%2 = "tf.Relu6"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

  %3 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = true, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>
  %4 = "tf.BiasAdd"(%3, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<10xf32>) -> tensor<*xf32>
  %5 = "tf.Relu"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %1 = "tfl.tanh"(%0) : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %2 = "quantfork.stats"(%1) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  func.return %2 : tensor<1x1xf32>

        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20x20xf32>, none,
        tensor<1x20xf32>, tensor<1x20xf32>,
        none, none, none, none) -> tensor<1x28x20xf32>
    %1 = "quantfork.stats"(%0) {layerStats = dense<[-1.0, 2.0]> : tensor<2xf32>} : (tensor<1x28x20xf32>) -> tensor<1x28x20xf32>

// CHECK:           [[VAL_5:%.*]] = call @func_1_GPU_FLOAT([[VAL_0]], [[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_1"} : (tensor<100xf32>, tensor<100xf32>) -> tensor<100xf32>

  %2 = func.call @fold_all_test(%arg0, %0, %1) : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %2 : tensor<256x30x30x16xf32>
}

// ALL-LABEL: @fold_all_test
func.func @fold_all_test(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>) -> tensor<256x30x30x16xf32> {

func.func @add(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Add"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0: tensor<1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexAdd", custom_option = #tfl<const_bytes : "0x03416464001412034164641A001A002A070A015412023001320000021B171414042801">}> : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  %b = arith.constant dense<0.0> : tensor<512xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<512x12xf32>, tensor<512xf32>) -> tensor<1x112x112x512xf32>
  func.return %fc : tensor<1x112x112x512xf32>

// CHECK: %[[b:.*]] = arith.constant dense<0.000000e+00> : tensor<512xf32>