Preparing new implementation of SOP/DSD balancing in 'if' mapper.

src/aig/gia/giaIf.c

@@ -691,7 +691,7 @@ Gia_Man_t * Gia_ManFromIfAig( If_Man_t * pIfMan )
     Vec_Int_t * vCover;
     int i, k;
     assert( pIfMan->pPars->pLutStruct == NULL );
-    assert( pIfMan->pPars->fDelayOpt || pIfMan->pPars->fUserRecLib );
+    assert( pIfMan->pPars->fDelayOpt || pIfMan->pPars->fDsdBalance || pIfMan->pPars->fUserRecLib );
     // create new manager
     pNew = Gia_ManStart( If_ManObjNum(pIfMan) );
     Gia_ManHashAlloc( pNew );
@@ -713,6 +713,8 @@ Gia_Man_t * Gia_ManFromIfAig( If_Man_t * pIfMan )
             // get the functionality
             if ( pIfMan->pPars->fDelayOpt )
                 pIfObj->iCopy = Gia_ManNodeIfSopToGia( pNew, pIfMan, pCutBest, vLeaves, fHash );
+            else if ( pIfMan->pPars->fDsdBalance )
+                pIfObj->iCopy = If_DsdCutBalance( pNew, pIfMan, pCutBest, vLeaves, fHash );
             else if ( pIfMan->pPars->fUserRecLib )
                 pIfObj->iCopy = Abc_RecToGia3( pNew, pIfMan, pCutBest, vLeaves, fHash );
             else assert( 0 );
@@ -1292,7 +1294,7 @@ Gia_Man_t * Gia_ManFromIfLogic( If_Man_t * pIfMan )
         {
             pCutBest = If_ObjCutBest( pIfObj );
             // perform sorting of cut leaves by delay, so that the slowest pin drives the fastest input of the LUT
-            if ( !pIfMan->pPars->fUseTtPerm && !pIfMan->pPars->fDelayOpt && !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserRecLib && !pIfMan->pPars->nGateSize && !pIfMan->pPars->fEnableCheck75 && !pIfMan->pPars->fEnableCheck75u && !pIfMan->pPars->fEnableCheck07 )
+            if ( !pIfMan->pPars->fUseTtPerm && !pIfMan->pPars->fDelayOpt && !pIfMan->pPars->fDsdBalance && !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserRecLib && !pIfMan->pPars->nGateSize && !pIfMan->pPars->fEnableCheck75 && !pIfMan->pPars->fEnableCheck75u && !pIfMan->pPars->fEnableCheck07 )
                 If_CutRotatePins( pIfMan, pCutBest );
             // collect leaves of the best cut
             Vec_IntClear( vLeaves );
@@ -1534,7 +1536,7 @@ Gia_Man_t * Gia_ManPerformMapping( Gia_Man_t * p, void * pp, int fNormalized )
     If_Man_t * pIfMan;
     If_Par_t * pPars = (If_Par_t *)pp;
     // disable cut minimization when GIA strucure is needed
-    if ( !pPars->fDelayOpt && !pPars->fUserRecLib && !pPars->fDeriveLuts && !pPars->fUseDsd && !pPars->fUseTtPerm )
+    if ( !pPars->fDelayOpt && !pPars->fDsdBalance && !pPars->fUserRecLib && !pPars->fDeriveLuts && !pPars->fUseDsd && !pPars->fUseTtPerm )
         pPars->fCutMin = 0;
 
     // reconstruct GIA according to the hierarchy manager
@@ -1593,7 +1595,7 @@ Gia_Man_t * Gia_ManPerformMapping( Gia_Man_t * p, void * pp, int fNormalized )
         return NULL;
     }
     // transform the result of mapping into the new network
-    if ( pIfMan->pPars->fDelayOpt || pIfMan->pPars->fUserRecLib )
+    if ( pIfMan->pPars->fDelayOpt || pIfMan->pPars->fDsdBalance || pIfMan->pPars->fUserRecLib )
         pNew = Gia_ManFromIfAig( pIfMan );
     else
         pNew = Gia_ManFromIfLogic( pIfMan );

