// Calc Grad
    AbstractTensorHandle* matmul_A_output;
    AbstractTensorHandle* matmul_B_output;
    std::string name_grad_A = "MatMul_Grad_A";
    std::string name_grad_B = "MatMul_Grad_B";
    if (!t_a && !t_b) {
      TF_RETURN_IF_ERROR(MatMul(ctx, upstream_grad, B.get(), &matmul_A_output,
                                /*transpose_a = */ false,