/**CFile**************************************************************** FileName [giaNf.c] SystemName [ABC: Logic synthesis and verification system.] PackageName [Scalable AIG package.] Synopsis [Standard-cell mapper.] Author [Alan Mishchenko] Affiliation [UC Berkeley] Date [Ver. 1.0. Started - June 20, 2005.] Revision [$Id: giaNf.c,v 1.00 2005/06/20 00:00:00 alanmi Exp $] ***********************************************************************/ #include #include "gia.h" #include "misc/st/st.h" #include "map/mio/mio.h" #include "misc/util/utilTruth.h" #include "misc/extra/extra.h" #include "base/main/main.h" #include "misc/vec/vecMem.h" #include "misc/vec/vecWec.h" #include "opt/dau/dau.h" ABC_NAMESPACE_IMPL_START //////////////////////////////////////////////////////////////////////// /// DECLARATIONS /// //////////////////////////////////////////////////////////////////////// #define NF_LEAF_MAX 6 #define NF_CUT_MAX 32 #define NF_NO_LEAF 31 #define NF_INFINITY FLT_MAX typedef struct Nf_Cut_t_ Nf_Cut_t; struct Nf_Cut_t_ { word Sign; // signature int Delay; // delay float Flow; // flow unsigned iFunc : 26; // function unsigned Useless : 1; // function unsigned nLeaves : 5; // leaf number (NF_NO_LEAF) int pLeaves[NF_LEAF_MAX+1]; // leaves }; typedef struct Pf_Mat_t_ Pf_Mat_t; struct Pf_Mat_t_ { unsigned fCompl : 8; // complemented unsigned Phase : 6; // match phase unsigned Perm : 18; // match permutation }; typedef struct Nf_Mat_t_ Nf_Mat_t; struct Nf_Mat_t_ { unsigned Gate : 20; // gate unsigned CutH : 10; // cut handle unsigned fCompl : 1; // complemented unsigned fBest : 1; // best cut int Conf; // input literals float D; // delay float A; // area }; typedef struct Nf_Obj_t_ Nf_Obj_t; struct Nf_Obj_t_ { Nf_Mat_t M[2][2]; // del/area (2x) }; typedef struct Nf_Man_t_ Nf_Man_t; struct Nf_Man_t_ { // user data Gia_Man_t * pGia; // derived manager Jf_Par_t * pPars; // parameters // matching Vec_Mem_t * vTtMem; // truth tables Vec_Wec_t * vTt2Match; // matches for truth tables Mio_Cell_t * pCells; // library gates int nCells; // library gate count // cut data Nf_Obj_t * pNfObjs; // best cuts Vec_Ptr_t vPages; // cut memory Vec_Int_t vCutSets; // cut offsets Vec_Int_t vMapRefs; // mapping refs (2x) Vec_Flt_t vFlowRefs; // flow refs (2x) Vec_Flt_t vRequired; // required times (2x) Vec_Flt_t vCutFlows; // temporary cut area Vec_Int_t vCutDelays; // temporary cut delay Vec_Int_t vBackup; // backup literals Vec_Int_t vBackup2; // backup literals int iCur; // current position int Iter; // mapping iterations int fUseEla; // use exact area int nInvs; // the inverter count float InvDelay; // inverter delay float InvArea; // inverter area // statistics abctime clkStart; // starting time double CutCount[6]; // cut counts int nCutUseAll; // objects with useful cuts }; static inline int Pf_Mat2Int( Pf_Mat_t Mat ) { union { int x; Pf_Mat_t y; } v; v.y = Mat; return v.x; } static inline Pf_Mat_t Pf_Int2Mat( int Int ) { union { int x; Pf_Mat_t y; } v; v.x = Int; return v.y; } static inline Nf_Obj_t * Nf_ManObj( Nf_Man_t * p, int i ) { return p->pNfObjs + i; } static inline Mio_Cell_t* Nf_ManCell( Nf_Man_t * p, int i ) { return p->pCells + i; } static inline int * Nf_ManCutSet( Nf_Man_t * p, int i ) { return (int *)Vec_PtrEntry(&p->vPages, i >> 16) + (i & 0xFFFF); } static inline int Nf_ObjCutSetId( Nf_Man_t * p, int i ) { return Vec_IntEntry( &p->vCutSets, i ); } static inline int * Nf_ObjCutSet( Nf_Man_t * p, int i ) { return Nf_ManCutSet(p, Nf_ObjCutSetId(p, i)); } static inline int Nf_ObjHasCuts( Nf_Man_t * p, int i ) { return (int)(Vec_IntEntry(&p->vCutSets, i) > 0); } static inline int * Nf_ObjCutBest( Nf_Man_t * p, int i ) { return NULL; } static inline int Nf_ObjCutUseless( Nf_Man_t * p, int TruthId ) { return (int)(TruthId >= Vec_WecSize(p->vTt2Match)); } static inline float Nf_ObjCutFlow( Nf_Man_t * p, int i ) { return Vec_FltEntry(&p->vCutFlows, i); } static inline int Nf_ObjCutDelay( Nf_Man_t * p, int i ) { return Vec_IntEntry(&p->vCutDelays, i); } static inline void Nf_ObjSetCutFlow( Nf_Man_t * p, int i, float a ) { Vec_FltWriteEntry(&p->vCutFlows, i, a); } static inline void Nf_ObjSetCutDelay( Nf_Man_t * p, int i, int d ) { Vec_IntWriteEntry(&p->vCutDelays, i, d); } static inline int Nf_ObjMapRefNum( Nf_Man_t * p, int i, int c ) { return Vec_IntEntry(&p->vMapRefs, Abc_Var2Lit(i,c)); } static inline int Nf_ObjMapRefInc( Nf_Man_t * p, int i, int c ) { return (*Vec_IntEntryP(&p->vMapRefs, Abc_Var2Lit(i,c)))++; } static inline int Nf_ObjMapRefDec( Nf_Man_t * p, int i, int c ) { return --(*Vec_IntEntryP(&p->vMapRefs, Abc_Var2Lit(i,c))); } static inline float Nf_ObjFlowRefs( Nf_Man_t * p, int i, int c ) { return Vec_FltEntry(&p->vFlowRefs, Abc_Var2Lit(i,c)); } static inline float Nf_ObjRequired( Nf_Man_t * p, int i, int c ) { return Vec_FltEntry(&p->vRequired, Abc_Var2Lit(i,c)); } static inline void Nf_ObjSetRequired(Nf_Man_t * p,int i, int c, float f) { Vec_FltWriteEntry(&p->vRequired, Abc_Var2Lit(i,c), f); } static inline void Nf_ObjUpdateRequired(Nf_Man_t * p,int i, int c, float f) { if (Nf_ObjRequired(p, i, c) > f) Nf_ObjSetRequired(p, i, c, f); } static inline Nf_Mat_t * Nf_ObjMatchD( Nf_Man_t * p, int i, int c ) { return &Nf_ManObj(p, i)->M[c][0]; } static inline Nf_Mat_t * Nf_ObjMatchA( Nf_Man_t * p, int i, int c ) { return &Nf_ManObj(p, i)->M[c][1]; } static inline Nf_Mat_t * Nf_ObjMatchBest( Nf_Man_t * p, int i, int c ) { Nf_Mat_t * pD = Nf_ObjMatchD(p, i, c); Nf_Mat_t * pA = Nf_ObjMatchA(p, i, c); assert( pD->fBest != pA->fBest ); assert( Nf_ObjMapRefNum(p, i, c) > 0 ); if ( pA->fBest ) return pA; if ( pD->fBest ) return pD; return NULL; } static inline int Nf_CutSize( int * pCut ) { return pCut[0] & NF_NO_LEAF; } static inline int Nf_CutFunc( int * pCut ) { return ((unsigned)pCut[0] >> 5); } static inline int * Nf_CutLeaves( int * pCut ) { return pCut + 1; } static inline int Nf_CutSetBoth( int n, int f ) { return n | (f << 5); } static inline int Nf_CutIsTriv( int * pCut, int i ) { return Nf_CutSize(pCut) == 1 && pCut[1] == i; } static inline int Nf_CutHandle( int * pCutSet, int * pCut ) { assert( pCut > pCutSet ); return pCut - pCutSet; } static inline int * Nf_CutFromHandle( int * pCutSet, int h ) { assert( h > 0 ); return pCutSet + h; } static inline int Nf_CutConfLit( int Conf, int i ) { return 15 & (Conf >> (i << 2)); } static inline int Nf_CutConfVar( int Conf, int i ) { return Abc_Lit2Var( Nf_CutConfLit(Conf, i) ); } static inline int Nf_CutConfC( int Conf, int i ) { return Abc_LitIsCompl( Nf_CutConfLit(Conf, i) ); } #define Nf_SetForEachCut( pList, pCut, i ) for ( i = 0, pCut = pList + 1; i < pList[0]; i++, pCut += Nf_CutSize(pCut) + 1 ) #define Nf_CutForEachLit( pCut, Conf, iLit, i ) for ( i = 0; i < Nf_CutSize(pCut) && (iLit = Abc_Lit2LitV(Nf_CutLeaves(pCut), Nf_CutConfLit(Conf, i))); i++ ) #define Nf_CutForEachVar( pCut, Conf, iVar, c, i ) for ( i = 0; i < Nf_CutSize(pCut) && (iVar = Nf_CutLeaves(pCut)[Nf_CutConfVar(Conf, i)]) && ((c = Nf_CutConfC(Conf, i)), 1); i++ ) //////////////////////////////////////////////////////////////////////// /// FUNCTION DEFINITIONS /// //////////////////////////////////////////////////////////////////////// /**Function************************************************************* Synopsis [Sort inputs by delay.] