%cst_3 = "tf.Const"() {value = dense<[[[[1.400000e+01]], [[-2.800000e+01]], [[4.200000e+01]]], [[[-5.600000e+01]], [[7.100000e+01]], [[-8.500000e+01]]], [[[9.900000e+01]], [[-1.130000e+02]], [[1.270000e+02]]]]> : tensor<3x3x1x1xf32>} : () -> tensor<3x3x1x1xf32>
  %cst_4 = "tf.Const"() {value = dense<-1.280000e+02> : tensor<f32>} : () -> tensor<f32>
  %cst_5 = "tf.Const"() {value = dense<0.00118110236> : tensor<1xf32>} : () -> tensor<1xf32>

  %0 = "tf.Const"() {value = dense<0.111111112> : tensor<3x3x1x1xf32>} : () -> tensor<3x3x1x1xf32>
  %1 = "tf.Const"() {value = dense<1.000000e+00> : tensor<1x520x520x1xf32>} : () -> tensor<1x520x520x1xf32>
  %2 = "tf.DepthwiseConv2dNative"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x520x520x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x520x520x1xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  %w = arith.constant dense<[[[[0.0]]], [[[127.0]]], [[[-127.0]]]]> : tensor<3x1x1x1xf32>
  %b = arith.constant dense<0.0> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x5x5x1xf32>, tensor<3x1x1x1xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>

  %9 = "tfl.quantize"(%8) {qtype = tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>, volatile} : (tensor<1x3x1x1xf32>) -> tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>
  %10 = "tfl.dequantize"(%9) : (tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>) -> tensor<1x3x1x1xf32>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<2.000000e+01> : tensor<3x3x4x4xf32>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  return %2 : tensor<?x2x2x1xi8>
}

// -----

// CHECK-LABEL: func @convolution_add_add_f32
func.func @convolution_add_add_f32(
    %lhs: tensor<?x3x2x1xf32>, %rhs: tensor<2x1x1x1xf32>,
    %zp_offset: tensor<?x2x2x1xf32>, %bias: tensor<1xf32>
  ) -> tensor<?x2x2x1xf32> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution

    return %5 : tensor<1x3x1x1xf32>
  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1x3x1x1xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x1x1xf32> attributes {tf_quant.composite_function} {

  %cst_0 = arith.constant dense<[1.0, 2.0, 3.0]> : tensor<3xf32>
  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>) -> tensor<3x3x3x3xf32>