Function lut_dists_12b

cpp/src/quantize/multi_codebook.hpp:433–476

template<int NBytes, class dist_t>
inline void lut_dists_12b(const uint8_t* codes, const dist_t* luts,
    dist_t* dists_out, int64_t N)
{
    static constexpr int ncodebook_bits = 12;
    static constexpr int total_bits = 8 * NBytes;
    static constexpr int ncodewords = total_bits / ncodebook_bits;
    static constexpr int codebook_sz = 1 << ncodebook_bits;
    static constexpr bool nbytes_multiple_of_8 = NBytes % 8 == 0;
    static constexpr uint16_t idx_mask = (1 << ncodebook_bits) - 1;
    static constexpr uint8_t low_4bits_mask = 0x0F;
    static constexpr uint8_t high_4bits_mask = 0xF0;
    static_assert(NBytes >= 2, "NBytes < 2 is not meaningful");

    for (int64_t i = 0; i < N; i++) {
        dists_out[i] = 0;
        for (int j = 0; j < ncodewords; j += 2) {
            auto lut_ptr0 = luts + codebook_sz * j;
            auto lut_ptr1 = lut_ptr0 + codebook_sz;

            // packed indices format:
            //
            // codes[j]             codes[j+1]           codes[j+2]
            // -------------------- ----------|---------- --------------------
            //       idx0[:8]       idx0[8:12] idx1[8:12]       idx1[:8]
            uint16_t low_bits = codes[3*j+1] & low_4bits_mask;
            uint16_t high_bits = codes[3*j+1] & high_4bits_mask;
            uint16_t idx0 = codes[3*j] + (low_bits << 8);
            uint16_t idx1 = codes[3*j+2] + (high_bits << 4);

            dists_out[i] += lut_ptr0[idx0];
            dists_out[i] += lut_ptr1[idx1];

            // codes += 3;
        }
        if (ncodewords % 2 != 0) {
            auto lut_ptr = luts + codebook_sz * (ncodewords - 1);
            uint16_t low_bits = codes[NBytes-1] & low_4bits_mask; // use last byte
            auto idx = codes[NBytes-2] + (low_bits << 8); // use 2nd to last byte
            dists_out[i] += lut_ptr[idx];
        }
        codes += NBytes;
        // codes += (NBytes % 3);
    }
}
template<int B, int NBytes, class dist_t>
inline void lut_dists_12b_vertical(const uint8_t* codes, const dist_t* luts,
    dist_t* dists_out, int64_t N)