diff --git a/src/base/abci/abc.c b/src/base/abci/abc.c
index 36c8f698b..778e7f319 100644
--- a/src/base/abci/abc.c
+++ b/src/base/abci/abc.c
@@ -19514,7 +19514,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
     If_ManSetDefaultPars( pPars );
     pPars->pLutLib = (If_LibLut_t *)Abc_FrameReadLibLut();
     Extra_UtilGetoptReset();
-    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRNTXYZDEWSJqaflepmrsdbgxyuojiktncvh" ) ) != EOF )
+    while ( ( c = Extra_UtilGetopt( argc, argv, "KCFAGRNTXYUZDEWSJqaflepmrsdbgxyuojiktncvh" ) ) != EOF )
     {
         switch ( c )
         {
@@ -19630,6 +19630,18 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
             if ( pPars->nAndDelay < 0 )
                 goto usage;
             break;
+        case 'U':
+            if ( globalUtilOptind >= argc )
+            {
+                Abc_Print( -1, "Command line switch \"-U\" should be followed by a positive integer 3, 4, 5, or 6.\n" );
+                goto usage;
+            }
+            pPars->nLutDecSize = atoi(argv[globalUtilOptind]);
+            pPars->fUserLut2D = 1;
+            globalUtilOptind++;
+            if ( pPars->nLutDecSize < 3 || pPars->nLutDecSize > 6 )
+                goto usage;
+            break;
         case 'Z':
             if ( globalUtilOptind >= argc )
             {
@@ -19892,7 +19904,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->fCutMin = 1;
     }
 
-    if ( pPars->fUserLutDec )
+    if ( pPars->fUserLutDec || pPars->fUserLut2D )
     {
         if ( pPars->nLutDecSize == 0 )
         {
@@ -19927,7 +19939,7 @@ int Abc_CommandIf( Abc_Frame_t * pAbc, int argc, char ** argv )
         pPars->pLutLib     =  NULL;
     }
     // modify for delay optimization
-    if ( pPars->fDelayOpt || pPars->fDsdBalance || pPars->fDelayOptLut || pPars->fUserLutDec )
+    if ( pPars->fDelayOpt || pPars->fDsdBalance || pPars->fDelayOptLut || pPars->fUserLutDec || pPars->fUserLut2D )
     {
         pPars->fTruth      =  1;
         pPars->fCutMin     =  1;
@@ -20073,7 +20085,7 @@ usage:
         sprintf(LutSize, "library" );
     else
         sprintf(LutSize, "%d", pPars->nLutSize );
-    Abc_Print( -2, "usage: if [-KCFAGRNTXYZ num] [-DEW float] [-S str] [-qarlepmsdbgxyuojiktncvh]\n" );
+    Abc_Print( -2, "usage: if [-KCFAGRNTXYUZ num] [-DEW float] [-SJ str] [-qarlepmsdbgxyuojiktncvh]\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 );
@@ -20085,6 +20097,7 @@ usage:
     Abc_Print( -2, "\t-T num   : the type of LUT structures [default = any]\n" );
     Abc_Print( -2, "\t-X num   : delay of AND-gate in LUT library units [default = %d]\n", pPars->nAndDelay );
     Abc_Print( -2, "\t-Y num   : area of AND-gate in LUT library units [default = %d]\n", pPars->nAndArea );
+    Abc_Print( -2, "\t-U num   : the number of LUT inputs for delay-driven LUT decomposition [default = not used]\n" );
     Abc_Print( -2, "\t-Z num   : the number of LUT inputs for delay-driven LUT decomposition [default = not used]\n" );
     Abc_Print( -2, "\t-D float : sets the delay constraint for the mapping [default = %s]\n", Buffer );
     Abc_Print( -2, "\t-E float : sets epsilon used for tie-breaking [default = %f]\n", pPars->Epsilon );
diff --git a/src/base/abci/abcIf.c b/src/base/abci/abcIf.c
index 357d7d83f..8c4ac1448 100644
--- a/src/base/abci/abcIf.c
+++ b/src/base/abci/abcIf.c
@@ -116,7 +116,7 @@ Abc_Ntk_t * Abc_NtkIf( Abc_Ntk_t * pNtk, If_Par_t * pPars )
     pPars->pTimesReq = Abc_NtkGetCoRequiredFloats(pNtk);
 
