// 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>

    %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>

    "tfl.yield"(%1#0, %1#1) : (tensor<*xi32>, tensor<*xf32>) -> ()
  }) : (tensor<i32>, tensor<1xf32>) -> (tensor<i32>, tensor<1xf32>) loc("WhileOp")
  // CHECK: (tensor<i32>, tensor<1xf32>, tensor<i32>) ->
  // CHECK-SAME: (tensor<i32>, tensor<1xf32>, tensor<i32>)
  func.return %0#1 : tensor<1xf32>
}

func.func private @WhileOp_cond(%arg0: tensor<*xi32>, %arg1: tensor<*xf32>, %arg2: tensor<i32>) -> tensor<i1> {

  func.return %0 : tensor<*xf32>
}

// CHECK-NOT: "tfl.reshape"
// CHECK: "tfl.pack"

// -----

func.func @squaredDifference(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
  %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  func.return %0 : tensor<4xf32>
}

    %4 = "tf.BiasAdd"(%3, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %5 = "tf.Relu6"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

    func.return %2, %5 : tensor<*xf32>, tensor<*xf32>
  }

  func.func @float_conv_1(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> (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>

  func.return %0 : tensor<? x ? x ? x f32>
}

  %cst_1 = arith.constant dense<0> : tensor<4xi32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<1xf32>
  %0 = "tfl.quantize"(%cst_2) {qtype = tensor<1x!quant.uniform<i32:f32, 1.0>>, volatile} : (tensor<1xf32>) -> tensor<1x!quant.uniform<i32:f32, 1.0>>
  %1 = "tfl.dequantize"(%0) : (tensor<1x!quant.uniform<i32:f32, 1.0>>) -> tensor<1xf32>
  %cst_3 = arith.constant dense<[[[[1.0]], [[1.0]], [[1.0]]]]> : tensor<1x3x1x1xf32>

  %7 = "tfl.add"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<  f32>) -> tensor<4xf32>
  %8 = "tfl.add"(%2, %3) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  %9 = "tfl.add"(%2, %3) {fused_activation_function = "SIGN_BIT"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

    tensor type, the shape shouldn't be changed during the conversion.

    Example:

    ```mlir
    %1 = tfr.constant_tensor(%0) : f32 -> tensor<f32>
    %3 = tfr.constant_tensor(%2) : vector<1xf32> -> tensor<1xf32>
    ```
  }];

  let arguments = (ins TFR_AllAttrTypes:$arg);

  let results = (outs TFR_singleTensorType:$out);

  let hasCanonicalizer = 1;

  let hasVerifier = 1;
}