Vectorized Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/sve/vec_qint.h lines 311–442

template <>
struct Vectorized<c10::qint8> : public VectorizedQuantizedConverter<
                                    c10::qint8,
                                    std::array<Vectorized<float>, 4>,
                                    std::array<Vectorized<c10::qint32>, 4>,
                                    VECTOR_WIDTH> {
  Vectorized()
      : VectorizedQuantizedConverter<
            c10::qint8,
            std::array<Vectorized<float>, 4>,
            std::array<Vectorized<c10::qint32>, 4>,
            VECTOR_WIDTH>() {}
  Vectorized(c10::qint8 val)
      : VectorizedQuantizedConverter<
            c10::qint8,
            std::array<Vectorized<float>, 4>,
            std::array<Vectorized<c10::qint32>, 4>,
            VECTOR_WIDTH>(val) {}
  Vectorized(const void* ptr)
      : VectorizedQuantizedConverter<
            c10::qint8,
            std::array<Vectorized<float>, 4>,
            std::array<Vectorized<c10::qint32>, 4>,
            VECTOR_WIDTH>(ptr) {}

  static Vectorized<c10::qint8> loadu(const void* ptr) {
    return Vectorized<c10::qint8>(ptr);
  }

  static Vectorized<c10::qint8> loadu(const void* ptr, int64_t count) {
    __at_align__ value_type tmp_values[size()];
    // Ensure uninitialized memory does not change the output value See
    // https://github.com/pytorch/pytorch/issues/32502 for more details. We do
    // not initialize arrays to zero using "={0}" because gcc would compile it
    // to two instructions while a loop would be compiled to one instruction.
    for (const auto i : c10::irange(size())) {
      tmp_values[i] = 0;
    }
    std::memcpy(
        tmp_values,
        reinterpret_cast<const value_type*>(ptr),
        count * sizeof(value_type));
    return loadu(tmp_values);
  }

  static Vectorized<c10::qint8> quantize(
      const float_vec_return_type& rhs,
      float scale,
      int32_t zero_point,
      float inverse_scale) {
    std::array<value_type, size()> qvals;
    std::array<float, float_num_vecs() * Vectorized<float>::size()> float_vals;

    for (int i = 0; i < float_num_vecs(); ++i) {
      rhs[i].store(
          &float_vals[i * Vectorized<float>::size()],
          Vectorized<float>::size());
    }

    at::native::quantize_vec<c10::qint8>(
        scale,
        zero_point,
        float_vals.data(),
        (c10::qint8*)qvals.data(),
        Vectorized<float>::size() * float_num_vecs());

    return Vectorized<c10::qint8>::loadu(qvals.data());
  }

  Vectorized<c10::qint8> maximum(Vectorized<c10::qint8> b) const {
    Vectorized<c10::qint8> retval;
    for (size_t i = 0; i < size(); ++i) {
      retval.vals[i] = std::max<value_type>(vals[i], b.vals[i]);
    }
    return retval;
  }

  Vectorized<c10::qint8> minimum(Vectorized<c10::qint8> b) const {
    Vectorized<c10::qint8> retval;
    for (size_t i = 0; i < size(); ++i) {
      retval.vals[i] = std::min<value_type>(vals[i], b.vals[i]);
    }
    return retval;
  }

  Vectorized<c10::qint8> relu(Vectorized<c10::qint8> zero_point) const {
    return maximum(zero_point);
  }

  Vectorized<c10::qint8> relu6(
      Vectorized<c10::qint8> zero_point,
      Vectorized<c10::qint8> q_six) {
    Vectorized<c10::qint8> retval;
    for (size_t i = 0; i < size(); ++i) {
      retval.vals[i] = std::min<value_type>(
          std::max<value_type>(vals[i], zero_point.vals[i]), q_six.vals[i]);
    }
    return retval;
  }

  int_vec_return_type widening_subtract(Vectorized<c10::qint8> b) const {
    int_vec_return_type retval;
    constexpr int elem_per_int_vec = size() / int_num_vecs();
    for (size_t i = 0; i < int_num_vecs(); ++i) {
      for (size_t j = 0; j < elem_per_int_vec; ++j) {
        retval[i].vals[j] =
            static_cast<int32_t>(vals[i * elem_per_int_vec + j]) -
            static_cast<int32_t>(b.vals[i * elem_per_int_vec + j]);
      }
    }
    return retval;
  }
  static Vectorized<c10::qint8> requantize_from_int(
      const int_vec_return_type& inp,
      float multiplier,
      int32_t zero_point) {
    constexpr int elem_per_int_vec = size() / int_num_vecs();
    constexpr auto min_val = std::numeric_limits<value_type>::min();
    constexpr auto max_val = std::numeric_limits<value_type>::max();
    Vectorized<c10::qint8> retval;
    for (size_t i = 0; i < int_num_vecs(); ++i) {
      for (size_t j = 0; j < elem_per_int_vec; ++j) {
        int32_t rounded =
            nearbyint(static_cast<float>(inp[i].vals[j]) * multiplier) +
            zero_point;
        retval.vals[i * elem_per_int_vec + j] =
            std::min<int32_t>(std::max<int32_t>(rounded, min_val), max_val);
      }
    }
    return retval;
  }
};