     // update timing info to reflect logic level
-    if ( (pPars->fDelayOpt || pPars->fDsdBalance || pPars->fUserRecLib || pPars->fUserSesLib || pPars->fUserLutDec) && pNtk->pManTime )
+    if ( (pPars->fDelayOpt || pPars->fDsdBalance || pPars->fUserRecLib || pPars->fUserSesLib || pPars->fUserLutDec || pPars->fUserLut2D ) && pNtk->pManTime )
     {
         int c;
         if ( pNtk->AndGateDelay == 0.0 )
@@ -433,8 +433,8 @@ Hop_Obj_t * Abc_NodeBuildFromMini( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pC
    SideEffects []
    SeeAlso     []
  ***********************************************************************/
- void Abc_DecRecordToHop( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Cut_t * pCutBest, If_Obj_t * pIfObj, Vec_Int_t * vCover, Abc_Obj_t * pNodeTop )
- {
+void Abc_DecRecordToHop( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Cut_t * pCutBest, If_Obj_t * pIfObj, Vec_Int_t * vCover, Abc_Obj_t * pNodeTop )
+{
     extern Hop_Obj_t * Kit_TruthToHop( Hop_Man_t * pMan, unsigned * pTruth, int nVars, Vec_Int_t * vMemory );
     assert( !pIfMan->pPars->fUseTtPerm );
 
@@ -460,7 +460,15 @@ Hop_Obj_t * Abc_NodeBuildFromMini( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pC
 
     // perform LUT-decomposition and return the LUT-structure
     unsigned char decompArray[92];
-    int val = acd_decompose( pTruth, pCutBest->nLeaves, pIfMan->pPars->nLutDecSize, &(delayProfile), decompArray );
+    int val;
+    if ( pIfMan->pPars->fUserLutDec )
+    {
+        val = acd_decompose( pTruth, pCutBest->nLeaves, pIfMan->pPars->nLutDecSize, &(delayProfile), decompArray );
+    }
+    else
+    {
+        val = acd2_decompose( pTruth, pCutBest->nLeaves, pIfMan->pPars->nLutDecSize, &(delayProfile), decompArray );
+    }
     assert( val == 0 );
 
     // convert the LUT-structure into a set of logic nodes in Abc_Ntk_t 
@@ -473,7 +481,7 @@ Hop_Obj_t * Abc_NodeBuildFromMini( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pC
     word *tt;
     Abc_Obj_t *pNewNodes[5];
 
-    /* create intermediate LUTs*/
+    /* create intermediate LUTs */
     assert( decompArray[1] <= 6 );
     Abc_Obj_t * pFanin;
     for ( i = 0; i < decompArray[1]; ++i )
@@ -537,7 +545,7 @@ Hop_Obj_t * Abc_NodeBuildFromMini( Hop_Man_t * pMan, If_Man_t * p, If_Cut_t * pC
 
