libdpf/include/dpf/doerner_shelat.hpp

/// @file dpf/doerner_shelat.hpp
/// @brief Doerner–Shelat generation of a dealer DPF key.
/// @details Two shares of the point are walked level by level — XOR shares by
///          default, or additive shares when tagged with `arith_input`.
///          Correction words, advice bits, seeds, and leaves are the ones
///          `make_dpf` would emit for the reconstructed point, the same roots,
///          and the same beaver coins. Beaver pads used to hide the path bit
///          cancel and are not part of the key. Pad randomness must not come
///          from `uniform_fill` if the beaver tape is being matched.
/// @copyright Copyright (c) 2019-2026 Ryan Henry and [others](@ref authors)
/// @license Released under a GNU General Public v2.0 (GPLv2) license;
///          see [LICENSE.md](@ref license) for details.

#ifndef LIBDPF_INCLUDE_DPF_DOERNER_SHELAT_HPP__
#define LIBDPF_INCLUDE_DPF_DOERNER_SHELAT_HPP__

#include <cstddef>
#include <cstdint>
#include <type_traits>
#include <utility>

#include "hedley/hedley.h"
#include "simde/simde/x86/avx2.h"

#include "dpf/dpf_key.hpp"
#include "dpf/random.hpp"
#include "dpf/dcf.hpp"

namespace dpf
{

/// Tag: Doerner–Shelat / geneval takes additive shares of the point
/// (`x0 + x1` in the input ring). Default calls take XOR shares.
struct arith_input_t
{
};

inline constexpr arith_input_t arith_input{};

/// Roots and the Beaver-pad stream for one Doerner–Shelat generation.
/// `root` is called twice, same as `make_dpf`: party 0 clears the low bit of
/// the first sample, party 1 sets the low bit of the second.
template <typename RootSampler, typename PadRng>
struct ds_randomness
{
    RootSampler root;
    PadRng pad;
};

namespace detail
{

struct urandom_pad_rng
{
    simde__m128i block()
    {
        return dpf::uniform_sample<simde__m128i>();
    }

    uint8_t bit()
    {
        return static_cast<uint8_t>(dpf::uniform_sample<unsigned char>() & 1u);
    }
};

struct ds_cw_party
{
    simde__m128i rand;
    simde__m128i gamma;
    uint8_t bit;
};

struct ds_cw_pads
{
    ds_cw_party p0;
    ds_cw_party p1;
};

struct ds_blind
{
    simde__m128i msg;
    uint8_t bit;
};

struct ds_and_pads
{
    uint8_t a0;
    uint8_t a1;
    simde__m128i b0_share, b1_share, c0_share, c1_share;
};

struct ds_and_shares
{
    simde__m128i z0;
    simde__m128i z1;
};

HEDLEY_NO_THROW
HEDLEY_ALWAYS_INLINE
simde__m128i ds_xor(simde__m128i a, simde__m128i b) noexcept
{
    return simde_mm_xor_si128(a, b);
}

HEDLEY_NO_THROW
HEDLEY_ALWAYS_INLINE
simde__m128i ds_gate(uint8_t bit, simde__m128i block) noexcept
{
    return dpf::get_if(block, bit & 1u);
}

template <typename PadRng>
ds_cw_pads ds_sample_cw(PadRng & pad)
{
    ds_cw_pads p{};
    const simde__m128i zero = simde_mm_setzero_si128();
    p.p0.rand = pad.block();
    p.p1.rand = pad.block();
    p.p0.bit = static_cast<uint8_t>(pad.bit() & 1u);
    p.p1.bit = static_cast<uint8_t>(pad.bit() & 1u);
    p.p0.gamma = p.p1.bit ? p.p0.rand : zero;
    p.p1.gamma = p.p0.bit ? p.p1.rand : zero;
    return p;
}

template <typename PadRng>
ds_and_pads ds_sample_and(PadRng & pad)
{
    ds_and_pads p{};
    const uint8_t a = static_cast<uint8_t>(pad.bit() & 1u);
    const simde__m128i B = pad.block();
    const simde__m128i C = ds_gate(a, B);
    p.a0 = static_cast<uint8_t>(pad.bit() & 1u);
    p.a1 = static_cast<uint8_t>(a ^ p.a0);
    p.b0_share = pad.block();
    p.b1_share = ds_xor(B, p.b0_share);
    p.c0_share = pad.block();
    p.c1_share = ds_xor(C, p.c0_share);
    return p;
}

