%0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<4xi32>) -> tensor<1x1x2x4xi32>
  // CHECK: %[[MUL:.*]] = mhlo.multiply %[[BROADCAST]], %[[ARG]] : tensor<1x1x2x4xi32>
  %1 = mhlo.multiply %0, %arg0 : tensor<1x1x2x4xi32>
  // CHECK:      return %[[MUL]] : tensor<1x1x2x4xi32>
  func.return %1 : tensor<1x1x2x4xi32>

// 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>) {

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_to_fp32

    %1 = "tf.Identity"(%0#3) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    func.return %1 : tensor<1x1024xf32>
  }

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>

      platforms = [], version = 4 : i64
    } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[0.000000e+00, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    return %2 : tensor<1x2x2x2xf32>
  }

  // CHECK-LABEL: composite_conv2d_with_bias_and_relu6_fn_10

// 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 @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>

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>
// CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>
    %19 = call @uniform_quantize_1(%18, %5, %6) : (tensor<1x3x3x4xf32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xi8>) -> tensor<1x3x3x4xi8>
    %20 = call @uniform_dequantize_0(%19, %5, %6) : (tensor<1x3x3x4xi8>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xi8>) -> tensor<1x3x3x4xf32>
    return %20 : tensor<1x3x3x4xf32>
  }