// CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) <{dilations = [1, 2, 2, 1], padding = "SAME", strides = [1, 1, 1, 1]}> : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
  // CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[CONV]], [[BIAS]]) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>

    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 {

  // CHECK: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL1]]) : (tensor<1x1x1x8x16x!quant.uniform<{{.+}}>>) -> tensor<1x1x1x8x16xf32>
  // CHECK: %[[VAL3:.+]] = "tfl.conv_3d"(%arg0, %[[VAL2]], %[[VAL0]]) <{dilation_d_factor = 1 : i32, dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_d = 1 : i32, stride_h = 1 : i32, stride_w = 1 : i32}>

    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 {

    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>

    }
  }

  // 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: %[[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>) {