diff options
author | Alan Mishchenko <alanmi@berkeley.edu> | 2015-10-24 16:13:19 -0700 |
---|---|---|
committer | Alan Mishchenko <alanmi@berkeley.edu> | 2015-10-24 16:13:19 -0700 |
commit | a43d8273b708c3e0dedc045dfa2b58ccea5978f5 (patch) | |
tree | a92d2e401a431cc3123747e184db908f0c57ccba | |
parent | 61d462320744107bb609878210d1058bbc585b91 (diff) | |
download | abc-a43d8273b708c3e0dedc045dfa2b58ccea5978f5.tar.gz abc-a43d8273b708c3e0dedc045dfa2b58ccea5978f5.tar.bz2 abc-a43d8273b708c3e0dedc045dfa2b58ccea5978f5.zip |
Changes for delay-oriented computation.
-rw-r--r-- | src/base/abci/abc.c | 16 | ||||
-rw-r--r-- | src/opt/sfm/sfm.h | 1 | ||||
-rw-r--r-- | src/opt/sfm/sfmDec.c | 174 | ||||
-rw-r--r-- | src/opt/sfm/sfmInt.h | 4 | ||||
-rw-r--r-- | src/opt/sfm/sfmLib.c | 16 | ||||
-rw-r--r-- | src/opt/sfm/sfmTime.c | 27 |
6 files changed, 128 insertions, 110 deletions
diff --git a/src/base/abci/abc.c b/src/base/abci/abc.c index c1f7eb5a..037a1580 100644 --- a/src/base/abci/abc.c +++ b/src/base/abci/abc.c @@ -5173,7 +5173,7 @@ int Abc_CommandMfs3( Abc_Frame_t * pAbc, int argc, char ** argv ) // set defaults Sfm_ParSetDefault3( pPars ); Extra_UtilGetoptReset(); - while ( ( c = Extra_UtilGetopt( argc, argv, "OIFLHDMCNPdamzospvwh" ) ) != EOF ) + while ( ( c = Extra_UtilGetopt( argc, argv, "OIFLHDMCNPWdamzospvwh" ) ) != EOF ) { switch ( c ) { @@ -5290,6 +5290,17 @@ int Abc_CommandMfs3( Abc_Frame_t * pAbc, int argc, char ** argv ) if ( pPars->iNodeOne < 0 ) goto usage; break; + case 'W': + if ( globalUtilOptind >= argc ) + { + Abc_Print( -1, "Command line switch \"-W\" should be followed by an integer.\n" ); + goto usage; + } + pPars->nTimeWin = atoi(argv[globalUtilOptind]); + globalUtilOptind++; + if ( pPars->nTimeWin < 0 || pPars->nTimeWin > 100 ) + goto usage; + break; case 'a': pPars->fArea ^= 1; break; @@ -5335,7 +5346,7 @@ int Abc_CommandMfs3( Abc_Frame_t * pAbc, int argc, char ** argv ) return 0; usage: - Abc_Print( -2, "usage: mfs3 [-OIFLHDMCNP <num>] [-amzospvwh]\n" ); + Abc_Print( -2, "usage: mfs3 [-OIFLHDMCNPW <num>] [-amzospvwh]\n" ); Abc_Print( -2, "\t performs don't-care-based optimization of mapped networks\n" ); Abc_Print( -2, "\t-O <num> : the number of levels in the TFO cone (0 <= num) [default = %d]\n", pPars->nTfoLevMax ); Abc_Print( -2, "\t-I <num> : the number of levels in the TFI cone (1 <= num) [default = %d]\n", pPars->nTfiLevMax ); @@ -5347,6 +5358,7 @@ usage: Abc_Print( -2, "\t-C <num> : the max number of conflicts in one SAT run (0 = no limit) [default = %d]\n", pPars->nBTLimit ); Abc_Print( -2, "\t-N <num> : the max number of nodes to try (0 = all) [default = %d]\n", pPars->nNodesMax ); Abc_Print( -2, "\t-P <num> : one particular node to try (0 = none) [default = %d]\n", pPars->iNodeOne ); + Abc_Print( -2, "\t-W <num> : size of timing window in percents (0 <= num <= 100) [default = %d]\n", pPars->nTimeWin ); Abc_Print( -2, "\t-a : toggle area minimization [default = %s]\n", pPars->fArea? "yes": "no" ); Abc_Print( -2, "\t-m : toggle detecting multi-input AND/OR gates [default = %s]\n", pPars->fUseAndOr? "yes": "no" ); Abc_Print( -2, "\t-z : toggle zero-cost replacements [default = %s]\n", pPars->fZeroCost? "yes": "no" ); diff --git a/src/opt/sfm/sfm.h b/src/opt/sfm/sfm.h index ea40f908..53cbe8e5 100644 --- a/src/opt/sfm/sfm.h +++ b/src/opt/sfm/sfm.h @@ -55,6 +55,7 @@ struct Sfm_Par_t_ int nNodesMax; // the maximum number of nodes to try int iNodeOne; // one particular node to try int nFirstFixed; // the number of first nodes to be treated as fixed + int nTimeWin; // the size of timing window in percents int fRrOnly; // perform redundance removal int fArea; // performs optimization for area int fMoreEffort; // performs high-affort minimization diff --git a/src/opt/sfm/sfmDec.c b/src/opt/sfm/sfmDec.c index fe806289..775b9f4c 100644 --- a/src/opt/sfm/sfmDec.c +++ b/src/opt/sfm/sfmDec.c @@ -56,6 +56,7 @@ struct Sfm_Dec_t_ int nDivs; // the number of divisors int nMffc; // the number of divisors int AreaMffc; // the area of gates in MFFC + int DelayMin; // temporary min delay int iTarget; // target node word uCareSet; // computed careset Vec_Int_t vObjRoots; // roots of the window @@ -113,6 +114,8 @@ struct Sfm_Dec_t_ int nMaxWin; word nAllDivs; word nAllWin; + int nLuckySizes[10]; + int nLuckyGates[10]; }; #define SFM_MASK_PI 1 // supp(node) is contained in supp(TFI(pivot)) @@ -152,6 +155,7 @@ void Sfm_ParSetDefault3( Sfm_Par_t * pPars ) pPars->nWinSizeMax = 0; // the maximum window size pPars->nGrowthLevel = 0; // the maximum allowed growth in level pPars->nBTLimit = 0; // the maximum number of conflicts in one SAT run + pPars->nTimeWin = 1; // the size of timing window in percents pPars->fUseAndOr = 0; // enable internal detection of AND/OR gates pPars->fZeroCost = 0; // enable zero-cost replacement pPars->fUseSim = 0; // enable simulation @@ -784,7 +788,7 @@ int Sfm_DecCombineDec( Sfm_Dec_t * p, word * pTruth0, word * pTruth1, int * pSup Abc_TtMux( pTruth, p->pTtElems[iSuppVar], pTruth1, pTruth0, Abc_TtWordNum(nSupp) ); return nSupp; } -int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssump, int nAssump, word Masks[2], int fCofactor ) +int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssump, int nAssump, word Masks[2], int fCofactor, int nSuppAdd ) { int nBTLimit = p->pPars->nBTLimit; // int fVerbose = p->pPars->fVeryVerbose; @@ -805,12 +809,6 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu printf( "\n" ); } } - if ( nAssump > p->nMffc ) - { - if ( p->pPars->fVeryVerbose ) - printf( "The number of assumption is more than MFFC size.\n" ); - return -2; - } // check constant for ( c = 0; c < 2; c++ ) { @@ -847,48 +845,6 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu *Vec_WrdEntryP(&p->vSets[c], i) |= ((word)1 << p->nPats[c]); p->uMask[c] |= ((word)1 << p->nPats[c]++); } -/* - // precompute blocking matrix - for ( c = 0; c < 2; c++ ) - { - for ( d = 0; d < p->nDivs; d += 5 ) - { - word MaskAll = p->uMask[c] & Masks[c]; - word MaskCur = Vec_WrdEntry(&p->vSets[c], d) & Masks[c]; - if ( MaskAll != 0 && MaskCur != 0 && MaskCur != MaskAll ) // has both ones and zeros - continue; - p->nSatCalls++; - pAssump[nAssump] = Abc_Var2Lit( p->iTarget, c ); - pAssump[nAssump+1] = Abc_Var2Lit( d, MaskCur != 0 ); - clk = Abc_Clock(); - status = sat_solver_solve( p->pSat, pAssump, pAssump + nAssump + 2, nBTLimit, 0, 0, 0 ); - if ( status == l_Undef ) - { - p->nTimeOuts++; - return -2; - } - if ( status == l_False ) - { - p->nSatCallsUnsat++; - p->timeSatUnsat += Abc_Clock() - clk; - continue; - } - assert( status == l_True ); - p->nSatCallsSat++; - p->timeSatSat += Abc_Clock() - clk; - if ( p->nPats[c] == 64 ) - { - p->nSatCallsOver++; - return -2;//continue; - } - // record this status - for ( i = 0; i < p->nDivs; i++ ) - if ( sat_solver_var_value(p->pSat, i) ) - *Vec_WrdEntryP(&p->vSets[c], i) |= ((word)1 << p->nPats[c]); - p->uMask[c] |= ((word)1 << p->nPats[c]++); - } - } -*/ // check implications Vec_IntClear( &p->vImpls[0] ); Vec_IntClear( &p->vImpls[1] ); @@ -949,7 +905,12 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu printf( "Found variable %s%d.