Vectorized Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec512/vec512_qint.h lines 1112–1230

template <>
struct Vectorized<c10::qint32> : public VectorizedQuantizedConverter<
                                     c10::qint32,
                                     std::array<Vectorized<float>, 1>,
                                     std::array<Vectorized<c10::qint32>, 1>,
                                     16> {
  Vectorized()
      : VectorizedQuantizedConverter<
            c10::qint32,
            std::array<Vectorized<float>, 1>,
            std::array<Vectorized<c10::qint32>, 1>,
            16>() {}
  Vectorized(c10::qint32 val)
      : VectorizedQuantizedConverter<
            c10::qint32,
            std::array<Vectorized<float>, 1>,
            std::array<Vectorized<c10::qint32>, 1>,
            16>(val) {}
  Vectorized(const void* ptr)
      : VectorizedQuantizedConverter<
            c10::qint32,
            std::array<Vectorized<float>, 1>,
            std::array<Vectorized<c10::qint32>, 1>,
            16>(ptr) {}

  static Vectorized<c10::qint32> loadu(const void* ptr) {
    return Vectorized<c10::qint32>(ptr);
  }

  static Vectorized<c10::qint32> 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::qint32> quantize(
      const float_vec_return_type& rhs,
      float scale,
      int32_t zero_point,
      float inverse_scale [[maybe_unused]]) {
    std::array<value_type, size()> qvals;
    std::array<float, float_num_vecs() * 16> float_vals;

    for (const auto i : c10::irange(float_num_vecs())) {
      rhs[i].store(&float_vals[i * 16], 16);
    }

    at::native::quantize_vec<c10::qint32, /*precision=*/32>(
        scale,
        zero_point,
        float_vals.data(),
        (c10::qint32*)qvals.data(),
        16 * float_num_vecs());

    return Vectorized<c10::qint32>::loadu(qvals.data());
  }

  Vectorized<c10::qint32> maximum(Vectorized<c10::qint32> b) const {
    Vectorized<c10::qint32> retval;
    for (const auto i : c10::irange(size())) {
      retval.vals[i] = std::max<value_type>(vals[i], b.vals[i]);
    }
    return retval;
  }

  Vectorized<c10::qint32> minimum(Vectorized<c10::qint32> b) const {
    Vectorized<c10::qint32> retval;
    for (const auto i : c10::irange(size())) {
      retval.vals[i] = std::min<value_type>(vals[i], b.vals[i]);
    }
    return retval;
  }

  Vectorized<c10::qint32> relu(Vectorized<c10::qint32> zero_point) const {
    return maximum(zero_point);
  }

  Vectorized<c10::qint32> relu6(
      Vectorized<c10::qint32> zero_point,
      Vectorized<c10::qint32> q_six) {
    Vectorized<c10::qint32> retval;
    for (const auto i : c10::irange(size())) {
      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::qint32> b) const {
    int_vec_return_type retval;
    for (const auto i : c10::irange(size())) {
      retval[0].vals[i] = vals[i] - b.vals[i];
    }
    return retval;
  }

  static Vectorized<c10::qint32> requantize_from_int(
      const int_vec_return_type& inp,
      float multiplier,
      int32_t zero_point) {
    Vectorized<c10::qint32> retval;
    for (const auto i : c10::irange(size())) {
      retval.vals[i] =
          std::nearbyint(static_cast<float>(inp[0].vals[i]) * multiplier) +
          zero_point;
    }
    return retval;
  }
};