fout
	)		`[virtual]`

Definition at line 1045 of file klass.cpp.

References fZone1Code, fZone2cCode, fZone2Code, fZone3Code, gTaskCount, gVecSize, printlines(), printLoopGraphScheduler(), START_TASK_INDEX, subst(), tab(), and xfloat().

Referenced by printComputeMethod().

{
    tab(n+1,fout); fout << "void display() {";
        tab(n+2,fout); fout << "fGraph.Display();";
    tab(n+1,fout); fout << "}";

    tab(n+1,fout); fout << subst("virtual void compute (int fullcount, $0** input, $0** output) {", xfloat());

        tab(n+2,fout); fout << "GetRealTime();";

        tab(n+2,fout); fout << "this->input = input;";
        tab(n+2,fout); fout << "this->output = output;";

        tab(n+2,fout); fout << "StartMeasure();";

        tab(n+2,fout); fout << "for (fIndex = 0; fIndex < fullcount; fIndex += " << gVecSize << ") {";

        tab(n+3,fout); fout << "fFullCount = min ("<< gVecSize << ", fullcount-fIndex);";
        tab(n+3,fout); fout << "TaskQueue::Init();";
        printlines (n+3, fZone2cCode, fout);

        tab(n+3,fout); fout << "fIsFinished = false;";
        tab(n+3,fout); fout << "fThreadPool->SignalAll(fDynamicNumThreads - 1, this);";
        tab(n+3,fout); fout << "computeThread(0);";
        tab(n+3,fout); fout << "while (!fThreadPool->IsFinished()) {}";

        tab(n+2,fout); fout << "}";

        tab(n+2,fout); fout << "StopMeasure(fStaticNumThreads, fDynamicNumThreads);";

    tab(n+1,fout); fout << "}";

    tab(n+1,fout); fout << "void computeThread(int cur_thread) {";
        printlines (n+2, fZone1Code, fout);
        printlines (n+2, fZone2Code, fout);

        tab(n+2,fout); fout << "// Init graph state";

        tab(n+2,fout); fout << "{";
            tab(n+3,fout); fout << "TaskQueue taskqueue(cur_thread);";
            tab(n+3,fout); fout << "int tasknum = -1;";
            tab(n+3,fout); fout << "int count = fFullCount;";

            // Init input and output
            tab(n+3,fout); fout << "// Init input and output";
            printlines (n+3, fZone3Code, fout);

            tab(n+3,fout); fout << "while (!fIsFinished) {";
                 tab(n+4,fout); fout << "switch (tasknum) {";

                    // Work stealing task
                    tab(n+5, fout); fout << "case WORK_STEALING_INDEX: { ";
                        tab(n+6, fout); fout << "tasknum = TaskQueue::GetNextTask(cur_thread, fDynamicNumThreads);";
                        tab(n+6, fout); fout << "break;";
                    tab(n+5, fout); fout << "} ";

                    // End task
                    tab(n+5, fout); fout << "case LAST_TASK_INDEX: { ";
                        tab(n+6, fout); fout << "fIsFinished = true;";
                        tab(n+6, fout); fout << "break;";
                    tab(n+5, fout); fout << "} ";

                    gTaskCount = START_TASK_INDEX;

                    // DSP tasks
                    printLoopGraphScheduler (n+5,fout);

                 tab(n+4,fout); fout << "}";
            tab(n+3,fout); fout << "}";
        tab(n+2,fout); fout << "}";
    tab(n+1,fout); fout << "}";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printComputeMethodVectorFaster	(	int	n,
		ostream &	fout
	)		`[virtual]`

Uses loops of constant gVecSize boundary in order to provide the C compiler with more optimisation opportunities.

Improves performances in general, but not always

Definition at line 846 of file klass.cpp.

References fZone1Code, fZone2bCode, fZone2Code, fZone3Code, gVecSize, printlines(), printLoopGraphVector(), subst(), tab(), and xfloat().

Referenced by printComputeMethod().

