auto reduce =
      internal::BroadcastGradientArgs(scope, x_batch_shape, y_batch_shape);
  (*grad_outputs)[0] =
      Reshape(scope, ReduceSum(scope, (*grad_outputs)[0], reduce.r0), sx);
  (*grad_outputs)[1] =
      Reshape(scope, ReduceSum(scope, (*grad_outputs)[1], reduce.r1), sy);
  return scope.status();
}
REGISTER_GRADIENT_OP("BatchMatMulV2", BatchMatMulV2Grad);

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a BiasAdd: A + bias, the gradients are:
     *
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

    // Recover data format from forward pass for gradient.
    std::string data_format;

  // all the gradients from the shape it fills.
  // We use ReduceSum to implement this, which needs an argument providing
  // the indices of all the dimensions of the incoming gradient.
  // grad(x) = reduce_sum(grad(y), [0..rank(grad(y))])
  auto all_dims = Range(scope, Const(scope, 0), Rank(scope, grad_inputs[0]),
                        Const(scope, 1));
  grad_outputs->push_back(ReduceSum(scope, grad_inputs[0], all_dims));
  return scope.status();

      scope, Slice1dHelper(scope, x_shape, bx_start, bx_end),
      Slice1dHelper(scope, y_shape, by_start, by_end));
  grad_x = Reshape(
      scope, ReduceSum(scope, grad_x, Add(scope, bx_start, args.r0)), x_shape);
  grad_y = Reshape(
      scope, ReduceSum(scope, grad_y, Add(scope, by_start, args.r1)), y_shape);
  grad_outputs->push_back(grad_x);
  grad_outputs->push_back(grad_y);
  return scope.status();
}