mpn_extras.h – support functions for limb arrays

Macros

MPN_NORM(a, an)

Normalise (a, an) so that either an is zero or a[an - 1] is nonzero.

MPN_SWAP(a, an, b, bn)

Swap (a, an) and (b, bn), i.e. swap pointers and sizes.

Utility functions

void flint_mpn_debug(mp_srcptr x, mp_size_t xsize)

Prints debug information about (x, xsize) to stdout. In particular, this will print binary representations of all the limbs.

char *_flint_mpn_get_str(mp_srcptr x, mp_size_t n)

Returns a string containing the decimal representation of (x, n).

int flint_mpn_zero_p(mp_srcptr x, mp_size_t xsize)

Returns \(1\) if all limbs of (x, xsize) are zero, otherwise \(0\).

int flint_mpn_equal_p(mp_srcptr x, mp_srcptr y, mp_size_t xsize)

Returns \(1\) if all limbs of (x, xsize) and (y, xsize) are equal, otherwise \(0\).

Addition and subtraction

mp_limb_t flint_mpn_sumdiff_n(mp_ptr s, mp_ptr d, mp_srcptr x, mp_srcptr y, mp_size_t n)

Simultaneously computes the sum s and difference d of (x, n) and (y, n), returning carry multiplied by two plus borrow.

void flint_mpn_negmod_n(mp_ptr res, mp_srcptr x, mp_srcptr m, mp_size_t n)
void flint_mpn_addmod_n(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_srcptr m, mp_size_t n)
void flint_mpn_submod_n(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_srcptr m, mp_size_t n)
void flint_mpn_addmod_n_m(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_size_t yn, mp_srcptr m, mp_size_t n)
void flint_mpn_submod_n_m(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_size_t yn, mp_srcptr m, mp_size_t n)

Arithmetic modulo (m, n). These functions assume that (x, n) and (y, n) are already reduced modulo (m, n). The n_m variants accept (y, yn) with yn <= n, where (y, yn) is already reduced modulo (m, n).

void flint_mpn_negmod_2(mp_ptr res, mp_srcptr x, mp_srcptr m)
void flint_mpn_addmod_2(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_srcptr m)
void _flint_mpn_addmod_2(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_srcptr m)
void flint_mpn_submod_2(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_srcptr m)

Modular arithmetic specialized for two limbs. The _flint_mpn_addmod_2 version assumes that the most significant bit of m[1] is not set.

int flint_mpn_signed_sub_n(mp_ptr res, mp_srcptr x, mp_srcptr y, mp_size_t n)

Sets res to \(|x - y|\), returning 0 if the result equals \(x - y\) and returning 1 if the result equals \(y - x\).

Multiplication

mp_limb_t flint_mpn_mul(mp_ptr z, mp_srcptr x, mp_size_t xn, mp_srcptr y, mp_size_t yn)

Sets (z, xn+yn) to the product of (x, xn) and (y, yn) and returns the top limb of the result. We require \(xn \ge yn \ge 1\) and that z is not aliased with either input operand. This function is intended for all operand sizes. It will automatically select an appropriate algorithm out of the following:

  • A hardcoded multiplication function for small sizes.

  • Karatsuba or Toom-Cook multiplication for intermediate sizes.

  • FFT multiplication for huge sizes.

  • A GMP fallback for cases where we do currently not have optimized code.

void flint_mpn_mul_n(mp_ptr z, mp_srcptr x, mp_srcptr y, mp_size_t n)

Sets z to the product of (x, n) and (y, n). We require \(n \ge 1\) and that z is not aliased with either input operand. The algorithm selection is similar to flint_mpn_mul().

void flint_mpn_sqr(mp_ptr z, mp_srcptr x, mp_size_t n)

Sets z to the square of (x, n). We require \(n \ge 1\) and that z is not aliased with the input operand. The algorithm selection is similar to flint_mpn_sqr().

mp_size_t flint_mpn_fmms1(mp_ptr y, mp_limb_t a1, mp_srcptr x1, mp_limb_t a2, mp_srcptr x2, mp_size_t n)

Given not-necessarily-normalized \(x_1\) and \(x_2\) of length \(n > 0\) and output \(y\) of length \(n\), try to compute \(y = a_1\cdot x_1 - a_2\cdot x_2\). Return the normalized length of \(y\) if \(y \ge 0\) and \(y\) fits into \(n\) limbs. Otherwise, return \(-1\). \(y\) may alias \(x_1\) but is not allowed to alias \(x_2\).

void flint_mpn_mul_toom22(mp_ptr pp, mp_srcptr ap, mp_size_t an, mp_srcptr bp, mp_size_t bn, mp_ptr scratch)

Toom-22 (Karatsuba) multiplication. The scratch space must have room for \(2 \text{an} + k\) limbs where \(k\) is the number of limbs. If NULL is passed, space will be allocated internally.

