*   Keras:

    *   `tf.keras.layers.Conv` now includes a public `convolution_op` method.
        This method can be used to simplify the implementation of Conv
        subclasses. There are two primary ways to use this new method. The first
        is to use the method directly in your own `call` method: `python class
        StandardizedConv2D(tf.keras.layers.Conv2D): def call(self, inputs):

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc(callsite("Model/conv2d@conv2d_with_valid_loc"("Conv2D") at "QuantizationUnit({{.*}})"))
}

    key: "narrow_range"
    value {
      b: true
    }
  }
  attr {
    key: "num_bits"
    value {
      i: 8
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/Conv2D"
  op: "Conv2D"
  input: "input"
  input: "BoxPredictor_4/ClassPredictor/weights_quant/FakeQuantWithMinMaxVarsPerChannel"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {

    key: "narrow_range"
    value {
      b: true
    }
  }
  attr {
    key: "num_bits"
    value {
      i: 4
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/Conv2D"
  op: "Conv2D"
  input: "input"
  input: "BoxPredictor_4/ClassPredictor/weights_quant/FakeQuantWithMinMaxVarsPerChannel"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {

// CHECK: %[[CONV_1:.*]] = "tf.Conv2D"(%[[GATHER]], %[[DEQUANTIZED_1]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}> {device = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1024xf32>) -> tensor<1x3x2x1024xf32>

// CHECK: %[[CONV2D:.*]] = "tf.Conv2D"(%arg0, %[[CONST]]) <{data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}> : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}
        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "Model/conv2d"])
    %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

    }
  }

  // Handle Conv2D input, stride and filter.
  HandleConv2DInput(conv2d, block_size);
  HandleConv2DStride(conv2d);
  HandleConv2DFilter(conv2d, block_size);

  // Book keeping new filter shape for backprop filter rewrite.
  // Filter shape is defined in HandleConv2DFilter, thus it is RankedTensorType.
  filter_shape =
      mlir::cast<RankedTensorType>(conv2d.getFilter().getType()).getShape();

  // Check that Conv2D computed in NCHW format, and all redundant transpose
  // operations removed from the function.

  // CHECK: %[[CONV:[0-9]*]] = "tf.Conv2D"(%arg0, %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: -> tensor<1x8x32x32xf32>

  // CHECK: return %[[CONV]]

  func.return %4 : tensor<1x8x32x32xf32>

// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: %[[CONV:.*]] = "tfl.conv_2d"(%arg0, %[[DEQUANTIZE]], %[[CONSTANT]])
// CHECK: return %[[CONV]]
}

// CHECK-LABEL: perChannelFakeQuantWithConv2D
func.func @perChannelFakeQuantWithConv2D(tensor<256x32x32x3xf32>) -> (tensor<256x8x7x16xf32>) {