Skip to content

A
abc
Commit 10c31c65 by Alan Mishchenko
Options

Experiments with LUT structure mapping.

parent bc1eae79
Showing with 275 additions and 88 deletions
src/aig/gia/giaIf.c
......@@ -560,7 +560,7 @@ void Gia_ManPrintPackingStats( Gia_Man_t * p )
int fVerbose = 0;
int nObjToShow = 200;
int nNumStr[5] = {0};
int i, k, Entry, nEntries, nEntries2, MaxSize = -1;
int i, k, Entry, nEntries, nEntries2, MaxSize = -1, Count = 0;
if ( p->vPacking == NULL )
return;
nEntries = Vec_IntEntry( p->vPacking, 0 );
......@@ -586,8 +586,12 @@ void Gia_ManPrintPackingStats( Gia_Man_t * p )
MaxSize = 1;
Abc_Print( 1, "Packing (N=%d) : ", MaxSize );
for ( i = 1; i <= MaxSize; i++ )
{
Abc_Print( 1, "%d x LUT = %d ", i, nNumStr[i] );
Abc_Print( 1, "Total = %d", nEntries2 );
Count += i * nNumStr[i];
}
Abc_Print( 1, "Total = %d ", nEntries2 );
Abc_Print( 1, "Total LUT = %d", Count );
Abc_Print( 1, "\n" );
}
......
src/aig/gia/giaOf.c
......@@ -38,7 +38,6 @@ ABC_NAMESPACE_IMPL_START
#define OF_CUT_MAX 32
#define OF_NO_LEAF 31
#define OF_NO_FUNC 0x7FFFFFF
#define OF_INFINITY FLT_MAX
#define OF_CUT_EXTRA 4 // size; delay1, delay2; area
typedef struct Of_Cut_t_ Of_Cut_t;
......@@ -55,11 +54,13 @@ typedef struct Of_Obj_t_ Of_Obj_t;
struct Of_Obj_t_
{
int iCutH; // best cut
int iCutH2; // best cut
int Delay1; // arrival time
int Delay2; // arrival time
int Required; // required
int nRefs; // references
int Flow; // area flow
int Temp; // unused
};
typedef struct Of_Man_t_ Of_Man_t;
struct Of_Man_t_
......@@ -116,11 +117,12 @@ static inline void Of_CutSetAreaFlow( int * pCut, int d )
static inline int Of_CutVar( int * pCut, int v ) { return Abc_Lit2Var(Of_CutLeaves(pCut)[v]); }
static inline int Of_CutFlag( int * pCut, int v ) { return Abc_LitIsCompl(Of_CutLeaves(pCut)[v]); }
static inline void Of_CutCleanFlag( int * pCut, int v ) { Of_CutLeaves(pCut)[v] = Abc_LitRegular(Of_CutLeaves(pCut)[v]); }
static inline void Of_CutSetFlag( int * pCut, int v ) { Of_CutLeaves(pCut)[v] |= 1; }
static inline void Of_CutSetFlag( int * pCut, int v, int x ) { Of_CutLeaves(pCut)[v] = Abc_Var2Lit(Of_CutVar(pCut, v), x); }
static inline Of_Obj_t * Of_ObjData( Of_Man_t * p, int i ) { return p->pObjs + i; }
static inline int Of_ObjCutBest( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->iCutH; }
static inline int Of_ObjCutBest2( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->iCutH2; }
static inline int Of_ObjDelay1( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->Delay1; }
static inline int Of_ObjDelay2( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->Delay2; }
static inline int Of_ObjRequired( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->Required; }
......@@ -128,6 +130,7 @@ static inline int Of_ObjRefNum( Of_Man_t * p, int i )
