%1 = "tf.Conv2D"(%0, %cst_0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1024xf32>) -> tensor<1x3x2x1024xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>) -> tensor<256x8x7x4xf32>
  func.return %rst : tensor<256x8x7x4xf32>

  %0 = "tf.TPUPartitionedInputV2"(%arg0) {_XlaSharding = "\08\03\1A\05\04\01\01\01\02\22\08\00\01\02\03\04\05\06\070\01", partition_dims = [4, 1, 1, 1, 2], is_packed = true} : (tensor<!tf_type.resource<tensor<1x4x4x4xf32>>>) -> tensor<!tf_type.resource<tensor<4x4x4x4xf32>>>
  %1 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<4x4x4x4xf32>>>) -> tensor<4x4x4x4xf32>

// %1 = stablehlo.constant dense<4>
// %2 = stablehlo.constant dense<2>
// %3 = stablehlo.convolution(%%arg0, %%arg1) :
//          (tensor<?x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<?x3x4x2xf32>
// %4 = stablehlo.get_dimension_size %3, dim = 0 :
//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :