%3 = "tf.Cast"(%2) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %4 = "tf.IdentityN"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %4 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_with_bias_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  %1 = "tf.Mul"(%0, %cst2) : (tensor<1x28x23x2xf32>, tensor<2xf32>) -> tensor<1x28x23x2xf32>

  func.return %1 : tensor<1x28x23x2xf32>
  // CHECK: %[[CST:.*]] = "tf.Const{{.*}} dense<
  // CHECK-SAME: [1.000000e+00, 4.000000e+00], [3.000000e+00, 8.000000e+00], [5.000000e+00, 1.200000e+01]

      platforms = [], version = 4 : i64
    } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[0.000000e+00, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    return %2 : tensor<1x2x2x2xf32>
  }

  // CHECK-LABEL: composite_conv2d_with_bias_and_relu6_fn_10

  // CHECK-DAG: %[[CST:.+]] = mhlo.constant dense<[1.000000e-01, 2.000000e-01]> : tensor<2xf32>
  // CHECK-DAG: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>
  // CHECK-DAG: %[[NEW_FILTER:.+]] =  mhlo.multiply %[[CST_BCAST]], %[[FILTER]] : tensor<1x1x3x2xf32>

  %output_1, %min_2, %max_3, %histogram_4 = "tf.CustomAggregator"(%0) <{calibration_method = 5 : i32, id = "keeping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x2x2x2xf32>) -> (tensor<1x2x2x2xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

  %perm = arith.constant dense<[3, 0, 1, 2]> : tensor<4xi32>
  %0 = "tfl.transpose"(%arg0, %perm) : (tensor<2x3x4x1xf32>, tensor<4xi32>) -> tensor<1x2x3x4xf32>
  %cst = arith.constant dense<1.0> : tensor<5x2x3x4xf32>
  %1 = "tfl.add"(%0, %cst) { fused_activation_function = "NONE" } : (tensor<1x2x3x4xf32>, tensor<5x2x3x4xf32>) -> tensor<5x2x3x4xf32>
  func.return %1 : tensor<5x2x3x4xf32>
}

       } : (tensor<1x32x32x3xf32>, tensor<4xi32>, tensor<1x32x32x8xf32>)
         -> tensor<1x1x3x8xf32>

  func.return %0 : tensor<1x1x3x8xf32>
}

// CHECK-LABEL: func @transposeConv2DBackpropInput
func.func @transposeConv2DBackpropInput(
  %input_sizes: tensor<4xi32>,
  %filter: tensor<1x1x3x8xf32>,
  %out_backprop: tensor<1x32x32x8xf32>
) -> tensor<1x32x32x3xf32> {

func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}

// CHECK-LABEL: transpose_simple_4d
func.func @transpose_simple_4d() -> tensor<5x2x3x4xf32> {
  %0 = stablehlo.constant dense<1.000000e+0> : tensor<2x3x4x5xf32>
  %1 = stablehlo.transpose %0, dims = [3, 0, 1, 2] : (tensor<2x3x4x5xf32>) -> tensor<5x2x3x4xf32>
  return %1 : tensor<5x2x3x4xf32>
}
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<5x2x3x4xf32>
// CHECK-NOT: transpose
// CHECK: return %[[CONST_0]] : tensor<5x2x3x4xf32>