} : (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> {

// 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>

      -> tensor<1x2x2x2xf32>
    %3 = "quantfork.stats"(%arg2) {layerStats = dense<[7.05456924, 7.11401462]> : tensor<2xf32>} : (tensor<2xf32>) -> tensor<2xf32>
    %4 = "quantfork.stats"(%2) {layerStats = dense<[-1.36523, 3.57373]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %5 = "chlo.broadcast_add"(%4, %3) : (tensor<1x2x2x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>

  %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>
}

  // 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>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2x2x6xbf16>) -> tensor<1x2x2x6xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2x2x6xf32>) -> tensor<1x2x2x6xf32>
  return %3 : tensor<1x2x2x6xf32>
}

// CHECK: func @cast_bf16_depthwise_conv_to_fp32

// RUN: tf-opt %s -tf-layout-optimization=force-data-format=NCHW -verify-diagnostics | FileCheck %s --dump-input=always

// CHECK-LABEL: func @transposeConv2D
func.func @transposeConv2D(%arg0: tensor<1x3x32x32xf32>, %arg1: tensor<1x1x3x8xf32>) -> tensor<1x8x32x32xf32> {

  // Convert input: NCHW -> NHWC
  %0 = "tf.Const"() {value = dense<[0, 2, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>

  func.return %17 : tensor<1x2x3xf32>

  // CHECK: %[[NONE:.*]] = "tfl.no_value"() <{value}> : () -> none
  // CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

// dilations, etc...). This test only verifies that changing convolution data
// layout will update all the attributes.

// CHECK-LABEL: func @transposeConv2D
func.func @transposeConv2D(%input: tensor<1x3x32x32xf32>, %filter: tensor<1x1x3x8xf32>) -> tensor<1x8x7x6xf32> {

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])