func.func @move_across_double_transpose(%arg0: tensor<1x4x4x8xf32>, %arg1: tensor<1x4x4x8xf32>) -> tensor<1x4x8x4xf32> {

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  // CHECK: %[[RES_TRANSPOSE_0:[0-9]*]] = "tf.Transpose"(%[[ADD]], %[[RES_PERM]])

  } : (tensor<6x7x8xf32>, tensor<4xi32>, tensor<4xi32>, tensor<4xi32>) -> tensor<1x4x1x8xf32>
  func.return %0 : tensor<1x4x1x8xf32>

  // CHECK-DAG: %[[NEW_DIMS:.*]] = arith.constant dense<[6, 1, 7, 1, 8]> : tensor<5xi32>
  // CHECK-DAG: %[[BEGIN:.*]] = arith.constant dense<0> : tensor<5xi32>

  return %2 : tensor<1x1x1x4xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_2(%arg0: tensor<1x1x1x8xf32>) -> tensor<1x1x1x4xf32>

// CHECK-LABEL: avg_pool2d_3
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>
// CHECK: %0 = "tfl.transpose"(%arg0, %cst) : (tensor<1x1x1x8xf32>, tensor<4xi32>) -> tensor<1x1x8x1xf32>

  %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

         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>
{

// 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>

  // 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 = 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.

// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:           %[[VAL_7:.*]] = "tfl.concatenation"(%[[VAL_5]], %[[VAL_6]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %1 = "tfl.reshape"(%arg1, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
    %2 = "tfl.concatenation"(%0, %1) {axis = 3 : i32, fused_activation_function = "NONE", tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>