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ void Nf_StoCreateGateAdd( Nf_Man_t * pMan, word uTruth, int * pFans, int nFans, int CellId ) { Vec_Int_t * vArray; Pf_Mat_t Mat = Pf_Int2Mat(0); int i, GateId, Entry, fCompl = (int)(uTruth & 1); word uFunc = fCompl ? ~uTruth : uTruth; int iFunc = Vec_MemHashInsert( pMan->vTtMem, &uFunc ); if ( iFunc == Vec_WecSize(pMan->vTt2Match) ) Vec_WecPushLevel( pMan->vTt2Match ); vArray = Vec_WecEntry( pMan->vTt2Match, iFunc ); Mat.fCompl = fCompl; assert( nFans < 7 ); for ( i = 0; i < nFans; i++ ) { Mat.Perm |= (unsigned)(Abc_Lit2Var(pFans[i]) << (3*i)); Mat.Phase |= (unsigned)(Abc_LitIsCompl(pFans[i]) << i); } if ( pMan->pPars->fPinPerm ) // use pin-permutation (slower but good for delay when pin-delays differ) { Vec_IntPush( vArray, CellId ); Vec_IntPush( vArray, Pf_Mat2Int(Mat) ); return; } // check if the same one exists Vec_IntForEachEntryDouble( vArray, GateId, Entry, i ) if ( GateId == CellId && Pf_Int2Mat(Entry).Phase == Mat.Phase ) break; if ( i == Vec_IntSize(vArray) ) { Vec_IntPush( vArray, CellId ); Vec_IntPush( vArray, Pf_Mat2Int(Mat) ); } } void Nf_StoCreateGateMaches( Nf_Man_t * pMan, Mio_Cell_t * pCell, int ** pComp, int ** pPerm, int * pnPerms ) { int Perm[NF_LEAF_MAX], * Perm1, * Perm2; int nPerms = pnPerms[pCell->nFanins]; int nMints = (1 << pCell->nFanins); word tCur, tTemp1, tTemp2; int i, p, c; for ( i = 0; i < (int)pCell->nFanins; i++ ) Perm[i] = Abc_Var2Lit( i, 0 ); tCur = tTemp1 = pCell->uTruth; for ( p = 0; p < nPerms; p++ ) { tTemp2 = tCur; for ( c = 0; c < nMints; c++ ) { Nf_StoCreateGateAdd( pMan, tCur, Perm, pCell->nFanins, pCell->Id ); // update tCur = Abc_Tt6Flip( tCur, pComp[pCell->nFanins][c] ); Perm1 = Perm + pComp[pCell->nFanins][c]; *Perm1 = Abc_LitNot( *Perm1 ); } assert( tTemp2 == tCur ); // update tCur = Abc_Tt6SwapAdjacent( tCur, pPerm[pCell->nFanins][p] ); Perm1 = Perm + pPerm[pCell->nFanins][p]; Perm2 = Perm1 + 1; ABC_SWAP( int, *Perm1, *Perm2 ); } assert( tTemp1 == tCur ); } void Nf_StoDeriveMatches( Nf_Man_t * p, int fVerbose ) { // abctime clk = Abc_Clock(); int * pComp[7]; int * pPerm[7]; int nPerms[7], i; for ( i = 2; i <= 6; i++ ) pComp[i] = Extra_GreyCodeSchedule( i ); for ( i = 2; i <= 6; i++ ) pPerm[i] = Extra_PermSchedule( i ); for ( i = 2; i <= 6; i++ ) nPerms[i] = Extra_Factorial( i ); p->pCells = Mio_CollectRootsNewDefault( 6, &p->nCells, fVerbose ); for ( i = 4; i < p->nCells; i++ ) Nf_StoCreateGateMaches( p, p->pCells + i, pComp, pPerm, nPerms ); for ( i = 2; i <= 6; i++ ) ABC_FREE( pComp[i] ); for ( i = 2; i <= 6; i++ ) ABC_FREE( pPerm[i] ); // Abc_PrintTime( 1, "Time", Abc_Clock() - clk ); } void Nf_StoPrintOne( Nf_Man_t * p, int Count, int t, int i, int GateId, Pf_Mat_t Mat ) { Mio_Cell_t * pC = p->pCells + GateId; word * pTruth = Vec_MemReadEntry(p->vTtMem, t); int k, nSuppSize = Abc_TtSupportSize(pTruth, 6); printf( "%6d : ", Count ); printf( "%6d : ", t ); printf( "%6d : ", i ); printf( "Gate %16s ", pC->pName ); printf( "Area =%8.2f ", pC->Area ); printf( "In = %d ", pC->nFanins ); if ( Mat.fCompl ) printf( " compl " ); else printf( " " ); for ( k = 0; k < (int)pC->nFanins; k++ ) { int fComplF = (Mat.Phase >> k) & 1; int iFanin = (Mat.Perm >> (3*k)) & 7; printf( "%c", 'a' + iFanin - fComplF * ('a' - 'A') ); } printf( " " ); Dau_DsdPrintFromTruth( pTruth, nSuppSize ); } void Nf_StoPrint( Nf_Man_t * p, int fVerbose ) { int t, i, GateId, Entry, Count = 0; for ( t = 2; t < Vec_WecSize(p->vTt2Match); t++ ) { Vec_Int_t * vArr = Vec_WecEntry( p->vTt2Match, t ); Vec_IntForEachEntryDouble( vArr, GateId, Entry, i ) { Count++; if ( !fVerbose ) continue; //if ( t < 10 ) // Nf_StoPrintOne( p, Count, t, i/2, GateId, Pf_Int2Mat(Entry) ); } } printf( "Gates = %d. Truths = %d. Matches = %d.\n", p->nCells, Vec_MemEntryNum(p->vTtMem), Count ); } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ Nf_Man_t * Nf_StoCreate( Gia_Man_t * pGia, Jf_Par_t * pPars ) { extern void Mf_ManSetFlowRefs( Gia_Man_t * p, Vec_Int_t * vRefs ); Vec_Int_t * vFlowRefs; Nf_Man_t * p; int i, Entry; assert( pPars->nCutNum > 1 && pPars->nCutNum <= NF_CUT_MAX ); assert( pPars->nLutSize > 1 && pPars->nLutSize <= NF_LEAF_MAX ); ABC_FREE( pGia->pRefs ); Vec_IntFreeP( &pGia->vCellMapping ); if ( Gia_ManHasChoices(pGia) ) Gia_ManSetPhase(pGia); // create p = ABC_CALLOC( Nf_Man_t, 1 ); p->clkStart = Abc_Clock(); p->pGia = pGia; p->pPars = pPars; p->pNfObjs = ABC_CALLOC( Nf_Obj_t, Gia_ManObjNum(pGia) ); p->iCur = 2; // other Vec_PtrGrow( &p->vPages, 256 ); // cut memory Vec_IntFill( &p->vMapRefs, 2*Gia_ManObjNum(pGia), 0 ); // mapping refs (2x) Vec_FltFill( &p->vFlowRefs, 2*Gia_ManObjNum(pGia), 0 ); // flow refs (2x) Vec_FltFill( &p->vRequired, 2*Gia_ManObjNum(pGia), NF_INFINITY ); // required times (2x) Vec_IntFill( &p->vCutSets, Gia_ManObjNum(pGia), 0 ); // cut offsets Vec_FltFill( &p->vCutFlows, Gia_ManObjNum(pGia), 0 ); // cut area Vec_IntFill( &p->vCutDelays,Gia_ManObjNum(pGia), 0 ); // cut delay Vec_IntGrow( &p->vBackup, 1000 ); Vec_IntGrow( &p->vBackup2, 1000 ); // references vFlowRefs = Vec_IntAlloc(0); Mf_ManSetFlowRefs( pGia, vFlowRefs ); Vec_IntForEachEntry( vFlowRefs, Entry, i ) { Vec_FltWriteEntry( &p->vFlowRefs, 2*i, /*0.5* */Entry ); Vec_FltWriteEntry( &p->vFlowRefs, 2*i+1, /*0.5* */Entry ); } Vec_IntFree(vFlowRefs); // matching p->vTtMem = Vec_MemAllocForTT( 6, 0 ); p->vTt2Match = Vec_WecAlloc( 1000 ); Vec_WecPushLevel( p->vTt2Match ); Vec_WecPushLevel( p->vTt2Match ); assert( Vec_WecSize(p->vTt2Match) == Vec_MemEntryNum(p->vTtMem) ); Nf_StoDeriveMatches( p, 0 );//pPars->fVerbose ); p->InvDelay = p->pCells[3].Delays[0]; p->InvArea = p->pCells[3].Area; Nf_ObjMatchD(p, 0, 0)->Gate = 0; Nf_ObjMatchD(p, 0, 1)->Gate = 1; // prepare cuts return p; } void Nf_StoDelete( Nf_Man_t * p ) { Vec_PtrFreeData( &p->vPages ); ABC_FREE( p->vPages.pArray ); ABC_FREE( p->vMapRefs.pArray ); ABC_FREE( p->vFlowRefs.pArray ); ABC_FREE( p->vRequired.pArray ); ABC_FREE( p->vCutSets.pArray ); ABC_FREE( p->vCutFlows.pArray ); ABC_FREE( p->vCutDelays.pArray ); ABC_FREE( p->vBackup.pArray ); ABC_FREE( p->vBackup2.pArray ); ABC_FREE( p->pNfObjs ); // matching Vec_WecFree( p->vTt2Match ); Vec_MemHashFree( p->vTtMem ); Vec_MemFree( p->vTtMem ); ABC_FREE( p->pCells ); ABC_FREE( p ); } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ static inline int Nf_CutComputeTruth6( Nf_Man_t * p, Nf_Cut_t * pCut0, Nf_Cut_t * pCut1, int fCompl0, int fCompl1, Nf_Cut_t * pCutR, int fIsXor ) { // extern int Nf_ManTruthCanonicize( word * t, int nVars ); int nOldSupp = pCutR->nLeaves, truthId, fCompl; word t; word t0 = *Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCut0->iFunc)); word t1 = *Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCut1->iFunc)); if ( Abc_LitIsCompl(pCut0->iFunc) ^ fCompl0 ) t0 = ~t0; if ( Abc_LitIsCompl(pCut1->iFunc) ^ fCompl1 ) t1 = ~t1; t0 = Abc_Tt6Expand( t0, pCut0->pLeaves, pCut0->nLeaves, pCutR->pLeaves, pCutR->nLeaves ); t1 = Abc_Tt6Expand( t1, pCut1->pLeaves, pCut1->nLeaves, pCutR->pLeaves, pCutR->nLeaves ); t = fIsXor ? t0 ^ t1 : t0 & t1; if ( (fCompl = (int)(t & 1)) ) t = ~t; pCutR->nLeaves = Abc_Tt6MinBase( &t, pCutR->pLeaves, pCutR->nLeaves ); assert( (int)(t & 1) == 0 ); truthId = Vec_MemHashInsert(p->vTtMem, &t); pCutR->iFunc = Abc_Var2Lit( truthId, fCompl ); pCutR->Useless = Nf_ObjCutUseless( p, truthId ); assert( (int)pCutR->nLeaves <= nOldSupp ); return (int)pCutR->nLeaves < nOldSupp; } static inline int Nf_CutComputeTruthMux6( Nf_Man_t * p, Nf_Cut_t * pCut0, Nf_Cut_t * pCut1, Nf_Cut_t * pCutC, int fCompl0, int fCompl1, int fComplC, Nf_Cut_t * pCutR ) { int nOldSupp = pCutR->nLeaves, truthId, fCompl; word t; word t0 = *Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCut0->iFunc)); word t1 = *Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCut1->iFunc)); word tC = *Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCutC->iFunc)); if ( Abc_LitIsCompl(pCut0->iFunc) ^ fCompl0 ) t0 = ~t0; if ( Abc_LitIsCompl(pCut1->iFunc) ^ fCompl1 ) t1 = ~t1; if ( Abc_LitIsCompl(pCutC->iFunc) ^ fComplC ) tC = ~tC; t0 = Abc_Tt6Expand( t0, pCut0->pLeaves, pCut0->nLeaves, pCutR->pLeaves, pCutR->nLeaves ); t1 = Abc_Tt6Expand( t1, pCut1->pLeaves, pCut1->nLeaves, pCutR->pLeaves, pCutR->nLeaves ); tC = Abc_Tt6Expand( tC, pCutC->pLeaves, pCutC->nLeaves, pCutR->pLeaves, pCutR->nLeaves ); t = (tC & t1) | (~tC & t0); if ( (fCompl = (int)(t & 1)) ) t = ~t; pCutR->nLeaves = Abc_Tt6MinBase( &t, pCutR->pLeaves, pCutR->nLeaves ); assert( (int)(t & 1) == 0 ); truthId = Vec_MemHashInsert(p->vTtMem, &t); pCutR->iFunc = Abc_Var2Lit( truthId, fCompl ); pCutR->Useless = Nf_ObjCutUseless( p, truthId ); assert( (int)pCutR->nLeaves <= nOldSupp ); return (int)pCutR->nLeaves < nOldSupp; } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ static inline int Nf_CutCountBits( word i ) { i = i - ((i >> 1) & 0x5555555555555555); i = (i & 0x3333333333333333) + ((i >> 2) & 0x3333333333333333); i = ((i + (i >> 4)) & 0x0F0F0F0F0F0F0F0F); return (i*(0x0101010101010101))>>56; } static inline word Nf_CutGetSign( int * pLeaves, int nLeaves ) { word Sign = 0; int i; for ( i = 0; i < nLeaves; i++ ) Sign |= ((word)1) << (pLeaves[i] & 0x3F); return Sign; } static inline int Nf_CutCreateUnit( Nf_Cut_t * p, int i ) { p->Delay = 