Matrix multiplication over Z/nZ for multiword n

mpn_mod_mat_mul is the matrix multiplication method of the mpn_mod ring: gr_mat_mul over the ring calls it, and so do the generic block algorithms built on matrix multiplication (recursive LU decomposition, triangular solving, inversion, determinants). It chooses between three algorithms, all of which return the exact product modulo n:

Here b is the bit size of n and \ell its number of limbs.

Choice of algorithm#

Matrices with at most 3 rows use classical multiplication. Otherwise two cutoffs c_W \le c_M are read from a table indexed by b in steps of 16 bits, tuned on x86-64:

• if m, k or p is less than c_W, the classical algorithm is used;
• otherwise, if one of them is less than c_M, Waksman's algorithm;
• otherwise the multimodular algorithm.

For moduli of at most 128 bits and of 177 to 192 bits the two cutoffs are equal, so Waksman's algorithm is not used: the classical algorithm runs up to dimension 88 to 230 (for b \le 128) or 74 (for 177 \le b \le 192). For other sizes Waksman's algorithm starts at dimension 4 to 17 and the multimodular algorithm takes over between dimensions 39 and 93; at 256 bits the cutoffs are 17 and 63, at 1024 bits 4 and 53. Strassen's algorithm is not used: at these sizes the multimodular algorithm wins first.

Classical multiplication#

Each entry of C is a dot product of a row of A and a column of B, computed by _mpn_mod_vec_dot_strided without intermediate reductions: the full products are added in \lceil (2b + \lfloor \log_2 k \rfloor + 1)/\mathtt{FLINT\_BITS} \rceil limbs, and the sum is reduced once (Dot products).

Waksman's algorithm#

Winograd [Win1968] observed that for two vectors of even length k,

In a matrix product, x is a row of A and y a column of B, so the last two sums depend only on the row or only on the column and can be computed once for each: the product costs mkp/2 + (m + p) k/2 multiplications instead of mkp. Waksman [Waksman1970] noticed that the correction terms can be obtained from the same kind of products. Writing R_l = \sum_i a_{l,2i-1} a_{l,2i} and S_t = \sum_i b_{2i,t} b_{2i-1,t}, the identity

gives R_0 + S_t for every column t from the products for row 0, and R_l + S_0 for every row l from those for column 0; then R_l + S_t = (R_l + S_0) + (R_0 + S_t) - (R_0 + S_0). The extra products (\alpha - \beta)(\alpha' - \beta') are only needed in the first row and column. For odd k the last term is added with ordinary products.

mpn_mod_mat_mul_waksman evaluates all sums exactly on integers and reduces each entry of C once. Accumulators have 2\ell + 1 limbs and hold signed values. The sums \alpha + \beta of two residues fit in \ell limbs unless the top bit of n is set, in which case they get an extra limb and the products are (\ell + 1) \times (\ell + 1)-limb products; the differences \alpha - \beta are formed as sign and magnitude. The halvings are exact divisions by 2 of signed multi-limb integers. Since a multiplication of \ell-limb numbers costs about \ell^2 word operations and an addition about \ell, halving the number of multiplications at the price of O(mkp) additions pays off from small dimensions on.

Multimodular multiplication#

mpn_mod_mat_mul_multi_mod computes the integer product AB of the matrices of residues, whose entries are less than k (n-1)^2 < 2^{c} with c = 2b + \lfloor \log_2 k \rfloor + 1, modulo r word-size moduli whose product exceeds 2^c, and reconstructs it by the Chinese remainder theorem:

1. The moduli are 2^{\beta} and the r - 1 smallest primes above 2^{\beta}, where \beta = NMOD_MAT_OPTIMAL_MODULUS_BITS (59 on 64-bit machines), with r = 1 + \lceil (c - \beta)/\beta \rceil. Using a power of two as the first modulus lets the fixed-length code of flint_mpn_multi_crt apply for small products.
2. Each row of A and of B is reduced modulo all moduli at once with flint_mpn_multi_mod_vec.
3. The r products are computed with nmod_mat_mul, which itself uses delayed reduction, Strassen–Winograd, SIMD or BLAS kernels (Multiplying matrices over Z/nZ).
4. Each row of C is reconstructed with flint_mpn_multi_crt_vec and each entry reduced modulo n with mpn_mod_set_mpn.

The reductions and the reconstruction are split over threads (by rows, when the dimensions are in the hundreds or larger), and nmod_mat_mul is multithreaded itself. The code follows the multimodular multiplication of fmpz_mat (Multimodular and p-adic linear algebra over the integers).

Example#

The program below multiplies two 30 \times 30 matrices modulo 2^{255} - 19 with the three algorithms and checks that they agree.

#include <flint/fmpz.h>
#include <flint/gr.h>
#include <flint/gr_mat.h>
#include <flint/mpn_mod.h>

int main(void)
{
    gr_ctx_t ctx;
    fmpz_t p;
    gr_mat_t A, B, C1, C2, C3;
    slong i, j, d = 30;
    int status = GR_SUCCESS;

    /* p = 2^255 - 19 */
    fmpz_init(p);
    fmpz_ui_pow_ui(p, 2, 255);
    fmpz_sub_ui(p, p, 19);
    GR_MUST_SUCCEED(gr_ctx_init_mpn_mod(ctx, p));

    gr_mat_init(A, d, d, ctx);
    gr_mat_init(B, d, d, ctx);
    gr_mat_init(C1, d, d, ctx);
    gr_mat_init(C2, d, d, ctx);
    gr_mat_init(C3, d, d, ctx);

    /* A[i][j] = ij - 7, B[i][j] = A[i][j]^40 */
    for (i = 0; i < d; i++)
        for (j = 0; j < d; j++)
        {
            status |= gr_set_si(gr_mat_entry_ptr(A, i, j, ctx), i * j - 7, ctx);
            status |= gr_pow_ui(gr_mat_entry_ptr(B, i, j, ctx),
                                gr_mat_entry_ptr(A, i, j, ctx), 40, ctx);
        }

    status |= gr_mat_mul_classical(C1, A, B, ctx);
    status |= mpn_mod_mat_mul_waksman(C2, A, B, ctx);
    status |= mpn_mod_mat_mul_multi_mod(C3, A, B, ctx);

    printf("waksman: %d, multi_mod: %d\n",
        gr_mat_equal(C1, C2, ctx) == T_TRUE, gr_mat_equal(C1, C3, ctx) == T_TRUE);
    status |= gr_println(gr_mat_entry_ptr(C1, 1, 2, ctx), ctx);

    gr_mat_clear(A, ctx);
    gr_mat_clear(B, ctx);
    gr_mat_clear(C1, ctx);
    gr_mat_clear(C2, ctx);
    gr_mat_clear(C3, ctx);
    gr_ctx_clear(ctx);
    fmpz_clear(p);
    printf("status %d\n", status);
    return 0;
}

waksman: 1, multi_mod: 1
5452489965307334725493791908297774450834361894610887641359429602394561
status 0