void FastMul(const BigInt &A,const BigInt &B, BigInt &C) {
ulong Len = 2;
while ( Len < A.Size + B.Size ) Len *=2;
if ( Len < 40 ) {
BigInt Atmp(A), Btmp(B);
Mul(Atmp,Btmp,C);
return;
}
ulong x;
const ushort *a=A.Coef, *b=B.Coef;
for (x = 0; x < A.Size; x++) LongNum1[x] = a[x];
for (; x < Len; x++) LongNum1[x] = 0.0;
FHT_F(LongNum1,Len);
if (a == b) {
FHTConvolution(LongNum1,LongNum1,Len);
} else {
for (x = 0; x < B.Size; x++) LongNum2[x] = b[x];
for (; x < Len; x++) LongNum2[x] = 0.0;
FHT_F(LongNum2,Len);
FHTConvolution(LongNum1,LongNum2,Len);
}
FHT_T(LongNum1,Len);
CarryNormalize(LongNum1, Len, C);
}
void RFHT_T(FFT_DATA_TYPE *Data,int Len)
/*
** recursive Decimation in Time style Fast Hartley Transform
*/
{int x,Len2,Len4;
TRIG_VARS
FFT_DATA_TYPE *Left,*Right;
Len/=2;Right=&Data[Len];Left=&Data[0];
if (Len<4) FatalError("OOps, how did rfht get called with %d\n",Len*2);
if (Len==4)
{FFT_DATA_TYPE d45,d67,sd0123,dd0123;
StartTimer(FFTRTime);
StartTimer(CRTTime);
{FFT_DATA_TYPE ss0123,ds0123,ss4567,ds4567;
{FFT_DATA_TYPE s01,s23,d01,d23;
DD_Sub(&Data[0],&Data[1],&d01);
DD_Add(&Data[0],&Data[1],&s01);
DD_Sub(&Data[2],&Data[3],&d23);
DD_Add(&Data[2],&Data[3],&s23);
DD_Sub(&s01,&s23,&ds0123);
DD_Add(&s01,&s23,&ss0123);
DD_Sub(&d01,&d23,&dd0123);
DD_Add(&d01,&d23,&sd0123);
}
{FFT_DATA_TYPE s45,s67;
DD_Add(&Data[4],&Data[5],&s45);
DD_Add(&Data[6],&Data[7],&s67);
DD_Sub(&Data[4],&Data[5],&d45);
DD_Sub(&Data[6],&Data[7],&d67);
DD_Sub(&s45,&s67,&ds4567);
DD_Add(&s45,&s67,&ss4567);
}
DD_Sub(&ss0123,&ss4567,&Data[4]);
DD_Add(&ss0123,&ss4567,&Data[0]);
DD_Sub(&ds0123,&ds4567,&Data[6]);
DD_Add(&ds0123,&ds4567,&Data[2]);
}
DD_Mul(&d45,&Sqrt2,&d45);
DD_Mul(&d67,&Sqrt2,&d67);
DD_Sub(&sd0123,&d45,&Data[5]);
DD_Add(&sd0123,&d45,&Data[1]);
DD_Sub(&dd0123,&d67,&Data[7]);
DD_Add(&dd0123,&d67,&Data[3]);
StopTimer(FFTRTime);
StopTimer(CRTTime);
return;
}
if (Len<=(CPU_CACHE/sizeof(FFT_DATA_TYPE))) {FHT_T(Data,Len*2);return;}
RFHT_T(&Left[0], Len);
RFHT_T(&Right[0],Len);
StartTimer(FFTRTime);
/* Do the special x=0 loop below. */
FHT_T1Butterfly(0,0,1.0,0.0);
INIT_TRIG(Len);
Len2=Len/2;
Len4=Len/4;
for (x=1;x<Len4;x++)
{
NEXT_TRIG_POW;
FHT_T2Butterfly(x,Len-x,Pow_r,Pow_i);
FHT_T2Butterfly(Len2-x,Len2+x,Pow_i,Pow_r);
}
/* Now do the two Len/4 points the loop missed */
if (Len4)
// {FFT_DATA_TYPE sq=MY_SQRT_2; /* variable allows optimizations */
// FHT_T2Butterfly(Len4,Len-Len4,sq,sq);
{
FHT_T2Butterfly(Len4,Len-Len4,Sqrt05,Sqrt05);
}
/* Now do the Len/2 point the loop couldn't do. */
if (Len2)
FHT_T1Butterfly(Len2,Len2,0.0,1.0);
StopTimer(FFTRTime);
}