{
    // in vector mode we need to split loops in smaller pieces not larger
    // than gVecSize
    tab(n+1,fout); fout << subst("virtual void compute (int fullcount, $0** input, $0** output) {", xfloat());
        printlines(n+2, fZone1Code, fout);
        printlines(n+2, fZone2Code, fout);
        printlines(n+2, fZone2bCode, fout);

        tab(n+2,fout); fout << "int index;";
        tab(n+2,fout); fout << "for (index = 0; index <= fullcount - " << gVecSize << "; index += " << gVecSize << ") {";
            tab(n+3,fout); fout << "// compute by blocks of " << gVecSize << " samples";
            tab(n+3,fout); fout << "const int count = " << gVecSize << ";";
            printlines (n+3, fZone3Code, fout);
            printLoopGraphVector(n+3,fout);
        tab(n+2,fout); fout << "}";

        tab(n+2,fout); fout << "if (index < fullcount) {";
            tab(n+3,fout); fout << "// compute the remaining samples if any";
            tab(n+3,fout); fout << "int count = fullcount-index;";
            printlines (n+3, fZone3Code, fout);
             printLoopGraphVector(n+3,fout);
        tab(n+2,fout); fout << "}";
    tab(n+1,fout); fout << "}";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printComputeMethodVectorSimple	(	int	n,
		ostream &	fout
	)		`[virtual]`

Simple loop layout, generally less efficient than printComputeMethodVectorFaster.

Definition at line 875 of file klass.cpp.

References fZone1Code, fZone2bCode, fZone2Code, fZone3Code, gVecSize, printlines(), printLoopGraphVector(), subst(), tab(), and xfloat().

Referenced by printComputeMethod().

{
    // in vector mode we need to split loops in smaller pieces not larger
    // than gVecSize
    tab(n+1,fout); fout << subst("virtual void compute (int fullcount, $0** input, $0** output) {", xfloat());
        printlines(n+2, fZone1Code, fout);
        printlines(n+2, fZone2Code, fout);
        printlines(n+2, fZone2bCode, fout);
        tab(n+2,fout); fout << "for (int index = 0; index < fullcount; index += " << gVecSize << ") {";
            tab(n+3,fout); fout << "int count = min("<< gVecSize << ", fullcount-index);";
            printlines (n+3, fZone3Code, fout);
            printLoopGraphVector(n+3,fout);
        tab(n+2,fout); fout << "}";
    tab(n+1,fout); fout << "}";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printGraphDotFormat ( ostream & fout ) [virtual]

Print the loop graph in dot format.

Definition at line 517 of file klass.cpp.

References fTopLoop, and sortGraph().

Referenced by main().

{
    lgraph G;
    sortGraph(fTopLoop, G);

    fout << "strict digraph loopgraph {" << endl;
    fout << '\t' << "rankdir=LR;" << endl;
    fout << '\t' << "node[color=blue, fillcolor=lightblue, style=filled, fontsize=9];" << endl;

    int lnum = 0;       // used for loop numbers
    // for each level of the graph
    for (int l=G.size()-1; l>=0; l--) {
        // for each task in the level
        for (lset::const_iterator t =G[l].begin(); t!=G[l].end(); t++) {
            // print task label "Lxxx : 0xffffff"
            fout << '\t' << 'L'<<(*t)<<"[label=<<font face=\"verdana,bold\">L"<<lnum++<<"</font> : "<<(*t)<<">];"<<endl;
            // for each source of the task
            for (lset::const_iterator src = (*t)->fBackwardLoopDependencies.begin(); src!=(*t)->fBackwardLoopDependencies.end(); src++) {
                // print the connection Lxxx -> Lyyy;
                fout << '\t' << 'L'<<(*src)<<"->"<<'L'<<(*t)<<';'<<endl;
            }
        }
    }
    fout << "}" << endl;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printIncludeFile ( ostream & fout ) [virtual]

Print the required include files.

Definition at line 198 of file klass.cpp.

References collectIncludeFile(), and gOpenMPSwitch.

Referenced by main().

{
    set<string> S;
    set<string>::iterator f;

    if (gOpenMPSwitch) {
        fout << "#include <omp.h>" << "\n";
    }

    collectIncludeFile(S);
    for (f = S.begin(); f != S.end(); f++)  {
        fout << "#include " << *f << "\n";
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLastLoopLevelScheduler	(	int	n,
		int	lnum,
		const lset &	L,
		ostream &	fout
	)		`[virtual]`

Print the 'level' of the loop graph as a set of parallel loops.

