#include "fmpz.h"
#include "fmpz_vec.h"
#include "fmpz_poly.h"
#include "fmpz_poly/impl.h"
static void _fmpz_poly_sqr_kara_recursive(fmpz * out, fmpz * rev,
fmpz * temp, slong bits)
{
slong len = (WORD(1) << bits);
slong m = len / 2;
if (len == 1)
{
fmpz_mul(out, rev, rev);
fmpz_zero(out + 1);
return;
}
_fmpz_vec_add(temp, rev, rev + m, m);
_fmpz_poly_sqr_kara_recursive(out, rev, temp + 2 * m, bits - 1);
_fmpz_poly_sqr_kara_recursive(out + len, temp, temp + 2 * m, bits - 1);
_fmpz_poly_sqr_kara_recursive(temp, rev + m, temp + 2 * m, bits - 1);
_fmpz_vec_sub(out + len, out + len, out, len);
_fmpz_vec_sub(out + len, out + len, temp, len);
_fmpz_vec_add_rev(out, temp, bits);
}
void _fmpz_poly_sqr_karatsuba(fmpz * res, const fmpz * poly, slong len)
{
fmpz *rev, *out, *temp;
slong length, loglen = 0;
if (len == 1)
{
fmpz_mul(res, poly, poly);
return;
}
while ((WORD(1) << loglen) < len)
loglen++;
length = (WORD(1) << loglen);
rev = flint_calloc(3 * length, sizeof(fmpz));
out = rev + length;
temp = _fmpz_vec_init(2 * length);
revbin1(rev, poly, len, loglen);
_fmpz_poly_sqr_kara_recursive(out, rev, temp, loglen);
_fmpz_vec_zero(res, 2 * len - 1);
revbin2(res, out, 2 * len - 1, loglen + 1);
_fmpz_vec_clear(temp, 2 * length);
flint_free(rev);
}
void fmpz_poly_sqr_karatsuba(fmpz_poly_t res, const fmpz_poly_t poly)
{
slong len;
if (poly->length == 0)
{
fmpz_poly_zero(res);
return;
}
len = 2 * poly->length - 1;
fmpz_poly_fit_length(res, len);
_fmpz_poly_sqr_karatsuba(res->coeffs, poly->coeffs, poly->length);
_fmpz_poly_set_length(res, len);
}