Truncating multiplication

Given two \(n\)-limb integers, a high product (or mulhigh) is an approximation of the leading \(n\) limbs of the full \(2n\)-limb product. In the basecase regime, a high product can be computed in roughly half the time of the full product, and in some fraction \(0.5 < c < 1\) of the time in the Toom-Cook regime. This speedup vanishes asymptotically in the FFT regime. Contrary to polynomial high products or integer low products, integer high products are not uniquely defined due to carry propagation. We make the following definitions:

  • Rough mulhigh accumulates at least \(n + 1\) limbs of partial products, outputting \(n\) limbs where the \(n - 1\) most significant limbs are essentially correct and the \(n\)-th most significant limb may have an error of \(O(n)\) ulp. This is the version of mulhigh used in [HZ2011].

  • Precise mulhigh accumulates at least \(n + 2\) limbs of partial products, outputting \(n + 1\) limbs where the \(n\) most significant limbs are essentially correct and the \((n+1)\)-th most significant limb may have an error of \(O(n)\) ulp.

  • Exact mulhigh is the exact truncation of the full product. This cannot be computed faster than the full product in the worst case, but it can be computed faster on average by performing a precise mulhigh, inspecting the low output limb, and correcting with a low product when necessary.

In all cases, a high product is either equal to or smaller than the high part of the full product.

More generally, we can define \(n\)-limb high products of \(m\)-limb and \(p\)-limb integers where \(m + p > n\), but this is not currently implemented.

void _flint_mpn_mulhigh_n_mulders_recursive(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
void _flint_mpn_sqrhigh_mulders_recursive(mp_ptr res, mp_srcptr u, mp_size_t n)

Rough mulhigh implemented using Mulders’ recursive algorithm as described in [HZ2011]. Puts in res[n], …, res[2n-1] an approximation of the \(n\) high limbs of {u, n} times {v, n}. The error is less than n ulps of res[n]. Assumes \(2n\) limbs are allocated at res; the low limbs will be used as scratch space. The sqrhigh version implements squaring.

mp_limb_t _flint_mpn_mulhigh_basecase(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t _flint_mpn_mulhigh_n_mulders(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t _flint_mpn_mulhigh_n_mul(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t flint_mpn_mulhigh_n(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)

Precise mulhigh. Puts in res[0], …, res[n-1] an approximation of the \(n\) high limbs of {u, n} times {v, n}. and returns the \((n+1)\)-th most significant limb. The error is at most n + 2 ulp in the returned limb.

  • The basecase version implements the \(O(n^2)\) schoolbook algorithm. On x86-64 machines with ADX, the basecase version currently assumes that \(n \ge 6\).

  • The mulders version computes a rough mulhigh with one extra limb of precision in temporary scratch space using _flint_mpn_mulhigh_n_mulders_recursive() and then copies the high limbs to the output.

  • The mul version computes a full product in temporary scratch space and copies the high limbs to the output. The output is actually the exact mulhigh.

  • The default version looks up a hardcoded basecase multiplication routine in a table for small n, and otherwise calls the basecase, mulders or mul implementations.

mp_limb_t _flint_mpn_sqrhigh_basecase(mp_ptr res, mp_srcptr u, mp_size_t n)
mp_limb_t _flint_mpn_sqrhigh_mulders(mp_ptr res, mp_srcptr u, mp_size_t n)
mp_limb_t _flint_mpn_sqrhigh_sqr(mp_ptr res, mp_srcptr u, mp_size_t n)
mp_limb_t flint_mpn_sqrhigh(mp_ptr res, mp_srcptr u, mp_size_t n)

Squaring counterparts of flint_mpn_mulhigh_n().

On x86-64 machines with ADX, the basecase version currently assumes that \(n \ge 8\).

void _flint_mpn_mullow_n_mulders_recursive(mp_ptr rp, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t flint_mpn_mullow_basecase(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t _flint_mpn_mullow_n_mulders(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t _flint_mpn_mullow_n_mul(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t _flint_mpn_mullow_n(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)
mp_limb_t flint_mpn_mullow_n(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)

Compute the low \(n\) limbs of the product.

