Function packet

tensorflow/core/kernels/mirror_pad_op.h:183–240

  template <int LoadMode>
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType
  packet(Index index) const {
    constexpr int kPacketSize =
        internal::unpacket_traits<PacketReturnType>::size;

    EIGEN_STATIC_ASSERT(kPacketSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
    eigen_assert(index + kPacketSize <= dimensions().TotalSize());

    // Find the effective inner-most dimension where padding actually happens.
    // NOTE: This is independent of index argument, and can be done in the
    // constructor to save computation. However, if packet access does not
    // happen, then moving to constructor will incur needless overhead.
    int dim = -1;
    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      for (int k = 0; k < Dims; ++k) {
        if (padding_[k].first != 0 || padding_[k].second != 0) {
          dim = k;
          break;
        }
      }
    } else {
      for (int k = Dims - 1; k >= 0; --k) {
        if (padding_[k].first != 0 || padding_[k].second != 0) {
          dim = k;
          break;
        }
      }
    }

    const Index input_index = ToInputIndex(index);

    // If dim < 0, this means there is no padding at all.
    if (dim < 0) {
      return impl_.template packet<Unaligned>(input_index);
    }

    // Check if the way from the begin of the packet to the end of the packet
    // is paved with contiguous road. That is, the indices must be between the
    // padded region in the effective inner-most dimension.
    const Index left = padding_[dim].first * output_strides_[dim];
    const Index right =
        (dimensions_[dim] - padding_[dim].second) * output_strides_[dim];

    const Index index_mod = index % (dimensions_[dim] * output_strides_[dim]);
    if (left <= index_mod && (index_mod + kPacketSize - 1) < right) {
      return impl_.template packet<Unaligned>(input_index);
    }

    // If the road is not contiguous, then fall back to coeff().
    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type
        values[kPacketSize];
    values[0] = impl_.coeff(input_index);
    for (int i = 1; i < kPacketSize; ++i) {
      values[i] = coeff(index + i);
    }
    PacketReturnType result = internal::pload<PacketReturnType>(values);
    return result;
  }