#include "mpn_extras.h"
#include "fmpz.h"
#include "fmpz_vec.h"
#include "fmpz_poly.h"
static void _fmpz_poly_sqrlow_kara_recursive(fmpz * out,
const fmpz * pol, fmpz * temp, slong len)
{
slong m1 = len / 2;
slong m2 = len - m1;
int odd = (len & 1);
if (len <= 6)
{
_fmpz_poly_sqrlow_classical(out, pol, len, len);
return;
}
_fmpz_vec_add(temp + m2, pol, pol + m1, m1);
if (odd)
fmpz_set(temp + m2 + m1, pol + 2 * m1);
_fmpz_poly_sqrlow_kara_recursive(temp, temp + m2, temp + 2 * m2, m2);
_fmpz_poly_sqrlow_kara_recursive(temp + m2, pol + m1, temp + 2 * m2, m2);
_fmpz_poly_sqr_karatsuba(out, pol, m1);
fmpz_zero(out + 2 * m1 - 1);
_fmpz_vec_sub(temp, temp, out, m2);
_fmpz_vec_sub(temp, temp, temp + m2, m2);
if (odd)
fmpz_set(out + 2 * m1, temp + m2);
_fmpz_vec_add(out + m1, out + m1, temp, m2);
}
void _fmpz_poly_sqrlow_karatsuba_n(fmpz * res, const fmpz * poly, slong n)
{
fmpz *temp;
slong len, loglen = 0;
if (n == 1)
{
fmpz_mul(res, poly, poly);
return;
}
while ((WORD(1) << loglen) < n)
loglen++;
len = (WORD(1) << loglen);
temp = _fmpz_vec_init(2 * len + 2);
_fmpz_poly_sqrlow_kara_recursive(res, poly, temp, n);
_fmpz_vec_clear(temp, 2 * len + 2);
}
void _fmpz_poly_sqrlow_karatsuba(fmpz * res, const fmpz * poly, slong len, slong n)
{
int clear = 0;
slong i;
fmpz *copy;
if (len >= n)
copy = (fmpz *) poly;
else
{
copy = flint_malloc(n * sizeof(fmpz));
for (i = 0; i < len; i++)
copy[i] = poly[i];
flint_mpn_zero((nn_ptr) copy + len, n - len);
clear = 1;
}
_fmpz_poly_sqrlow_karatsuba_n(res, copy, n);
if (clear)
flint_free(copy);
}
void
fmpz_poly_sqrlow_karatsuba_n(fmpz_poly_t res, const fmpz_poly_t poly, slong n)
{
const slong len = FLINT_MIN(poly->length, n);
slong lenr;
if (len == 0)
{
fmpz_poly_zero(res);
return;
}
lenr = 2 * len - 1;
n = FLINT_MIN(n, lenr);
if (res != poly)
{
fmpz_poly_fit_length(res, n);
_fmpz_poly_sqrlow_karatsuba(res->coeffs, poly->coeffs, len, n);
}
else
{
fmpz_poly_t t;
fmpz_poly_init2(t, n);
_fmpz_poly_sqrlow_karatsuba(t->coeffs, poly->coeffs, len, n);
fmpz_poly_swap(res, t);
fmpz_poly_clear(t);
}
_fmpz_poly_set_length(res, n);
_fmpz_poly_normalise(res);
}