     /* check correct read */
     assert( byte_p == decompArray[0] );
- }
+}
 
 /**Function*************************************************************
 
@@ -577,14 +585,15 @@ Abc_Obj_t * Abc_NodeFromIf_rec( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Obj_t
     pNodeNew = Abc_NtkCreateNode( pNtkNew );
 //    if ( pIfMan->pPars->pLutLib && pIfMan->pPars->pLutLib->fVarPinDelays )
     if ( !pIfMan->pPars->fDelayOpt && !pIfMan->pPars->fDelayOptLut && !pIfMan->pPars->fDsdBalance && !pIfMan->pPars->fUseTtPerm && 
-         !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserLutDec && !pIfMan->pPars->fUserRecLib && !pIfMan->pPars->fUserSesLib && !pIfMan->pPars->nGateSize )
+         !pIfMan->pPars->pLutStruct && !pIfMan->pPars->fUserLutDec && !pIfMan->pPars->fUserLut2D && !pIfMan->pPars->fUserRecLib &&
+         !pIfMan->pPars->fUserSesLib && !pIfMan->pPars->nGateSize )
         If_CutRotatePins( pIfMan, pCutBest );
     if ( pIfMan->pPars->fUseCnfs || pIfMan->pPars->fUseMv )
     {
         If_CutForEachLeafReverse( pIfMan, pCutBest, pIfLeaf, i )
             Abc_ObjAddFanin( pNodeNew, Abc_NodeFromIf_rec(pNtkNew, pIfMan, pIfLeaf, vCover) );
     }
-    else if ( pIfMan->pPars->fUserLutDec )
+    else if ( pIfMan->pPars->fUserLutDec || pIfMan->pPars->fUserLut2D )
     {
         If_CutForEachLeaf( pIfMan, pCutBest, pIfLeaf, i )
             Abc_NodeFromIf_rec(pNtkNew, pIfMan, pIfLeaf, vCover);
@@ -642,7 +651,7 @@ Abc_Obj_t * Abc_NodeFromIf_rec( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Obj_t
             extern Hop_Obj_t * Abc_RecToHop3( Hop_Man_t * pMan, If_Man_t * pIfMan, If_Cut_t * pCut, If_Obj_t * pIfObj );
             pNodeNew->pData = Abc_RecToHop3( (Hop_Man_t *)pNtkNew->pManFunc, pIfMan, pCutBest, pIfObj ); 
         }
-        else if ( pIfMan->pPars->fUserLutDec )
+        else if ( pIfMan->pPars->fUserLutDec || pIfMan->pPars->fUserLut2D )
         {
             extern void Abc_DecRecordToHop( Abc_Ntk_t * pNtkNew, If_Man_t * pIfMan, If_Cut_t * pCut, If_Obj_t * pIfObj, Vec_Int_t * vMemory, Abc_Obj_t * pNodeTop );
             Abc_DecRecordToHop( pNtkNew, pIfMan, pCutBest, pIfObj, vCover, pNodeNew ); 
diff --git a/src/map/if/if.h b/src/map/if/if.h
index 47fe109ca..723fda780 100644
--- a/src/map/if/if.h
+++ b/src/map/if/if.h
@@ -147,7 +147,8 @@ struct If_Par_t_
     int                fDeriveLuts;   // enables deriving LUT structures
     int                fDoAverage;    // optimize average rather than maximum level
     int                fHashMapping;  // perform AIG hashing after mapping
-    int                fUserLutDec;  // perform AIG hashing after mapping
+    int                fUserLutDec;   // perform Boolean decomposition during mapping
+    int                fUserLut2D;    // perform Boolean decomposition during mapping
     int                fVerbose;      // the verbosity flag
     int                fVerboseTrace; // the verbosity flag
     char *             pLutStruct;    // LUT structure
@@ -573,6 +574,7 @@ extern int             If_CutSopBalancePinDelays( If_Man_t * p, If_Cut_t * pCut,
 extern int             If_CutLutBalanceEval( If_Man_t * p, If_Cut_t * pCut );
 extern int             If_CutLutBalancePinDelays( If_Man_t * p, If_Cut_t * pCut, char * pPerm );
 extern int             If_LutDecEval( If_Man_t * p, If_Cut_t * pCut, If_Obj_t * pObj, int optDelay, int fFirst );
+extern int             If_Lut2DecEval( If_Man_t * p, If_Cut_t * pCut, If_Obj_t * pObj, int optDelay, int fFirst );
 extern int             If_LutDecReEval( If_Man_t * p, If_Cut_t * pCut );
 extern float           If_LutDecPinRequired( If_Man_t * p, If_Cut_t * pCut, int i, float required );
 /*=== ifDsd.c =============================================================*/
@@ -704,6 +706,8 @@ extern void            If_ObjPrint( If_Obj_t * pObj );
 
 extern int             acd_evaluate( word * pTruth, unsigned nVars, int lutSize, unsigned *pdelay, unsigned *cost, int try_no_late_arrival );
 extern int             acd_decompose( word * pTruth, unsigned nVars, int lutSize, unsigned *pdelay, unsigned char *decomposition );
+extern int             acd2_evaluate( word * pTruth, unsigned nVars, int lutSize, unsigned *pdelay, unsigned *cost, int try_no_late_arrival );
+extern int             acd2_decompose( word * pTruth, unsigned nVars, int lutSize, unsigned *pdelay, unsigned char *decomposition );
 