src/base/abci/abc.c

@@ -14762,7 +14762,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
 
     fLutMux = 0;
     Extra_UtilGetoptReset();
-    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGNDEWSTqaflepmrsdbugyojikncvh" ) ) != EOF )
+    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGNDEWSTqaflepmrsdbugxyojikncvh" ) ) != EOF )
     {
         switch ( c )
         {
@@ -14931,6 +14931,9 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         case 'g':
             pPars->fDelayOpt ^= 1;
             break;
+        case 'x':
+            pPars->fDsdBalance ^= 1;
+            break;
         case 'y':
             pPars->fUserRecLib ^= 1;
             break;
@@ -15093,8 +15096,8 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->fUsePerm    =  1;
         pPars->pLutLib     =  NULL;
     }
-    // modify for global delay optimization
-    if ( pPars->fDelayOpt )
+    // modify for delay optimization
+    if ( pPars->fDelayOpt || pPars->fDsdBalance )
     {
         pPars->fTruth      =  1;
         pPars->fCutMin     =  1;
@@ -15102,7 +15105,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->fUsePerm    =  1;
         pPars->pLutLib     =  NULL;
     }
-    // modify for global delay optimization
+    // modify for delay optimization
     if ( pPars->nGateSize > 0 )
     {
         pPars->fTruth      =  1;
@@ -15216,7 +15219,7 @@ usage:
         sprintf(LutSize, "library" );
     else
         sprintf(LutSize, "%d", pPars->nLutSize );
-    Abc_Print( -2, "usage: if [-KCFANGT num] [-DEW float] [-S str] [-qarlepmsdbugyojikncvh]\n" );
+    Abc_Print( -2, "usage: if [-KCFANGT num] [-DEW float] [-S str] [-qarlepmsdbugxyojikncvh]\n" );
     Abc_Print( -2, "\t           performs FPGA technology mapping of the network\n" );
     Abc_Print( -2, "\t-K num   : the number of LUT inputs (2 < num < %d) [default = %s]\n", IF_MAX_LUTSIZE+1, LutSize );
     Abc_Print( -2, "\t-C num   : the max number of priority cuts (0 < num < 2^12) [default = %d]\n", pPars->nCutsMax );
@@ -15242,6 +15245,7 @@ usage:
     Abc_Print( -2, "\t-b       : toggles the use of one special feature [default = %s]\n", pPars->fUseBat? "yes": "no" );
     Abc_Print( -2, "\t-u       : toggles the use of MUXes along with LUTs [default = %s]\n", fLutMux? "yes": "no" );
     Abc_Print( -2, "\t-g       : toggles delay optimization by SOP balancing [default = %s]\n", pPars->fDelayOpt? "yes": "no" );
+    Abc_Print( -2, "\t-x       : toggles delay optimization by DSD balancing [default = %s]\n", pPars->fDsdBalance? "yes": "no" );
     Abc_Print( -2, "\t-y       : toggles delay optimization with recorded library [default = %s]\n", pPars->fUserRecLib? "yes": "no" );
     Abc_Print( -2, "\t-o       : toggles using buffers to decouple combinational outputs [default = %s]\n", pPars->fUseBuffs? "yes": "no" );
     Abc_Print( -2, "\t-j       : toggles enabling additional check [default = %s]\n", pPars->fEnableCheck07? "yes": "no" );
@@ -29503,7 +29507,7 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
     }
     pPars->pLutLib = (If_LibLut_t *)pAbc->pLibLut;
     Extra_UtilGetoptReset();
-    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRDEWSTqalepmrsdbgyojikfuztncvh" ) ) != EOF )
+    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRDEWSTqalepmrsdbgxyojikfuztncvh" ) ) != EOF )
     {
         switch ( c )
         {
@@ -29666,6 +29670,9 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
         case 'g':
             pPars->fDelayOpt ^= 1;
             break;
+        case 'x':
+            pPars->fDsdBalance ^= 1;
+            break;
         case 'y':
             pPars->fUserRecLib ^= 1;
             break;
@@ -29842,8 +29849,8 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->fUsePerm    =  1;
         pPars->pLutLib     =  NULL;
     }
-    // modify for global delay optimization
-    if ( pPars->fDelayOpt )
+    // modify for delay optimization
+    if ( pPars->fDelayOpt || pPars->fDsdBalance )
     {
         pPars->fTruth      =  1;
         pPars->fCutMin     =  1;
@@ -29851,7 +29858,7 @@ int Abc_CommandAbc9If( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->fUsePerm    =  1;
         pPars->pLutLib     =  NULL;
     }
-    // modify for global delay optimization
+    // modify for delay optimization
     if ( pPars->nGateSize > 0 )
     {
         pPars->fTruth      =  1;
@@ -29924,7 +29931,7 @@ usage:
         sprintf(LutSize, "library" );
     else
         sprintf(LutSize, "%d", pPars->nLutSize );
-    Abc_Print( -2, "usage: &if [-KCFAGRT num] [-DEW float] [-S str] [-qarlepmsdbgyojikfuztncvh]\n" );
+    Abc_Print( -2, "usage: &if [-KCFAGRT num] [-DEW float] [-S str] [-qarlepmsdbgxyojikfuztncvh]\n" );
     Abc_Print( -2, "\t           performs FPGA technology mapping of the network\n" );
     Abc_Print( -2, "\t-K num   : the number of LUT inputs (2 < num < %d) [default = %s]\n", IF_MAX_LUTSIZE+1, LutSize );
     Abc_Print( -2, "\t-C num   : the max number of priority cuts (0 < num < 2^12) [default = %d]\n", pPars->nCutsMax );
@@ -29948,6 +29955,7 @@ usage:
     Abc_Print( -2, "\t-d       : toggles deriving local AIGs using bi-decomposition [default = %s]\n", pPars->fBidec? "yes": "no" );
     Abc_Print( -2, "\t-b       : toggles the use of one special feature [default = %s]\n", pPars->fUseBat? "yes": "no" );
     Abc_Print( -2, "\t-g       : toggles delay optimization by SOP balancing [default = %s]\n", pPars->fDelayOpt? "yes": "no" );
+    Abc_Print( -2, "\t-x       : toggles delay optimization by DSD balancing [default = %s]\n", pPars->fDsdBalance? "yes": "no" );
     Abc_Print( -2, "\t-y       : toggles delay optimization with recorded library [default = %s]\n", pPars->fUserRecLib? "yes": "no" );
     Abc_Print( -2, "\t-o       : toggles using buffers to decouple combinational outputs [default = %s]\n", pPars->fUseBuffs? "yes": "no" );
     Abc_Print( -2, "\t-j       : toggles enabling additional check [default = %s]\n", pPars->fEnableCheck07? "yes": "no" );