Definition at line 623 of file klass.cpp.

References gTaskCount, nonRecursiveLevel(), and tab().

Referenced by printLoopGraphScheduler().

{
    if (nonRecursiveLevel(L) && L.size() == 1 && !(*L.begin())->isEmpty()) {

        lset::const_iterator p =L.begin();
        tab(n, fout); fout << "case " << gTaskCount++ << ": { ";
        (*p)->println(n+1, fout);
        tab(n+1, fout); fout << "tasknum = LAST_TASK_INDEX;";
        tab(n+1, fout); fout << "break;";
        tab(n, fout); fout << "} ";

    } else if (L.size() > 1) {

        for (lset::const_iterator p =L.begin(); p!=L.end(); p++) {
            tab(n, fout); fout << "case " << gTaskCount++ << ": { ";
            (*p)->println(n+1, fout);
            tab(n+1, fout); fout << "fGraph.ActivateOneOutputTask(taskqueue, LAST_TASK_INDEX, tasknum);";
            tab(n+1, fout); fout << "break;";
            tab(n, fout); fout << "} ";
        }

    } else if (L.size() == 1 && !(*L.begin())->isEmpty()) {

        lset::const_iterator p =L.begin();
        tab(n, fout); fout << "case " << gTaskCount++ << ": { ";
        (*p)->println(n+1, fout);
        tab(n+1, fout); fout << "tasknum = LAST_TASK_INDEX;";
        tab(n+1, fout); fout << "break;";
        tab(n, fout); fout << "} ";

    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLibrary ( ostream & fout ) [virtual]

Print the required C++ libraries as comments in source code.

Definition at line 181 of file klass.cpp.

References collectLibrary().

Referenced by main().

{
    set<string> S;
    set<string>::iterator f;

    string sep;
    collectLibrary(S);
    fout << "/* link with ";
    for (f = S.begin(), sep =": "; f != S.end(); f++, sep = ", ")   {
        fout << sep << *f;
    }
    fout << " */\n";
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::println	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print a full C++ class corresponding to a Faust dsp.

Reimplemented in SigFloatGenKlass, and SigIntGenKlass.

Definition at line 732 of file klass.cpp.

References fDeclCode, fInitCode, fKlassName, fNumActives, fNumInputs, fNumOutputs, fNumPassives, fStaticFields, fStaticInitCode, fSubClassList, fSuperKlassName, fUICode, fUIMacro, gMetaDataSet, gSchedulerSwitch, gUIMacroSwitch, printComputeMethod(), printlines(), printMetadata(), and tab().

Referenced by main().

{
    list<Klass* >::iterator k;

    tab(n,fout); fout << "#define FAUSTCLASS "<< fKlassName << endl;

    if (gSchedulerSwitch) {
        tab(n,fout); fout << "class " << fKlassName << " : public " << fSuperKlassName << ", public Runnable {";
    } else {
        tab(n,fout); fout << "class " << fKlassName << " : public " << fSuperKlassName << " {";
    }

    if (gUIMacroSwitch) {
        tab(n,fout); fout << "  public:";
    } else {
        tab(n,fout); fout << "  private:";
    }

    for (k = fSubClassList.begin(); k != fSubClassList.end(); k++)  (*k)->println(n+1, fout);

    printlines(n+1, fDeclCode, fout);

    tab(n,fout); fout << "  public:";

    printMetadata(n+1, gMetaDataSet, fout);

    if (gSchedulerSwitch) {
        tab(n+1,fout); fout << "virtual ~" << fKlassName << "() \t{ "
                            << "DSPThreadPool::Destroy()"
                            << "; }";
    }

    tab(n+1,fout); fout     << "virtual int getNumInputs() \t{ "
                    << "return " << fNumInputs
                    << "; }";
    tab(n+1,fout); fout     << "virtual int getNumOutputs() \t{ "
                    << "return " << fNumOutputs
                    << "; }";

    tab(n+1,fout); fout << "static void classInit(int samplingFreq) {";
        printlines (n+2, fStaticInitCode, fout);
    tab(n+1,fout); fout << "}";

