func.func @broadcast_to_to_reshape(%arg0: tensor<4x4x4xf32>, %arg1 : tensor<4xi32>) -> tensor<1x4x4x4xf32> {
  %0 = "tfl.broadcast_to"(%arg0, %arg1) : (tensor<4x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x4xf32>
  // CHECK: "tfl.reshape"
  // CHECK-SAME: (tensor<4x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x4xf32>
  func.return %0 : tensor<1x4x4x4xf32>
}

// Converts tfl.broadcast_to to tfl.reshape if input and output have the same

// RUN: tf-quant-opt %s -quant-cast-bf16-ops-to-f32 | FileCheck %s

func.func @cast_bf16_conv_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<3xf32>) -> tensor<*xf32>
    %2 = "tf.PartitionedCall"(%arg0, %cst_2) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_1} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

// RUN: FileCheck %s

// Remove TPU related ops.
func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

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>
// CHECK: return %[[CUSTOM_0]] : tensor<1x2x2x1xf32>
}

}
// CHECK: func.func @main(%[[ARG:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
// CHECK-DAG: %[[CONST_0:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2xf32>
// CHECK-DAG: %[[CONST_1:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2x3x3x2xf32>
// CHECK-DAG: %[[CONV:.*]] = stablehlo.convolution(%[[ARG]], %[[CONST_1]]) {{.*}} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

    %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.

  return %2 : tensor<1x8x4x4xf32>
}
// CHECK-DAG: %[[CONST:.+]] = stablehlo.constant {{.*}} : tensor<3x3x8x8xf32>
// CHECK-DAG: %[[TRANSPOSE_0:.+]] = stablehlo.transpose %[[ARG]], dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>