%16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", device = "", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    %2 = "quantfork.stats"(%1) {layerStats = dense<[4.00000000e-6, 9.80000000e-1]> : tensor<2xf32>} : (tensor<2x3x2x2xf32>) -> tensor<2x3x2x2xf32>
    return %2 : tensor<2x3x2x2xf32>
  }
// Checks that the quantized XlaCallModule has been replaced by a CallOp, which
// calls the quantized entry function.

    %cst_0 = "tf.Const"() {value = dense<[-2.000000e+00, 3.000000e+00]> : tensor<2xf32>} : () -> tensor<2xf32>
    %0 = "quantfork.qcast"(%cst) : (tensor<2x2x3x2xf32>) -> tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>

    %6 = "tf.Div"(%arg1, %cst_2) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>
    %7 = "tf.AddV2"(%6, %cst_1) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>
    %8 = "tf.Maximum"(%7, %cst_1) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>
    %9 = "tf.Minimum"(%8, %cst_4) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>
  %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<3x3x3x4xf32>) -> tensor<256x8x7x4xf32>
  func.return %rst : tensor<256x8x7x4xf32>

  %w = arith.constant dense<127.0> : tensor<3x3x3x3xf32>
  %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<1x5x5x3xf32>, tensor<3x3x3x3xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>

func.func @FoldTrivialReshapeIntoTranspose(%arg: tensor<2x1x3x3xf32>) -> tensor<1x3x3x2xf32> {
  %cst = arith.constant dense<[1, 3, 3, 2]> : tensor<4xi32>
  %cst_2 = arith.constant dense<[2, 3, 0, 1]> : tensor<4xi32>
  %2 = "tf.Transpose"(%arg, %cst_2) : (tensor<2x1x3x3xf32>, tensor<4xi32>) -> tensor<3x3x2x1xf32>
  %3 = "tf.Reshape"(%2, %cst) : (tensor<3x3x2x1xf32>, tensor<4xi32>) -> tensor<1x3x3x2xf32>
  return %3: tensor<1x3x3x2xf32>

  %cst_1 = "tf.Const"() {value = dense<[0.1, 0.2]> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_2 = "tf.Const"() {value = dense<[0.3, 0.4]> : tensor<2xf32>} : () -> tensor<2xf32>