HEDLEY_NO_THROW
HEDLEY_ALWAYS_INLINE
simde__m128i ds_cw_share(simde__m128i L, simde__m128i R, uint8_t my_bit,
    const ds_cw_party & mine, const ds_blind & their) noexcept
{
    simde__m128i out = ds_xor(R, mine.gamma);
    if (my_bit & 1u)
    {
        out = ds_xor(out, ds_xor(ds_xor(L, R), their.msg));
    }
    if (their.bit & 1u)
    {
        out = ds_xor(out, mine.rand);
    }
    return out;
}

HEDLEY_NO_THROW
inline void ds_cw_blinds(const ds_cw_pads & p,
    simde__m128i L0, simde__m128i R0, uint8_t bit0,
    simde__m128i L1, simde__m128i R1, uint8_t bit1,
    ds_blind & b0, ds_blind & b1) noexcept
{
    b0.bit = static_cast<uint8_t>(bit0 ^ p.p0.bit);
    b1.bit = static_cast<uint8_t>(bit1 ^ p.p1.bit);
    b0.msg = ds_xor(ds_xor(L0, R0), p.p0.rand);
    b1.msg = ds_xor(ds_xor(L1, R1), p.p1.rand);
}

HEDLEY_NO_THROW
inline simde__m128i ds_cw_outs(const ds_cw_pads & p,
    simde__m128i L0, simde__m128i R0, uint8_t bit0,
    simde__m128i L1, simde__m128i R1, uint8_t bit1,
    const ds_blind & b0, const ds_blind & b1) noexcept
{
    return ds_xor(
        ds_cw_share(L0, R0, bit0, p.p0, b1),
        ds_cw_share(L1, R1, bit1, p.p1, b0));
}

HEDLEY_NO_THROW
inline uint8_t ds_open_advice(simde__m128i L0, simde__m128i R0, uint8_t bit0,
    simde__m128i L1, simde__m128i R1, uint8_t bit1) noexcept
{
    const uint8_t a00 = static_cast<uint8_t>(dpf::get_lo_bit(L0) ^ bit0);
    const uint8_t a01 = static_cast<uint8_t>(dpf::get_lo_bit(R0) ^ bit0);
    const uint8_t a10 = static_cast<uint8_t>(dpf::get_lo_bit(L1) ^ bit1);
    const uint8_t a11 = static_cast<uint8_t>(dpf::get_lo_bit(R1) ^ bit1);
    const uint8_t t0 = static_cast<uint8_t>(a00 ^ a10 ^ 1u);
    const uint8_t t1 = static_cast<uint8_t>(a01 ^ a11);
    return static_cast<uint8_t>((t1 << 1) | (t0 & 1u));
}

HEDLEY_NO_THROW
inline void ds_next_terms(simde__m128i L, simde__m128i R, uint8_t advice,
    simde__m128i cw, uint8_t tpack, simde__m128i & M, simde__m128i & base) noexcept
{
    const simde__m128i D = ds_xor(L, R);
    const uint8_t t0 = static_cast<uint8_t>(tpack & 1u);
    const uint8_t t1 = static_cast<uint8_t>((tpack >> 1) & 1u);
    const simde__m128i lo = dpf::set_lo_bit(simde_mm_setzero_si128(), 1);
    const simde__m128i DT = ds_gate(static_cast<uint8_t>(t0 ^ t1), lo);
    const simde__m128i cw_base = ds_xor(dpf::unset_lo_bit(cw), ds_gate(t0, lo));
    M = (advice & 1u) ? ds_xor(D, DT) : D;
    base = (advice & 1u) ? ds_xor(L, cw_base) : L;
}