\n", Abc_LitIsCompl(Impls[0]) ? "!":"", pSupp[0] ); return 1; } - + if ( nSuppAdd > 4 ) + { + if ( p->pPars->fVeryVerbose ) + printf( "The number of assumption is more than MFFC size.\n" ); + return -2; + } // try using all implications at once if ( p->pPars->fUseAndOr ) for ( c = 0; c < 2; c++ ) @@ -973,7 +934,7 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu int * pFinal, nFinal = sat_solver_final( p->pSat, &pFinal ); p->nSatCallsUnsat++; p->timeSatUnsat += Abc_Clock() - clk; - if ( nFinal - nAssump - 0 > p->nMffc ) + if ( nFinal + nSuppAdd > 6 ) continue; // collect only relevant literals for ( i = d = 0; i < nFinal; i++ ) @@ -1021,6 +982,7 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu // find the best cofactoring variable Var = -1, CostMin = ABC_INFINITY; +// if ( !fCofactor || Vec_IntSize(&p->vImpls[0]) + Vec_IntSize(&p->vImpls[1]) > 2 ) for ( c = 0; c < 2; c++ ) { Vec_IntForEachEntry( &p->vImpls[c], iLit, i ) @@ -1038,10 +1000,11 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu if ( Var == -1 && fCofactor ) { //for ( Var = p->nDivs - 1; Var >= 0; Var-- ) - Vec_IntForEachEntry( &p->vObjInMffc, Var, i ) + Vec_IntForEachEntryReverse( &p->vObjInMffc, Var, i ) if ( Vec_IntFind(&p->vObjDec, Var) == -1 ) break; - if ( i == Vec_IntSize(&p->vObjInMffc) ) +// if ( i == Vec_IntSize(&p->vObjInMffc) ) + if ( i == -1 ) Var = -1; fCofactor = 0; } @@ -1051,15 +1014,13 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu Sfm_DecPrint( p, Masks ); printf( "Best var %d with weight %d. Cofactored = %s\n", Var, CostMin, Var == p->nDivs - 1 ? "yes" : "no" ); printf( "\n" ); - //if ( Var == 14 ) - // Var = 13; } // cofactor the problem if ( Var >= 0 ) { word uTruth[2][SFM_WORD_MAX], MasksNext[2]; - int Supp[2][2*SFM_SUPP_MAX], nSupp[2], nSuppAll; + int Supp[2][2*SFM_SUPP_MAX], nSupp[2] = {0}, nSuppAll; Vec_IntPush( &p->vObjDec, Var ); for ( i = 0; i < 2; i++ ) { @@ -1069,19 +1030,14 @@ int Sfm_DecPeformDec_rec( Sfm_Dec_t * p, word * pTruth, int * pSupp, int * pAssu MasksNext[c] = Masks[c] & (i ? MaskVar | ~p->uMask[c] : ~MaskVar); } pAssump[nAssump] = Abc_Var2Lit( Var, !i ); - nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], Supp[i], pAssump, nAssump+1, MasksNext, fCofactor ); + nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], Supp[i], pAssump, nAssump+1, MasksNext, fCofactor, (i ? nSupp[0] : 0) + nSuppAdd + 1 ); if ( nSupp[i] == -2 ) return -2; } // combine solutions nSuppAll = Sfm_DecCombineDec( p, uTruth[0], uTruth[1], Supp[0], Supp[1], nSupp[0], nSupp[1], pTruth, pSupp, Var ); - //if ( nSuppAll > p->nMffc ) - // return -2; -// if ( p->pPars->fVeryVerbose ) -// { -// int s = 0; -// ABC_SWAP( int, pSupp[0], pSupp[1] ); -// } + if ( nSuppAll > 6 ) + return -2; return nSuppAll; } return -2; @@ -1108,6 +1064,7 @@ int Sfm_DecPeformDec2( Sfm_Dec_t * p, Abc_Obj_t * pObj ) if ( fVeryVerbose ) printf( "\nNode %4d : MFFC %2d\n", p->iTarget, p->nMffc ); assert( p->pPars->nDecMax <= SFM_DEC_MAX ); + Vec_IntClear( &p->vObjDec ); for ( i = 0; i < nDecs; i++ ) { // reduce the variable array @@ -1116,7 +1073,7 @@ int Sfm_DecPeformDec2( Sfm_Dec_t * p, Abc_Obj_t * pObj ) Prev = Vec_IntSize(&p->vObjDec) + 1; // perform decomposition Masks[0] = Masks[1] = ~(word)0; - nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], pSupp[i], pAssump, 0, Masks, 1 ); + nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], pSupp[i], pAssump, 0, Masks, 1, 0 ); if ( nSupp[i] == -2 ) { if ( fVeryVerbose ) @@ -1152,6 +1109,10 @@ int Sfm_DecPeformDec2( Sfm_Dec_t * p, Abc_Obj_t * pObj ) RetValue = Sfm_LibImplement( p->pLib, uTruth[iBest][0], pSupp[iBest], nSupp[iBest], p->AreaMffc, &p->vObjGates, &p->vObjFanins, p->pPars->fZeroCost ); if ( fVeryVerbose ) printf( "Area-reducing implementation %sfound.\n", RetValue < 0 ? "NOT " : "" ); + if ( RetValue >= 0 ) + p->nLuckySizes[nSupp[iBest]]++; + if ( RetValue >= 0 ) + p->nLuckyGates[RetValue]++; return RetValue; } int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) @@ -1161,10 +1122,11 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) int nSupp[SFM_DEC_MAX], pAssump[SFM_WIN_MAX]; int fVeryVerbose = p->pPars->fPrintDecs || p->pPars->fVeryVerbose; int nDecs = Abc_MaxInt(p->pPars->nDecMax, 1); - int i, k, DelayMin, nMatches, iBest = -1, RetValue, Prev = 0; + int i, k, DelayOrig = 0, DelayMin, nMatches, iBest = -1, RetValue, Prev = 0; Mio_Gate_t * pGate1Best = NULL, * pGate2Best = NULL; char * pFans1Best = NULL, * pFans2Best = NULL; assert( p->pPars->fArea == 0 ); + p->DelayMin = 0; if ( p->pPars->fUseSim ) Sfm_ObjSetupSimInfo( pObj ); else @@ -1177,7 +1139,9 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) //Sfm_DecPrint( p, NULL ); if ( fVeryVerbose ) printf( "\nNode %4d : MFFC %2d\n", p->iTarget, p->nMffc ); + // set limit on search for decompositions in delay-model assert( p->pPars->nDecMax <= SFM_DEC_MAX ); + Vec_IntClear( &p->vObjDec ); for ( i = 0; i < nDecs; i++ ) { // reduce the variable array @@ -1186,7 +1150,7 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) Prev = Vec_IntSize(&p->vObjDec) + 1; // perform decomposition Masks[0] = Masks[1] = ~(word)0; - nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], pSupp[i], pAssump, 0, Masks, 1 ); + nSupp[i] = Sfm_DecPeformDec_rec( p, uTruth[i], pSupp[i], pAssump, 0, Masks, 1, 0 ); if ( nSupp[i] == -2 ) { if ( fVeryVerbose ) @@ -1197,14 +1161,21 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) printf( "Dec %d: Pat0 = %2d Pat1 = %2d Supp = %d ", i, p->nPats[0], p->nPats[1], nSupp[i] ); if ( fVeryVerbose ) Dau_DsdPrintFromTruth( uTruth[i], nSupp[i] ); - if ( nSupp[iBest] < 2 ) + if ( nSupp[i] < 2 ) { RetValue = Sfm_LibImplement( p->pLib, uTruth[i][0], pSupp[i], nSupp[i], p->AreaMffc, &p->vObjGates, &p->vObjFanins, p->pPars->fZeroCost ); + p->nLuckySizes[nSupp[i]]++; + p->nLuckyGates[RetValue]++; return RetValue; } + //if ( pObj->Id == 689 ) + // { + // int s = 0; + //} + // try the delay nMatches = Sfm_LibFindMatches( p->pLib, uTruth[i][0], pSupp[i], nSupp[i], &p->vMatchGates, &p->vMatchFans ); - DelayMin = Sfm_TimReadObjDelay( p->pTim, Abc_ObjId(pObj) ); + DelayMin = DelayOrig = Sfm_TimReadObjDelay( p->pTim, Abc_ObjId(pObj) ); for ( k = 0; k < nMatches; k++ ) { Mio_Gate_t * pGate1 = (Mio_Gate_t *)Vec_PtrEntry( &p->vMatchGates, 2*k+0 ); @@ -1212,14 +1183,15 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) char * pFans1 = (char *)Vec_PtrEntry( &p->vMatchFans, 2*k+0 ); char * pFans2 = (char *)Vec_PtrEntry( &p->vMatchFans, 2*k+1 ); Vec_Int_t vFanins = { nSupp[i], nSupp[i], pSupp[i] }; - int Delay = Sfm_TimEvalRemapping( p->pTim, &vFanins, pGate1, pFans1, pGate2, pFans2 ); - if ( Delay < DelayMin ) + int Delay = Sfm_TimEvalRemapping( p->pTim, &vFanins, &p->vObjMap, pGate1, pFans1, pGate2, pFans2 ); + if ( DelayMin > Delay ) { + DelayMin = Delay; pGate1Best = pGate1; pGate2Best = pGate2; pFans1Best = pFans1; pFans2Best = pFans2; - iBest = i; + iBest = i; } } } @@ -1232,17 +1204,14 @@ int Sfm_DecPeformDec3( Sfm_Dec_t * p, Abc_Obj_t * pObj ) p->nNoDecs++; return -2; } - else - { - if ( fVeryVerbose ) - printf( "Best %d: %d ", iBest, nSupp[iBest] ); -// if ( fVeryVerbose ) -// Dau_DsdPrintFromTruth( uTruth[iBest], nSupp[iBest] ); - Sfm_LibAddNewGates( p->pLib, pSupp[iBest], pGate1Best, pGate2Best, pFans1Best, pFans2Best, &p->vObjGates, &p->vObjFanins ); - } -// return -1; if ( fVeryVerbose ) - printf( "Delay-reducing implementation found.\n" ); + printf( "Best %d: %d ", iBest, nSupp[iBest] ); +// if ( fVeryVerbose ) +// Dau_DsdPrintFromTruth( uTruth[iBest], nSupp[iBest] ); + RetValue = Sfm_LibAddNewGates( p->pLib, pSupp[iBest], pGate1Best, pGate2Best, pFans1Best, pFans2Best, &p->vObjGates, &p->vObjFanins ); + p->nLuckySizes[nSupp[iBest]]++; + p->nLuckyGates[RetValue]++; + p->DelayMin = DelayMin; return 1; } @@ -1351,7 +1320,7 @@ static inline int Sfm_DecNodeIsMffc( Abc_Obj_t * p, int nLevelMin ) } static inline int Sfm_DecNodeIsMffcInput( Abc_Obj_t * p, int nLevelMin, Sfm_Tim_t * pTim, Abc_Obj_t * pPivot ) { - return Abc_NodeIsTravIdCurrent(p) && (Abc_ObjLevel(p) >= nLevelMin || Abc_ObjFaninNum(p) == 0) && Sfm_TimNodeIsNonCritical(pTim, pPivot, p); + return Abc_NodeIsTravIdCurrent(p) && Sfm_TimNodeIsNonCritical(pTim, pPivot, p); } void Sfm_DecMarkMffc( Abc_Obj_t * pPivot, int nLevelMin, int nMffcMax, int fVeryVerbose, Vec_Int_t * vMffc, Vec_Int_t * vInMffc, Sfm_Tim_t * pTim ) { @@ -1620,6 +1589,7 @@ void Sfm_DecInsert( Abc_Ntk_t * pNtk, Abc_Obj_t * pPivot, int Limit, Vec_Int_t * Vec_IntForEachEntry( vLevel, iObj, k ) Abc_ObjAddFanin( pObjNew, Abc_NtkObj(pNtk, Vec_IntEntry(vMap, iObj)) ); pObjNew->pData = Vec_PtrEntry( vGateHandles, Gate ); + assert( Abc_ObjFaninNum(pObjNew) == Mio_GateReadPinNum((Mio_Gate_t *)pObjNew->pData) ); Vec_IntPush( vMap, Abc_ObjId(pObjNew) ); if ( vTimeNodes ) Vec_IntPush( vTimeNodes, Abc_ObjId(pObjNew) ); @@ -1631,6 +1601,7 @@ void Sfm_DecInsert( Abc_Ntk_t * pNtk, Abc_Obj_t * pPivot, int Limit, Vec_Int_t * } void Sfm_DecPrintStats( Sfm_Dec_t * p ) { + int i; printf( "Node = %d. Try = %d. Change = %d. Const0 = %d. Const1 = %d. Buf = %d. Inv = %d. Gate = %d. AndOr = %d. NoDec = %d.\n", p->nTotalNodesBeg, p->nNodesTried, p->nNodesChanged, p->nNodesConst0, p->nNodesConst1, p->nNodesBuf, p->nNodesInv, p->nNodesResyn, p->nNodesAndOr, p->nNoDecs ); printf( "MaxDiv = %d. MaxWin = %d. AveDiv = %d. AveWin = %d. Calls = %d. (Sat = %d. Unsat = %d.) Over = %d. T/O = %d.