%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}>

// 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>>)

  %0 = mhlo.constant dense<0.000000e+00> : tensor<f32>
  %1 = "mhlo.pad"(%arg0, %0) <{edge_padding_high = dense<0> : tensor<3xi64>, edge_padding_low = dense<1> : tensor<3xi64>, interior_padding = dense<0> : tensor<3xi64>}> : (tensor<10x10x10xf32>, tensor<f32>) -> tensor<11x11x11xf32>
  %2 = mhlo.constant dense<0.000000e+00> : tensor<f32>

}

// -----

// CHECK-LABEL: lift_float_conv
func.func @lift_float_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst_1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],

    %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.

  func.return %0 : tensor<384x128x!quant.uniform<i8:f32, 0.09:-4>>
}

// PARTIAL:   func @quantization_test(%[[VAL_0:.*]]: tensor<384x512x!quant.uniform<i8:f32, 1.000000e-01>>, %[[VAL_1:.*]]: tensor<128x512x!quant.uniform<i8<-127:127>:f32, 1.000000e-02>>, %[[VAL_2:.*]]: tensor<128x!quant.uniform<i32:f32, 0.69999999999999996>>) -> tensor<384x128x!quant.uniform<i8:f32, 0.089999999999999996:-4>> {

  %cst = arith.constant dense<0.000000e+00> : tensor<4x2xf32>
  %cst_0 = arith.constant dense<0.000000e+00> : tensor<8xf32>
  %cst_1 = arith.constant dense<[1, 0, 2]> : tensor<3xi32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<4x4x2xf32>
  %cst_3 = arith.constant dense<4> : tensor<i32>
  %cst_4 = arith.constant dense<1.000000e+00> : tensor<f32>
  %cst_5 = arith.constant dense<1> : tensor<i32>

func.func @nchw_conv_with_bias_add_relu(%arg0: tensor<1x2x5x5xf32>) -> tensor<1x4x5x5xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<4x2x3x3xf32>
  %5 = stablehlo.constant dense<0.000000e+00> : tensor<1x4x5x5xf32>
  %1 = stablehlo.constant dense<3.000000e+00> : tensor<1x4x5x5xf32>