HEDLEY_NO_THROW
inline ds_and_shares ds_and_open(const ds_and_pads & p, simde__m128i M,
    uint8_t b_recv) noexcept
{
    const simde__m128i e = ds_xor(ds_xor(M, p.b0_share), p.b1_share);
    const uint8_t d = static_cast<uint8_t>((b_recv ^ p.a1) ^ p.a0);
    ds_and_shares z;
    z.z0 = ds_xor(ds_xor(ds_xor(ds_gate(d, e), ds_gate(d, p.b0_share)),
        ds_gate(p.a0, e)), p.c0_share);
    z.z1 = ds_xor(ds_xor(ds_gate(d, p.b1_share), ds_gate(p.a1, e)), p.c1_share);
    return z;
}

HEDLEY_NO_THROW
inline simde__m128i ds_deliver(uint8_t b_exp, simde__m128i base, simde__m128i M,
    const ds_and_shares & z) noexcept
{
    const simde__m128i local = ds_xor(base, ds_gate(b_exp, M));
    return ds_xor(ds_xor(local, z.z0), z.z1);
}

/// Per-level messages prepared before the CW protocol runs (blinds + pads).
struct ds_level_blinds
{
    ds_cw_pads cwp;
    ds_and_pads and0;
    ds_and_pads and1;
    ds_blind b0;
    ds_blind b1;
    simde__m128i L0, R0, L1, R1;
    uint8_t bit0;
    uint8_t bit1;
};

/// Opened CW, advice, and AND products delivered by a `CwProtocol`.
struct ds_level_open
{
    simde__m128i cw;
    uint8_t advice;
    ds_and_shares z0;
    ds_and_shares z1;
    uint64_t value_cw = 0;  // public after open when cmp is active at this level
};

/// Running comparison-gen state shared across DS levels (Va residual).
/// When `track_coeff` is set (wildcard cmp payload), a parallel β = 1
/// accumulator `Va1` is advanced alongside `Va` so the gen can stash
/// `value_cw(1) − value_cw(β)` coefficients for a later `assign_cmp`.
struct ds_cmp_gen_state
{
    bool active = false;
    std::size_t nbits = 0;
    uint64_t mask = 0;
    uint64_t beta = 0;
    bool include_eq = false;
    cmp_trivial trivial = cmp_trivial::none;
    uint64_t Va = 0;
    unsigned __int128 thresh = 0;
    bool track_coeff = false;
    uint64_t Va1 = 0;
    uint64_t last_vcw_coeff = 0;
    cmp_kind kind = cmp_kind::lt;
    bool paint = false;
    std::size_t length_bits = 0;
    paint_callback paint_cb = nullptr;
    const void * paint_ctx = nullptr;
};

/// Local joint simulation: today's `ds_cw_outs` / `ds_open_advice` / `ds_and_open`.
/// An MPC backend would send `blinds` and return the same `ds_level_open` shape.
template <typename PadRng>
struct local_cw_protocol
{
    PadRng & pads;

    ds_level_blinds prepare_level(simde__m128i L0, simde__m128i R0, uint8_t bit0,
        simde__m128i L1, simde__m128i R1, uint8_t bit1)
    {
        ds_level_blinds b;
        b.cwp = ds_sample_cw(pads);
        b.and0 = ds_sample_and(pads);
        b.and1 = ds_sample_and(pads);
        b.L0 = L0;
        b.R0 = R0;
        b.L1 = L1;
        b.R1 = R1;
        b.bit0 = bit0;
        b.bit1 = bit1;
        ds_cw_blinds(b.cwp, L0, R0, bit0, L1, R1, bit1, b.b0, b.b1);
        return b;
    }

    ds_level_open complete_level(std::size_t /*level*/, const ds_level_blinds & b,
        simde__m128i M0, simde__m128i M1, uint8_t rec0, uint8_t rec1)
    {
        ds_level_open out;
        out.advice = ds_open_advice(b.L0, b.R0, b.bit0, b.L1, b.R1, b.bit1);
        out.cw = ds_cw_outs(b.cwp, b.L0, b.R0, b.bit0, b.L1, b.R1, b.bit1,
            b.b0, b.b1);
        out.z0 = ds_and_open(b.and0, M0, rec0);
        out.z1 = ds_and_open(b.and1, M1, rec1);
        return out;
    }