\n", @@ -1647,6 +1618,18 @@ void Sfm_DecPrintStats( Sfm_Dec_t * p ) ABC_PRTP( "Other ", p->timeOther, p->timeTotal ); ABC_PRTP( "ALL ", p->timeTotal, p->timeTotal ); + printf( "Cone sizes: " ); + for ( i = 0; i < 10; i++ ) + if ( p->nLuckySizes[i] ) + printf( "%d=%d ", i, p->nLuckySizes[i] ); + printf( " " ); + + printf( "Gate sizes: " ); + for ( i = 0; i < 10; i++ ) + if ( p->nLuckyGates[i] ) + printf( "%d=%d ", i, p->nLuckyGates[i] ); + printf( "\n" ); + printf( "Reduction: " ); printf( "Nodes %6d out of %6d (%6.2f %%) ", p->nTotalNodesBeg-p->nTotalNodesEnd, p->nTotalNodesBeg, 100.0*(p->nTotalNodesBeg-p->nTotalNodesEnd)/Abc_MaxInt(1, p->nTotalNodesBeg) ); printf( "Edges %6d out of %6d (%6.2f %%) ", p->nTotalEdgesBeg-p->nTotalEdgesEnd, p->nTotalEdgesBeg, 100.0*(p->nTotalEdgesBeg-p->nTotalEdgesEnd)/Abc_MaxInt(1, p->nTotalEdgesBeg) ); @@ -1740,15 +1723,22 @@ void Abc_NtkDelayOpt( Sfm_Dec_t * p ) Sfm_Par_t * pPars = p->pPars; printf( "Initial delay = %8.2f.\n", MIO_NUMINV*Sfm_TimReadNtkDelay(p->pTim) ); + Abc_NtkCleanMarkABC( pNtk ); while ( 1 ) { - Abc_Obj_t * pObj; abctime clk; + Abc_Obj_t * pObj, * pObjNew; abctime clk; int i = 0, Limit, RetValue; - // try improving delay for the nodes according to the priority - if ( !Sfm_TimPriorityNodes(p->pTim, &p->vCands) ) + // collect nodes + if ( pPars->iNodeOne ) + Vec_IntFill( &p->vCands, 1, pPars->iNodeOne ); + else if ( !Sfm_TimPriorityNodes(p->pTim, &p->vCands, p->pPars->nTimeWin) ) break; + // try improving delay for the nodes according to the priority Abc_NtkForEachObjVec( &p->vCands, p->pNtk, pObj, i ) { + int OldId = Abc_ObjId(pObj); + int DelayOld = Sfm_TimReadObjDelay(p->pTim, OldId); + p->nNodesTried++; clk = Abc_Clock(); p->nDivs = Sfm_DecExtract( pNtk, pPars, pObj, &p->vObjRoots, &p->vObjGates, &p->vObjFanins, &p->vObjMap, &p->vTemp, &p->vTemp2, &p->vObjMffc, &p->vObjInMffc, p->pTim ); @@ -1794,10 +1784,17 @@ p->timeSat += Abc_Clock() - clk; Sfm_DecInsert( pNtk, pObj, Limit, &p->vObjGates, &p->vObjFanins, &p->vObjMap, &p->vGateHands, p->GateBuffer, p->GateInvert, &p->vGateFuncs, &p->vCands ); Sfm_TimUpdateTiming( p->pTim, &p->vCands ); - printf( "Node %5d delay = %8.2f.\n", Abc_ObjId(pObj), MIO_NUMINV*Sfm_TimReadNtkDelay(p->pTim) ); + pObjNew = Abc_NtkObj( pNtk, Abc_NtkObjNumMax(pNtk)-1 ); + printf( "Node %5d : Old =%8.2f. Predicted =%8.2f. New =%8.2f. Final =%8.2f\n", + OldId, MIO_NUMINV*DelayOld, MIO_NUMINV*p->DelayMin, + MIO_NUMINV*Sfm_TimReadObjDelay(p->pTim, Abc_ObjId(pObjNew)), + MIO_NUMINV*Sfm_TimReadNtkDelay(p->pTim) ); break; } + if ( pPars->iNodeOne ) + break; } + Abc_NtkCleanMarkABC( pNtk ); } void Abc_NtkPerformMfs3( Abc_Ntk_t * pNtk, Sfm_Par_t * pPars ) { @@ -1823,13 +1820,14 @@ void Abc_NtkPerformMfs3( Abc_Ntk_t * pNtk, Sfm_Par_t * pPars ) printf( "DecMax = %d. ", pPars->nDecMax ); if ( pPars->iNodeOne ) printf( "Pivot = %d. ", pPars->iNodeOne ); + if ( !pPars->fArea ) + printf( "Win = %d. ", pPars->nTimeWin ); printf( "Sim = %s. ", pPars->fUseSim ? "yes" : "no" ); printf( "0-cost = %s. ", pPars->fZeroCost ? "yes" : "no" ); printf( "\n" ); } // preparation steps Abc_NtkLevel( pNtk ); - Abc_NtkCleanMarkABC( pNtk ); if ( p->pPars->fUseSim ) Sfm_NtkSimulate( pNtk ); // record statistics diff --git a/src/opt/sfm/sfmInt.h b/src/opt/sfm/sfmInt.h index 81438466..1b84a6fb 100644 --- a/src/opt/sfm/sfmInt.h +++ b/src/opt/sfm/sfmInt.h @@ -218,9 +218,9 @@ extern int Sfm_TimReadNtkDelay( Sfm_Tim_t * p ); extern int Sfm_TimReadObjDelay( Sfm_Tim_t * p, int iObj ); extern void Sfm_TimUpdateTiming( Sfm_Tim_t * p, Vec_Int_t * vTimeNodes ); extern int Sfm_TimSortArrayByArrival( Sfm_Tim_t * p, Vec_Int_t * vNodes, int iPivot ); -extern int Sfm_TimPriorityNodes( Sfm_Tim_t * p, Vec_Int_t * vCands ); +extern int Sfm_TimPriorityNodes( Sfm_Tim_t * p, Vec_Int_t * vCands, int Window ); extern int Sfm_TimNodeIsNonCritical( Sfm_Tim_t * p, Abc_Obj_t * pPivot, Abc_Obj_t * pNode ); -extern int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Mio_Gate_t * pGate1, char * pFans1, Mio_Gate_t * pGate2, char * pFans2 ); +extern int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Vec_Int_t * vMap, Mio_Gate_t * pGate1, char * pFans1, Mio_Gate_t * pGate2, char * pFans2 ); /*=== sfmWin.c ==========================================================*/ extern int Sfm_ObjMffcSize( Sfm_Ntk_t * p, int iObj ); extern int Sfm_NtkCreateWindow( Sfm_Ntk_t * p, int iNode, int fVerbose ); diff --git a/src/opt/sfm/sfmLib.c b/src/opt/sfm/sfmLib.c index 97f4c71b..bf4f636a 100644 --- a/src/opt/sfm/sfmLib.c +++ b/src/opt/sfm/sfmLib.c @@ -509,15 +509,18 @@ void Sfm_LibPrintObj( Sfm_Lib_t * p, Sfm_Fun_t * pObj ) } void Sfm_LibPrint( Sfm_Lib_t * p ) { - word * pTruth; Sfm_Fun_t * pObj; int iFunc; + word * pTruth; Sfm_Fun_t * pObj; int iFunc, nSupp; Vec_MemForEachEntry( p->vTtMem, pTruth, iFunc ) { if ( iFunc < 2 ) continue; + nSupp = Abc_TtSupportSize(pTruth, 6); + if ( nSupp > 3 ) + continue; //if ( iFunc % 10000 ) // continue; printf( "%d : Count = %d ", iFunc, Vec_IntEntry(&p->vCounts, iFunc) ); - Dau_DsdPrintFromTruth( pTruth, Abc_TtSupportSize(pTruth, 6) ); + Dau_DsdPrintFromTruth( pTruth, nSupp ); Sfm_LibForEachSuper( p, pObj, iFunc ) Sfm_LibPrintObj( p, pObj ); } @@ -565,8 +568,8 @@ int Sfm_LibFindMatches( Sfm_Lib_t * p, word uTruth, int * pFanins, int nFanins, { pCellB = p->pCells + (int)pObj->pFansB[0]; pCellT = p->pCells + (int)pObj->pFansT[0]; - Vec_PtrPush( vGates, pCellB ); - Vec_PtrPush( vGates, pCellT == p->pCells ? NULL : pCellT ); + Vec_PtrPush( vGates, pCellB->pMioGate ); + Vec_PtrPush( vGates, pCellT == p->pCells ? NULL : pCellT->pMioGate ); Vec_PtrPush( vFans, pObj->pFansB + 1 ); Vec_PtrPush( vFans, pCellT == p->pCells ? NULL : pObj->pFansT + 1 ); } @@ -590,7 +593,7 @@ int Sfm_LibAddNewGates( Sfm_Lib_t * p, int * pFanins, Mio_Gate_t * pGateB, Mio_G int i, nFanins; // create bottom gate Vec_IntPush( vGates, Mio_GateReadValue(pGateB) ); - vLevel = Vec_WecPushLevel( vFanins ); + vLevel = Vec_WecPushLevel( vFanins ); nFanins = Mio_GateReadPinNum( pGateB ); for ( i = 0; i < nFanins; i++ ) Vec_IntPush( vLevel, pFanins[(int)pFansB[i]] ); @@ -598,7 +601,8 @@ int Sfm_LibAddNewGates( Sfm_Lib_t * p, int * pFanins, Mio_Gate_t * pGateB, Mio_G return 1; // create top gate Vec_IntPush( vGates, Mio_GateReadValue(pGateT) ); - vLevel = Vec_WecPushLevel( vFanins ); + vLevel = Vec_WecPushLevel( vFanins ); + nFanins = Mio_GateReadPinNum( pGateT ); for ( i = 0; i < nFanins; i++ ) if ( pFansT[i] == (char)16 ) Vec_IntPush( vLevel, Vec_WecSize(vFanins)-2 ); diff --git a/src/opt/sfm/sfmTime.c b/src/opt/sfm/sfmTime.c index 5cca7477..b58a4bf2 100644 --- a/src/opt/sfm/sfmTime.c +++ b/src/opt/sfm/sfmTime.c @@ -242,11 +242,11 @@ Sfm_Tim_t * Sfm_TimStart( Mio_Library_t * pLib, Scl_Con_t * pExt, Abc_Ntk_t * pN p->pLib = pLib; p->pExt = pExt; p->pNtk = pNtk; - Vec_IntFill( &p->vTimArrs, 2*Abc_NtkObjNumMax(pNtk), 0 ); - Vec_IntFill( &p->vTimReqs, 2*Abc_NtkObjNumMax(pNtk), 0 ); -// Vec_IntFill( &p->vTimSlews, 