src/base/abci/abcIf.c

@@ -356,7 +356,7 @@ Abc_Ntk_t * Abc_NtkFromIf( If_Man_t * pIfMan, Abc_Ntk_t * pNtk )
   SeeAlso     []
 
 ***********************************************************************/
-Hop_Obj_t * Abc_NodeTruthToHopInt( Hop_Man_t * pMan, Vec_Wrd_t * vAnds, int nVars )
+Hop_Obj_t * Abc_NodeFromSopBalanceInt( Hop_Man_t * pMan, Vec_Wrd_t * vAnds, int nVars )
 {
     Vec_Ptr_t * vResults;
     Hop_Obj_t * pRes0, * pRes1, * pRes = NULL;
@@ -387,24 +387,64 @@ Hop_Obj_t * Abc_NodeTruthToHopInt( Hop_Man_t * pMan, Vec_Wrd_t * vAnds, int nVar
     Vec_PtrFree( vResults );
     return Hop_NotCond( pRes, This.fCompl );
 }
+Hop_Obj_t * Abc_NodeFromSopBalance( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pCut )
+{
+    Hop_Obj_t * pResult;
+    Vec_Wrd_t * vArray;
+    vArray  = If_CutDelaySopArray( p, pCut );
+    pResult = Abc_NodeFromSopBalanceInt( pMan, vArray, If_CutLeaveNum(pCut) );
+//    Vec_WrdFree( vArray );
+    return pResult;
+}
 
 /**Function*************************************************************
 
-  Synopsis    [Creates the mapped network.]
+  Synopsis    [Inserts the entry while sorting them by delay.]
 
-  Description [Assuming the copy field of the mapped nodes are NULL.]
+  Description []
                
   SideEffects []
 
   SeeAlso     []
 
