padding = "EXPLICIT",
         strides = [5, 6, 7, 8]
       } : (tensor<1x32x32x3xf32>, tensor<1x1x3x8xf32>) -> tensor<1x7x7x8xf32>

  func.return %0 : tensor<1x7x7x8xf32>
}

// CHECK-LABEL: func @transposeConv2DWithDefaultAttr
func.func @transposeConv2DWithDefaultAttr(%input: tensor<1x32x32x3xf32>, %filter: tensor<1x1x3x8xf32>) -> tensor<?x?x?x?xf32>
{

// 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>

  // CHECK-SAME: explicit_paddings = [1, 2, 5, 6, 7, 8, 3, 4]
  // CHECK-SAME: padding = "EXPLICIT"
  // CHECK-SAME: strides = [5, 7, 8, 6]
  // CHECK-SAME: (tensor<1x32x32x3xf32>, tensor<1x1x3x8xf32>) -> tensor<1x7x6x8xf32>

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%[[CONV2D]], %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE]]

  %1 = "tf.AddV2"(%0, %0) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Transpose"(%1, %2) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

  func.return %3 : tensor<1x8x4x4xf32>
}

// CHECK-LABEL: move_transpose_handle_broadcast

// 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>

func.func @move_across_broadcastable_op(%arg0: tensor<1x4x1x8xf32>, %arg1: tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32> {

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x1x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%[[ADD]], %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE]]

  // 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>

// CHECK: %[[CUSTOM_AGGREGATOR_1:.+]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%[[XLA_CALL_MODULE]]) {{.*}} : (tensor<1x4x4x8xf32>) -> (tensor<1x4x4x8xf32>, tensor<f32>, tensor<f32>, tensor<0xi64>)

// CHECK: %[[TRANSPOSE_2:.+]] = stablehlo.transpose %[[CUSTOM_AGGREGATOR_1]], dims = [0, 3, 1, 2] : (tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32>

  %1 = mhlo.multiply %0, %cst1 : tensor<1x1x2x4xf32>
  // CHECK:      return %[[RES]] : tensor<1x1x2x4xf32>
  func.return %1 : tensor<1x1x2x4xf32>
}

// CHECK-LABEL: @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float
func.func @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float() -> (tensor<1x2x2x3xf32>) {

  %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>
}