// 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>) {

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:}

// 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 @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>

func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>

  %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>) {

    %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.

       } : (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> {

// MinElement-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w_1 = arith.constant dense<127.0> : tensor<4096x1x1x1xf32>

// 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>