 ***********************************************************************/
-Hop_Obj_t * Abc_NodeTruthToHop( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pCut )
+Hop_Obj_t * Abc_NodeFromDsdBalanceInt( Hop_Man_t * pMan, Vec_Wrd_t * vAnds, int nVars )
+{
+    Vec_Ptr_t * vResults;
+    Hop_Obj_t * pRes0, * pRes1, * pRes = NULL;
+    If_And_t This;
+    word Entry;
+    int i;
+    if ( Vec_WrdSize(vAnds) == 0 )
+        return Hop_ManConst0(pMan);
+    if ( Vec_WrdSize(vAnds) == 1 && Vec_WrdEntry(vAnds,0) == 0 )
+        return Hop_ManConst1(pMan);
+    vResults = Vec_PtrAlloc( Vec_WrdSize(vAnds) );
+    for ( i = 0; i < nVars; i++ )
+        Vec_PtrPush( vResults, Hop_IthVar(pMan, i) );
+    Vec_WrdForEachEntryStart( vAnds, Entry, i, nVars )
+    {
+        This  = If_WrdToAnd( Entry );
+        pRes0 = Hop_NotCond( (Hop_Obj_t *)Vec_PtrEntry(vResults, This.iFan0), This.fCompl0 ); 
+        pRes1 = Hop_NotCond( (Hop_Obj_t *)Vec_PtrEntry(vResults, This.iFan1), This.fCompl1 ); 
+        pRes  = Hop_And( pMan, pRes0, pRes1 );
+        Vec_PtrPush( vResults, pRes );
+/*
+        printf( "fan0 = %c%d  fan1 = %c%d  Del = %d\n", 
+            This.fCompl0? '-':'+', This.iFan0, 
+            This.fCompl1? '-':'+', This.iFan1, 
+            This.Delay );
+*/
+    }
+    Vec_PtrFree( vResults );
+    return Hop_NotCond( pRes, This.fCompl );
+}
+Hop_Obj_t * Abc_NodeFromDsdBalance( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pCut )
 {
     Hop_Obj_t * pResult;
     Vec_Wrd_t * vArray;
     vArray  = If_CutDelaySopArray( p, pCut );
-    pResult = Abc_NodeTruthToHopInt( pMan, vArray, If_CutLeaveNum(pCut) );
+    pResult = Abc_NodeFromDsdBalanceInt( pMan, vArray, If_CutLeaveNum(pCut) );
 //    Vec_WrdFree( vArray );
     return pResult;
 }
@@ -437,7 +477,7 @@ Abc_Obj_t * Abc_NodeFromIf_rec( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Obj_t
     pCutBest = If_ObjCutBest( pIfObj );
 //    printf( "%d 0x%02X %d\n", pCutBest->nLeaves, 0xff & *If_CutTruth(pCutBest), pIfMan->pPars->pFuncCost(pCutBest) );
 //    if ( pIfMan->pPars->pLutLib && pIfMan->pPars->pLutLib->fVarPinDelays )
-    if ( !pIfMan->pPars->fUseTtPerm && !pIfMan->pPars->fDelayOpt && !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserRecLib && !pIfMan->pPars->nGateSize )
+    if ( !pIfMan->pPars->fUseTtPerm && !pIfMan->pPars->fDelayOpt && !pIfMan->pPars->fDsdBalance && !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserRecLib && !pIfMan->pPars->nGateSize )
         If_CutRotatePins( pIfMan, pCutBest );
     if ( pIfMan->pPars->fUseCnfs || pIfMan->pPars->fUseMv )
     {
@@ -486,8 +526,13 @@ Abc_Obj_t * Abc_NodeFromIf_rec( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Obj_t
         }
         else if ( pIfMan->pPars->fDelayOpt )
         {
-            extern Hop_Obj_t * Abc_NodeTruthToHop( Hop_Man_t * pMan, If_Man_t * pIfMan, If_Cut_t * pCut );
-            pNodeNew->pData = Abc_NodeTruthToHop( (Hop_Man_t *)pNtkNew->pManFunc, pIfMan, pCutBest );
+            extern Hop_Obj_t * Abc_NodeFromSopBalance( Hop_Man_t * pMan, If_Man_t * pIfMan, If_Cut_t * pCut );
+            pNodeNew->pData = Abc_NodeFromSopBalance( (Hop_Man_t *)pNtkNew->pManFunc, pIfMan, pCutBest );
+        }
+        else if ( pIfMan->pPars->fDsdBalance )
+        {
+            extern Hop_Obj_t * Abc_NodeFromDsdBalance( Hop_Man_t * pMan, If_Man_t * pIfMan, If_Cut_t * pCut );
+            pNodeNew->pData = Abc_NodeFromDsdBalance( (Hop_Man_t *)pNtkNew->pManFunc, pIfMan, pCutBest );
         }
         else if ( pIfMan->pPars->fUserRecLib )
         {