0; p->Flow = 0; p->iFunc = 2; p->nLeaves = 1; p->pLeaves[0] = i; p->Sign = ((word)1) << (i & 0x3F); return 1; } static inline void Nf_Cutprintf( Nf_Man_t * p, Nf_Cut_t * pCut ) { int i, nDigits = Abc_Base10Log(Gia_ManObjNum(p->pGia)); printf( "%d {", pCut->nLeaves ); for ( i = 0; i < (int)pCut->nLeaves; i++ ) printf( " %*d", nDigits, pCut->pLeaves[i] ); for ( ; i < (int)p->pPars->nLutSize; i++ ) printf( " %*s", nDigits, " " ); printf( " } Useless = %d. D = %4d A = %9.4f F = %6d ", pCut->Useless, pCut->Delay, pCut->Flow, pCut->iFunc ); if ( p->vTtMem ) Dau_DsdPrintFromTruth( Vec_MemReadEntry(p->vTtMem, Abc_Lit2Var(pCut->iFunc)), pCut->nLeaves ); else printf( "\n" ); } static inline int Nf_ManPrepareCuts( Nf_Cut_t * pCuts, Nf_Man_t * p, int iObj, int fAddUnit ) { if ( Nf_ObjHasCuts(p, iObj) ) { Nf_Cut_t * pMfCut = pCuts; int i, * pCut, * pList = Nf_ObjCutSet(p, iObj); Nf_SetForEachCut( pList, pCut, i ) { pMfCut->Delay = 0; pMfCut->Flow = 0; pMfCut->iFunc = Nf_CutFunc( pCut ); pMfCut->nLeaves = Nf_CutSize( pCut ); pMfCut->Sign = Nf_CutGetSign( pCut+1, Nf_CutSize(pCut) ); pMfCut->Useless = Nf_ObjCutUseless( p, Abc_Lit2Var(pMfCut->iFunc) ); memcpy( pMfCut->pLeaves, pCut+1, sizeof(int) * Nf_CutSize(pCut) ); pMfCut++; } if ( fAddUnit && pCuts->nLeaves > 1 ) return pList[0] + Nf_CutCreateUnit( pMfCut, iObj ); return pList[0]; } return Nf_CutCreateUnit( pCuts, iObj ); } static inline int Nf_ManSaveCuts( Nf_Man_t * p, Nf_Cut_t ** pCuts, int nCuts, int fUseful ) { int i, * pPlace, iCur, nInts = 1, nCutsNew = 0; for ( i = 0; i < nCuts; i++ ) if ( !fUseful || !pCuts[i]->Useless ) nInts += pCuts[i]->nLeaves + 1, nCutsNew++; if ( (p->iCur & 0xFFFF) + nInts > 0xFFFF ) p->iCur = ((p->iCur >> 16) + 1) << 16; if ( Vec_PtrSize(&p->vPages) == (p->iCur >> 16) ) Vec_PtrPush( &p->vPages, ABC_ALLOC(int, (1<<16)) ); iCur = p->iCur; p->iCur += nInts; pPlace = Nf_ManCutSet( p, iCur ); *pPlace++ = nCutsNew; for ( i = 0; i < nCuts; i++ ) if ( !fUseful || !pCuts[i]->Useless ) { *pPlace++ = Nf_CutSetBoth( pCuts[i]->nLeaves, pCuts[i]->iFunc ); memcpy( pPlace, pCuts[i]->pLeaves, sizeof(int) * pCuts[i]->nLeaves ); pPlace += pCuts[i]->nLeaves; } return iCur; } static inline int Nf_ManCountUseful( Nf_Cut_t ** pCuts, int nCuts ) { int i, Count = 0; for ( i = 0; i < nCuts; i++ ) Count += !pCuts[i]->Useless; return Count; } static inline int Nf_ManCountMatches( Nf_Man_t * p, Nf_Cut_t ** pCuts, int nCuts ) { int i, Count = 0; for ( i = 0; i < nCuts; i++ ) if ( !pCuts[i]->Useless ) Count += Vec_IntSize(Vec_WecEntry(p->vTt2Match, Abc_Lit2Var(pCuts[i]->iFunc))) / 2; return Count; } /**Function************************************************************* Synopsis [Check correctness of cuts.] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ static inline int Nf_CutCheck( Nf_Cut_t * pBase, Nf_Cut_t * pCut ) // check if pCut is contained in pBase { int nSizeB = pBase->nLeaves; int nSizeC = pCut->nLeaves; int i, * pB = pBase->pLeaves; int k, * pC = pCut->pLeaves; for ( i = 0; i < nSizeC; i++ ) { for ( k = 0; k < nSizeB; k++ ) if ( pC[i] == pB[k] ) break; if ( k == nSizeB ) return 0; } return 1; } static inline int Nf_SetCheckArray( Nf_Cut_t ** ppCuts, int nCuts ) { Nf_Cut_t * pCut0, * pCut1; int i, k, m, n, Value; assert( nCuts > 0 ); for ( i = 0; i < nCuts; i++ ) { pCut0 = ppCuts[i]; assert( pCut0->nLeaves <= NF_LEAF_MAX ); assert( pCut0->Sign == Nf_CutGetSign(pCut0->pLeaves, pCut0->nLeaves) ); // check duplicates for ( m = 0; m < (int)pCut0->nLeaves; m++ ) for ( n = m + 1; n < (int)pCut0->nLeaves; n++ ) assert( pCut0->pLeaves[m] < pCut0->pLeaves[n] ); // check pairs for ( k = 0; k < nCuts; k++ ) { pCut1 = ppCuts[k]; if ( pCut0 == pCut1 ) continue; // check containments Value = Nf_CutCheck( pCut0, pCut1 ); assert( Value == 0 ); } } return 1; } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ static inline int Nf_CutMergeOrder( Nf_Cut_t * pCut0, Nf_Cut_t * pCut1, Nf_Cut_t * pCut, int nLutSize ) { int nSize0 = pCut0->nLeaves; int nSize1 = pCut1->nLeaves; int i, * pC0 = pCut0->pLeaves; int k, * pC1 = pCut1->pLeaves; int c, * pC = pCut->pLeaves; // the case of the largest cut sizes if ( nSize0 == nLutSize && nSize1 == nLutSize ) { for ( i = 0; i < nSize0; i++ ) { if ( pC0[i] != pC1[i] ) return 0; pC[i] = pC0[i]; } pCut->nLeaves = nLutSize; pCut->iFunc = -1; pCut->Sign = pCut0->Sign | pCut1->Sign; return 1; } // compare two cuts with different numbers i = k = c = 0; if ( nSize0 == 0 ) goto FlushCut1; if ( nSize1 == 0 ) goto FlushCut0; while ( 1 ) { if ( c == nLutSize ) return 0; if ( pC0[i] < pC1[k] ) { pC[c++] = pC0[i++]; if ( i >= nSize0 ) goto FlushCut1; } else if ( pC0[i] > pC1[k] ) { pC[c++] = pC1[k++]; if ( k >= nSize1 ) goto FlushCut0; } else { pC[c++] = pC0[i++]; k++; if ( i >= nSize0 ) goto FlushCut1; if ( k >= nSize1 ) goto FlushCut0; } } FlushCut0: if ( c + nSize0 > nLutSize + i ) return 0; while ( i < nSize0 ) pC[c++] = pC0[i++]; pCut->nLeaves = c; pCut->iFunc = -1; pCut->Sign = pCut0->Sign | pCut1->Sign; return 1; FlushCut1: if ( c + nSize1 > nLutSize + k ) return 0; while ( k < nSize1 ) pC[c++] = pC1[k++]; pCut->nLeaves = c; pCut->iFunc = -1; pCut->Sign = pCut0->Sign | pCut1->Sign; return 1; } static inline int Nf_CutMergeOrderMux( Nf_Cut_t * pCut0, Nf_Cut_t * pCut1, Nf_Cut_t * pCut2, Nf_Cut_t * pCut, int nLutSize ) { int x0, i0 = 0, nSize0 = pCut0->nLeaves, * pC0 = pCut0->pLeaves; int x1, i1 = 0, nSize1 = pCut1->nLeaves, * pC1 = pCut1->pLeaves; int x2, i2 = 0, nSize2 = pCut2->nLeaves, * pC2 = pCut2->pLeaves; int xMin, c = 0, * pC = pCut->pLeaves; while ( 1 ) { x0 = (i0 == nSize0) ? ABC_INFINITY : pC0[i0]; x1 = (i1 == nSize1) ? ABC_INFINITY : pC1[i1]; x2 = (i2 == nSize2) ? ABC_INFINITY : pC2[i2]; xMin = Abc_MinInt( Abc_MinInt(x0, x1), x2 ); if ( xMin == ABC_INFINITY ) break; if ( c == nLutSize ) return 0; pC[c++] = xMin; if (x0 == xMin) i0++; if (x1 == xMin) i1++; if (x2 == xMin) i2++; } pCut->nLeaves = c; pCut->iFunc = -1; pCut->Sign = pCut0->Sign | pCut1->Sign | pCut2->Sign; return 1; } static inline int Nf_SetCutIsContainedOrder( Nf_Cut_t * pBase, Nf_Cut_t * pCut ) // check if pCut is contained in pBase { int i, nSizeB = pBase->nLeaves; int k, nSizeC = pCut->nLeaves; if ( nSizeB == nSizeC ) { for ( i = 0; i < nSizeB; i++ ) if ( pBase->pLeaves[i] != pCut->pLeaves[i] ) return 0; return 1; } assert( nSizeB > nSizeC ); if ( nSizeC == 0 ) return 1; for ( i = k = 0; i < nSizeB; i++ ) { if ( pBase->pLeaves[i] > pCut->pLeaves[k] ) return 0; if ( pBase->pLeaves[i] == pCut->pLeaves[k] ) { if ( ++k == nSizeC ) return 1; } } return 0; } static inline int Nf_SetLastCutIsContained( Nf_Cut_t ** pCuts, int nCuts ) { int i; for ( i = 0; i < nCuts; i++ ) if ( pCuts[i]->nLeaves <= pCuts[nCuts]->nLeaves && (pCuts[i]->Sign & pCuts[nCuts]->Sign) == pCuts[i]->Sign && Nf_SetCutIsContainedOrder(pCuts[nCuts], pCuts[i]) ) return 1; return 0; } static inline int Nf_SetLastCutContainsArea( Nf_Cut_t ** pCuts, int nCuts ) { int i, k, fChanges = 0; for ( i = 0; i < nCuts; i++ ) if ( pCuts[nCuts]->nLeaves < pCuts[i]->nLeaves && (pCuts[nCuts]->Sign & pCuts[i]->Sign) == pCuts[nCuts]->Sign && Nf_SetCutIsContainedOrder(pCuts[i], pCuts[nCuts]) ) pCuts[i]->nLeaves = NF_NO_LEAF, fChanges = 1; if ( !fChanges ) return nCuts; for ( i = k = 0; i <= nCuts; i++ ) { if ( pCuts[i]->nLeaves == NF_NO_LEAF ) continue; if ( k < i ) ABC_SWAP( Nf_Cut_t *, pCuts[k], pCuts[i] ); k++; } return k - 1; } static inline int Nf_CutCompareArea( Nf_Cut_t * pCut0, Nf_Cut_t * pCut1 ) { if ( pCut0->Useless < pCut1->Useless ) return -1; if ( pCut0->Useless > pCut1->Useless ) return 1; if ( pCut0->Flow < pCut1->Flow ) return -1; if ( pCut0->Flow > pCut1->Flow ) return 1; if ( pCut0->Delay < pCut1->Delay ) return -1; if ( pCut0->Delay > pCut1->Delay ) return 1; if ( pCut0->nLeaves < pCut1->nLeaves ) return -1; if ( pCut0->nLeaves > pCut1->nLeaves ) return 1; return 0; } static inline void Nf_SetSortByArea( Nf_Cut_t ** pCuts, int nCuts ) { int i; for ( i = nCuts; i > 0; i-- ) { if ( Nf_CutCompareArea(pCuts[i - 1], pCuts[i]) < 0 )//!