%output, %min, %max, %histogram = "tf.CustomAggregator"(%arg0) <{calibration_method = 5 : i32, id = "skipping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

// CHECK-LABEL: conv2d_backprop_input_with_add
func.func @conv2d_backprop_input_with_add(%arg0: tensor<4xi32>, %arg1: tensor<3x3x1x32xf32>, %arg2: tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32> {
  %0 = "tf.Conv2DBackpropInput"(%arg0, %arg1, %arg2) {strides = [1, 2, 2, 1], padding="SAME", dilations=[1, 1, 1, 1]}: (tensor<4xi32>, tensor<3x3x1x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>

// CHECK: %[[q:.*]] = "tfl.quantize"(%[[w]]) <{qtype = tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.003937007874015748:1>>}> {volatile} : (tensor<32x3x3x3xf32>) -> tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.003937007874015748:1>>
// CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[q]]) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.003937007874015748:1>>) -> tensor<32x3x3x3xf32>

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
    %cst_2 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>

    return %2 : tensor<1x3x2x2xf32>
  }
  func.func private @quantize_i8(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<f32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xi8> {
    %0 = "tf.Div"(%arg0, %arg1) : (tensor<1x3x4x3xf32>, tensor<f32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.Round"(%0) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>
    %2 = "tf.Cast"(%1) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xi32>

    %0 = "tf.GatherV2"(%arg0, %arg1, %arg2) {attr_map = "0:batch_dims", batch_dims = 0 : i64, device = ""} : (tensor<1024x3x4x3xf32>, tensor<1xi32>, tensor<i32>) -> tensor<1x3x4x3xf32>
    return %0 : tensor<1x3x4x3xf32>
  }

  %3 = "tfl.depthwise_conv_2d"(%arg0, %2, %0) {depth_multiplier = 4 : i32, dilation_h_factor = 2 : i32, dilation_w_factor = 2 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 4 : i32, stride_w = 5 : i32} : (tensor<1x224x224x3xf32>, tensor<32x3x3x3xf32>, tensor<32xf32>) -> tensor<1x112x112x32xf32>

  func.return %3 : tensor<1x112x112x32xf32>

    %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst_0) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<1x2x2x3xf32>

// CHECK: %[[CONV:.+]] = stablehlo.convolution(%[[ARG0]], %[[DQ]])
// CHECK{LITERAL}: dim_numbers = [b, 0, 1, f]x[o, 0, 1, i]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64}
// CHECK-SAME: (tensor<1x3x3x4xf32>, tensor<2x3x3x4xf32>) -> tensor<1x3x3x2xf32>
// CHECK: return %[[CONV]]

// -----

}

// -----

// 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) {