DCHECK(while_ctx != nullptr);

  // Record 'summed_grads' as the backprop input associated with 'exit_node'
  std::map<Node*, Output>& backprops = while_backprops_[while_ctx];
  DCHECK(backprops.find(exit_node) == backprops.end());
  backprops[exit_node] = summed_grads;

  // Wait until we have all exit nodes' backprops collected before processing
  // the while loop.

tScrollbar(),t(e._element).trigger(h.HIDDEN)})},p._removeBackdrop=function(){this._backdrop&&(t(this._backdrop).remove(),this._backdrop=null)},p._showBackdrop=function(e){var n=this,i=t(this._element).hasClass(d)?d:"";if(this._isShown&&this._config.backdrop){var s=P.supportsTransitionEnd()&&i;if(this._backdrop=document.createElement("div"),this._backdrop.className=u,i&&t(this._backdrop).addClass(i),t(this._backdrop).appendTo(document.body),t(this._element).on(h.CLICK_DISMISS,function(t){n._ignor...

  // Build new BackPropFilterOp.
  auto loc = backprop.getLoc();
  auto new_backprop = builder.create<TF::Conv2DBackpropFilterOp>(
      loc, new_result_type, input, new_filter_sizes, backprop.getOutBackprop(),
      strides, backprop.getUseCudnnOnGpu(), backprop.getPadding(),
      backprop.getExplicitPaddings(), backprop.getDataFormat(),
      backprop.getDilations());

  // We multiply the backprop for cost with the gradients - op.output[1].
  // There is no gradient for labels.

  // The outputs of the network are at input index 0.
  auto logits = op.input(0);
  // The "truth" labels are at index 1.
  auto softmax_grad = op.output(1);

  // The loss is the output at index 0, and backprop is the output at index 1.
  auto grad_loss = grad_inputs[0];

// computes loss and backprop of the loss with respect to 'features'.
//
// Softmax cross entropy loss is defined as follows:
//
//  loss = Sum(-labels * Log(Exp(features) / Sum(Exp(features)))
//  loss = Sum(-labels * LogSoftmax(features))
//
// Computing gradient of the loss with respect to features gives us,
//
//  backprop = (Exp(features) / Sum(Exp(features))) - labels

// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of
// targets (not all since targets which have outputs in the tape will not have
// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update

    // CHECK: %[[offset_backprop:.*]] = mhlo.convert %[[red2]] : tensor<8xf32>

    // CHECK: %[[x_backprop:.*]] = mhlo.convert %[[mul3]] : tensor<8x8x8x8xf32>
    // CHECK: return %[[x_backprop]] : tensor<8x8x8x8xf32>

  auto y =
      tensorflow::ops::SoftmaxCrossEntropyWithLogits(scope_, logits, labels);
  // Note the reversal of the backprop and loss orders. Issue #18734 has been
  // opened for this.
  RunTest({logits, labels}, {logits_shape, logits_shape}, {y.backprop, y.loss},
          {logits_shape, loss_shape});
}

TEST_F(NNGradTest, LogSoftmaxGrad) {
  TensorShape shape({5, 3});

    // CHECK: %[[BACKPROP:.*]] = "tf.Conv2DBackpropFilter"
    // CHECK-SAME: strides = [1, 1, 1, 1]
    // CHECK-SAME: (tensor<2x115x115x12xf32>, tensor<4xi32>, tensor<2x112x112x64xf32>) -> tensor<4x4x12x64xf32>

                                        {dx, dy, dz}, &grad_outputs));
    }
  }
  CompareTestAndExpectedGraphs();
}

TEST_F(GradientsTest, StackUnstack_StopBackprop) {
  // Tests that backprop stops before calculating gradients for Stack (because
  // only gradients w.r.t the output of Stack are requested).
  for (const bool expected : {false, true}) {
    const Scope& scope = expected ? scope_expected_ : scope_test_;