= 1 ) return; ABC_SWAP( Nf_Cut_t *, pCuts[i - 1], pCuts[i] ); } } static inline int Nf_SetAddCut( Nf_Cut_t ** pCuts, int nCuts, int nCutNum ) { if ( nCuts == 0 ) return 1; nCuts = Nf_SetLastCutContainsArea(pCuts, nCuts); Nf_SetSortByArea( pCuts, nCuts ); return Abc_MinInt( nCuts + 1, nCutNum - 1 ); } static inline int Nf_CutArea( Nf_Man_t * p, int nLeaves ) { if ( nLeaves < 2 ) return 0; return nLeaves + p->pPars->nAreaTuner; } static inline void Nf_CutParams( Nf_Man_t * p, Nf_Cut_t * pCut, float FlowRefs ) { int i, nLeaves = pCut->nLeaves; assert( nLeaves <= p->pPars->nLutSize ); pCut->Delay = 0; pCut->Flow = 0; for ( i = 0; i < nLeaves; i++ ) { pCut->Delay = Abc_MaxInt( pCut->Delay, Nf_ObjCutDelay(p, pCut->pLeaves[i]) ); pCut->Flow += Nf_ObjCutFlow(p, pCut->pLeaves[i]); } pCut->Delay += (int)(nLeaves > 1); pCut->Flow = (pCut->Flow + Nf_CutArea(p, nLeaves)) / FlowRefs; } void Nf_ObjMergeOrder( Nf_Man_t * p, int iObj ) { Nf_Cut_t pCuts0[NF_CUT_MAX], pCuts1[NF_CUT_MAX], pCuts[NF_CUT_MAX], * pCutsR[NF_CUT_MAX]; Gia_Obj_t * pObj = Gia_ManObj(p->pGia, iObj); //Nf_Obj_t * pBest = Nf_ManObj(p, iObj); float dFlowRefs = Nf_ObjFlowRefs(p, iObj, 0) + Nf_ObjFlowRefs(p, iObj, 1); int nLutSize = p->pPars->nLutSize; int nCutNum = p->pPars->nCutNum; int nCuts0 = Nf_ManPrepareCuts(pCuts0, p, Gia_ObjFaninId0(pObj, iObj), 1); int nCuts1 = Nf_ManPrepareCuts(pCuts1, p, Gia_ObjFaninId1(pObj, iObj), 1); int fComp0 = Gia_ObjFaninC0(pObj); int fComp1 = Gia_ObjFaninC1(pObj); int iSibl = Gia_ObjSibl(p->pGia, iObj); Nf_Cut_t * pCut0, * pCut1, * pCut0Lim = pCuts0 + nCuts0, * pCut1Lim = pCuts1 + nCuts1; int i, nCutsUse, nCutsR = 0; assert( !Gia_ObjIsBuf(pObj) ); for ( i = 0; i < nCutNum; i++ ) pCutsR[i] = pCuts + i; if ( iSibl ) { Nf_Cut_t pCuts2[NF_CUT_MAX]; Gia_Obj_t * pObjE = Gia_ObjSiblObj(p->pGia, iObj); int fCompE = Gia_ObjPhase(pObj) ^ Gia_ObjPhase(pObjE); int nCuts2 = Nf_ManPrepareCuts(pCuts2, p, iSibl, 0); Nf_Cut_t * pCut2, * pCut2Lim = pCuts2 + nCuts2; for ( pCut2 = pCuts2; pCut2 < pCut2Lim; pCut2++ ) { *pCutsR[nCutsR] = *pCut2; pCutsR[nCutsR]->iFunc = Abc_LitNotCond( pCutsR[nCutsR]->iFunc, fCompE ); Nf_CutParams( p, pCutsR[nCutsR], dFlowRefs ); nCutsR = Nf_SetAddCut( pCutsR, nCutsR, nCutNum ); } } if ( Gia_ObjIsMuxId(p->pGia, iObj) ) { Nf_Cut_t pCuts2[NF_CUT_MAX]; int nCuts2 = Nf_ManPrepareCuts(pCuts2, p, Gia_ObjFaninId2(p->pGia, iObj), 1); int fComp2 = Gia_ObjFaninC2(p->pGia, pObj); Nf_Cut_t * pCut2, * pCut2Lim = pCuts2 + nCuts2; p->CutCount[0] += nCuts0 * nCuts1 * nCuts2; for ( pCut0 = pCuts0; pCut0 < pCut0Lim; pCut0++ ) for ( pCut1 = pCuts1; pCut1 < pCut1Lim; pCut1++ ) for ( pCut2 = pCuts2; pCut2 < pCut2Lim; pCut2++ ) { if ( Nf_CutCountBits(pCut0->Sign | pCut1->Sign | pCut2->Sign) > nLutSize ) continue; p->CutCount[1]++; if ( !Nf_CutMergeOrderMux(pCut0, pCut1, pCut2, pCutsR[nCutsR], nLutSize) ) continue; if ( Nf_SetLastCutIsContained(pCutsR, nCutsR) ) continue; p->CutCount[2]++; if ( Nf_CutComputeTruthMux6(p, pCut0, pCut1, pCut2, fComp0, fComp1, fComp2, pCutsR[nCutsR]) ) pCutsR[nCutsR]->Sign = Nf_CutGetSign(pCutsR[nCutsR]->pLeaves, pCutsR[nCutsR]->nLeaves); Nf_CutParams( p, pCutsR[nCutsR], dFlowRefs ); nCutsR = Nf_SetAddCut( pCutsR, nCutsR, nCutNum ); } } else { int fIsXor = Gia_ObjIsXor(pObj); p->CutCount[0] += nCuts0 * nCuts1; for ( pCut0 = pCuts0; pCut0 < pCut0Lim; pCut0++ ) for ( pCut1 = pCuts1; pCut1 < pCut1Lim; pCut1++ ) { if ( (int)(pCut0->nLeaves + pCut1->nLeaves) > nLutSize && Nf_CutCountBits(pCut0->Sign | pCut1->Sign) > nLutSize ) continue; p->CutCount[1]++; if ( !Nf_CutMergeOrder(pCut0, pCut1, pCutsR[nCutsR], nLutSize) ) continue; if ( Nf_SetLastCutIsContained(pCutsR, nCutsR) ) continue; p->CutCount[2]++; if ( Nf_CutComputeTruth6(p, pCut0, pCut1, fComp0, fComp1, pCutsR[nCutsR], fIsXor) ) pCutsR[nCutsR]->Sign = Nf_CutGetSign(pCutsR[nCutsR]->pLeaves, pCutsR[nCutsR]->nLeaves); Nf_CutParams( p, pCutsR[nCutsR], dFlowRefs ); nCutsR = Nf_SetAddCut( pCutsR, nCutsR, nCutNum ); } } // debug printout if ( 0 ) // if ( iObj % 10000 == 0 ) // if ( iObj == 1090 ) { printf( "*** Obj = %d Useful = %d\n", iObj, Nf_ManCountUseful(pCutsR, nCutsR) ); for ( i = 0; i < nCutsR; i++ ) Nf_Cutprintf( p, pCutsR[i] ); printf( "\n" ); } // verify assert( nCutsR > 0 && nCutsR < nCutNum ); // assert( Nf_SetCheckArray(pCutsR, nCutsR) ); // store the cutset Nf_ObjSetCutFlow( p, iObj, pCutsR[0]->Flow ); Nf_ObjSetCutDelay( p, iObj, pCutsR[0]->Delay ); *Vec_IntEntryP(&p->vCutSets, iObj) = Nf_ManSaveCuts(p, pCutsR, nCutsR, 0); p->CutCount[3] += nCutsR; nCutsUse = Nf_ManCountUseful(pCutsR, nCutsR); p->CutCount[4] += nCutsUse; p->nCutUseAll += nCutsUse == nCutsR; p->CutCount[5] += Nf_ManCountMatches(p, pCutsR, nCutsR); } void Nf_ManComputeCuts( Nf_Man_t * p ) { Gia_Obj_t * pObj; int i, iFanin; Gia_ManForEachAnd( p->pGia, pObj, i ) if ( Gia_ObjIsBuf(pObj) ) { iFanin = Gia_ObjFaninId0(pObj, i); Nf_ObjSetCutFlow( p, i, Nf_ObjCutFlow(p, iFanin) ); Nf_ObjSetCutDelay( p, i, Nf_ObjCutDelay(p, iFanin) ); } else Nf_ObjMergeOrder( p, i ); } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ void Nf_ManPrintStats( Nf_Man_t * p, char * pTitle ) { if ( !p->pPars->fVerbose ) return; printf( "%s : ", pTitle ); printf( "Delay =%8.2f ", p->pPars->MapDelay ); printf( "Area =%12.2f ", p->pPars->MapArea ); printf( "Gate =%6d ", (int)p->pPars->Area ); printf( "Inv =%6d ", (int)p->nInvs ); printf( "Edge =%7d ", (int)p->pPars->Edge ); Abc_PrintTime( 1, "Time", Abc_Clock() - p->clkStart ); fflush( stdout ); } void Nf_ManPrintInit( Nf_Man_t * p ) { int nChoices; if ( !p->pPars->fVerbose ) return; printf( "LutSize = %d ", p->pPars->nLutSize ); printf( "CutNum = %d ", p->pPars->nCutNum ); printf( "Iter = %d ", p->pPars->nRounds );//+ p->pPars->nRoundsEla ); printf( "Coarse = %d ", p->pPars->fCoarsen ); printf( "Cells = %d ", p->nCells ); printf( "Funcs = %d ", Vec_MemEntryNum(p->vTtMem) ); printf( "Matches = %d ", Vec_WecSizeSize(p->vTt2Match)/2 ); nChoices = Gia_ManChoiceNum( p->pGia ); if ( nChoices ) printf( "Choices = %d ", nChoices ); printf( "\n" ); printf( "Computing cuts...\r" ); fflush( stdout ); } void Nf_ManPrintQuit( Nf_Man_t * p ) { float MemGia = Gia_ManMemory(p->pGia) / (1<<20); float MemMan =(1.0 * sizeof(Nf_Obj_t) + 8.0 * sizeof(int)) * Gia_ManObjNum(p->pGia) / (1<<20); float MemCuts = 1.0 * sizeof(int) * (1 << 16) * Vec_PtrSize(&p->vPages) / (1<<20); float MemTt = p->vTtMem ? Vec_MemMemory(p->vTtMem) / (1<<20) : 0; if ( p->CutCount[0] == 0 ) p->CutCount[0] = 1; if ( !p->pPars->fVerbose ) return; printf( "CutPair = %.0f ", p->CutCount[0] ); printf( "Merge = %.0f (%.1f) ", p->CutCount[1], 1.0*p->CutCount[1]/Gia_ManAndNum(p->pGia) ); printf( "Eval = %.0f (%.1f) ", p->CutCount[2], 1.0*p->CutCount[2]/Gia_ManAndNum(p->pGia) ); printf( "Cut = %.0f (%.1f) ", p->CutCount[3], 1.0*p->CutCount[3]/Gia_ManAndNum(p->pGia) ); printf( "Use = %.0f (%.1f) ", p->CutCount[4], 1.0*p->CutCount[4]/Gia_ManAndNum(p->pGia) ); printf( "Mat = %.0f (%.1f) ", p->CutCount[5], 1.0*p->CutCount[5]/Gia_ManAndNum(p->pGia) ); // printf( "Equ = %d (%.2f %%) ", p->nCutUseAll, 100.0*p->nCutUseAll /p->CutCount[0] ); printf( "\n" ); printf( "Gia = %.2f MB ", MemGia ); printf( "Man = %.2f MB ", MemMan ); printf( "Cut = %.2f MB ", MemCuts ); printf( "TT = %.2f MB ", MemTt ); printf( "Total = %.2f MB ", MemGia + MemMan + MemCuts + MemTt ); // printf( "\n" ); Abc_PrintTime( 1, "Time", Abc_Clock() - p->clkStart ); fflush( stdout ); } /**Function************************************************************* Synopsis [Technology mappping.] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ float Nf_MatchDeref2_rec( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM ) { int k, iVar, fCompl, * pCut; float Area = 0; if ( pM->fCompl ) { assert( Nf_ObjMapRefNum(p, i, !c) > 0 ); if ( !Nf_ObjMapRefDec(p, i, !c) ) Area += Nf_MatchDeref2_rec( p, i, !c, Nf_ObjMatchBest(p, i, !c) ); return Area + p->InvArea; } if ( Nf_ObjCutSetId(p, i) == 0 ) return 0; pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); Nf_CutForEachVar( pCut, pM->Conf, iVar, fCompl, k ) { assert( Nf_ObjMapRefNum(p, iVar, fCompl) > 0 ); if ( !Nf_ObjMapRefDec(p, iVar, fCompl) ) Area += Nf_MatchDeref2_rec( p, iVar, fCompl, Nf_ObjMatchBest(p, iVar, fCompl) ); } return Area + Nf_ManCell(p, pM->Gate)->Area; } float Nf_MatchRef2_rec( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM, Vec_Int_t * vBackup ) { int k, iVar, fCompl, * pCut; float Area = 0; if ( pM->fCompl ) { if ( vBackup ) Vec_IntPush( vBackup, Abc_Var2Lit(i, !c) ); assert( Nf_ObjMapRefNum(p, i, !c) >= 0 ); if ( !Nf_ObjMapRefInc(p, i, !c) ) Area += Nf_MatchRef2_rec( p, i, !c, Nf_ObjMatchBest(p, i, !c), vBackup ); return Area + p->InvArea; } if ( Nf_ObjCutSetId(p, i) == 0 ) return 0; pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); Nf_CutForEachVar( pCut, pM->Conf, iVar, fCompl, k ) { if ( vBackup ) Vec_IntPush( vBackup, Abc_Var2Lit(iVar, fCompl) ); assert( Nf_ObjMapRefNum(p, iVar, fCompl) >= 0 ); if ( !Nf_ObjMapRefInc(p, iVar, fCompl) ) Area += Nf_MatchRef2_rec( p, iVar, fCompl, Nf_ObjMatchBest(p, iVar, fCompl), vBackup ); } return Area + Nf_ManCell(p, pM->Gate)->Area; } float Nf_MatchRef2Area( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM ) { float Area; int iLit, k; Vec_IntClear( &p->vBackup ); Area = Nf_MatchRef2_rec( p, i, c, pM, &p->vBackup ); Vec_IntForEachEntry( &p->vBackup, iLit, k ) { assert( Nf_ObjMapRefNum(p, Abc_Lit2Var(iLit), Abc_LitIsCompl(iLit)) > 0 ); Nf_ObjMapRefDec( p, Abc_Lit2Var(iLit), Abc_LitIsCompl(iLit) ); } return Area; } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ void Nf_ManCutMatchPrint( Nf_Man_t * p, int iObj, int fCompl, Nf_Mat_t * pM ) { Mio_Cell_t * pCell; int i, * pCut; printf( "%5d %d : ", iObj, fCompl ); if ( pM->CutH == 0 ) { printf( "Unassigned\n" ); return; } pCell = Nf_ManCell( p, pM->Gate ); pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, iObj), pM->CutH ); printf( "D = %8.2f ", pM->D ); printf( "A = %8.2f ", pM->A ); printf( "C = %d ", pM->fCompl ); // printf( "B = %d ", pM->fBest ); printf( " " ); printf( "Cut = {" ); for ( i = 0; i < (int)pCell->nFanins; i++ ) printf( "%5d ", Nf_CutLeaves(pCut)[i] ); for ( ; i < 6; i++ ) printf( " " ); printf( "} " ); printf( "%12s ", pCell->pName ); printf( "%d ", pCell->nFanins ); printf( "{" ); for ( i = 0; i < (int)pCell->nFanins; i++ ) printf( "%7.2f ", pCell->Delays[i] ); for ( ; i < 6; i++ ) printf( " " ); printf( " } " ); for ( i = 0; i < (int)pCell->nFanins; i++ ) printf( "%2d ", Nf_CutConfLit(pM->Conf, i) ); for ( ; i < 6; i++ ) printf( " " ); Dau_DsdPrintFromTruth( &pCell->uTruth, pCell->nFanins ); } void Nf_ManCutMatchOne( Nf_Man_t * p, int iObj, int * pCut, int * pCutSet ) { Nf_Obj_t * pBest = Nf_ManObj(p, iObj); int * pFans = Nf_CutLeaves(pCut); int nFans = Nf_CutSize(pCut); int iFuncLit = Nf_CutFunc(pCut); int fComplExt = Abc_LitIsCompl(iFuncLit); float Epsilon = p->pPars->Epsilon; Vec_Int_t * vArr = Vec_WecEntry( p->vTt2Match, Abc_Lit2Var(iFuncLit) ); int i, k, c, Info, Offset, iFanin, fComplF; float ArrivalD, ArrivalA; Nf_Mat_t * pD, * pA; // assign fanins matches Nf_Obj_t * pBestF[NF_LEAF_MAX]; for ( i = 0; i < nFans; i++ ) pBestF[i] = Nf_ManObj( p, pFans[i] ); // special cases if ( nFans == 0 ) { int Const = (iFuncLit == 1); assert( iFuncLit == 0 || iFuncLit == 1 ); for ( c = 0; c < 2; c++ ) { pD = Nf_ObjMatchD( p, iObj, c ); pA = Nf_ObjMatchA( p, iObj, c ); pD->D = pA->D = 0; pD->A = pA->A = p->pCells[c ^ Const].Area; pD->CutH = pA->CutH = Nf_CutHandle(pCutSet, pCut); pD->Gate = pA->Gate = c ^ Const; pD->Conf = pA->Conf = 0; } return; } if ( nFans == 1 ) { int Const = (iFuncLit == 3); assert( iFuncLit == 2 || iFuncLit == 3 ); for ( c = 0; c < 2; c++ ) { pD = Nf_ObjMatchD( p, iObj, c ); pA = Nf_ObjMatchA( p, iObj, c ); pD->D = pA->D = pBestF[0]->M[c ^ !Const][0].D + p->pCells[2 + (c ^ Const)].Delays[0]; pD->A = pA->A = pBestF[0]->M[c ^ !Const][0].A + p->pCells[2 + (c ^ Const)].Area; pD->CutH = pA->CutH = Nf_CutHandle(pCutSet, pCut); pD->Gate = pA->Gate = 2 + (c ^ Const); pD->Conf = pA->Conf = 0; } return; } // consider matches of this function Vec_IntForEachEntryDouble( vArr, Info, Offset, i ) { Pf_Mat_t Mat = Pf_Int2Mat(Offset); Mio_Cell_t* pC = Nf_ManCell( p, Info ); int fCompl = Mat.fCompl ^ fComplExt; float Required = Nf_ObjRequired( p, iObj, fCompl ); Nf_Mat_t * pD = &pBest->M[fCompl][0]; Nf_Mat_t * pA = &pBest->M[fCompl][1]; float Area = pC->Area, Delay = 0; assert( nFans == (int)pC->nFanins ); //char * pInfo = Vec_StrEntryP( p->vMemStore, Offset ); // for ( k = 0; k < nFans; k++ ) // pInfo[k] = (char)Abc_Var2Lit( (Mat.Perm >> (3*k)) & 7, (Mat.Phase >> k) & 1 ); for ( k = 0; k < nFans; k++ ) { // iFanin = Abc_Lit2Var((int)pInfo[k]); // fComplF = Abc_LitIsCompl((int)pInfo[k]); iFanin = (Mat.Perm >> (3*k)) & 7; fComplF = (Mat.Phase >> k) & 1; ArrivalD = pBestF[k]->M[fComplF][0].D; ArrivalA = pBestF[k]->M[fComplF][1].D; if ( ArrivalA + pC->Delays[iFanin] < Required + Epsilon && Required != NF_INFINITY ) { Delay = Abc_MaxFloat( Delay, ArrivalA + pC->Delays[iFanin] ); Area += pBestF[k]->M[fComplF][1].A; } else { // assert( ArrivalD + pC->Delays[iFanin] < Required + Epsilon ); if ( pD->D < NF_INFINITY && pA->D < NF_INFINITY && ArrivalD + pC->Delays[iFanin] >= Required + Epsilon ) break; Delay = Abc_MaxFloat( Delay, ArrivalD + pC->Delays[iFanin] ); Area += pBestF[k]->M[fComplF][0].A; } } if ( k < nFans ) continue; // select best match if ( pD->D > Delay )//+ Epsilon ) { pD->D = Delay; pD->A = Area; pD->CutH = Nf_CutHandle(pCutSet, pCut); pD->Gate = pC->Id; pD->Conf = 0; for ( k = 0; k < nFans; k++ ) // pD->Conf |= ((int)pInfo[k] << (k << 2)); // pD->Conf |= (Abc_Var2Lit(k, Abc_LitIsCompl((int)pInfo[k])) << (Abc_Lit2Var((int)pInfo[k]) << 2)); pD->Conf |= (Abc_Var2Lit(k, (Mat.Phase >> k) & 1) << (((Mat.Perm >> (3*k)) & 7) << 2)); } if ( pA->A > Area )//+ Epsilon ) { pA->D = Delay; pA->A = Area; pA->CutH = Nf_CutHandle(pCutSet, pCut); pA->Gate = pC->Id; pA->Conf = 0; for ( k = 0; k < nFans; k++ ) // pA->Conf |= ((int)pInfo[k] << (k << 2)); // pA->Conf |= (Abc_Var2Lit(k, Abc_LitIsCompl((int)pInfo[k])) << (Abc_Lit2Var((int)pInfo[k]) << 2)); pA->Conf |= (Abc_Var2Lit(k, (Mat.Phase >> k) & 1) << (((Mat.Perm >> (3*k)) & 7) << 2)); } } /* Nf_ManCutMatchPrint( p, iObj, 0, &pBest->M[0][0] ); Nf_ManCutMatchPrint( p, iObj, 0, &pBest->M[0][1] ); Nf_ManCutMatchPrint( p, iObj, 1, &pBest->M[1][0] ); Nf_ManCutMatchPrint( p, iObj, 1, &pBest->M[1][1] ); */ } static inline void Nf_ObjPrepareCi( Nf_Man_t * p, int iObj ) { Nf_Mat_t * pD = Nf_ObjMatchD( p, iObj, 1 ); Nf_Mat_t * pA = Nf_ObjMatchA( p, iObj, 1 ); pD->fCompl = 1; pD->D = p->InvDelay; pD->A = p->InvArea; pA->fCompl = 1; pA->D = p->InvDelay; pA->A = p->InvArea; Nf_ObjMatchD( p, iObj, 0 )->fBest = 1; Nf_ObjMatchD( p, iObj, 1 )->fBest = 1; } static inline void Nf_ObjPrepareBuf( Nf_Man_t * p, Gia_Obj_t * pObj ) { // get fanin info int iObj = Gia_ObjId( p->pGia, pObj ); int iFanin = Gia_ObjFaninId0( pObj, iObj ); Nf_Mat_t * pDf = Nf_ObjMatchD( p, iFanin, Gia_ObjFaninC0(pObj) ); //Nf_Mat_t * pAf = Nf_ObjMatchA( p, iFanin, Gia_ObjFaninC0(pObj) ); // set the direct phase Nf_Mat_t * pDp = Nf_ObjMatchD( p, iObj, 0 ); Nf_Mat_t * pAp = Nf_ObjMatchA( p, iObj, 0 ); Nf_Mat_t * pDn = Nf_ObjMatchD( p, iObj, 1 ); Nf_Mat_t * pAn = Nf_ObjMatchA( p, iObj, 1 ); assert( Gia_ObjIsBuf(pObj) ); memset( Nf_ManObj(p, iObj), 0, sizeof(Nf_Obj_t) ); // set the direct phase pDp->D = pAp->D = pDf->D; pDp->A = pAp->A = pDf->A; // do not pass flow??? pDp->fBest = 1; // set the inverted phase pDn->D = pAn->D = pDf->D + p->InvDelay; pDn->A = pAn->A = pDf->A + p->InvArea; pDn->fCompl = pAn->fCompl = 1; pDn->fBest = 1; } static inline float Nf_CutRequired( Nf_Man_t * p, Nf_Mat_t * pM, int * pCutSet ) { Mio_Cell_t * pCell = Nf_ManCell( p, pM->Gate ); int * pCut = Nf_CutFromHandle( pCutSet, pM->CutH ); int * pFans = Nf_CutLeaves(pCut); int i, nFans = Nf_CutSize(pCut); float Arrival = 0, Required = 0; for ( i = 0; i < nFans; i++ ) { int iLit = Nf_CutConfLit( pM->Conf, i ); int iFanin = pFans[ Abc_Lit2Var(iLit) ]; int fCompl = Abc_LitIsCompl( iLit ); float Arr = Nf_ManObj(p, iFanin)->M[fCompl][0].D + pCell->Delays[i]; float Req = Nf_ObjRequired(p, iFanin, fCompl); Arrival = Abc_MaxInt( Arrival, Arr ); if ( Req < NF_INFINITY ) Required = Abc_MaxInt( Required, Req + pCell->Delays[i] ); } return Abc_MaxFloat( Required + 2*p->InvDelay, Arrival ); } static inline void Nf_ObjComputeRequired( Nf_Man_t * p, int iObj ) { Nf_Obj_t * pBest = Nf_ManObj(p, iObj); int c, * pCutSet = Nf_ObjCutSet( p, iObj ); for ( c = 0; c < 2; c++ ) if ( Nf_ObjRequired(p, iObj, c) == NF_INFINITY ) Nf_ObjSetRequired( p, iObj, c, Nf_CutRequired(p, &pBest->M[c][0], pCutSet) ); } void Nf_ManCutMatch( Nf_Man_t * p, int iObj ) { Nf_Obj_t * pBest = Nf_ManObj(p, iObj); Nf_Mat_t * pDp = &pBest->M[0][0]; Nf_Mat_t * pDn = &pBest->M[1][0]; Nf_Mat_t * pAp = &pBest->M[0][1]; Nf_Mat_t * pAn = &pBest->M[1][1]; float FlowRefP = Nf_ObjFlowRefs(p, iObj, 0); float FlowRefN = Nf_ObjFlowRefs(p, iObj, 1); float Epsilon = p->pPars->Epsilon; int i, Index, * pCut, * pCutSet = Nf_ObjCutSet( p, iObj ); float ValueBeg[2] = {0}, ValueEnd[2] = {0}, Required[2] = {0}; if ( p->Iter ) { Nf_ObjComputeRequired( p, iObj ); Required[0] = Nf_ObjRequired( p, iObj, 0 ); Required[1] = Nf_ObjRequired( p, iObj, 1 ); } if ( p->fUseEla && Nf_ObjMapRefNum(p, iObj, 0) > 0 ) ValueBeg[0] = Nf_MatchDeref2_rec( p, iObj, 0, Nf_ObjMatchBest(p, iObj, 0) ); if ( p->fUseEla && Nf_ObjMapRefNum(p, iObj, 1) > 0 ) ValueBeg[1] = Nf_MatchDeref2_rec( p, iObj, 1, Nf_ObjMatchBest(p, iObj, 1) ); memset( pBest, 0, sizeof(Nf_Obj_t) ); pDp->D = pDp->A = NF_INFINITY; pDn->D = pDn->A = NF_INFINITY; pAp->D = pAp->A = NF_INFINITY; pAn->D = pAn->A = NF_INFINITY; Nf_SetForEachCut( pCutSet, pCut, i ) { if ( Abc_Lit2Var(Nf_CutFunc(pCut)) >= Vec_WecSize(p->vTt2Match) ) continue; assert( !Nf_CutIsTriv(pCut, iObj) ); assert( Nf_CutSize(pCut) <= p->pPars->nLutSize ); assert( Abc_Lit2Var(Nf_CutFunc(pCut)) < Vec_WecSize(p->vTt2Match) ); Nf_ManCutMatchOne( p, iObj, pCut, pCutSet ); } /* if ( 18687 == iObj ) { printf( "Obj %6d (%f %f):\n", iObj, Required[0], Required[1] ); Nf_ManCutMatchPrint( p, iObj, 0, &pBest->M[0][0] ); Nf_ManCutMatchPrint( p, iObj, 0, &pBest->M[0][1] ); Nf_ManCutMatchPrint( p, iObj, 1, &pBest->M[1][0] ); Nf_ManCutMatchPrint( p, iObj, 1, &pBest->M[1][1] ); printf( "\n" ); } */ // divide by ref count pDp->A /= FlowRefP; pAp->A /= FlowRefP; pDn->A /= FlowRefN; pAn->A /= FlowRefN; // add the inverters //assert( pDp->D < NF_INFINITY || pDn->D < NF_INFINITY ); if ( pDp->D > pDn->D + p->InvDelay + Epsilon ) { *pDp = *pDn; pDp->D += p->InvDelay; pDp->A += p->InvArea; pDp->fCompl = 1; if ( pAp->D == NF_INFINITY ) *pAp = *pDp; //printf( "Using inverter to improve delay at node %d in phase %d.\n", iObj, 1 ); } else if ( pDn->D > pDp->D + p->InvDelay + Epsilon ) { *pDn = *pDp; pDn->D += p->InvDelay; pDn->A += p->InvArea; pDn->fCompl = 1; if ( pAn->D == NF_INFINITY ) *pAn = *pDn; //printf( "Using inverter to improve delay at node %d in phase %d.\n", iObj, 0 ); } //assert( pAp->A < NF_INFINITY || pAn->A < NF_INFINITY ); // try replacing pos with neg if ( pAp->D == NF_INFINITY || (pAp->A > pAn->A + p->InvArea + Epsilon && pAn->D + p->InvDelay + Epsilon < Required[1]) ) { assert( p->Iter > 0 ); *pAp = *pAn; pAp->D += p->InvDelay; pAp->A += p->InvArea; pAp->fCompl = 1; if ( pDp->D == NF_INFINITY ) *pDp = *pAp; //printf( "Using inverter to improve area at node %d in phase %d.\n", iObj, 1 ); } // try replacing neg with pos else if ( pAn->D == NF_INFINITY || (pAn->A > pAp->A + p->InvArea + Epsilon && pAp->D + p->InvDelay + Epsilon < Required[0]) ) { assert( p->Iter > 0 ); *pAn = *pAp; pAn->D += p->InvDelay; pAn->A += p->InvArea; pAn->fCompl = 1; if ( pDn->D == NF_INFINITY ) *pDn = *pAn; //printf( "Using inverter to improve area at node %d in phase %d.\n", iObj, 0 ); } if ( pDp->D == NF_INFINITY ) printf( "Object %d has pDp unassigned.\n", iObj ); if ( pDn->D == NF_INFINITY ) printf( "Object %d has pDn unassigned.\n", iObj ); if ( pAp->D == NF_INFINITY ) printf( "Object %d has pAp unassigned.\n", iObj ); if ( pAn->D == NF_INFINITY ) printf( "Object %d has pAn unassigned.\n", iObj ); pDp->A = Abc_MinFloat( pDp->A, NF_INFINITY/1000000 ); pDn->A = Abc_MinFloat( pDn->A, NF_INFINITY/1000000 ); pAp->A = Abc_MinFloat( pAp->A, NF_INFINITY/1000000 ); pAn->A = Abc_MinFloat( pAn->A, NF_INFINITY/1000000 ); assert( pDp->D < NF_INFINITY ); assert( pDn->D < NF_INFINITY ); assert( pAp->D < NF_INFINITY ); assert( pAn->D < NF_INFINITY ); assert( pDp->A < NF_INFINITY ); assert( pDn->A < NF_INFINITY ); assert( pAp->A < NF_INFINITY ); assert( pAn->A < NF_INFINITY ); //printf( "%16f %16f %16f %16f\n", pDp->A, pDn->A, pAp->A, pAn->A ); // assert ( pDp->A < 1000 ); if ( p->fUseEla ) { // set the first good cut Index = (pAp->D != NF_INFINITY && pAp->D < Nf_ObjRequired(p, iObj, 0) + Epsilon); assert( !pDp->fBest && !pAp->fBest ); pBest->M[0][Index].fBest = 1; assert( pDp->fBest != pAp->fBest ); // set the second good cut Index = (pAn->D != NF_INFINITY && pAn->D < Nf_ObjRequired(p, iObj, 1) + Epsilon); assert( !pDn->fBest && !pAn->fBest ); pBest->M[1][Index].fBest = 1; assert( pDn->fBest != pAn->fBest ); // reference if needed if ( Nf_ObjMapRefNum(p, iObj, 0) > 0 ) ValueEnd[0] = Nf_MatchRef2_rec( p, iObj, 0, Nf_ObjMatchBest(p, iObj, 0), NULL ); if ( Nf_ObjMapRefNum(p, iObj, 1) > 0 ) ValueEnd[1] = Nf_MatchRef2_rec( p, iObj, 1, Nf_ObjMatchBest(p, iObj, 1), NULL ); // assert( ValueBeg[0] > ValueEnd[0] - Epsilon ); // assert( ValueBeg[1] > ValueEnd[1] - Epsilon ); } } void Nf_ManComputeMapping( Nf_Man_t * p ) { Gia_Obj_t * pObj; int i; Gia_ManForEachAnd( p->pGia, pObj, i ) if ( Gia_ObjIsBuf(pObj) ) Nf_ObjPrepareBuf( p, pObj ); else Nf_ManCutMatch( p, i ); } /**Function************************************************************* Synopsis [] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ void Nf_ManSetMapRefsGate( Nf_Man_t * p, int iObj, float Required, Nf_Mat_t * pM ) { int k, iVar, fCompl; Mio_Cell_t * pCell = Nf_ManCell( p, pM->Gate ); int * pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, iObj), pM->CutH ); Nf_CutForEachVar( pCut, pM->Conf, iVar, fCompl, k ) { Nf_ObjMapRefInc( p, iVar, fCompl ); Nf_ObjUpdateRequired( p, iVar, fCompl, Required - pCell->Delays[k] ); } assert( Nf_CutSize(pCut) == (int)pCell->nFanins ); // update global stats p->pPars->MapArea += pCell->Area; p->pPars->Edge += Nf_CutSize(pCut); p->pPars->Area++; // update status of the gate assert( pM->fBest == 0 ); pM->fBest = 1; } int Nf_ManSetMapRefs( Nf_Man_t * p ) { float Coef = 1.0 / (1.0 + (p->Iter + 1) * (p->Iter + 1)); float * pFlowRefs = Vec_FltArray( &p->vFlowRefs ); int * pMapRefs = Vec_IntArray( &p->vMapRefs ); float Epsilon = p->pPars->Epsilon; int nLits = 2*Gia_ManObjNum(p->pGia); int i, c, Id, nRefs[2]; Nf_Mat_t * pD, * pA, * pM; Nf_Mat_t * pDs[2], * pAs[2], * pMs[2]; Gia_Obj_t * pObj; float Required = 0, Requireds[2]; // check references assert( !p->fUseEla ); memset( pMapRefs, 0, sizeof(int) * nLits ); Vec_FltFill( &p->vRequired, nLits, NF_INFINITY ); // for ( i = 0; i < Gia_ManObjNum(p->pGia); i++ ) // assert( !Nf_ObjMapRefNum(p, i, 0) && !Nf_ObjMapRefNum(p, i, 1) ); // compute delay p->pPars->MapDelay = 0; Gia_ManForEachCo( p->pGia, pObj, i ) { Required = Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) )->D; if ( Required == NF_INFINITY ) { Nf_ManCutMatchPrint( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj), Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) ) ); } p->pPars->MapDelay = Abc_MaxFloat( p->pPars->MapDelay, Required ); } // check delay target if ( p->pPars->MapDelayTarget == -1 && p->pPars->nRelaxRatio ) p->pPars->MapDelayTarget = (int)((float)p->pPars->MapDelay * (100.0 + p->pPars->nRelaxRatio) / 100.0); if ( p->pPars->MapDelayTarget != -1 ) { if ( p->pPars->MapDelay < p->pPars->MapDelayTarget + Epsilon ) p->pPars->MapDelay = p->pPars->MapDelayTarget; else if ( p->pPars->nRelaxRatio == 0 ) Abc_Print( 0, "Relaxing user-specified delay target from %.2f to %.2f.\n", p->pPars->MapDelayTarget, p->pPars->MapDelay ); } // set required times Gia_ManForEachCo( p->pGia, pObj, i ) { Required = Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) )->D; Required = p->pPars->fDoAverage ? Required * (100.0 + p->pPars->nRelaxRatio) / 100.0 : p->pPars->MapDelay; Nf_ObjUpdateRequired( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj), Required ); Nf_ObjMapRefInc( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj)); } // compute area and edges p->nInvs = 0; p->pPars->MapArea = 0; p->pPars->Area = p->pPars->Edge = 0; Gia_ManForEachAndReverse( p->pGia, pObj, i ) { if ( Gia_ObjIsBuf(pObj) ) { if ( Nf_ObjMapRefNum(p, i, 1) ) { Nf_ObjMapRefInc( p, i, 0 ); Nf_ObjUpdateRequired( p, i, 0, Nf_ObjRequired(p, i, 1) - p->InvDelay ); p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; p->nInvs++; } Nf_ObjUpdateRequired( p, Gia_ObjFaninId0(pObj, i), Gia_ObjFaninC0(pObj), Nf_ObjRequired(p, i, 0) ); Nf_ObjMapRefInc( p, Gia_ObjFaninId0(pObj, i), Gia_ObjFaninC0(pObj)); continue; } // skip if this node is not used for ( c = 0; c < 2; c++ ) { nRefs[c] = Nf_ObjMapRefNum(p, i, c); //if ( Nf_ObjMatchD( p, i, c )->fCompl ) // printf( "Match D of node %d has inv in phase %d.