src/map/if/if.h

@@ -134,6 +134,7 @@ struct If_Par_t_
     float              WireDelay;     // wire delay
     // internal parameters
     int                fDelayOpt;     // special delay optimization
+    int                fDsdBalance;   // special delay optimization
     int                fUserRecLib;   // use recorded library
     int                fSkipCutFilter;// skip cut filter
     int                fAreaOnly;     // area only mode
@@ -546,6 +547,8 @@ extern int             If_DsdManLutSize( If_DsdMan_t * p );
 extern int             If_DsdManSuppSize( If_DsdMan_t * p, int iDsd );
 extern int             If_DsdManCheckDec( If_DsdMan_t * p, int iDsd );
 extern unsigned        If_DsdManCheckXY( If_DsdMan_t * p, int iDsd, int LutSize, int fDerive, int fHighEffort, int fVerbose );
+extern int             If_DsdCutBalanceCost( If_Man_t * pIfMan, If_Cut_t * pCut );
+extern int             If_DsdCutBalance( void * pGia, If_Man_t * pIfMan, If_Cut_t * pCut, Vec_Int_t * vLeaves, int fHash );
 /*=== ifLib.c =============================================================*/
 extern If_LibLut_t *   If_LibLutRead( char * FileName );
 extern If_LibLut_t *   If_LibLutDup( If_LibLut_t * p );

src/map/if/ifDsd.c

@@ -1849,6 +1849,28 @@ void If_DsdManTest()
   SeeAlso     []
 
 ***********************************************************************/
+int If_DsdCutBalanceCost( If_Man_t * pIfMan, If_Cut_t * pCut )
+{
+    return 0;
+}
+int If_DsdCutBalance( void * pGia, If_Man_t * pIfMan, If_Cut_t * pCut, Vec_Int_t * vLeaves, int fHash )
+{
+    Gia_Man_t * p = (Gia_Man_t *)pGia;
+    return 0;
+}
+
+
+/**Function*************************************************************
+
+  Synopsis    []
+
+  Description []
+               
+  SideEffects []
+
+  SeeAlso     []
+
+***********************************************************************/
 void If_DsdManTune( If_DsdMan_t * p, int LutSize, int fFast, int fAdd, int fSpec, int fVerbose )
 {
     ProgressBar * pProgress = NULL;

src/map/if/ifMap.c

@@ -99,6 +99,7 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
     If_Cut_t * pCut0R, * pCut1R;
     int fFunc0R, fFunc1R;
     int i, k, v, fChange;
+    int fSave0 = p->pPars->fDelayOpt || p->pPars->fDsdBalance || p->pPars->fUserRecLib;
     assert( p->pPars->fSeqMap || !If_ObjIsAnd(pObj->pFanin0) || pObj->pFanin0->pCutSet->nCuts > 0 );
     assert( p->pPars->fSeqMap || !If_ObjIsAnd(pObj->pFanin1) || pObj->pFanin1->pCutSet->nCuts > 0 );
 
