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>

        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,

// CHECK:           }) : (tensor<1x?xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<3x1xf32>, tensor<3xf32>, tensor<1x3xf32>, tensor<1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x3xf32>

  %cst1 = arith.constant dense<2.0> : tensor<2xf32>
  %cst2 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %cst3 = arith.constant dense<[1, 8]> : tensor<2xi32>

  %0 = "tfl.fully_connected"(%arg0, %cst0, %cst1) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<4x2xf32>

  // CHECK: tac.cost = 1.010000e+02
  %1 = "tfl.reshape"(%0, %cst) {tac.device = "CPU"} : (tensor<200xf32>, tensor<2xi64>) -> tensor<2x100xf32>
  func.return %1 : tensor<2x100xf32>
}

// -----

func.func @testConv2DCPU(tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x32x32x16xf32> {

  %1 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1x3xf32>) -> tensor<1x1x3xf32>
  %2 = "tfl.pseudo_const"() {value = dense<[-1, 3]> : tensor<2xi32>} : () -> tensor<2xi32>
  %3 = "tfl.reshape"(%1, %2) : (tensor<1x1x3xf32>, tensor<2xi32>) -> tensor<1x3xf32>
  %4 = "quantfork.stats"(%3) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %4 : tensor<1x3xf32>

  }

  // CHECK-LABEL: @body(%arg0: tensor<0xf32>) -> tensor<0xf32>
  func.func private @body(%arg0:tensor<0xf32>, %arg1:tensor<*x!tf_type.resource>) -> (tensor<0xf32>, tensor<*x!tf_type.resource>) {
    %val = "tf.ReadVariableOp"(%arg1) : (tensor<*x!tf_type.resource>) -> (tensor<0xf32>)
    %0 = "tf.Mul"(%val, %arg0) : (tensor<0xf32>, tensor<0xf32>) -> tensor<0xf32>
    func.return %0, %arg1 : tensor<0xf32>, tensor<*x!tf_type.resource>
  }

  %7 = "tf.BiasAdd"(%6, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  func.return %2, %5, %7 : tensor<*xf32>, tensor<*xf32>, tensor<*xf32>

// CHECK: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %arg1, %[[CONST_0]])

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DDynamicShape(tensor<?x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<?x32x32x16xf32> {
^bb0(%arg0: tensor<?x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: _arithmetic_count = -1 : i64

  // CHECK:  %[[MUL1:.*]] = "tf.Mul"(%arg1, %[[RSQRT]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  // CHECK:  %[[MUL2:.*]] = "tf.Mul"(%arg0, %[[MUL1]]) : (tensor<1x1x6x2xf32>, tensor<2xf32>) -> tensor<1x1x6x2xf32>
  // CHECK:  %[[MUL3:.*]] = "tf.Mul"(%[[MEAN]], %[[MUL1]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  // CHECK:  %[[SUB:.*]] = "tf.Sub"(%arg2, %[[MUL3]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>