// perform convolution with the taps
FINL
vcs Convolute ( const vcs& input, // The input data
const vcs *Taps,
vcs *temp, // The Temp Table for next convolution
int Tap_num )
{
register vcs ans[4];
register vcs * pTap = (vcs*)Taps;
register vcs * pTemp1 = temp;
register vcs * pTemp2 = temp;
for ( int i=0; i<4; i ++ ) {
ans[i] = add((vcs)(mul_high((vs)(input),(vs)(*pTap))),*pTemp1);
pTap ++; pTemp1 ++;
}
vcs ret = hadd4 (ans[0], ans[1], ans[2], ans[3] );
for (int i = 4; i < Tap_num; i++)
{
*pTemp2 = add((vcs)(mul_high((vs)(input),(vs)(*pTap))), *pTemp1);
pTemp1 ++; pTap ++; pTemp2++;
}
return ret;
}
void xasm_mulm192(ulong *c, ulong *a, ulong *b)
{
ulong carry, inv, i, tmp[3+1], l, j, x;
ulong chk[3], a0, a1, a2, b0, b1, b2;
a0=a[0]; b0=b[0];
a1=a[1]; b1=b[1];
a2=a[2]; b2=b[2];
asm_mulm192(chk,a,b);
for (i=0; i<3+1; i++) tmp[i]=0;
for (i=0; i<3; i++) {
x=a[i]; carry=0;
for (j=0; j<3; j++) {
tmp[j]+=carry; if (tmp[j]<carry) carry=1; else carry=0;
l=x*b[j];
tmp[j]+=l; if (tmp[j]<l) carry++;
carry+=mul_high(x,b[j]);
}
tmp[3]+=carry; /* no carry possible */
inv=tmp[0]*montgomery_inv_n; carry=0;
for (j=0; j<3; j++) {
tmp[j]+=carry; if (tmp[j]<carry) carry=1; else carry=0;
l=inv*montgomery_modulo_n[j];
tmp[j]+=l; if (tmp[j]<l) carry++;
carry+=mul_high(inv,montgomery_modulo_n[j]);
}
tmp[3]+=carry;
for (j=0; j<3; j++) tmp[j]=tmp[j+1];
if (tmp[2]<carry) tmp[3]=1; else tmp[3]=0;
}
if (tmp[3]) asm_sub_n192(tmp,montgomery_modulo_n);
asm_copy192(c,tmp);
asm_add192_ui(c,0);
if ((c[0]!=chk[0]) || (c[1]!=chk[1]) || (c[2]!=chk[2])) {
printf("mul failed: %lu %lu %lu\n",(ulong)a,(ulong)b,(ulong)c);
complain("%lu %lu %lu * %lu %lu %lu\n%lu %lu %lu!= %lu %lu %lu\n%lu %lu %lu %lu\n",a0,a1,a2,b0,b1,b2,c[0],c[1],c[2],chk[0],chk[1],chk[2],montgomery_modulo_n[0],montgomery_modulo_n[1],montgomery_modulo_n[2],montgomery_inv_n);
} // else printf("ok");
}
void casm_mulm64(ulong *c, ulong *a, ulong *b)
{
ulong carry, inv, i, tmp[1+1], l, j, x;
ulong chk, a1, b1;
a1=a[0]; b1=b[0];
asm_mulm64(&chk,a,b);
for (i=0; i<1+1; i++) tmp[i]=0;
for (i=0; i<1; i++) {
x=a[i]; carry=0;
for (j=0; j<1; j++) {
tmp[j]+=carry; if (tmp[j]<carry) carry=1; else carry=0;
l=x*b[j];
tmp[j]+=l; if (tmp[j]<l) carry++;
carry+=mul_high(x,b[j]);
}
tmp[1]+=carry; /* no carry possible */
inv=tmp[0]*montgomery_inv_n; carry=0;
for (j=0; j<1; j++) {
tmp[j]+=carry; if (tmp[j]<carry) carry=1; else carry=0;
l=inv*montgomery_modulo_n[j];
tmp[j]+=l; if (tmp[j]<l) carry++;
carry+=mul_high(inv,montgomery_modulo_n[j]);
}
tmp[1]+=carry;
for (j=0; j<1; j++) tmp[j]=tmp[j+1];
if (tmp[0]<carry) tmp[1]=1; else tmp[1]=0;
}
if (tmp[1]) asm_sub_n64(tmp,montgomery_modulo_n);
asm_copy64(c,tmp);
asm_add64_ui(c,0);
if (*c!=chk) {
printf("mul failed: %lu %lu %lu\n",(ulong)a,(ulong)b,(ulong)c);
complain("%lu %lu %lu %lu\n%lu %lu\n",a1,b1,c[0],chk,montgomery_modulo_n[0],montgomery_inv_n);
}
}
