%4 = "quantfork.stats"(%3) {
    layerStats = dense<[0.0, 4.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %5 = "tfl.sqrt"(%4) : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %6 = "quantfork.stats"(%5) {
    layerStats = dense<[0.0, 2.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>

// CHECK: %cst_0 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK: %2 = "tfl.transpose"(%1, %cst_0) : (tensor<1x1x4x1xf32>, tensor<4xi32>) -> tensor<1x1x1x4xf32>
// CHECK: return %2 : tensor<1x1x1x4xf32>


func.func @avg_pool2d_5(%arg0: tensor<1x1x3x3xf32>) -> (tensor<1x1x2x2xf32>) {

  %res_p3 = "tfl.cumsum"(%arg, %axis_p3) {exclusive = false, reverse = false} : (tensor<1x2x1x3xf32>, tensor<i32>) -> tensor<1x2x1x3xf32>
  func.return %res_m4, %res_m3, %res_m2, %res_m1, %res_00, %res_p1, %res_p2, %res_p3 : tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>, tensor<1x2x1x3xf32>

func.func @dont_move_transpose_different_ranks(%arg0:tensor<1x1x2x3xf32>, %arg1:tensor<2x3xf32>) -> tensor<1x2x1x3xf32> {
  %cst = "tf.Const"() {value = dense<[0, 2, 1, 3]> : tensor<4xi32>} : () -> tensor<4xi32>
  %0 = "tf.AddV2"(%arg0, %arg1) {device = ""} : (tensor<1x1x2x3xf32>, tensor<2x3xf32>) -> tensor<1x1x2x3xf32>
  %1 = "tf.Transpose"(%0, %cst) {device = ""} : (tensor<1x1x2x3xf32>, tensor<4xi32>) -> tensor<1x2x1x3xf32>

  func.return %1 : tensor<1x2x1x3xf32>

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x3x1x1xf32>) -> tensor<1x3x1x1xf32>
    return %3 : tensor<1x3x1x1xf32>
  }

// CHECK-LABEL: func @multiple_quantizable_ops_in_graph

func.func @exp(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
  %0 = "vhlo.exponential_v1" (%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
  func.return %0 : tensor<1x1x1x96xf32>
}

//CHECK:func.func private @exp(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
//CHECK-NEXT: %0 = "vhlo.exponential_v1"(%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
//CHECK-NEXT: return %0 : tensor<1x1x1x96xf32>
//CHECK-NEXT:}

    %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<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

    %output_0, %min_1, %max_2, %histogram_3 = "tf.CustomAggregator"(%0) <{calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32}> : (tensor<10x1x3xf32>) -> (tensor<10x1x3xf32>, tensor<f32>, tensor<f32>, tensor<0xi64>)
    return %output_0 : tensor<10x1x3xf32>
  }
  // CHECK-LABEL: @main

// RUN: tf-quant-opt %s -quant-cast-bf16-ops-to-f32 | FileCheck %s

func.func @cast_bf16_conv_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>