Function complex_mul_and_reduce

tests/validation/reference/DFT.cpp:261–311

 */
template <typename T>
SimpleTensor<T> complex_mul_and_reduce(const SimpleTensor<T> &input, const SimpleTensor<T> &weights)
{
    const uint32_t W  = input.shape().x();
    const uint32_t H  = input.shape().y();
    const uint32_t Ci = input.shape().z();
    const uint32_t Co = weights.shape()[3];
    const uint32_t N  = input.shape().total_size() / (W * H * Ci);

    TensorShape output_shape = input.shape();
    output_shape.set(2, Co);
    SimpleTensor<T> dst(output_shape, input.data_type(), input.num_channels());

    // dst memory to zero
    const auto total_element_count = dst.num_channels() * dst.num_elements();
    std::fill_n(dst.data(), total_element_count, 0);

    for (uint32_t b = 0; b < N; ++b)
    {
        for (uint32_t co = 0; co < Co; ++co)
        {
            for (uint32_t ci = 0; ci < Ci; ++ci)
            {
                for (uint32_t h = 0; h < H; ++h)
                {
                    for (uint32_t w = 0; w < W; ++w)
                    {
                        const uint32_t    i_index  = w + h * W + ci * H * W + b * H * W * Ci;
                        const uint32_t    w_index  = w + h * W + ci * H * W + co * H * W * Ci;
                        const uint32_t    o_index  = w + h * W + co * H * W + b * H * W * Co;
                        const Coordinates i_coords = index2coords(input.shape(), i_index);
                        const Coordinates w_coords = index2coords(weights.shape(), w_index);
                        const Coordinates o_coords = index2coords(dst.shape(), o_index);

                        auto i_ptr = static_cast<const T *>(input(i_coords));
                        auto w_ptr = static_cast<const T *>(weights(w_coords));
                        auto o_ptr = static_cast<T *>(dst(o_coords));

                        const T Rin = i_ptr[0];
                        const T Iin = i_ptr[1];
                        const T Rw  = w_ptr[0];
                        const T Iw  = w_ptr[1];

                        o_ptr[0] += Rin * Rw - Iin * Iw;
                        o_ptr[1] += Rin * Iw + Rw * Iin;
                    }
                }
            }
        }
    }
    return dst;
}
} // namespace

template <typename T>