    tab(n+1,fout); fout << "virtual void instanceInit(int samplingFreq) {";
        tab(n+2,fout); fout << "fSamplingFreq = samplingFreq;";
        printlines (n+2, fInitCode, fout);
    tab(n+1,fout); fout << "}";

    tab(n+1,fout); fout << "virtual void init(int samplingFreq) {";
        tab(n+2,fout); fout << "classInit(samplingFreq);";
         tab(n+2,fout); fout << "instanceInit(samplingFreq);";
    tab(n+1,fout); fout << "}";


    tab(n+1,fout); fout << "virtual void buildUserInterface(UI* interface) {";
        printlines (n+2, fUICode, fout);
    tab(n+1,fout); fout << "}";

    printComputeMethod(n, fout);

    tab(n,fout); fout << "};\n" << endl;

    printlines(n, fStaticFields, fout);

    // generate user interface macros if needed
    if (gUIMacroSwitch) {
        tab(n, fout); fout << "#ifdef FAUST_UIMACROS";
            tab(n+1,fout); fout << "#define FAUST_INPUTS " << fNumInputs;
            tab(n+1,fout); fout << "#define FAUST_OUTPUTS " << fNumOutputs;
            tab(n+1,fout); fout << "#define FAUST_ACTIVES " << fNumActives;
            tab(n+1,fout); fout << "#define FAUST_PASSIVES " << fNumPassives;
            printlines(n+1, fUIMacro, fout);
        tab(n, fout); fout << "#endif";
    }

    fout << endl;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopDeepFirst	(	int	n,
		ostream &	fout,
		Loop *	l,
		set< Loop * > &	visited
	)		`[virtual]`

Print a loop graph deep first.

Definition at line 262 of file klass.cpp.

References Loop::fBackwardLoopDependencies, Loop::fOrder, isElement(), Loop::println(), and tab().

Referenced by printLoopGraphVector().

{
    // avoid printing already printed loops
    if (isElement(visited, l)) return;

    // remember we have printed this loop
    visited.insert(l);

    // print the dependencies loops (that need to be computed before this one)
    for (lset::const_iterator p =l->fBackwardLoopDependencies.begin(); p!=l->fBackwardLoopDependencies.end(); p++) {
        printLoopDeepFirst(n, fout, *p, visited);
    }
    // the print the loop itself
    tab(n, fout);
    tab(n, fout); fout << "// LOOP " << l << ", ORDER " << l->fOrder << endl;
    l->println(n+1, fout);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopGraphInternal	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print the loop graph (used for internals classes)

Definition at line 546 of file klass.cpp.

References fTopLoop, gVectorSwitch, sortGraph(), and tab().

Referenced by SigIntGenKlass::println(), and SigFloatGenKlass::println().

{
    lgraph G;
    sortGraph(fTopLoop, G);

    // normal mode
    for (int l=G.size()-1; l>=0; l--) {
        if (gVectorSwitch) { tab(n, fout); fout << "// SECTION : " << G.size() - l; }
        for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
            (*p)->printoneln(n, fout);
        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopGraphOpenMP	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print the loop graph as a serie of parallel loops.

Definition at line 474 of file klass.cpp.

References computeUseCount(), fTopLoop, gGroupTaskSwitch, groupSeqLoops(), printLoopLevelOpenMP(), sortGraph(), and tab().

Referenced by printComputeMethodOpenMP().

{
    if (gGroupTaskSwitch) {
        computeUseCount(fTopLoop);
        groupSeqLoops(fTopLoop);
    }

    lgraph G;
    sortGraph(fTopLoop, G);

    // OpenMP mode : add OpenMP directives
    for (int l=G.size()-1; l>=0; l--) {
        tab(n, fout); fout << "// SECTION : " << G.size() - l;
        printLoopLevelOpenMP(n, G.size() - l, G[l], fout);
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopGraphScalar	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print the loop graph (scalar mode)

Definition at line 563 of file klass.cpp.

References fTopLoop, and Loop::printoneln().

Referenced by printComputeMethodScalar().

{
    fTopLoop->printoneln(n, fout);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopGraphScheduler	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print the loop graph as a serie of parallel loops.

