%4 = "quantfork.dcast"(%3) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>) -> tensor<*xf32>
  %5 = "tf.Reshape"(%4, %cst) {device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>
  %6 = "quantfork.qcast"(%5) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>

  // CKECK-SAME: <{calibration_method = 1 : i32, id = "0", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32}>
  // CHECK: %[[CUSTOM_AGGREGATOR_1:.*]], %[[MIN_1:.*]], %[[MAX_1:.*]], %[[HISTOGRAM_1:.*]] = "tf.CustomAggregator"
  // CKECK-SAME: <{calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32}>

// CHECK-DAG: %[[CST:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<4x3xf32>

// MIN-MAX-CHECK-NEXT:  [[lhs:%.*]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%arg0) <{calibration_method = 1 : i32, id = "composite_conv2d_with_relu6_fn_arg_0_calibration_method_1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32}>...

// QDQ-NEXT: %[[split:.*]]:4 = "tfl.split"(%arg1, %[[q]]) <{num_splits = 4 : i32}> : (tensor<i32>, tensor<4x!quant.uniform<u8:f32, 1.000000e+00>>) -> (tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>, tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>, tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>, tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>)

// CHECK: stablehlo.constant dense<[2.000000e+00, 3.000000e+00]> : tensor<2xf32>

// -----

func.func @foldDivBothSplat() -> tensor<2xf32> {
  %0 = stablehlo.constant dense<4.0> : tensor<2xf32>
  %1 = stablehlo.constant dense<2.0> : tensor<2xf32>
  %2 = stablehlo.divide %0, %1 : tensor<2xf32>
  return %2 : tensor<2xf32>
}

// CHECK-LABEL: foldDivBothSplat
// CHECK: stablehlo.constant dense<2.000000e+00> : tensor<2xf32>

// -----

00000120  2b 01 92 8d f9 4f 5a 3a  53 11 fc 32 52 cc af cd  |+....OZ:S..2R...|
00000130  7b 94 0e 76 10 c2 16 36  2d a4 64 69 1c 05 70 20  |{..v...6-.di..p |
00000140  0d 23 cd 4a 33 c5 c7 db  db 0f f8 b6 42 0c 83 0a  |.#.J3.......B...|
00000150  a1 73 68 fb 87 2c 9d d2  d3 cf d7 3a bb 36 7e 83  |.sh..,.....:.6~.|
00000160  c3 3f bc e2 61 d9 c2 8b  15 a2 cc bf 14 a0 69 f4  |.?..a.........i.|
00000170  22 02 a9 ff 5e 55 37 6e  61 86 71 73 94 2f 7e 50  |"...^U7na.qs./~P|

  // CHECK-DAG: %[[LHS1:.*]] = mhlo.bitcast_convert %[[LHS]] : (tensor<3x2x!quant.uniform<i32:f32, 2.000000e+00:4>>) -> tensor<3x2xi32>
  // CHECK-DAG: %[[LHS2:.*]] = mhlo.bitcast_convert %[[LHS1]] : (tensor<3x2xi32>) -> tensor<3x2x!quant.uniform<i32:f32, 2.000000e+00:4>>
  %0 = "tf.UniformQuantize"(%input, %input_scales, %input_zps) {

  %10 = "tfl.pseudo_const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %11 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %recurrent_input = "tfl.pseudo_const"() {value = dense<0.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  %cell_input = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>

}

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x512x2xf32>} : () -> tensor<2x3x512x2xf32>