@@ -122,12 +123,14 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
     if ( !fFirst )
     {
         // recompute the parameters of the best cut
-        if ( p->pPars->fUserRecLib )
-            pCut->Delay = If_CutDelayRecCost3(p, pCut, pObj); 
-        else if(p->pPars->fDelayOpt)
-            pCut->Delay = If_CutDelaySopCost(p,pCut);
-        else if(p->pPars->nGateSize > 0)
-            pCut->Delay = If_CutDelaySop(p,pCut);
+        if ( p->pPars->fDelayOpt )
+            pCut->Delay = If_CutDelaySopCost( p, pCut );
+        else if ( p->pPars->fDsdBalance )
+            pCut->Delay = If_DsdCutBalanceCost( p, pCut );
+        else if ( p->pPars->fUserRecLib )
+            pCut->Delay = If_CutDelayRecCost3( p, pCut, pObj ); 
+        else if( p->pPars->nGateSize > 0 )
+            pCut->Delay = If_CutDelaySop( p, pCut );
         else
             pCut->Delay = If_CutDelay( p, pObj, pCut );
 //        assert( pCut->Delay <= pObj->Required + p->fEpsilon );
@@ -184,7 +187,7 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
         p->nCutsMerged++;
         p->nCutsTotal++;
         // check if this cut is contained in any of the available cuts
-        if ( !p->pPars->fSkipCutFilter && If_CutFilter( pCutSet, pCut, p->pPars->fUserRecLib || p->pPars->fDelayOpt ) )
+        if ( !p->pPars->fSkipCutFilter && If_CutFilter( pCutSet, pCut, fSave0 ) )
             continue;
         // compute the truth table
         pCut->iCutFunc = -1;
@@ -202,7 +205,7 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
                 fChange = If_CutComputeTruth( p, pCut, pCut0, pCut1, pObj->fCompl0, pObj->fCompl1 );
             if ( p->pPars->fVerbose )
                 p->timeCache[4] += Abc_Clock() - clk;
-            if ( !p->pPars->fSkipCutFilter && fChange && If_CutFilter( pCutSet, pCut, p->pPars->fUserRecLib || p->pPars->fDelayOpt ) )
+            if ( !p->pPars->fSkipCutFilter && fChange && If_CutFilter( pCutSet, pCut, fSave0 ) )
                 continue;
             if ( p->pPars->fUseDsd )
             {
@@ -276,19 +279,16 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
         if ( pCut->Cost == IF_COST_MAX )
             continue;
         // check if the cut satisfies the required times
-///        if ( p->pPars->pLutStruct )
-///            pCut->Delay = If_CutDelayLutStruct( p, pCut, p->pPars->pLutStruct, p->pPars->WireDelay );
-        if ( p->pPars->fUserRecLib )
-            pCut->Delay = If_CutDelayRecCost3(p, pCut, pObj); 
-        else if (p->pPars->fDelayOpt)
-            pCut->Delay = If_CutDelaySopCost(p, pCut);  
-        else if(p->pPars->nGateSize > 0)
-            pCut->Delay = If_CutDelaySop(p,pCut);
-        else
+        if ( p->pPars->fDelayOpt )
+            pCut->Delay = If_CutDelaySopCost( p, pCut );  
+        else if ( p->pPars->fDsdBalance )
+            pCut->Delay = If_DsdCutBalanceCost( p, pCut );
+        else if ( p->pPars->fUserRecLib )
+            pCut->Delay = If_CutDelayRecCost3( p, pCut, pObj ); 
+        else if( p->pPars->nGateSize > 0 )
+            pCut->Delay = If_CutDelaySop( p, pCut );
+        else 
             pCut->Delay = If_CutDelay( p, pObj, pCut );
-        //if ( pCut->Cost == IF_COST_MAX )
-       //     continue;
-//        Abc_Print( 1, "%.2f ", pCut->Delay );
         if ( Mode && pCut->Delay > pObj->Required + p->fEpsilon )
             continue;
         // compute area of the cut (this area may depend on the application specific cost)
@@ -352,7 +352,7 @@ void If_ObjPerformMappingChoice( If_Man_t * p, If_Obj_t * pObj, int Mode, int fP
     If_Set_t * pCutSet;
     If_Obj_t * pTemp;
     If_Cut_t * pCutTemp, * pCut;
-    int i;
+    int i, fSave0 = p->pPars->fDelayOpt || p->pPars->fDsdBalance || p->pPars->fUserRecLib;
     assert( pObj->pEquiv != NULL );
 
     // prepare
@@ -369,8 +369,6 @@ void If_ObjPerformMappingChoice( If_Man_t * p, If_Obj_t * pObj, int Mode, int fP
     // generate cuts
     for ( pTemp = pObj->pEquiv; pTemp; pTemp = pTemp->pEquiv )
     {
-//        assert( pTemp->nRefs == 0 );
-//        assert( p->pPars->fSeqMap || pTemp->pCutSet->nCuts > 0 ); // June 9, 2009
         if ( pTemp->pCutSet->nCuts == 0 )
             continue;
         // go through the cuts of this node
@@ -385,10 +383,10 @@ void If_ObjPerformMappingChoice( If_Man_t * p, If_Obj_t * pObj, int Mode, int fP
             // copy the cut into storage
             If_CutCopy( p, pCut, pCutTemp );
             // check if this cut is contained in any of the available cuts
-            if ( If_CutFilter( pCutSet, pCut, p->pPars->fUserRecLib || p->pPars->fDelayOpt ) )
+            if ( If_CutFilter( pCutSet, pCut, fSave0 ) )
                 continue;
             // check if the cut satisfies the required times
-            assert( pCut->Delay == If_CutDelay( p, pTemp, pCut ) );
+//            assert( pCut->Delay == If_CutDelay( p, pTemp, pCut ) );
             if ( Mode && pCut->Delay > pObj->Required + p->fEpsilon )
                 continue;
             // set the phase attribute