 ABC_NAMESPACE_HEADER_END
 
diff --git a/src/map/if/ifCore.c b/src/map/if/ifCore.c
index f7fcbca66..eed6a55e8 100644
--- a/src/map/if/ifCore.c
+++ b/src/map/if/ifCore.c
@@ -63,6 +63,7 @@ void If_ManSetDefaultPars( If_Par_t * pPars )
     pPars->fCutMin     =  0;
     pPars->fBidec      =  0;
     pPars->fUserLutDec =  0;
+    pPars->fUserLut2D  =  0;
     pPars->fVerbose    =  0;
 }
 
diff --git a/src/map/if/ifCut.c b/src/map/if/ifCut.c
index 49850d313..c9354b339 100644
--- a/src/map/if/ifCut.c
+++ b/src/map/if/ifCut.c
@@ -761,7 +761,7 @@ void If_CutSort( If_Man_t * p, If_Set_t * pCutSet, If_Cut_t * pCut )
 
     if ( !pCut->fUseless && 
          (p->pPars->fUseDsd || p->pPars->pFuncCell2 || p->pPars->fUseBat || 
-          p->pPars->pLutStruct || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUserLutDec ||
+          p->pPars->pLutStruct || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUserLutDec || p->pPars->fUserLut2D ||
           p->pPars->fEnableCheck07 || p->pPars->fUseCofVars || p->pPars->fUseAndVars || p->pPars->fUse34Spec || 
           p->pPars->fUseDsdTune || p->pPars->fEnableCheck75 || p->pPars->fEnableCheck75u || p->pPars->fUseCheck1 || p->pPars->fUseCheck2) )
     {
diff --git a/src/map/if/ifDelay.c b/src/map/if/ifDelay.c
index 3514327c1..0f94de49a 100644
--- a/src/map/if/ifDelay.c
+++ b/src/map/if/ifDelay.c
@@ -505,6 +505,95 @@ int If_LutDecEval( If_Man_t * p, If_Cut_t * pCut, If_Obj_t * pObj, int optDelay,
     return DelayMax + val;
 }
 