\n", i, c ); //if ( Nf_ObjMatchA( p, i, c )->fCompl ) // printf( "Match A of node %d has inv in phase %d.\n", i, c ); } if ( !nRefs[0] && !nRefs[1] ) continue; // consider two cases if ( nRefs[0] && nRefs[1] ) { // find best matches for both phases for ( c = 0; c < 2; c++ ) { Requireds[c] = Nf_ObjRequired( p, i, c ); //assert( Requireds[c] < NF_INFINITY ); pDs[c] = Nf_ObjMatchD( p, i, c ); pAs[c] = Nf_ObjMatchA( p, i, c ); pMs[c] = (pAs[c]->D < Requireds[c] + Epsilon) ? pAs[c] : pDs[c]; } // swap complemented matches if ( pMs[0]->fCompl && pMs[1]->fCompl ) { pMs[0]->fCompl = pMs[1]->fCompl = 0; ABC_SWAP( Nf_Mat_t *, pMs[0], pMs[1] ); } // check if intervers are involved if ( !pMs[0]->fCompl && !pMs[1]->fCompl ) { // no inverters for ( c = 0; c < 2; c++ ) Nf_ManSetMapRefsGate( p, i, Requireds[c], pMs[c] ); } else { // one interver assert( !pMs[0]->fCompl || !pMs[1]->fCompl ); c = pMs[1]->fCompl; assert( pMs[c]->fCompl && !pMs[!c]->fCompl ); //printf( "Using inverter at node %d in phase %d\n", i, c ); // update this phase phase pM = pMs[c]; pM->fBest = 1; Required = Requireds[c]; // update opposite phase Nf_ObjMapRefInc( p, i, !c ); Nf_ObjUpdateRequired( p, i, !c, Required - p->InvDelay ); // select oppositve phase Required = Nf_ObjRequired( p, i, !c ); //assert( Required < NF_INFINITY ); pD = Nf_ObjMatchD( p, i, !c ); pA = Nf_ObjMatchA( p, i, !c ); pM = (pA->D < Required + Epsilon) ? pA : pD; assert( !pM->fCompl ); // account for the inverter p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; p->nInvs++; // create gate Nf_ManSetMapRefsGate( p, i, Required, pM ); } } else { c = (int)(nRefs[1] > 0); assert( nRefs[c] && !nRefs[!c] ); // consider this phase Required = Nf_ObjRequired( p, i, c ); //assert( Required < NF_INFINITY ); pD = Nf_ObjMatchD( p, i, c ); pA = Nf_ObjMatchA( p, i, c ); pM = (pA->D < Required + Epsilon) ? pA : pD; if ( pM->fCompl ) // use inverter { p->nInvs++; //printf( "Using inverter at node %d in phase %d\n", i, c ); pM->fBest = 1; // update opposite phase Nf_ObjMapRefInc( p, i, !c ); Nf_ObjUpdateRequired( p, i, !c, Required - p->InvDelay ); // select oppositve phase Required = Nf_ObjRequired( p, i, !c ); //assert( Required < NF_INFINITY ); pD = Nf_ObjMatchD( p, i, !c ); pA = Nf_ObjMatchA( p, i, !c ); pM = (pA->D < Required + Epsilon) ? pA : pD; assert( !pM->fCompl ); // account for the inverter p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; } // create gate Nf_ManSetMapRefsGate( p, i, Required, pM ); } // the result of this: // - only one phase can be implemented as inverter of the other phase // - required times are propagated correctly // - references are set correctly } Gia_ManForEachCiId( p->pGia, Id, i ) if ( Nf_ObjMapRefNum(p, Id, 1) ) { Nf_ObjMapRefInc( p, Id, 0 ); Nf_ObjUpdateRequired( p, Id, 0, Required - p->InvDelay ); p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; p->nInvs++; } // blend references for ( i = 0; i < nLits; i++ ) // pFlowRefs[i] = Abc_MaxFloat(1.0, pMapRefs[i]); pFlowRefs[i] = Abc_MaxFloat(1.0, Coef * pFlowRefs[i] + (1.0 - Coef) * Abc_MaxFloat(1, pMapRefs[i])); // pFlowRefs[i] = 0.2 * pFlowRefs[i] + 0.8 * Abc_MaxFloat(1, pMapRefs[i]); // memset( pMapRefs, 0, sizeof(int) * nLits ); return p->pPars->Area; } Gia_Man_t * Nf_ManDeriveMapping( Nf_Man_t * p ) { Vec_Int_t * vMapping; Nf_Mat_t * pM; int i, k, c, Id, iLit, * pCut; assert( p->pGia->vCellMapping == NULL ); vMapping = Vec_IntAlloc( 2*Gia_ManObjNum(p->pGia) + (int)p->pPars->Edge + (int)p->pPars->Area * 2 ); Vec_IntFill( vMapping, 2*Gia_ManObjNum(p->pGia), 0 ); // create CI inverters Gia_ManForEachCiId( p->pGia, Id, i ) if ( Nf_ObjMapRefNum(p, Id, 1) ) Vec_IntWriteEntry( vMapping, Abc_Var2Lit(Id, 1), -1 ); // create internal nodes Gia_ManForEachAndId( p->pGia, i ) { Gia_Obj_t * pObj = Gia_ManObj(p->pGia, i); if ( Gia_ObjIsBuf(pObj) ) { if ( Nf_ObjMapRefNum(p, i, 1) ) Vec_IntWriteEntry( vMapping, Abc_Var2Lit(i, 1), -1 ); Vec_IntWriteEntry( vMapping, Abc_Var2Lit(i, 0), -2 ); continue; } for ( c = 0; c < 2; c++ ) if ( Nf_ObjMapRefNum(p, i, c) ) { // printf( "Using %d %d\n", i, c ); pM = Nf_ObjMatchBest( p, i, c ); // remember inverter if ( pM->fCompl ) { Vec_IntWriteEntry( vMapping, Abc_Var2Lit(i, c), -1 ); continue; } // Nf_ManCutMatchPrint( p, i, c, pM ); pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); // create mapping Vec_IntWriteEntry( vMapping, Abc_Var2Lit(i, c), Vec_IntSize(vMapping) ); Vec_IntPush( vMapping, Nf_CutSize(pCut) ); Nf_CutForEachLit( pCut, pM->Conf, iLit, k ) Vec_IntPush( vMapping, iLit ); Vec_IntPush( vMapping, pM->Gate ); } } // assert( Vec_IntCap(vMapping) == 16 || Vec_IntSize(vMapping) == Vec_IntCap(vMapping) ); p->pGia->vCellMapping = vMapping; return p->pGia; } void Nf_ManUpdateStats( Nf_Man_t * p ) { Nf_Mat_t * pM; Gia_Obj_t * pObj; Mio_Cell_t * pCell; int i, c, Id, * pCut; p->pPars->MapDelay = 0; Gia_ManForEachCo( p->pGia, pObj, i ) { float Delay = Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) )->D; p->pPars->MapDelay = Abc_MaxFloat( p->pPars->MapDelay, Delay ); } p->pPars->MapArea = 0; p->pPars->Area = p->pPars->Edge = 0; Gia_ManForEachAndId( p->pGia, i ) for ( c = 0; c < 2; c++ ) if ( Nf_ObjMapRefNum(p, i, c) ) { pM = Nf_ObjMatchBest( p, i, c ); pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); pCell = Nf_ManCell( p, pM->Gate ); assert( Nf_CutSize(pCut) == (int)pCell->nFanins ); p->pPars->MapArea += pCell->Area; p->pPars->Edge += Nf_CutSize(pCut); p->pPars->Area++; } Gia_ManForEachCiId( p->pGia, Id, i ) if ( Nf_ObjMapRefNum(p, Id, 1) ) { p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; } } /**Function************************************************************* Synopsis [Technology mappping.] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ /* static inline Nf_Mat_t * Nf_ObjMatchBestReq( Nf_Man_t * p, int i, int c, float r ) { Nf_Mat_t * pD = Nf_ObjMatchD(p, i, c); Nf_Mat_t * pA = Nf_ObjMatchA(p, i, c); assert( !pD->fBest && !pA->fBest ); assert( Nf_ObjMapRefNum(p, i, c) == 0 ); if ( pA->D < r + p->pPars->Epsilon ) return pA; return pD; } float Nf_MatchDeref_rec( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM ) { int k, iVar, fCompl, * pCut; float Area = 0; int Value = pM->fBest; pM->fBest = 0; if ( pM->fCompl ) { assert( Nf_ObjMapRefNum(p, i, !c) > 0 ); if ( !Nf_ObjMapRefDec(p, i, !c) ) Area += Nf_MatchDeref_rec( p, i, !c, Nf_ObjMatchBest(p, i, !c) ); return Area + p->InvArea; } if ( Nf_ObjCutSetId(p, i) == 0 ) return 0; assert( Value == 1 ); pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); Nf_CutForEachVar( pCut, pM->Conf, iVar, fCompl, k ) { assert( Nf_ObjMapRefNum(p, iVar, fCompl) > 0 ); if ( !Nf_ObjMapRefDec(p, iVar, fCompl) ) Area += Nf_MatchDeref_rec( p, iVar, fCompl, Nf_ObjMatchBest(p, iVar, fCompl) ); } return Area + Nf_ManCell(p, pM->Gate)->Area; } float Nf_MatchRef_rec( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM, float Required, Vec_Int_t * vBackup ) { int k, iVar, fCompl, * pCut; float ReqFanin, Area = 0; assert( pM->fBest == 0 ); if ( vBackup == NULL ) pM->fBest = 1; if ( pM->fCompl ) { ReqFanin = Required - p->InvDelay; if ( vBackup ) Vec_IntPush( vBackup, Abc_Var2Lit(i, !c) ); assert( Nf_ObjMapRefNum(p, i, !c) >= 0 ); if ( !Nf_ObjMapRefInc(p, i, !c) ) Area += Nf_MatchRef_rec( p, i, !c, Nf_ObjMatchBestReq(p, i, !c, ReqFanin), ReqFanin, vBackup ); return Area + p->InvArea; } if ( Nf_ObjCutSetId(p, i) == 0 ) return 0; pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pM->CutH ); Nf_CutForEachVar( pCut, pM->Conf, iVar, fCompl, k ) { ReqFanin = Required - Nf_ManCell(p, pM->Gate)->Delays[k]; if ( vBackup ) Vec_IntPush( vBackup, Abc_Var2Lit(iVar, fCompl) ); assert( Nf_ObjMapRefNum(p, iVar, fCompl) >= 0 ); if ( !Nf_ObjMapRefInc(p, iVar, fCompl) ) Area += Nf_MatchRef_rec( p, iVar, fCompl, Nf_ObjMatchBestReq(p, iVar, fCompl, ReqFanin), ReqFanin, vBackup ); } return Area + Nf_ManCell(p, pM->Gate)->Area; } float Nf_MatchRefArea( Nf_Man_t * p, int i, int c, Nf_Mat_t * pM, float Required ) { float Area; int iLit, k; Vec_IntClear( &p->vBackup ); Area = Nf_MatchRef_rec( p, i, c, pM, Required, &p->vBackup ); Vec_IntForEachEntry( &p->vBackup, iLit, k ) { assert( Nf_ObjMapRefNum(p, Abc_Lit2Var(iLit), Abc_LitIsCompl(iLit)) > 