static inline int Of_ObjFlow( Of_Man_t * p, int i ) { return Of_ObjData(p, i)->Flow; }
static inline void Of_ObjSetCutBest( Of_Man_t * p, int i, int x ) { Of_ObjData(p, i)->iCutH = x; }
static inline void Of_ObjSetCutBest2( Of_Man_t * p, int i, int x ) { Of_ObjData(p, i)->iCutH2 = x; }
static inline void Of_ObjSetDelay1( Of_Man_t * p, int i, int x ) { Of_ObjData(p, i)->Delay1 = x; }
static inline void Of_ObjSetDelay2( Of_Man_t * p, int i, int x ) { Of_ObjData(p, i)->Delay2 = x; }
static inline void Of_ObjSetRequired( Of_Man_t * p, int i, int x ) { Of_ObjData(p, i)->Required = x; }
......@@ -140,6 +143,9 @@ static inline int Of_ObjRefDec( Of_Man_t * p, int i )
static inline int * Of_ObjCutBestP( Of_Man_t * p, int iObj ) { assert(iObj>0 && iObj<Gia_ManObjNum(p->pGia));return Of_ManCutSet( p, Of_ObjCutBest(p, iObj) ); }
static inline void Of_ObjSetCutBestP( Of_Man_t * p, int * pCutSet, int iObj, int * pCut ) { Of_ObjSetCutBest( p, iObj, Of_ObjCutSetId(p, iObj) + Of_CutHandle(pCutSet, pCut) ); }
static inline int * Of_ObjCutBestP2( Of_Man_t * p, int iObj ) { assert(iObj>0 && iObj<Gia_ManObjNum(p->pGia));return Of_ManCutSet( p, Of_ObjCutBest2(p, iObj) ); }
static inline void Of_ObjSetCutBestP2( Of_Man_t * p, int * pCutSet, int iObj, int * pCut ) { Of_ObjSetCutBest2( p, iObj, Of_ObjCutSetId(p, iObj) + Of_CutHandle(pCutSet, pCut) ); }
#define Of_SetForEachCut( pList, pCut, i ) for ( i = 0, pCut = pList + 1; i < pList[0]; i++, pCut += Of_CutSize(pCut) + OF_CUT_EXTRA )
#define Of_ObjForEachCut( pCuts, i, nCuts ) for ( i = 0, i < nCuts; i++ )
#define Of_CutForEachVar( pCut, iVar, i ) for ( i = 0; i < Of_CutSize(pCut) && (iVar = Of_CutVar(pCut,i)); i++ )
......@@ -871,82 +877,6 @@ void Of_ManPrintQuit( Of_Man_t * p )
SeeAlso []
***********************************************************************/
void Of_ManCutMatch( Of_Man_t * p, int iObj, int * pCut, int * pDelay1, int * pDelay2 )
{
// Delay1 - main delay; Delay2 - precomputed LUT delay in terms of Delay1 for the fanins
int Delays[6], Perm[6];
int DelayLut1 = p->pPars->nDelayLut1;
int DelayLut2 = p->pPars->nDelayLut2;
int k, iVar, Flag, Delay;
Of_CutForEachVarFlag( pCut, iVar, Flag, k )
{
Delays[k] = Of_ObjDelay1(p, iVar) + DelayLut1;
Perm[k] = iVar;
// printf( "%3d%s ", iVar, Flag ? "*" : " " );
}
for ( ; k < p->pPars->nLutSize; k++ )
{
Delays[k] = -ABC_INFINITY;
Perm[k] = -1;
// printf( " " );
}
Vec_IntSelectSortCost2Reverse( Perm, Of_CutSize(pCut), Delays );
*pDelay1 = *pDelay2 = 0;
for ( k = 0; k < Of_CutSize(pCut); k++ )
{
Delay = (k < p->pPars->nFastEdges && Gia_ObjIsAndNotBuf(Gia_ManObj(p->pGia, Perm[k]))) ? Of_ObjDelay2(p, Perm[k]) + DelayLut2 : Delays[k];// + DelayLut2;
*pDelay1 = Abc_MaxInt( *pDelay1, Delay );
*pDelay2 = Abc_MaxInt( *pDelay2, Delays[k] );
}
// printf( " %5.2f", Of_Int2Flt(*pDelay1) );