Definition at line 494 of file klass.cpp.

References computeUseCount(), fTopLoop, gGroupTaskSwitch, groupSeqLoops(), printLastLoopLevelScheduler(), printLoopLevelScheduler(), sortGraph(), and tab().

Referenced by printComputeMethodScheduler().

{
    if (gGroupTaskSwitch) {
        computeUseCount(fTopLoop);
        groupSeqLoops(fTopLoop);
    }

    lgraph G;
    sortGraph(fTopLoop, G);

    // OpenMP mode : add OpenMP directives
    for (int l=G.size()-1; l>0; l--) {
        tab(n, fout); fout << "// SECTION : " << G.size() - l;
        printLoopLevelScheduler(n, G.size() - l, G[l], fout);
    }

    printLastLoopLevelScheduler(n, G.size(), G[0], fout);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopGraphVector	(	int	n,
		ostream &	fout
	)		`[virtual]`

Print the loop graph (used for vector code)

Definition at line 443 of file klass.cpp.

References computeUseCount(), fTopLoop, gDeepFirstSwitch, gGroupTaskSwitch, groupSeqLoops(), gVectorSwitch, printLoopDeepFirst(), sortGraph(), and tab().

Referenced by printComputeMethodVectorFaster(), and printComputeMethodVectorSimple().

{
    if (gGroupTaskSwitch) {
        computeUseCount(fTopLoop);
        groupSeqLoops(fTopLoop);
    }

    lgraph G;
    sortGraph(fTopLoop, G);

#if 1
    // EXPERIMENTAL
    if (gVectorSwitch && gDeepFirstSwitch) {
        set<Loop*> visited;
        printLoopDeepFirst(n, fout, fTopLoop, visited);
        return;
    }
#endif

    // normal mode
    for (int l=G.size()-1; l>=0; l--) {
        if (gVectorSwitch) { tab(n, fout); fout << "// SECTION : " << G.size() - l; }
        for (lset::const_iterator p =G[l].begin(); p!=G[l].end(); p++) {
            (*p)->println(n, fout);
        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopLevelOpenMP	(	int	n,
		int	lnum,
		const lset &	L,
		ostream &	fout
	)		`[virtual]`

Print the 'level' of the loop graph as a set of parallel loops.

Definition at line 583 of file klass.cpp.

References gOpenMPLoop, nonRecursiveLevel(), and tab().

Referenced by printLoopGraphOpenMP().

{
    if (nonRecursiveLevel(L) && L.size()==1) {
        for (lset::const_iterator p =L.begin(); p!=L.end(); p++) {
            if ((*p)->isEmpty() == false) {
                if (gOpenMPLoop) {
                    (*p)->printParLoopln(n, fout);
                } else {
                    tab(n, fout); fout << "#pragma omp single ";
                    tab(n, fout); fout << "{ ";
                    (*p)->println(n+1, fout);
                    tab(n, fout); fout << "} ";
                }
            }
        }

    } else if (L.size() > 1) {
        tab(n, fout); fout << "#pragma omp sections ";
        tab(n, fout); fout << "{ ";
        for (lset::const_iterator p =L.begin(); p!=L.end(); p++) {
            tab(n+1, fout); fout << "#pragma omp section ";
            tab(n+1, fout); fout << "{";
            (*p)->println(n+2, fout);
            tab(n+1, fout); fout << "} ";
        }
        tab(n, fout); fout << "} ";
    } else if (L.size() == 1 && !(*L.begin())->isEmpty()) {
        tab(n, fout); fout << "#pragma omp single ";
        tab(n, fout); fout << "{ ";
            for (lset::const_iterator p =L.begin(); p!=L.end(); p++) {
                (*p)->println(n+1, fout);
            }
        tab(n, fout); fout << "} ";
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printLoopLevelScheduler	(	int	n,
		int	lnum,
		const lset &	L,
		ostream &	fout
	)		`[virtual]`

Print the 'level' of the loop graph as a set of parallel loops.

Definition at line 716 of file klass.cpp.

References nonRecursiveLevel(), and printOneLoopScheduler().

Referenced by printLoopGraphScheduler().

