});
    }
    filter_value = filter.getValue();
    mul_value = multiplier.getValue();
    // In MHLO, Conv filter is in HWIO format, Depthwise conv filter is in HW1O
    // format and backprop input conv filter is in HWOI format.
    // Only fuses multiplier if all dimensions other than the out channel
    // dimension are equal to 1.
    if (!TFL::IsDimensionsDegenerateExceptLastOne(

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

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

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

                                        {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_;

      Add(params_->body_inputs[0], {one, 0}, params_->body_graph, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  params_->body_outputs[0] = {add, 0};

  ExpectOK();

  // Create backprop graph
  TF_Output grad_output;
  TF_AddGradients(graph_, outputs_.data(), outputs_.size(), inputs_.data(), 1,
                  nullptr, s_, &grad_output);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

      Value scratch2 =
          ApplyReduction(loc, weighted_grad, reduce_dims, &rewriter);

      // x_backprop = y_backprop * (scale * scratch1)
      auto scaled_grad =
          rewriter.create<mhlo::MulOp>(loc, op.getScale(), scratch1);
      x_backprop = rewriter.create<mhlo::MulOp>(
          loc, grad,
          Broadcast1DToFeatureDim(loc, act, scaled_grad, feature_dim,

example, suppose y = f(x) and we wish to apply a custom function g for backprop
such that dx = g(dy). In Python,

```python
with tf.get_default_graph().gradient_override_map(
    {'IdentityN': 'OverrideGradientWithG'}):
  y, _ = identity_n([f(x), x])

@tf.RegisterGradient('OverrideGradientWithG')
def ApplyG(op, dy, _):
  return [None, g(dy)]  # Do not backprop to f(x).
```
  }];

  let arguments = (ins

  // CHECK-DAG: %[[SOFTMAX:.*]] = "tf.Div"(%[[SOFTMAX_EXP]], %[[SOFTMAX_SUM]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>

  // CHECK-DAG: %[[BACKPROP:.*]] = "tf.Sub"(%[[SOFTMAX]], %[[LABELS]]) : (tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  // CHECK: return %[[LOSS]], %[[BACKPROP]]

  %0:2 = "tf.SoftmaxCrossEntropyWithLogits"(%features, %labels) : (tensor<2x3xf32>, tensor<2x3xf32>) -> (tensor<2xf32>, tensor<2x3xf32>)

        *   `tf.compat.v1.nn.fused_batch_norm` backprop to `offset` when
            `is_training=False`
        *   `tf.image.adjust_contrast` forward
        *   `tf.nn.depthwise_conv2d` backprop to `filter` when not using cuDNN
            convolution
        *   `tf.image.resize` with `method=ResizeMethod.NEAREST` backprop
        *   `tf.math.bincount` - TODO: confirm exception added

on:column;flex-direction:column;width:100%;pointer-events:auto;background-color:#fff;background-clip:padding-box;border:1px solid rgba(0,0,0,.2);border-radius:.3rem;outline:0}.modal-backdrop{position:fixed;top:0;right:0;bottom:0;left:0;z-index:1040;background-color:#000}.modal-backdrop.fade{opacity:0}.modal-backdrop.show{opacity:.5}.modal-header{display:-webkit-box;display:-ms-flexbox;display:flex;-webkit-box-align:start;-ms-flex-align:start;align-items:flex-start;-webkit-box-pack:justify;-ms-fl...