// printf( " %5.2f\n", Of_Int2Flt(*pDelay2) );
*pDelay1 = Abc_MinInt( *pDelay1, *pDelay2 );
assert( *pDelay1 <= *pDelay2 );
Of_CutSetDelay1( pCut, *pDelay1 );
Of_CutSetDelay2( pCut, *pDelay2 );
}
int Of_ManObjMatch( Of_Man_t * p, int iObj )
{
int Delay1 = ABC_INFINITY, Delay2 = ABC_INFINITY;
int Delay1This, Delay2This;
int i, * pCut, * pList = Of_ObjCutSet(p, iObj);
Of_SetForEachCut( pList, pCut, i )
{
Of_ManCutMatch( p, iObj, pCut, &Delay1This, &Delay2This );
Delay1 = Abc_MinInt( Delay1, Delay1This );
Delay2 = Abc_MinInt( Delay2, Delay2This );
}
Of_ObjSetDelay1( p, iObj, Delay1 );
Of_ObjSetDelay2( p, iObj, Delay2 );
return Delay1;
}
void Of_ManComputeMapping( Of_Man_t * p )
{
int Time = 0;
Gia_Obj_t * pObj; int i;
Gia_ManForEachAnd( p->pGia, pObj, i )
if ( Gia_ObjIsBuf(pObj) )
{
Of_ObjSetDelay1( p, i, Of_ObjDelay1(p, Gia_ObjFaninId0(pObj, i)) );
Of_ObjSetDelay2( p, i, Of_ObjDelay2(p, Gia_ObjFaninId0(pObj, i)) );
}
else
Time = Abc_MaxInt( Time, Of_ManObjMatch(p, i) );
printf( "Best delay = %.2f\n", Of_Int2Flt(Time) );
}
/**Function*************************************************************
Synopsis [Technology mappping.]
Description []
SideEffects []
SeeAlso []
***********************************************************************/
/*
static inline int Of_ManComputeForwardCut( Of_Man_t * p, int iObj, int * pCut )
{
......@@ -1102,7 +1032,7 @@ static inline void Of_ManComputeForwardObj( Of_Man_t * p, int iObj )
if ( p->Iter )
Of_ObjSetFlow( p, iObj, Of_ManComputeForwardCutArea(p, iObj, pCutMin) );
}
void Of_ManComputeForward( Of_Man_t * p )
void Of_ManComputeForward1( Of_Man_t * p )
{
Gia_Obj_t * pObj; int i;
Gia_ManForEachAnd( p->pGia, pObj, i )
......@@ -1370,6 +1300,214 @@ void Of_ManComputeBackward3( Of_Man_t * p )
SeeAlso []
***********************************************************************/
void Of_ManComputeForwardDirconCut( Of_Man_t * p, int iObj, int * pCut, int * pDelay1, int * pDelay2 )
{
// Delay1 - main delay; Delay2 - precomputed LUT delay in terms of Delay1 for the fanins
int Delays[6], Perm[6] = {0, 1, 2, 3, 4, 5};
int DelayLut1 = p->pPars->nDelayLut1;
int DelayLut2 = p->pPars->nDelayLut2;
int nSize = Of_CutSize(pCut);
int k, iVar, Flag, SlowCon, Delay, DelayAfter, fDirConWorks;
Of_CutForEachVar( pCut, iVar, k )
{
Delays[k] = Of_ObjDelay1(p, iVar) + DelayLut1;
// printf( "%3d%s ", iVar, Flag ? "*" : " " );
}
for ( ; k < p->pPars->nLutSize; k++ )
{
Delays[k] = -ABC_INFINITY;
// printf( " " );
}
Vec_IntSelectSortCost2Reverse( Perm, nSize, Delays );
assert( nSize < 2 || Delays[0] >= Delays[nSize-1] );
assert( Delays[0] >= 0 && Delays[nSize-1] >= 0 );
// consider speedup due to dircons
fDirConWorks = 1;
*pDelay1 = *pDelay2 = 0;
SlowCon = p->pPars->nFastEdges < nSize ? Delays[p->pPars->nFastEdges] : 0;
for ( k = 0; k < nSize; k++ )
{
