gate_separator.hpp

Go to the documentation of this file.

// === AUDIT STATUS ===
// internal:    { status: Complete, auditors: [Khashayar], commit: }
// external_1:  { status: not started, auditors: [], commit: }
// external_2:  { status: not started, auditors: [], commit: }
// =====================
 
#pragma once
#include "barretenberg/common/assert.hpp"
#include "barretenberg/common/bb_bench.hpp"
#include "barretenberg/common/compiler_hints.hpp"
#include "barretenberg/common/thread.hpp"
#include "barretenberg/numeric/bitop/get_msb.hpp"
#include "barretenberg/stdlib/primitives/bool/bool.hpp"
 
#include <cstddef>
#include <vector>
namespace bb {
 
template <typename FF> struct GateSeparatorPolynomial {
    std::vector<FF> betas;
 
    Polynomial<FF> beta_products;
    size_t current_element_idx = 0;
    size_t periodicity = 2;
    FF partial_evaluation_result = FF(1);
 
    GateSeparatorPolynomial(const std::vector<FF>& betas, const size_t log_num_monomials)
        : betas(betas)
        , beta_products(compute_beta_products(betas, log_num_monomials))
    {}
 
    GateSeparatorPolynomial(const std::vector<FF>& betas)
        : betas(betas)
    {}
 
    GateSeparatorPolynomial(const std::vector<FF>& betas, const std::vector<FF>& challenge)
        : betas(betas)
    {
        if (!betas.empty()) {
            for (const auto& u_k : challenge) {
                partially_evaluate(u_k);
            }
        }
    }
 
    FF const& operator[](size_t idx) const
    {
        // At round i (periodicity == 2^{i+1}), the idx-th surviving evaluation lives at beta_products index
        // (idx >> 1) * periodicity, i.e. the (idx * 2^i)-th element. Sumcheck consumes edges pairwise, so only
        // even indices are meaningful; an odd idx would silently alias to the idx - 1 slot.
        BB_ASSERT_DEBUG(idx % 2 == 0, "GateSeparatorPolynomial: edge index must be even");
        return beta_products.at((idx >> 1) * periodicity);
    }
 
    template <typename Element> Element gather(size_t edge_idx) const
    {
        return Element::from_lanes([&](size_t j) { return (*this)[edge_idx + (2 * j)]; });
    }
    FF current_element() const
    {
        if (betas.empty()) {
            return FF(1);
        };
        return betas[current_element_idx];
    }
 
    static FF univariate_factor(const FF& challenge, const FF& beta) { return FF(1) + (challenge * (beta - FF(1))); }
 
    FF univariate_eval(FF challenge) const { return univariate_factor(challenge, betas[current_element_idx]); };
 
    void partially_evaluate(FF challenge)
    {
        if (!betas.empty()) {
            FF current_univariate_eval = univariate_eval(challenge);
            partial_evaluation_result *= current_univariate_eval;
            current_element_idx++;
            periodicity *= 2;
        }
    }
 
    BB_PROFILE static Polynomial<FF> compute_beta_products(const std::vector<FF>& betas,
                                                           const size_t log_num_monomials,
                                                           const FF& scaling_factor = FF(1))
    {
        if (betas.empty()) {
            Polynomial<FF> out(1);
            return out;
        }
 
        BB_BENCH_NAME("GateSeparatorPolynomial::compute_beta_products");
        size_t pow_size = static_cast<size_t>(1) << log_num_monomials;
        Polynomial<FF> beta_products(pow_size, Polynomial<FF>::DontZeroMemory::FLAG);
 
        // Explanations of the algorithm:
        // The product of the betas at index i (beta_products[i]) contains the multiplicative factor betas[j] if and
        // only if the jth bit of i is 1 (j starting with 0 for the least significant bit). For instance, i = 13 = 1101
        // in binary, so the product is betas[0] * betas[2] * betas[3].
        //
        // Key insight: beta_products[i] = beta_products[predecessor] * betas[lsb_position], where predecessor is i
        // with the least significant bit cleared. For example:
        //   - i = 6 (binary 110): LSB is at position 1, predecessor = 4 (binary 100)
        //     beta_products[6] = beta_products[4] * betas[1]
        //   - i = 12 (binary 1100): LSB is at position 2, predecessor = 8 (binary 1000)
        //     beta_products[12] = beta_products[8] * betas[2]
        //
        // For each index i, if the predecessor falls within our thread's range [start, start + chunk_size), we use
        // this O(1) recurrence. Otherwise, we compute directly by iterating over all set bits in i, which requires
        // O(popcount(i)) multiplications. This direct computation handles boundary cases between thread chunks.
        //
        // This algorithm works with any number of threads (not just powers of 2), unlike the previous prefix/suffix
        // approach which required power-of-2 thread counts to ensure even work distribution.
 
        // Cost per iteration: typically 1 multiplication (when predecessor is in range),
        // occasionally O(popcount) multiplications at chunk boundaries
        constexpr size_t iteration_cost = thread_heuristics::FF_MULTIPLICATION_COST;
        parallel_for_heuristic(
            pow_size,
            [&](size_t start, size_t end, BB_UNUSED size_t chunk_index) {
                BB_BENCH_TRACY_NAME("GateSeparator::beta_products/chunk");
                for (size_t i = start; i < end; i++) {
                    if (i == 0) {
                        beta_products.at(0) = scaling_factor;
                        continue;
                    }
 
                    // Find the lowest set bit position and the predecessor index
                    size_t lsb_pos = numeric::get_lsb(i);
                    size_t predecessor = i ^ (static_cast<size_t>(1) << lsb_pos); // clear the lowest set bit
 
                    if (predecessor >= start) {
                        // Predecessor is in our range, O(1) computation
                        beta_products.at(i) = beta_products.at(predecessor) * betas[lsb_pos];
                    } else {
                        // Predecessor is not in our range, compute directly from set bits only
                        FF result = scaling_factor;
                        size_t remaining = i;
                        while (remaining != 0) {
                            size_t bit = numeric::get_lsb(remaining);
                            result *= betas[bit];
                            remaining ^= static_cast<size_t>(1) << bit; // clear this bit
                        }
                        beta_products.at(i) = result;
                    }
                }
            },
            iteration_cost);
 
        return beta_products;
    }
};
} // namespace bb