    /// Open CW + advice only (AND pads stay in `blinds` for a later open).
    HEDLEY_NO_THROW
    std::pair<simde__m128i, uint8_t> open_cw(const ds_level_blinds & b) noexcept
    {
        return {ds_cw_outs(b.cwp, b.L0, b.R0, b.bit0, b.L1, b.R1, b.bit1,
                    b.b0, b.b1),
            ds_open_advice(b.L0, b.R0, b.bit0, b.L1, b.R1, b.bit1)};
    }

    /// Open the public value CW for this level (local: clear convert+make_value_cw).
    /// MPC backends open additive shares of the same word.
    HEDLEY_NO_THROW
    uint64_t open_value_cw(const ds_level_blinds & b, uint8_t adv0, uint8_t adv1,
        int ai, uint64_t & Va, uint64_t beta, uint64_t mask) noexcept
    {
        return dcf_impl::make_value_cw(b.L0, b.R0, b.L1, b.R1, adv0,
            adv1, ai, Va, beta, mask);
    }

    /// Open a path-paint value CW. `plant` is the scaled lose-subtree constant.
    HEDLEY_NO_THROW
    uint64_t open_planted_cw(const ds_level_blinds & b, uint8_t adv0, uint8_t adv1,
        int ai, uint64_t & Va, uint64_t plant, uint64_t mask) noexcept
    {
        return dcf_impl::make_value_cw_planted(b.L0, b.R0, b.L1, b.R1, adv0,
            adv1, ai, Va, plant, mask);
    }

    HEDLEY_NO_THROW
    ds_and_shares open_and(const ds_and_pads & p, simde__m128i M,
        uint8_t b_recv) noexcept
    {
        return ds_and_open(p, M, b_recv);
    }

    /// Open the final comparison leaf CW. Wraps `make_final_cw` so the
    /// Doerner–Shelat gen does not call it directly on reconstructed seeds;
    /// an MPC backend would open additive shares of the same word.
    HEDLEY_NO_THROW
    uint64_t open_final_cw(simde__m128i s0, simde__m128i s1, uint8_t t1,
        uint64_t Va, uint64_t mask, uint64_t on_path) noexcept
    {
        return dcf_impl::make_final_cw(s0, s1, t1, Va, mask, on_path);
    }

    /// Draw the group-width `cmp_addend` blind. Local joint simulation reuses
    /// the shared root sampler so the blind matches the dealer's; an MPC
    /// backend would instead pull a group-width element from the pad stream.
    template <typename BlockSampler>
    HEDLEY_NO_THROW
    uint64_t sample_addend_blind(uint64_t mask, BlockSampler && sample) noexcept
    {
        return dcf_impl::sample_addend_blind(mask,
            std::forward<BlockSampler>(sample));
    }

    /// Majority of three bits (next carry of a full adder).
    HEDLEY_NO_THROW
    static constexpr uint8_t majority(uint8_t a, uint8_t b, uint8_t c) noexcept
    {
        return static_cast<uint8_t>((a & b) | (a & c) | (b & c));
    }

    /// One additive digit: sum bit `a XOR b XOR cin`, carry out = majority.
    HEDLEY_NO_THROW
    static constexpr uint8_t open_sum_bit(uint8_t a, uint8_t b, uint8_t cin,
        uint8_t & cout) noexcept
    {
        cout = majority(a, b, cin);
        return static_cast<uint8_t>(a ^ b ^ cin);
    }