// use dircon if the following is true
// - the input is eligible for dircon (does not exceed the limit)
// - there is an expected gain in delay, compared the largest delay without dircon
// - the dircon delay is indeed lower than the largest delay without dircon
// - all previous dircons worked out well
// - the node is an AND-gate
iVar = Of_CutVar( pCut, Perm[k] );
assert( Delays[k] == Of_ObjDelay1(p, iVar) + DelayLut1 );
DelayAfter = Of_ObjDelay2(p, iVar) + DelayLut2;
if ( k < p->pPars->nFastEdges && Delays[k] > SlowCon && DelayAfter < Delays[k] && fDirConWorks && Gia_ObjIsAndNotBuf(Gia_ManObj(p->pGia, iVar)) )
{
Delay = DelayAfter;
Of_CutSetFlag( pCut, Perm[k], 1 );
}
else
{
Delay = Delays[k];// + DelayLut2;
Of_CutSetFlag( pCut, Perm[k], 0 );
fDirConWorks = 0;
}
*pDelay1 = Abc_MaxInt( *pDelay1, Delay );
*pDelay2 = Abc_MaxInt( *pDelay2, Delays[k] );
}
// printf( " %5.2f", Of_Int2Flt(*pDelay1) );
// printf( " %5.2f\n", Of_Int2Flt(*pDelay2) );
// do not use the structure if simple LUT is better
if ( *pDelay1 > *pDelay2 )
{
for ( k = 0; k < nSize; k++ )
Of_CutSetFlag( pCut, k, 0 );
*pDelay1 = *pDelay2;
}
assert( *pDelay1 <= *pDelay2 );
Of_CutSetDelay1( pCut, *pDelay1 );
Of_CutSetDelay2( pCut, *pDelay2 );
// verify
Of_CutForEachVarFlag( pCut, iVar, Flag, k )
{
if ( Flag )
assert( Of_ObjDelay2(p, iVar) + DelayLut2 <= *pDelay1 );
else
assert( Of_ObjDelay1(p, iVar) + DelayLut1 <= *pDelay1 );
assert( Of_ObjDelay1(p, iVar) + DelayLut1 <= *pDelay2 );
}
}
int Of_ManComputeForwardDirconObj( Of_Man_t * p, int iObj )
{
int Delay1 = ABC_INFINITY, Delay2 = ABC_INFINITY;
int i, * pCut, * pCutMin = NULL, * pCutMin2 = NULL, * pList = Of_ObjCutSet(p, iObj);
Of_SetForEachCut( pList, pCut, i )
{
int Delay1This, Delay2This;
Of_ManComputeForwardDirconCut( p, iObj, pCut, &Delay1This, &Delay2This );
if ( Delay1 > Delay1This )
pCutMin = pCut;
if ( Delay2 > Delay2This )
pCutMin2 = pCut;
Delay1 = Abc_MinInt( Delay1, Delay1This );
Delay2 = Abc_MinInt( Delay2, Delay2This );
}
Of_ObjSetDelay1( p, iObj, Delay1 );
Of_ObjSetDelay2( p, iObj, Delay2 );
Of_ObjSetCutBestP( p, pList, iObj, pCutMin );
Of_ObjSetCutBestP2( p, pList, iObj, pCutMin2 );
return Delay1;
}
void Of_ManComputeForwardDircon1( Of_Man_t * p )
{
Gia_Obj_t * pObj; int i;
Gia_ManForEachAnd( p->pGia, pObj, i )
if ( Gia_ObjIsBuf(pObj) )
{
Of_ObjSetDelay1( p, i, Of_ObjDelay1(p, Gia_ObjFaninId0(pObj, i)) );
Of_ObjSetDelay2( p, i, Of_ObjDelay2(p, Gia_ObjFaninId0(pObj, i)) );
}
else
Of_ManComputeForwardDirconObj( p, i );
}
void Of_ManComputeBackwardDircon1( Of_Man_t * p )
{
Gia_Obj_t * pObj;
Vec_Bit_t * vPointed;
int DelayLut1 = p->pPars->nDelayLut1;
int DelayLut2 = p->pPars->nDelayLut2;
int i, k, iVar, Flag, * pList, * pCutMin;
int CountNodes = 0, CountEdges = 0;
Of_ManComputeOutputRequired( p, 1 );
