Function mul_saturate_quantized_8

src/cpu/kernels/CpuMulKernel.cpp:191–337

template <typename T>
void mul_saturate_quantized_8(const ITensor *src1, const ITensor *src2, ITensor *out, const Window &window, float scale)
{
    // Create input windows
    Window win        = window;
    Window input1_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
    Window input2_win = window.broadcast_if_dimension_le_one(src2->info()->tensor_shape());

    // Clear X Dimension on execution window as we handle manually
    win.set(Window::DimX, Window::Dimension(0, 1, 1));

    const int  window_step_x         = 16 / sizeof(T);
    const auto window_start_x        = static_cast<int>(window.x().start());
    const auto window_end_x          = static_cast<int>(window.x().end());
    const bool is_broadcast_across_x = src1->info()->tensor_shape().x() != src2->info()->tensor_shape().x();

    const UniformQuantizationInfo output_qua_info = out->info()->quantization_info().uniform();
    const UniformQuantizationInfo tmp_qua_info    = {output_qua_info.scale / scale, output_qua_info.offset};

    if (is_broadcast_across_x)
    {
        const bool                    is_broadcast_input_2 = input2_win.x().step() == 0;
        Window                        broadcast_win        = is_broadcast_input_2 ? input2_win : input1_win;
        Window                        non_broadcast_win    = !is_broadcast_input_2 ? input2_win : input1_win;
        const ITensor                *broadcast_tensor     = is_broadcast_input_2 ? src2 : src1;
        const ITensor                *non_broadcast_tensor = !is_broadcast_input_2 ? src2 : src1;
        const UniformQuantizationInfo broadcast_qinfo      = broadcast_tensor->info()->quantization_info().uniform();
        const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();

        // Clear X Dimension on execution window as we handle manually
        non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));

        Iterator broadcast_input(broadcast_tensor, broadcast_win);
        Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
        Iterator dst(out, win);

        using ExactTagType = typename wrapper::traits::neon_vector<T, window_step_x>::tag_type;

        execute_window_loop(
            win,
            [&](const Coordinates &)
            {
                const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
                const auto output_ptr              = reinterpret_cast<T *>(dst.ptr());

                const auto broadcast_value     = *reinterpret_cast<const T *>(broadcast_input.ptr());
                const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});

                // Compute window_step_x elements per iteration
                int x = window_start_x;
                for (; x <= (window_end_x - window_step_x); x += window_step_x)
                {
                    const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);

                    // Dequantize inputs
                    const float32x4x4_t in1_f32x4x4 = vdequantize(non_broadcast_v, non_broadcast_qinfo);
                    const float32x4x4_t in2_f32x4x4 = vdequantize(broadcast_value_vec, broadcast_qinfo);

                    const float32x4x4_t out_f32x4x4 = {