+int If_Lut2DecEval( If_Man_t * p, If_Cut_t * pCut, If_Obj_t * pObj, int optDelay, int fFirst )
+{
+    pCut->fUser = 1;
+    pCut->Cost = pCut->nLeaves > 1 ? 1 : 0;
+    pCut->decDelay = 0;
+    if ( pCut->nLeaves == 0 ) // const
+    {
+        assert( Abc_Lit2Var(If_CutTruthLit(pCut)) == 0 );
+        return 0;
+    }
+    if ( pCut->nLeaves == 1 ) // variable
+    {
+        assert( Abc_Lit2Var(If_CutTruthLit(pCut)) == 1 );
+        return (int)If_ObjCutBest(If_CutLeaf(p, pCut, 0))->Delay;
+    }
+
+    int LutSize = p->pPars->nLutDecSize;
+    int i, leaf_delay;
+    int DelayMax = -1, nLeafMax = 0;
+    unsigned uLeafMask = 0;
+    for ( i = 0; i < If_CutLeaveNum(pCut); i++ )
+    {
+        leaf_delay = If_ObjCutBest(If_CutLeaf(p, pCut, i))->Delay;
+
+        if ( DelayMax < leaf_delay )
+        {
+            DelayMax = leaf_delay;
+            nLeafMax = 1;
+            uLeafMask = (1 << i);
+        }
+        else if ( DelayMax == leaf_delay )
+        {
+            nLeafMax++;
+            uLeafMask |= (1 << i);
+        }
+    }
+    if ( If_CutLeaveNum(pCut) <= LutSize )
+    {
+        pCut->decDelay = ( 1 << LutSize ) - 1;
+        return DelayMax + 1;
+    }
+
+    /* compute the decomposition */
+    int use_late_arrival = 0;
+    unsigned cost = 1;
+
+    if ( !fFirst )
+    {
+        if ( optDelay )
+        {
+            /* checks based on delay: must be better than the previous best cut */
+            use_late_arrival = DelayMax + 2 >= If_ObjCutBest(pObj)->Delay;
+        }
+        else
+        {
+            /* checks based on delay: look at the required time */
+            use_late_arrival = DelayMax + 2 > pObj->Required + p->fEpsilon;
+        }
+    }
+
+    /* Too many late-arriving signals */
+    if ( nLeafMax == LutSize && use_late_arrival )
+    {
+        /* unfeasible decomposition */
+        pCut->Cost = IF_COST_MAX;
+        return ABC_INFINITY;
+    }
+
+    if ( !use_late_arrival )
+    {
+        uLeafMask = 0;
+    }
+
+    /* returns the delay of the decomposition */
+    word *pTruth = If_CutTruthW( p, pCut );
+    int val = acd2_evaluate( pTruth, pCut->nLeaves, LutSize, &uLeafMask, &cost, !use_late_arrival );
+
+    /* not feasible decomposition */
+    pCut->decDelay = uLeafMask;
+    if ( val < 0 )
+    {
+        pCut->Cost = IF_COST_MAX;
+        return ABC_INFINITY;
+    }
+
+    pCut->Cost = 2;
+    return DelayMax + val;
+}
+
 int If_LutDecReEval( If_Man_t * p, If_Cut_t * pCut )
 {
     // pCut->fUser = 1;
diff --git a/src/map/if/ifMap.c b/src/map/if/ifMap.c
index bdd3ae439..0ad63f050 100644
--- a/src/map/if/ifMap.c
+++ b/src/map/if/ifMap.c
@@ -166,7 +166,7 @@ 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, iCutDsd, fChange;
-    int fSave0 = p->pPars->fDelayOpt || p->pPars->fDelayOptLut || p->pPars->fDsdBalance || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUserLutDec ||
+    int fSave0 = p->pPars->fDelayOpt || p->pPars->fDelayOptLut || p->pPars->fDsdBalance || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUserLutDec || p->pPars->fUserLut2D ||
         p->pPars->fUseDsdTune || p->pPars->fUseCofVars || p->pPars->fUseAndVars || p->pPars->fUse34Spec || p->pPars->pLutStruct || p->pPars->pFuncCell2 || p->pPars->fUseCheck1 || p->pPars->fUseCheck2;
     int fUseAndCut = (p->pPars->nAndDelay > 0) || (p->pPars->nAndArea > 0);
     assert( !If_ObjIsAnd(pObj->pFanin0) || pObj->pFanin0->pCutSet->nCuts > 0 );
@@ -208,7 +208,7 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
                 pCut->fUseless = 1;
             }
         }
-        else if ( p->pPars->fUserLutDec )
+        else if ( p->pPars->fUserLutDec || p->pPars->fUserLut2D )
         {
             pCut->Delay = If_LutDecReEval( p, pCut ); 
         }
@@ -434,6 +434,11 @@ void If_ObjPerformMappingAnd( If_Man_t * p, If_Obj_t * pObj, int Mode, int fPrep
             pCut->Delay = If_LutDecEval( p, pCut, pObj, Mode == 0, fFirst );
             pCut->fUseless = pCut->Delay == ABC_INFINITY;
         }
+        else if ( p->pPars->fUserLut2D )
+        {
+            pCut->Delay = If_Lut2DecEval( p, pCut, pObj, Mode == 0, fFirst );
+            pCut->fUseless = pCut->Delay == ABC_INFINITY;
+        }
         else if ( p->pPars->fUserSesLib )
         {
             int Cost = 0;
@@ -518,7 +523,7 @@ 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, fSave0 = p->pPars->fDelayOpt || p->pPars->fDelayOptLut || p->pPars->fDsdBalance || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUse34Spec || p->pPars->fUserLutDec;
+    int i, fSave0 = p->pPars->fDelayOpt || p->pPars->fDelayOptLut || p->pPars->fDsdBalance || p->pPars->fUserRecLib || p->pPars->fUserSesLib || p->pPars->fUse34Spec || p->pPars->fUserLutDec || p->pPars->fUserLut2D;
     assert( pObj->pEquiv != NULL );
 
     // prepare
diff --git a/src/map/if/ifTime.c b/src/map/if/ifTime.c
index f20842384..a1b51d6ad 100644
--- a/src/map/if/ifTime.c
+++ b/src/map/if/ifTime.c
@@ -211,7 +211,7 @@ void If_CutPropagateRequired( If_Man_t * p, If_Obj_t * pObj, If_Cut_t * pCut, fl
                 pLeaf->Required = IF_MIN( pLeaf->Required, Required - pLutDelays[0] );
         }
     }
-    else if ( p->pPars->fUserLutDec )
+    else if ( p->pPars->fUserLutDec || p->pPars->fUserLut2D )
     {
         Required = ObjRequired;
         If_CutForEachLeaf( p, pCut, pLeaf, i )