Method Run

tensorflow/core/kernels/matmul_op_impl.h:174–213

  }

  static void Run(const Tensor& in_x, const Tensor& in_y, bool adj_x,
                  bool adj_y, bool trans_x, bool trans_y,
                  const MatMulBCast& bcast, Tensor* out, int start,
                  int limit) {
    const bool should_bcast = bcast.IsBroadcastingRequired();
    const auto& x_batch_indices = bcast.x_batch_indices();
    const auto& y_batch_indices = bcast.y_batch_indices();
    for (int64 i = start; i < limit; ++i) {
      const int64 x_batch_index = should_bcast ? x_batch_indices[i] : i;
      const int64 y_batch_index = should_bcast ? y_batch_indices[i] : i;
      auto x = ConstTensorSliceToEigenMatrix(in_x, x_batch_index);
      auto y = ConstTensorSliceToEigenMatrix(in_y, y_batch_index);
      auto z = TensorSliceToEigenMatrix(out, i);
      if (!adj_x && !trans_x) {
        if (!adj_y && !trans_y) {
          z.noalias() = x * y;
        } else if (adj_y) {
          z.noalias() = x * y.adjoint();
        } else {
          z.noalias() = x * y.transpose();
        }
      } else if (adj_x) {
        if (!adj_y && !trans_y) {
          z.noalias() = x.adjoint() * y;
        } else if (adj_y) {
          z.noalias() = x.adjoint() * y.adjoint();
        } else {  // trans_y == true
          z.noalias() = x.adjoint() * y.transpose();
        }
      } else { // trans_x == true
        if (!adj_y && !trans_y) {
          z.noalias() = x.transpose() * y;
        } else if (adj_y) {
          z.noalias() = x.transpose() * y.adjoint();
        } else {  // trans_y == true
          z.noalias() = x.transpose() * y.transpose();
        }
      }
    }
  }
};

}  // namespace