printf( "Global delay =%8.2f\n", Of_Int2Flt((int)p->pPars->Delay) );
//return;
// compute area and edges
vPointed = Vec_BitStart( Gia_ManObjNum(p->pGia) );
p->pPars->Area = p->pPars->Edge = 0;
Gia_ManForEachAndReverse( p->pGia, pObj, i )
{
int CostMin, fPointed, Required = Of_ObjRequired(p, i);
if ( Gia_ObjIsBuf(pObj) )
{
int FaninId = Gia_ObjFaninId0(pObj, i);
Of_ObjUpdateRequired( p, FaninId, Required );
Of_ObjRefInc( p, FaninId );
continue;
}
if ( !Of_ObjRefNum(p, i) )
continue;
// check if the LUT is has an outgoing dircon edge
fPointed = Vec_BitEntry(vPointed, i);
CountNodes += fPointed;
/*
// select the best cut
{
int * pCut;
pCutMin = NULL;
CostMin = ABC_INFINITY;
pList = Of_ObjCutSet( p, i );
Of_SetForEachCut( pList, pCut, k )
{
int Cost;
if ( (fPointed ? Of_CutDelay2(pCut) : Of_CutDelay1(pCut)) > Required )
continue;
Cost = Of_ManComputeBackwardCut( p, pCut );
if ( CostMin > Cost )
{
CostMin = Cost;
pCutMin = pCut;
}
}
}
*/
if ( fPointed )
{
pCutMin = Of_ObjCutBestP2( p, i );
CostMin = Of_CutDelay2(pCutMin);
//assert( Of_CutDelay2(pCutMin) <= Required );
}
else
{
pCutMin = Of_ObjCutBestP( p, i );
CostMin = Of_CutDelay1(pCutMin);
//assert( Of_CutDelay1(pCutMin) <= Required );
}
// remove dircon markers
//if ( fPointed )
// Of_CutForEachVarFlag( pCutMin, iVar, Flag, k )
// Of_CutSetFlag( pCutMin, k, 0 );
// the cut is selected
assert( pCutMin != NULL );
pList = Of_ObjCutSet( p, i );
Of_ObjSetCutBestP( p, pList, i, pCutMin ); ///// SET THE BEST CUT
Of_CutForEachVarFlag( pCutMin, iVar, Flag, k )
{
Of_ObjUpdateRequired( p, iVar, Required - ((Flag && !fPointed) ? DelayLut2 : DelayLut1) );
Of_ObjRefInc( p, iVar );
if ( Flag && !fPointed )
{
Vec_BitWriteEntry( vPointed, iVar, 1 );
CountEdges++;
}
}
// update parameters
p->pPars->Edge += Of_CutSize(pCutMin);
p->pPars->Area++;
}
Vec_BitFree( vPointed );
//printf( "Dircon nodes = %d. Dircon edges = %d.\n", CountNodes, CountEdges );
}
/**Function*************************************************************
Synopsis [Technology mappping.]
Description []
SideEffects []
SeeAlso []
***********************************************************************/
void Of_ManSetDefaultPars( Jf_Par_t * pPars )
{
memset( pPars, 0, sizeof(Jf_Par_t) );
......@@ -1384,7 +1522,7 @@ void Of_ManSetDefaultPars( Jf_Par_t * pPars )
pPars->DelayTarget = -1;
pPars->nDelayLut1 = 10;
pPars->nDelayLut2 = 2;
pPars->nFastEdges = 1;
pPars->nFastEdges = 0; //
pPars->fAreaOnly = 0;
pPars->fOptEdge = 1;
pPars->fCoarsen = 0;
......@@ -1399,11 +1537,18 @@ void Of_ManSetDefaultPars( Jf_Par_t * pPars )
}
Gia_Man_t * Of_ManDeriveMapping( Of_Man_t * p )
{
Vec_Int_t * vMapping;
int i, k, iVar, * pCut;
Vec_Int_t * vMapping, * vPacking = NULL;
Vec_Bit_t * vPointed;
int i, k, iVar, * pCut, Place, Flag;
assert( !p->pPars->fCutMin && p->pGia->vMapping == NULL );