    /// Reconstruct `a0 + a1` via a LSB→MSB carry chain, then flip the MSB
    /// when the domain is signed — matching `make_dpf` on the sum. The call
    /// site never forms the sum; an MPC backend would open the same bits.
    template <typename InputT>
    InputT open_arith_point(InputT a0, InputT a1) const
    {
        constexpr auto to_int = utils::to_integral_type<InputT>{};
        using FromI = typename utils::make_from_integral_value<InputT>::integral_type;
        using U = std::make_unsigned_t<FromI>;
        const U u0 = static_cast<U>(to_int(a0));
        const U u1 = static_cast<U>(to_int(a1));
        U sum = 0;
        uint8_t carry = 0;
        constexpr std::size_t nbits = utils::bitlength_of_v<InputT>;
        for (std::size_t i = 0; i < nbits; ++i)
        {
            const uint8_t b0 = static_cast<uint8_t>((u0 >> i) & U{1});
            const uint8_t b1 = static_cast<uint8_t>((u1 >> i) & U{1});
            const uint8_t s = open_sum_bit(b0, b1, carry, carry);
            sum = static_cast<U>(sum | (static_cast<U>(s) << i));
        }
        InputT out = utils::make_from_integral_value<InputT>{}(
            static_cast<FromI>(sum));
        utils::flip_msb_if_signed_integral(out);
        return out;
    }

    /// Encode shares for the XOR-style CW walk. XOR mode flips party 0's MSB
    /// (linear over XOR). Arithmetic mode opens the sum (carry + signed MSB)
    /// and returns `(alpha, 0)` so the walk matches `make_dpf(alpha)`.
    template <typename InputT>
    void encode_walk_shares(InputT & x0, InputT & x1, bool arith) const
    {
        if (arith)
        {
            const InputT alpha = open_arith_point(x0, x1);
            x0 = alpha;
            x1 = InputT{};
        }
        else
        {
            utils::flip_msb_if_signed_integral(x0);
        }
    }

    /// Open a group of leaf correction words for one prefix group. In this
    /// local joint simulation both XOR shares of the point are present, so the
    /// point is reconstructed *inside* the protocol and handed to `leaf_fn`
    /// (which runs `make_leaves` for the group). The Doerner–Shelat gen never
    /// forms `x = x0 ^ x1` at its own call site; an MPC backend would instead
    /// run a per-group leaf CW exchange that never reveals `x`. After
    /// `encode_walk_shares`, arithmetic inputs are already `(alpha, 0)`.
    template <typename InputT, typename LeafFn>
    void open_leaf_group(InputT x0, InputT x1, LeafFn && leaf_fn)
    {
        std::forward<LeafFn>(leaf_fn)(utils::xor_input_shares(x0, x1));
    }
};

/// Generation-side level state (seeds / home bits). Not an eval path memoizer.
template <typename NodeT>
struct ds_gen_state
{
    NodeT inbox[2];
    int home[2];
    NodeT root0;
    NodeT root1;

    HEDLEY_NO_THROW
    void init(NodeT r0, NodeT r1) noexcept
    {
        root0 = r0;
        root1 = r1;
        inbox[0] = r0;
        inbox[1] = r1;
        home[0] = 0;
        home[1] = 1;
    }

    HEDLEY_NO_THROW
    NodeT & seed0() noexcept { return inbox[home[0]]; }
    HEDLEY_NO_THROW
    NodeT & seed1() noexcept { return inbox[home[1]]; }
    HEDLEY_NO_THROW
    const NodeT & seed0() const noexcept { return inbox[home[0]]; }
    HEDLEY_NO_THROW
    const NodeT & seed1() const noexcept { return inbox[home[1]]; }
};

/// One interior level: expand, protocol open, advance both party seeds.
/// When `cmp` is non-null and active for `level`, also opens `value_cw` via
/// the protocol (no second PRG expand outside).
template <typename InteriorPRG, typename CwProtocol, typename NodeT,
          typename InputT, typename MaskT, typename AdviceT>
void ds_advance_level(ds_gen_state<NodeT> & st, InputT x0, InputT x1,
    MaskT mask, std::size_t level, CwProtocol & proto, NodeT & cw_out,
    AdviceT & advice_out, uint64_t * value_cw_out = nullptr,
    ds_cmp_gen_state * cmp = nullptr)
{
    // Integral bridge so bit extraction works for `keyword` / `modint` /
    // signed / bitstring the same way dealer gen does via `mask & x`.
    // `msb_mask` is the unsigned bit pattern; a signed input must not be
    // required to have that same type.
    constexpr auto to_int = utils::to_integral_type<InputT>{};
    constexpr auto to_mask = utils::to_integral_type<MaskT>{};
    const auto mi = to_mask(mask);
    const uint8_t bit0 = static_cast<uint8_t>(!!(mi & to_int(x0)));
    const uint8_t bit1 = static_cast<uint8_t>(!!(mi & to_int(x1)));