The \((n + 1)\)-th limb is also computed and returned. Warning: this extra limb of output may be removed in the future.

void flint_mpn_mul_or_mullow_n(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)

Write the low \(n + 1\) limbs of the product \(uv\) to res. The output is assumed to have space for \(2n\) limbs so that the high limbs can be used as scratch space or to write the whole product when this is the fastest method.

Warning: the one extra limb of output may be removed in the future.

void flint_mpn_mul_or_mulhigh_n(mp_ptr res, mp_srcptr u, mp_srcptr v, mp_size_t n)

Write the high \(n + 1\) limbs of the product \(uv\) to res + (n - 1) (with possible error of a few ulps as for flint_mpn_mulhigh_n()). The low \(n - 1\) limbs of the output may be used as scratch space or to write the whole product when this is the fastest method.

Divisibility

int flint_mpn_divisible_1_odd(mp_srcptr x, mp_size_t xsize, mp_limb_t d)

Expression determining whether (x, xsize) is divisible by the mp_limb_t d which is assumed to be odd-valued and at least \(3\).

This function is implemented as a macro.

mp_size_t flint_mpn_remove_2exp(mp_ptr x, mp_size_t xsize, flint_bitcnt_t *bits)

Divides (x, xsize) by \(2^n\) where \(n\) is the number of trailing zero bits in \(x\). The new size of \(x\) is returned, and \(n\) is stored in the bits argument. \(x\) may not be zero.

mp_size_t flint_mpn_remove_power_ascending(mp_ptr x, mp_size_t xsize, mp_ptr p, mp_size_t psize, ulong *exp)

Divides (x, xsize) by the largest power \(n\) of (p, psize) that is an exact divisor of \(x\). The new size of \(x\) is returned, and \(n\) is stored in the exp argument. \(x\) may not be zero, and \(p\) must be greater than \(2\).

This function works by testing divisibility by ascending squares \(p, p^2, p^4, p^8, \dotsc\), making it efficient for removing potentially large powers. Because of its high overhead, it should not be used as the first stage of trial division.

int flint_mpn_factor_trial(mp_srcptr x, mp_size_t xsize, slong start, slong stop)

Searches for a factor of (x, xsize) among the primes in positions start, ..., stop-1 of flint_primes. Returns \(i\) if flint_primes[i] is a factor, otherwise returns \(0\) if no factor is found. It is assumed that start >= 1.

int flint_mpn_factor_trial_tree(slong *factors, mp_srcptr x, mp_size_t xsize, slong num_primes)

Searches for a factor of (x, xsize) among the primes in positions approximately in the range 0, ..., num_primes - 1 of flint_primes.

Returns the number of prime factors found and fills factors with their indices in flint_primes. It is assumed that num_primes is in the range 0, ..., 3512.

If the input fits in a small fmpz the number is fully factored instead.

The algorithm used is a tree based gcd with a product of primes, the tree for which is cached globally (it is threadsafe).

Division

void flint_mpn_signed_div2(mp_ptr res, mp_srcptr x, mp_size_t n)

Sets res to (x, n) divided by two, where x is viewed as a signed integer in two’s complement form.

int flint_mpn_divides(mp_ptr q, mp_srcptr array1, mp_size_t limbs1, mp_srcptr arrayg, mp_size_t limbsg, mp_ptr temp)

If (arrayg, limbsg) divides (array1, limbs1) then (q, limbs1 - limbsg + 1) is set to the quotient and 1 is returned, otherwise 0 is returned. The temporary space temp must have space for limbsg limbs.

Assumes limbs1 >= limbsg > 0.

mp_limb_t flint_mpn_preinv1(mp_limb_t d, mp_limb_t d2)

Computes a precomputed inverse from the leading two limbs of the divisor b, n to be used with the preinv1 functions. We require the most significant bit of b, n to be 1.

mp_limb_t flint_mpn_divrem_preinv1(mp_ptr q, mp_ptr a, mp_size_t m, mp_srcptr b, mp_size_t n, mp_limb_t dinv)

Divide a, m by b, n, returning the high limb of the quotient (which will either be 0 or 1), storing the remainder in-place in a, n and the rest of the quotient in q, m - n. We require the most significant bit of b, n to be 1. dinv must be computed from b[n - 1], b[n - 2] by flint_mpn_preinv1. We also require m >= n >= 2.

void flint_mpn_mulmod_preinv1(mp_ptr r, mp_srcptr a, mp_srcptr b, mp_size_t n, mp_srcptr d, mp_limb_t dinv, ulong norm)

Given a normalised integer \(d\) with precomputed inverse dinv provided by flint_mpn_preinv1, computes \(ab \pmod{d}\) and stores the result in \(r\). Each of \(a\), \(b\) and \(r\) is expected to have \(n\) limbs of space, with zero padding if necessary.