0 ); Nf_ObjMapRefDec( p, Abc_Lit2Var(iLit), Abc_LitIsCompl(iLit) ); } return Area; } void Nf_ManElaBestMatchOne( Nf_Man_t * p, int iObj, int c, int * pCut, int * pCutSet, Nf_Mat_t * pRes, float Required ) { Nf_Mat_t Mb, * pMb = &Mb; Nf_Obj_t * pBest = Nf_ManObj(p, iObj); int * pFans = Nf_CutLeaves(pCut); int nFans = Nf_CutSize(pCut); int iFuncLit = Nf_CutFunc(pCut); int fComplExt = Abc_LitIsCompl(iFuncLit); float Epsilon = p->pPars->Epsilon; Vec_Int_t * vArr = Vec_WecEntry( p->vTt2Match, Abc_Lit2Var(iFuncLit) ); int i, k, Info, Offset, iFanin, fComplF; float ArrivalD, ArrivalA; // assign fanins matches Nf_Obj_t * pBestF[NF_LEAF_MAX]; for ( i = 0; i < nFans; i++ ) pBestF[i] = Nf_ManObj( p, pFans[i] ); // special cases if ( nFans < 2 ) { *pRes = *Nf_ObjMatchBestReq( p, iObj, c, Required ); return; } // consider matches of this function memset( pMb, 0, sizeof(Nf_Mat_t) ); pMb->D = pMb->A = NF_INFINITY; Vec_IntForEachEntryDouble( vArr, Info, Offset, i ) { Mio_Cell_t* pC = Nf_ManCell( p, Info >> 8 ); int Type = (Info >> 4) & 15; int fCompl = (Info & 1) ^ fComplExt; char * pInfo = Vec_StrEntryP( p->vMemStore, Offset ); Nf_Mat_t * pD = &pBest->M[fCompl][0]; Nf_Mat_t * pA = &pBest->M[fCompl][1]; assert( nFans == (int)pC->nFanins ); if ( fCompl != c ) continue; if ( Type == NF_PRIME ) { float Delay = 0; for ( k = 0; k < nFans; k++ ) { iFanin = Abc_Lit2Var((int)pInfo[k]); fComplF = Abc_LitIsCompl((int)pInfo[k]); ArrivalD = pBestF[k]->M[fComplF][0].D; ArrivalA = pBestF[k]->M[fComplF][1].D; if ( ArrivalA + pC->Delays[iFanin] < Required + Epsilon && Required != NF_INFINITY ) Delay = Abc_MaxFloat( Delay, ArrivalA + pC->Delays[iFanin] ); else Delay = Abc_MaxFloat( Delay, ArrivalD + pC->Delays[iFanin] ); if ( Delay > Required + Epsilon ) break; } if ( k < nFans ) continue; // create match pMb->D = Delay; pMb->A = -1; pMb->CutH = Nf_CutHandle(pCutSet, pCut); pMb->Gate = pC->Id; pMb->Conf = 0; for ( k = 0; k < nFans; k++ ) // pD->Conf |= ((int)pInfo[k] << (k << 2)); pMb->Conf |= (Abc_Var2Lit(k, Abc_LitIsCompl((int)pInfo[k])) << (Abc_Lit2Var((int)pInfo[k]) << 2)); // compute area pMb->A = Nf_MatchRefArea( p, iObj, c, pMb, Required ); // compare if ( pRes->A > pMb->A + Epsilon || (pRes->A == pMb->A && pRes->D > pMb->D + Epsilon) ) *pRes = *pMb; } } } void Nf_ManElaBestMatch( Nf_Man_t * p, int iObj, int c, Nf_Mat_t * pRes, float Required ) { int k, * pCut, * pCutSet = Nf_ObjCutSet( p, iObj ); memset( pRes, 0, sizeof(Nf_Mat_t) ); pRes->D = pRes->A = NF_INFINITY; Nf_SetForEachCut( pCutSet, pCut, k ) { if ( Abc_Lit2Var(Nf_CutFunc(pCut)) >= Vec_WecSize(p->vTt2Match) ) continue; Nf_ManElaBestMatchOne( p, iObj, c, pCut, pCutSet, pRes, Required ); } } // the best match is stored in pA provided that it satisfies pA->D < req // area is never compared void Nf_ManComputeMappingEla( Nf_Man_t * p ) { Gia_Obj_t * pObj; Mio_Cell_t * pCell; Nf_Mat_t Mb, * pMb = &Mb, * pM; float Epsilon = p->pPars->Epsilon; float AreaBef, AreaAft, Required, MapArea; int nLits = 2*Gia_ManObjNum(p->pGia); int i, c, iVar, Id, fCompl, k, * pCut; Vec_FltFill( &p->vRequired, nLits, NF_INFINITY ); // compute delay p->pPars->MapDelay = 0; Gia_ManForEachCo( p->pGia, pObj, i ) { Required = Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) )->D; p->pPars->MapDelay = Abc_MaxFloat( p->pPars->MapDelay, Required ); } // check delay target if ( p->pPars->MapDelayTarget == -1 && p->pPars->nRelaxRatio ) p->pPars->MapDelayTarget = (int)((float)p->pPars->MapDelay * (100.0 + p->pPars->nRelaxRatio) / 100.0); if ( p->pPars->MapDelayTarget != -1 ) { if ( p->pPars->MapDelay < p->pPars->MapDelayTarget + Epsilon ) p->pPars->MapDelay = p->pPars->MapDelayTarget; else if ( p->pPars->nRelaxRatio == 0 ) Abc_Print( 0, "Relaxing user-specified delay target from %.2f to %.2f.\n", p->pPars->MapDelayTarget, p->pPars->MapDelay ); } // set required times Gia_ManForEachCo( p->pGia, pObj, i ) { Required = Nf_ObjMatchD( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj) )->D; Required = p->pPars->fDoAverage ? Required * (100.0 + p->pPars->nRelaxRatio) / 100.0 : p->pPars->MapDelay; Nf_ObjUpdateRequired( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj), Required ); Nf_ObjMapRefInc( p, Gia_ObjFaninId0p(p->pGia, pObj), Gia_ObjFaninC0(pObj)); } // compute area and edges MapArea = p->pPars->MapArea; p->pPars->MapArea = 0; p->pPars->Area = p->pPars->Edge = 0; Gia_ManForEachAndReverseId( p->pGia, i ) for ( c = 0; c < 2; c++ ) if ( Nf_ObjMapRefNum(p, i, c) ) { pM = Nf_ObjMatchBest( p, i, c ); Required = Nf_ObjRequired( p, i, c ); assert( pM->D < Required + Epsilon ); // try different cuts at this node and find best match Vec_IntClear( &p->vBackup2 ); AreaBef = Nf_MatchDeref_rec( p, i, c, pM ); Nf_ManElaBestMatch( p, i, c, pMb, Required ); AreaAft = Nf_MatchRef_rec( p, i, c, pMb, Required, NULL ); assert( pMb->A == AreaAft ); assert( AreaBef + Epsilon > AreaAft ); MapArea += AreaAft - AreaBef; // printf( "%8.2f %8.2f\n", AreaBef, AreaAft ); // set match assert( pMb->D < Required + Epsilon ); assert( pMb->fBest == 0 ); *Nf_ObjMatchA(p, i, c) = *pMb; assert( Nf_ObjMatchA(p, i, c) == Nf_ObjMatchBest( p, i, c ) ); // count status pCell = Nf_ManCell( p, pMb->Gate ); pCut = Nf_CutFromHandle( Nf_ObjCutSet(p, i), pMb->CutH ); Nf_CutForEachVar( pCut, pMb->Conf, iVar, fCompl, k ) Nf_ObjUpdateRequired( p, iVar, fCompl, Required - pCell->Delays[k] ); p->pPars->MapArea += pCell->Area; p->pPars->Edge += Nf_CutSize(pCut); p->pPars->Area++; } Gia_ManForEachCiId( p->pGia, Id, i ) if ( Nf_ObjMapRefNum(p, Id, 1) ) { Nf_ObjMapRefInc( p, Id, 0 ); Nf_ObjUpdateRequired( p, Id, 0, Required - p->InvDelay ); p->pPars->MapArea += p->InvArea; p->pPars->Edge++; p->pPars->Area++; } // Nf_ManUpdateStats( p ); if ( !(MapArea < p->pPars->MapArea + Epsilon && MapArea + Epsilon > p->pPars->MapArea) ) printf( "Mismatch: Estimated = %.2f Real = %.2f\n", MapArea, p->pPars->MapArea ); // assert( MapArea < p->pPars->MapArea + Epsilon && MapArea + Epsilon > p->pPars->MapArea ); Nf_ManPrintStats( p, "Ela " ); } */ /**Function************************************************************* Synopsis [Technology mappping.] Description [] SideEffects [] SeeAlso [] ***********************************************************************/ void Nf_ManSetDefaultPars( Jf_Par_t * pPars ) { memset( pPars, 0, sizeof(Jf_Par_t) ); pPars->nLutSize = 6; pPars->nCutNum = 16; pPars->nProcNum = 0; pPars->nRounds = 3; pPars->nRoundsEla = 3; pPars->nRelaxRatio = 0; pPars->nCoarseLimit = 3; pPars->nAreaTuner = 1; pPars->nVerbLimit = 5; pPars->DelayTarget = -1; pPars->fAreaOnly = 0; pPars->fPinPerm = 0; pPars->fOptEdge = 1; pPars->fCoarsen = 0; pPars->fCutMin = 1; pPars->fGenCnf = 0; pPars->fPureAig = 0; pPars->fVerbose = 0; pPars->fVeryVerbose = 0; pPars->nLutSizeMax = NF_LEAF_MAX; pPars->nCutNumMax = NF_CUT_MAX; pPars->MapDelayTarget = -1; pPars->Epsilon = (float)0.01; } Gia_Man_t * Nf_ManPerformMapping( Gia_Man_t * pGia, Jf_Par_t * pPars ) { Gia_Man_t * pNew = NULL, * pCls; Nf_Man_t * p; int i, Id; if ( Gia_ManHasChoices(pGia) ) pPars->fCoarsen = 0; pCls = pPars->fCoarsen ? Gia_ManDupMuxes(pGia, pPars->nCoarseLimit) : pGia; p = Nf_StoCreate( pCls, pPars ); // if ( pPars->fVeryVerbose ) // Nf_StoPrint( p, pPars->fVeryVerbose ); if ( pPars->fVerbose && pPars->fCoarsen ) { printf( "Initial " ); Gia_ManPrintMuxStats( pGia ); printf( "\n" ); printf( "Derived " ); Gia_ManPrintMuxStats( pCls ); printf( "\n" ); } Nf_ManPrintInit( p ); Nf_ManComputeCuts( p ); Nf_ManPrintQuit( p ); Gia_ManForEachCiId( p->pGia, Id, i ) Nf_ObjPrepareCi( p, Id ); for ( p->Iter = 0; p->Iter < p->pPars->nRounds; p->Iter++ ) { Nf_ManComputeMapping( p ); Nf_ManSetMapRefs( p ); Nf_ManPrintStats( p, p->Iter ? "Area " : "Delay" ); } /* p->fUseEla = 1; for ( ; p->Iter < p->pPars->nRounds + pPars->nRoundsEla; p->Iter++ ) { Nf_ManComputeMapping( p ); Nf_ManUpdateStats( p ); Nf_ManPrintStats( p, "Ela " ); } */ pNew = Nf_ManDeriveMapping( p ); // Gia_ManMappingVerify( pNew ); Nf_StoDelete( p ); if ( pCls != pGia ) Gia_ManStop( pCls ); if ( pNew == NULL ) return Gia_ManDup( pGia ); return pNew; } //////////////////////////////////////////////////////////////////////// /// END OF FILE /// //////////////////////////////////////////////////////////////////////// ABC_NAMESPACE_IMPL_END