    NodeT s0 = st.seed0();
    NodeT s1 = st.seed1();
    const uint8_t adv0 = static_cast<uint8_t>(
        dpf::get_lo_bit_and_clear_lo_2bits(s0));
    const uint8_t adv1 = static_cast<uint8_t>(
        dpf::get_lo_bit_and_clear_lo_2bits(s1));
    const auto c0 = InteriorPRG::eval01(s0);
    const auto c1 = InteriorPRG::eval01(s1);

    auto blinds = proto.prepare_level(c0[0], c0[1], bit0, c1[0], c1[1], bit1);

    if (value_cw_out != nullptr && cmp != nullptr && cmp->active
        && cmp->trivial == cmp_trivial::none && level < cmp->nbits)
    {
        const int ai = static_cast<int>(
            (cmp->thresh >> (cmp->nbits - 1 - level)) & 1);
        if (cmp->paint)
        {
            const uint64_t unit = dcf_impl::paint_unit(cmp->kind, level,
                cmp->thresh, cmp->nbits, cmp->length_bits, false,
                cmp->paint_cb, cmp->paint_ctx);
            const uint64_t plant = dcf_impl::scale_plant(unit, cmp->beta,
                cmp->mask);
            *value_cw_out = proto.open_planted_cw(blinds, adv0, adv1, ai,
                cmp->Va, plant, cmp->mask);
            if (cmp->track_coeff)
            {
                const uint64_t plant1 = dcf_impl::scale_plant(unit, 1ULL,
                    cmp->mask);
                const uint64_t v1 = proto.open_planted_cw(blinds, adv0, adv1,
                    ai, cmp->Va1, plant1, cmp->mask);
                cmp->last_vcw_coeff =
                    (v1 + dcf_impl::neg_m(*value_cw_out, cmp->mask)) & cmp->mask;
            }
        }
        else
        {
            *value_cw_out = proto.open_value_cw(blinds, adv0, adv1, ai, cmp->Va,
                cmp->beta, cmp->mask);
            if (cmp->track_coeff)
            {
                // Affine coefficient: same level with β = 1 on a parallel Va.
                const uint64_t v1 = proto.open_value_cw(blinds, adv0, adv1, ai,
                    cmp->Va1, 1ULL, cmp->mask);
                cmp->last_vcw_coeff =
                    (v1 + dcf_impl::neg_m(*value_cw_out, cmp->mask)) & cmp->mask;
            }
        }
    }

    auto [cw, tpack] = proto.open_cw(blinds);

    const uint8_t exp0 = st.home[0] == 0 ? bit0 : bit1;
    const uint8_t rec0 = st.home[0] == 0 ? bit1 : bit0;
    const uint8_t exp1 = st.home[1] == 0 ? bit0 : bit1;
    const uint8_t rec1 = st.home[1] == 0 ? bit1 : bit0;

    NodeT M0, base0, M1, base1;
    ds_next_terms(c0[0], c0[1], adv0, cw, tpack, M0, base0);
    ds_next_terms(c1[0], c1[1], adv1, cw, tpack, M1, base1);
    const NodeT nxt0 =
        ds_deliver(exp0, base0, M0, proto.open_and(blinds.and0, M0, rec0));
    const NodeT nxt1 =
        ds_deliver(exp1, base1, M1, proto.open_and(blinds.and1, M1, rec1));

    st.home[0] ^= 1;
    st.home[1] ^= 1;
    st.inbox[st.home[0]] = nxt0;
    st.inbox[st.home[1]] = nxt1;

    cw_out = cw;
    advice_out = tpack;
}

template <typename T>
struct is_ds_randomness : std::false_type {};

template <typename RootSampler, typename PadRng>
struct is_ds_randomness<ds_randomness<RootSampler, PadRng>> : std::true_type {};

