%cst_0 = arith.constant dense<1.0> : tensor<3x1x1x3xf32>
      %cst_1 = arith.constant dense<2.0> : tensor<3xf32>
      %0 = "tf.XlaCallModule"(%arg0, %cst_0, %cst_1) <{Sout = [#tf_type.shape<1x4x4x3>], module = "", version = 9 : i64}> {_entry_function = @composite_fn_1, _original_entry_function = "composite_fn_1", _tfl_quant_trait = "fully_quantizable"} : (tensor<1x4x4x3xf32>, tensor<3x1x1x3xf32>, tensor<3xf32>) -> tensor<1x4x4x3xf32>

      stablehlo.return %2 : tensor<f32>
  }) {padding = dense<[[0, 0], [1, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>, window_dimensions = array<i64: 1, 3, 3, 1>} : (tensor<2x3x1x3xf32>, tensor<f32>) -> tensor<2x3x1x3xf32>
  return %1 : tensor<2x3x1x3xf32>
}

// -----

// CHECK-LABEL: @bitcast_convert_i32_f32(%arg0: tensor<1x256128xi32>) -> tensor<1x256128xbf16>

// PerChannel-NOT: tensor<2x3x3x2xf32>
// PerChannel-SAME: tensor<2x3x1x6xf32>
// PerChannel: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %[[CONST_1:.*]]) <{config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_0}> {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x3x4x3xf32>, tensor<2x3x1x6xf32>) -> tensor<*xf32>

// CHECK-LABEL: conv2d_backprop_input_with_add
func.func @conv2d_backprop_input_with_add(%arg0: tensor<4xi32>, %arg1: tensor<3x3x1x32xf32>, %arg2: tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32> {
  %0 = "tf.Conv2DBackpropInput"(%arg0, %arg1, %arg2) {strides = [1, 2, 2, 1], padding="SAME", dilations=[1, 1, 1, 1]}: (tensor<4xi32>, tensor<3x3x1x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>

// CHECK: %[[CONST_1:.*]] = arith.constant dense
// CHECK-NOT: tensor<2x3x3x2xf32>
// CHECK-SAME: tensor<2x3x1x6xf32>
// CHECK: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %[[CONST_1:.*]]) <{config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn_0}> {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x3x4x3xf32>, tensor<2x3x1x6xf32>) -> tensor<*xf32>

  %cst_3 = "tf.Const"() {value = dense<[[[[1.400000e+01]], [[-2.800000e+01]], [[4.200000e+01]]], [[[-5.600000e+01]], [[7.100000e+01]], [[-8.500000e+01]]], [[[9.900000e+01]], [[-1.130000e+02]], [[1.270000e+02]]]]> : tensor<3x3x1x1xf32>} : () -> tensor<3x3x1x1xf32>
  %cst_4 = "tf.Const"() {value = dense<-1.280000e+02> : tensor<f32>} : () -> tensor<f32>
  %cst_5 = "tf.Const"() {value = dense<0.00118110236> : tensor<1xf32>} : () -> tensor<1xf32>

  // CHECK-SAME: %[[ARG1:.*]]: tensor<*x!tf_type.resource<tensor<f32>>> {tf.device = "/device:TPU:1"},
  // CHECK-SAME: %[[ARG2:.*]]: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>> {tf.device = "/device:TPU:0"},
  // CHECK-SAME: %[[ARG3:.*]]: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>> {tf.device = "/device:TPU:1"})
  func.func @main(%arg0: !tf_res_f32 {tf.device = "/device:TPU:0"},
             %arg1: !tf_res_f32 {tf.device = "/device:TPU:1"},

}

func.func @depthwiseConv2D(tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>, tensor<256x3x32x32xf32>) -> (tensor<256x30x30x12xf32>, tensor<256x12x30x30xf32>, tensor<256x30x30x12xf32>, tensor<256x30x30x12xf32>) {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<3x3x3x4xf32>, %arg2: tensor<256x3x32x32xf32>) :
   // OK

// RUN: tf-mlir-translate -mlir-to-graphdef %s -tf-graph-as-function -o - | FileCheck %s

func.func @main(%arg0: tensor<*x!tf_type.resource>, %arg1: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>>, %arg2: tensor<*xf32>, %arg3: tensor<2x4x6x8xi32>) -> (tensor<f32>, tensor<f32>)
attributes {tf.entry_function = {inputs = "args_0,args_1,args_2,args_3", outputs = "rets_0_RetVal,rets_1_RetVal"}} {
  %graph:2 = tf_executor.graph {

func.func @conv_fn(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x3x3x4xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<3x3x4x4xf32>
  %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x4xf32>) -> tensor<1x3x3x4xf32>
  func.return %1: tensor<1x3x3x4xf32>
}