The value norm is provided for convenience. If \(a\), \(b\) and \(d\) have been shifted left by norm bits so that \(d\) is normalised, then \(r\) will be shifted right by norm bits so that it has the same shift as all the inputs.

We require \(a\) and \(b\) to be reduced modulo \(n\) before calling the function.

void flint_mpn_preinvn(mp_ptr dinv, mp_srcptr d, mp_size_t n)

Compute an \(n\) limb precomputed inverse dinv of the \(n\) limb integer \(d\).

We require that \(d\) is normalised, i.e. with the most significant bit of the most significant limb set.

void flint_mpn_mod_preinvn(mp_ptr r, mp_srcptr a, mp_size_t m, mp_srcptr d, mp_size_t n, mp_srcptr dinv)

Given a normalised integer \(d\) of \(n\) limbs, with precomputed inverse dinv provided by flint_mpn_preinvn and integer \(a\) of \(m\) limbs, computes \(a \pmod{d}\) and stores the result in-place in the lower \(n\) limbs of \(a\). The remaining limbs of \(a\) are destroyed.

We require \(m \geq n\). No aliasing of \(a\) with any of the other operands is permitted.

Note that this function is not always as fast as ordinary division.

mp_limb_t flint_mpn_divrem_preinvn(mp_ptr q, mp_ptr r, mp_srcptr a, mp_size_t m, mp_srcptr d, mp_size_t n, mp_srcptr dinv)

Given a normalised integer \(d\) with precomputed inverse dinv provided by flint_mpn_preinvn, computes the quotient of \(a\) by \(d\) and stores the result in \(q\) and the remainder in the lower \(n\) limbs of \(a\). The remaining limbs of \(a\) are destroyed.

The value \(q\) is expected to have space for \(m - n\) limbs and we require \(m \ge n\). No aliasing is permitted between \(q\) and \(a\) or between these and any of the other operands.

Note that this function is not always as fast as ordinary division.

void flint_mpn_mulmod_preinvn(mp_ptr r, mp_srcptr a, mp_srcptr b, mp_size_t n, mp_srcptr d, mp_srcptr dinv, ulong norm)

Given a normalised integer \(d\) with precomputed inverse dinv provided by flint_mpn_preinvn, computes \(ab \pmod{d}\) and stores the result in \(r\). Each of \(a\), \(b\) and \(r\) is expected to have \(n\) limbs of space, with zero padding if necessary.

The value norm is provided for convenience. If \(a\), \(b\) and \(d\) have been shifted left by norm bits so that \(d\) is normalised, then \(r\) will be shifted right by norm bits so that it has the same shift as all the inputs.

We require \(a\) and \(b\) to be reduced modulo \(n\) before calling the function.

void flint_mpn_mulmod_preinvn_2(mp_ptr r, mp_srcptr a, mp_srcptr b, mp_srcptr d, mp_srcptr dinv, ulong norm)

Version of flint_mpn_mulmod_preinv1() specialized for two limbs. The behavior is not exactly the same: \(a\) and \(b\) are assumed to be unshifted, and the output is unshifted.

GCD

mp_size_t flint_mpn_gcd_full2(mp_ptr arrayg, mp_srcptr array1, mp_size_t limbs1, mp_srcptr array2, mp_size_t limbs2, mp_ptr temp)
Sets (arrayg, retvalue) to the gcd of (array1, limbs1) and

(array2, limbs2).

The only assumption is that neither limbs1 nor limbs2 is zero.

The function must be supplied with limbs1 + limbs2 limbs of temporary space, or NULL must be passed to temp if the function should allocate its own space.

mp_size_t flint_mpn_gcd_full(mp_ptr arrayg, mp_srcptr array1, mp_size_t limbs1, mp_srcptr array2, mp_size_t limbs2)

Sets (arrayg, retvalue) to the gcd of (array1, limbs1) and (array2, limbs2).

The only assumption is that neither limbs1 nor limbs2 is zero.

Random Number Generation

void flint_mpn_rrandom(mp_ptr rp, flint_rand_t state, mp_size_t n)

Generates a random number with n limbs and stores it on rp. The number it generates will tend to have long strings of zeros and ones in the binary representation.

Useful for testing functions and algorithms, since this kind of random numbers have proven to be more likely to trigger corner-case bugs.

void flint_mpn_urandomb(mp_ptr rp, flint_rand_t state, flint_bitcnt_t n)

Generates a uniform random number of n bits and stores it on rp.