{
    if (nonRecursiveLevel(L) && L.size() == 1 && !(*L.begin())->isEmpty()) {
        printOneLoopScheduler(L.begin(), n, fout);
    } else if (L.size() > 1) {
        for (lset::const_iterator p = L.begin(); p != L.end(); p++) {
            printOneLoopScheduler(p, n, fout);
        }
    } else if (L.size() == 1 && !(*L.begin())->isEmpty()) {
        printOneLoopScheduler(L.begin(), n, fout);
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printMetadata	(	int	n,
		const map< Tree, set< Tree > > &	S,
		ostream &	fout
	)		`[virtual]`

Print metadata declaration.

Definition at line 233 of file klass.cpp.

References gMetaDataSet, tab(), and tree().

Referenced by println().

{
    tab(n,fout); fout   << "static void metadata(Meta* m) \t{ ";

    for (map<Tree, set<Tree> >::iterator i = gMetaDataSet.begin(); i != gMetaDataSet.end(); i++) {
        if (i->first != tree("author")) {
            tab(n+1,fout); fout << "m->declare(\"" << *(i->first) << "\", " << **(i->second.begin()) << ");";
        } else {
            for (set<Tree>::iterator j = i->second.begin(); j != i->second.end(); j++) {
                if (j == i->second.begin()) {
                     tab(n+1,fout); fout << "m->declare(\"" << *(i->first) << "\", " << **j << ");" ;
                } else {
                     tab(n+1,fout); fout << "m->declare(\"" << "contributor" << "\", " << **j << ");";
                }
            }
        }
    }

    tab(n,fout); fout << "}" << endl;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void Klass::printOneLoopScheduler	(	lset::const_iterator	p,
		int	n,
		ostream &	fout
	)		`[virtual]`

Definition at line 656 of file klass.cpp.

References Loop::fIndex, gTaskCount, subst(), T(), and tab().

Referenced by printLoopLevelScheduler().

{
    tab(n, fout); fout << "case " << gTaskCount++ << ": { ";
    (*p)->println(n+1, fout);

    // One output only
    if ((*p)->fForwardLoopDependencies.size() == 1) {

        lset::const_iterator p1 = (*p)->fForwardLoopDependencies.begin();
        if ((*p1)->fBackwardLoopDependencies.size () == 1) {
            tab(n+1, fout); fout << subst("tasknum = $0;", T((*p1)->fIndex));
        } else {
            tab(n+1, fout); fout << subst("fGraph.ActivateOneOutputTask(taskqueue, $0, tasknum);", T((*p1)->fIndex));
        }

    } else {

        Loop* keep = NULL;
        // Find one output with only one backward dependencies
        for (lset::const_iterator p1 = (*p)->fForwardLoopDependencies.begin(); p1!=(*p)->fForwardLoopDependencies.end(); p1++) {
            if ((*p1)->fBackwardLoopDependencies.size () == 1) {
                keep = *p1;
                break;
            }
        }

        if (keep == NULL) {
            tab(n+1, fout); fout << "tasknum = WORK_STEALING_INDEX;";
        }

        for (lset::const_iterator p1 = (*p)->fForwardLoopDependencies.begin(); p1!=(*p)->fForwardLoopDependencies.end(); p1++) {
            if ((*p1)->fBackwardLoopDependencies.size () == 1) {  // Task is the only input
                if (*p1 != keep) {
                    tab(n+1, fout); fout << subst("taskqueue.PushHead($0);", T((*p1)->fIndex));
                }
            } else {
                if (keep == NULL) {
                    tab(n+1, fout); fout << subst("fGraph.ActivateOutputTask(taskqueue, $0, tasknum);", T((*p1)->fIndex));
                } else {
                    tab(n+1, fout); fout << subst("fGraph.ActivateOutputTask(taskqueue, $0);", T((*p1)->fIndex));
                }
            }
        }

        if (keep != NULL) {
            tab(n+1, fout); fout << subst("tasknum = $0;", T(keep->fIndex)); // Last one
        } else {
            tab(n+1, fout); fout << "fGraph.GetReadyTask(taskqueue, tasknum);"; // Last one
        }
    }

    tab(n+1, fout); fout << "break;";
    tab(n, fout); fout << "} ";
}