device = ""
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel_default(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

func.func @testConv(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>) -> tensor<256x30x30x16xf32> {
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.conv_2d"(%arg0, %arg1, %arg2) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DDynamicShape(tensor<?x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<?x32x32x16xf32> {
^bb0(%arg0: tensor<?x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: _arithmetic_count = -1 : i64

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-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>
  %1 = "tf.Transpose"(%arg0, %0) : (tensor<1x3x32x32xf32>, tensor<4xi32>) -> tensor<1x32x32x3xf32>

  // Compute in NHWC
  %2 = "tf.Conv2D"(%1, %arg1)
    {

    return %2 : tensor<1x3x2x2xf32>
  }
  func.func private @quantize_i8(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<f32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xi8> {
    %0 = "tf.Div"(%arg0, %arg1) : (tensor<1x3x4x3xf32>, tensor<f32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.Round"(%0) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  %2 = "tfl.select"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  %3 = "tfl.select_v2"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  func.return %0, %1, %2, %3 : tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>

func.func @FoldTrivialReshapeIntoTranspose(%arg: tensor<2x1x3x3xf32>) -> tensor<1x3x3x2xf32> {
  %cst = arith.constant dense<[1, 3, 3, 2]> : tensor<4xi32>
  %cst_2 = arith.constant dense<[2, 3, 0, 1]> : tensor<4xi32>
  %2 = "tf.Transpose"(%arg, %cst_2) : (tensor<2x1x3x3xf32>, tensor<4xi32>) -> tensor<3x3x2x1xf32>
  %3 = "tf.Reshape"(%2, %cst) : (tensor<3x3x2x1xf32>, tensor<4xi32>) -> tensor<1x3x3x2xf32>
  return %3: tensor<1x3x3x2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

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

    %16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>