vMapping = Vec_IntAlloc( Gia_ManObjNum(p->pGia) + (int)p->pPars->Edge + (int)p->pPars->Area * 2 );
Vec_IntFill( vMapping, Gia_ManObjNum(p->pGia), 0 );
if ( p->pPars->nFastEdges )
{
vPacking = Vec_IntAlloc( 1000 );
Vec_IntPush( vPacking, 0 );
}
vPointed = Vec_BitStart( Gia_ManObjNum(p->pGia) );
Gia_ManForEachAndId( p->pGia, i )
{
if ( !Of_ObjRefNum(p, i) )
......@@ -1415,9 +1560,24 @@ Gia_Man_t * Of_ManDeriveMapping( Of_Man_t * p )
Of_CutForEachVar( pCut, iVar, k )
Vec_IntPush( vMapping, iVar );
Vec_IntPush( vMapping, i );
if ( vPacking == NULL || Vec_BitEntry(vPointed, i) )
continue;
Place = Vec_IntSize( vPacking );
Vec_IntPush( vPacking, 0 );
Vec_IntPush( vPacking, i );
Of_CutForEachVarFlag( pCut, iVar, Flag, k )
if ( Flag )
{
Vec_IntPush( vPacking, iVar );
Vec_BitWriteEntry( vPointed, iVar, 1 );
}
Vec_IntAddToEntry( vPacking, Place, Vec_IntSize(vPacking)-Place-1 );
Vec_IntAddToEntry( vPacking, 0, 1 );
}
assert( Vec_IntCap(vMapping) == 16 || Vec_IntSize(vMapping) == Vec_IntCap(vMapping) );
p->pGia->vMapping = vMapping;
p->pGia->vPacking = vPacking;
Vec_BitFree( vPointed );
return p->pGia;
}
Gia_Man_t * Of_ManPerformMapping( Gia_Man_t * pGia, Jf_Par_t * pPars )
......@@ -1444,17 +1604,38 @@ Gia_Man_t * Of_ManPerformMapping( Gia_Man_t * pGia, Jf_Par_t * pPars )
Of_ObjSetDelay2( p, Id, Time );
}
if ( p->pPars->nFastEdges )
{
p->pPars->nRounds = 1;
for ( p->Iter = 0; p->Iter < p->pPars->nRounds; p->Iter++ )
{
if ( p->Iter == 0 )
{
Of_ManComputeForward( p );
Of_ManComputeForwardDircon1( p );
Of_ManComputeBackwardDircon1( p );
Of_ManPrintStats( p, "Delay" );
}
else
{
Of_ManComputeForwardDircon1( p );
Of_ManComputeBackwardDircon1( p );
Of_ManPrintStats( p, "Flow " );
}
}
}
else
{
for ( p->Iter = 0; p->Iter < p->pPars->nRounds; p->Iter++ )
{
if ( p->Iter == 0 )
{
Of_ManComputeForward1( p );
Of_ManComputeBackward1( p );
Of_ManPrintStats( p, "Delay" );
}
else
{
Of_ManComputeForward( p );
Of_ManComputeForward1( p );
Of_ManComputeBackward1( p );
Of_ManPrintStats( p, "Flow " );
}
......@@ -1464,16 +1645,17 @@ Gia_Man_t * Of_ManPerformMapping( Gia_Man_t * pGia, Jf_Par_t * pPars )
if ( p->Iter < p->pPars->nRounds + p->pPars->nRoundsEla - 1 )
{
Of_ManComputeForward2( p );
Of_ManComputeBackward2( p );
Of_ManComputeBackward3( p );
Of_ManPrintStats( p, "Area " );
}
else
{
Of_ManComputeForward( p );
Of_ManComputeForward1( p );
Of_ManComputeBackward3( p );
Of_ManPrintStats( p, "Area " );
}
}
}
pNew = Of_ManDeriveMapping( p );
Gia_ManMappingVerify( pNew );
......
src/base/abci/abc.c
......@@ -34282,6 +34282,7 @@ int Abc_CommandAbc9Unmap( Abc_Frame_t * pAbc, int argc, char ** argv )
}
Vec_IntFreeP( &pAbc->pGia->vMapping );
Vec_IntFreeP( &pAbc->pGia->vPacking );
Vec_IntFreeP( &pAbc->pGia->vCellMapping );
return 0;
usage:
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment