quantize_val Class — pytorch Architecture

Source Code

aten/src/ATen/test/vec_test_all_types.h lines 1502–1512

template <typename T>
T quantize_val(float scale, int64_t zero_point, float value) {
    int64_t qvalue;
    constexpr int64_t qmin = std::numeric_limits<T>::min();
    constexpr int64_t qmax = std::numeric_limits<T>::max();
    float inv_scale = 1.0f / scale;
    qvalue = static_cast<int64_t>(zero_point + at::native::round_impl<float>(value * inv_scale));
    qvalue = std::max<int64_t>(qvalue, qmin);
    qvalue = std::min<int64_t>(qvalue, qmax);
    return static_cast<T>(qvalue);
}