template <typename T, typename = void>
struct is_cw_protocol : std::false_type {};

template <typename PadRng>
struct is_cw_protocol<local_cw_protocol<PadRng>, void> : std::true_type {};

template <typename T>
struct is_cw_protocol<T,
    std::void_t<decltype(std::declval<T &>().prepare_level(
                    simde_mm_setzero_si128(), simde_mm_setzero_si128(),
                    uint8_t{}, simde_mm_setzero_si128(),
                    simde_mm_setzero_si128(), uint8_t{})),
                decltype(std::declval<T &>().open_cw(
                    std::declval<const ds_level_blinds &>()))>>
    : std::true_type {};

template <typename ...Ts>
struct first_is_cw_protocol : std::false_type {};

template <typename T, typename ...Rest>
struct first_is_cw_protocol<T, Rest...>
    : is_cw_protocol<std::decay_t<T>> {};

template <typename InteriorPRG,
          typename ExteriorPRG,
          typename InputT,
          typename OutputT,
          typename ...OutputTs,
          typename RootSampler,
          typename CwProtocol>
auto make_dpf_doerner_shelat_impl(bool arith, InputT x0, InputT x1,
    RootSampler & root_sampler, CwProtocol & proto, OutputT && y,
    OutputTs && ...ys)
{
    static_assert(!dpf::is_wildcard_v<InputT>,
        "Doerner–Shelat gen takes shares of a concrete point");
    static_assert(!dpf::is_secret_share_v<InputT>,
        "Doerner–Shelat: pass additive_share of xor_wrapper, or raw shares");
    static_assert(sizeof(typename InteriorPRG::block_type) == sizeof(simde__m128i),
        "Doerner–Shelat gen uses the AES-block interior node");

    using dpf_type = utils::dpf_type_t<InteriorPRG, ExteriorPRG, InputT,
                                       OutputT, OutputTs...>;
    using node = typename dpf_type::interior_node;
    using input_type = typename dpf_type::input_type;
    constexpr auto depth = dpf_type::depth;

    proto.encode_walk_shares(x0, x1, arith);

    const node root0 = dpf::unset_lo_bit(static_cast<node>(root_sampler()));
    const node root1 = dpf::set_lo_bit(static_cast<node>(root_sampler()));

    ds_gen_state<node> st;
    st.init(root0, root1);

    typename dpf_type::correction_words_array correction_words{};
    typename dpf_type::correction_advice_array correction_advice{};

    auto mask = dpf_type::msb_mask;
    for (std::size_t level = 0; level < depth; ++level, mask >>= 1)
    {
        ds_advance_level<InteriorPRG>(st, x0, x1, mask, level, proto,
            correction_words[level], correction_advice[level]);
    }

    const node parent0 = st.seed0();
    const node parent1 = st.seed1();
    const bool sign0 = dpf::get_lo_bit(parent0);

    input_type x = utils::xor_input_shares(x0, x1);
    auto built = dpf::make_leaves<ExteriorPRG>(x,
        dpf::unset_lo_2bits(parent0), dpf::unset_lo_2bits(parent1), sign0,
        std::size_t{0}, std::forward<OutputT>(y), std::forward<OutputTs>(ys)...);

    input_type off0{};
    input_type off1{};
    return dpf::make_party_key_pair(
        dpf_type{root0, correction_words, correction_advice,
            built.first.first, built.first.second, off0},
        dpf_type{root1, correction_words, correction_advice,
            built.second.first, built.second.second, off1});
}

}  // namespace detail

/// Local CW protocol (pads cancel; same keys as dealer when roots match).
template <typename PadRng>
using local_cw_protocol = detail::local_cw_protocol<PadRng>;

template <typename NodeT>
using ds_gen_state = detail::ds_gen_state<NodeT>;

// Public `make_dpf_doerner_shelat(x0, x1, ...)` lives in incremental.hpp so
// classic and `at<>` / mixed-width packs share one entry point.

}  // namespace dpf

#endif  // LIBDPF_INCLUDE_DPF_DOERNER_SHELAT_HPP__