2*Abc_NtkObjNumMax(pNtk), 0 ); -// Vec_IntFill( &p->vTimLoads, 2*Abc_NtkObjNumMax(pNtk), 0 ); -// Vec_IntFill( &p->vObjOffs, Abc_NtkObjNumMax(pNtk), 0 ); + Vec_IntFill( &p->vTimArrs, 4*Abc_NtkObjNumMax(pNtk), 0 ); + Vec_IntFill( &p->vTimReqs, 4*Abc_NtkObjNumMax(pNtk), 0 ); +// Vec_IntFill( &p->vTimSlews, 4*Abc_NtkObjNumMax(pNtk), 0 ); +// Vec_IntFill( &p->vTimLoads, 4*Abc_NtkObjNumMax(pNtk), 0 ); +// Vec_IntFill( &p->vObjOffs, 2*Abc_NtkObjNumMax(pNtk), 0 ); // Abc_NtkForEachNode( pNtk, pObj, i ) // { // Vec_IntWriteEntry( &p->vObjOffs, i, Vec_IntSize(Vec_IntSize(&p->vTimEdges)) ); @@ -264,7 +264,9 @@ void Sfm_TimStop( Sfm_Tim_t * p ) Vec_IntErase( &p->vTimLoads ); Vec_IntErase( &p->vObjOffs ); Vec_IntErase( &p->vTimEdges ); + Vec_WecErase( &p->vLevels ); Vec_IntErase( &p->vPath ); + Vec_WrdErase( &p->vSortData ); ABC_FREE( p ); } int Sfm_TimReadNtkDelay( Sfm_Tim_t * p ) @@ -384,13 +386,14 @@ int Sfm_TimSortArrayByArrival( Sfm_Tim_t * p, Vec_Int_t * vNodes, int iPivot ) SeeAlso [] ***********************************************************************/ -int Sfm_TimPriorityNodes( Sfm_Tim_t * p, Vec_Int_t * vCands ) +int Sfm_TimPriorityNodes( Sfm_Tim_t * p, Vec_Int_t * vCands, int Window ) { Vec_Int_t * vLevel; Abc_Obj_t * pObj; - int i; + int i, k; + assert( Window >= 0 && Window <= 100 ); // collect critical path - Sfm_TimCriticalPath(p, 1); + Sfm_TimCriticalPath( p, Window ); // add nodes to the levelized structure Sfm_TimUpdateClean( p ); Abc_NtkForEachObjVec( &p->vPath, p->pNtk, pObj, i ) @@ -403,7 +406,7 @@ int Sfm_TimPriorityNodes( Sfm_Tim_t * p, Vec_Int_t * vCands ) Vec_WecSort( &p->vLevels, 0 ); Vec_IntClear( vCands ); Vec_WecForEachLevel( &p->vLevels, vLevel, i ) - Abc_NtkForEachObjVec( vLevel, p->pNtk, pObj, i ) + Abc_NtkForEachObjVec( vLevel, p->pNtk, pObj, k ) if ( !pObj->fMarkA ) Vec_IntPush( vCands, Abc_ObjId(pObj) ); return Vec_IntSize(vCands) > 0; @@ -436,7 +439,7 @@ int Sfm_TimNodeIsNonCritical( Sfm_Tim_t * p, Abc_Obj_t * pPivot, Abc_Obj_t * pNo SeeAlso [] ***********************************************************************/ -int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Mio_Gate_t * pGate1, char * pFans1, Mio_Gate_t * pGate2, char * pFans2 ) +int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Vec_Int_t * vMap, Mio_Gate_t * pGate1, char * pFans1, Mio_Gate_t * pGate2, char * pFans2 ) { int TimeOut[2][2]; int * pTimesIn1[6], * pTimesIn2[6]; @@ -444,7 +447,7 @@ int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Mio_Gate_t * pGate // process the first gate nFanins1 = Mio_GateReadPinNum( pGate1 ); for ( i = 0; i < nFanins1; i++ ) - pTimesIn1[i] = Sfm_TimArrId( p, Vec_IntEntry(vFanins, (int)pFans1[i]) ); + pTimesIn1[i] = Sfm_TimArrId( p, Vec_IntEntry(vMap, Vec_IntEntry(vFanins, (int)pFans1[i])) ); Sfm_TimGateArrival( p, pGate1, pTimesIn1, TimeOut[0] ); if ( pGate2 == NULL ) return Abc_MaxInt(TimeOut[0][0], TimeOut[0][1]); @@ -454,7 +457,7 @@ int Sfm_TimEvalRemapping( Sfm_Tim_t * p, Vec_Int_t * vFanins, Mio_Gate_t * pGate if ( (int)pFans2[i] == 16 ) pTimesIn2[i] = TimeOut[0]; else - pTimesIn2[i] = Sfm_TimArrId( p, Vec_IntEntry(vFanins, (int)pFans2[i]) ); + pTimesIn2[i] = Sfm_TimArrId( p, Vec_IntEntry(vMap, Vec_IntEntry(vFanins, (int)pFans2[i])) ); Sfm_TimGateArrival( p, pGate2, pTimesIn2, TimeOut[1] ); return Abc_MaxInt(TimeOut[1][0], TimeOut[1][1]); } |