return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

func.func @reshape_removeIdentity(tensor<4x4x4xf32>) -> tensor<4x4x4xf32> {
^bb0(%arg0: tensor<4x4x4xf32>) :
  %cst = arith.constant dense<[4, 4, 4]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<4x4x4xf32>, tensor<3xi32>) -> tensor<4x4x4xf32>
  func.return %0 : tensor<4x4x4xf32>

// CHECK-LABEL: func @reshape_removeIdentity
// CHECK:  return %arg0 : tensor<4x4x4xf32>
}

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x1x2xbf16>) -> tensor<1x1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x1x2xf32>) -> tensor<1x1x2xf32>
  return %3 : tensor<1x1x2xf32>
}

// CHECK: func @cast_bf16_batch_matmul_v2_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>

    %2 = "quantfork.stats"(%1) {layerStats = dense<[3.50919247, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    %4 = "quantfork.stats"(%3) {layerStats = dense<[3.50919247, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    func.return %4 : tensor<1x3x4x2xf32>

  func.func @batchmatmulv2(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
    // CHECK: mhlo.reduce
    // CHECK: mhlo.dot_general
    // CHECK: mhlo.transpose
    %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
    func.return %0 : tensor<3x4x4xf32>
  }

  %43 = "quantfork.stats"(%42) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x4x3xf32>) -> tensor<1x4x3xf32>
  %44 = "tfl.strided_slice"(%43, %1, %2, %3) {begin_mask = 7 : i32, ellipsis_mask = 0 : i32, end_mask = 5 : i32, new_axis_mask = 0 : i32, shrink_axis_mask = 0 : i32, offset = false} : (tensor<1x4x3xf32>, tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x2x3xf32>

  %0:3 = "tf.Split"(%arg0, %arg1) : (tensor<i32>, tensor<1x4x3x3xf32>) -> (tensor<1x4x3xf32>, tensor<1x4x3xf32>, tensor<1x4x3xf32>)
  func.return %0#0 : tensor<1x4x3xf32>

  // CHECK-LABEL: split
  // CHECK: "tfl.split"(%arg0, %arg1) <{num_splits = 3 : i32}> : (tensor<i32>, tensor<1x4x3x3xf32>) -> (tensor<1x4x3xf32>, tensor<1x4x3xf32>, tensor<1x4x3xf32>)
}

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

    %12 = "tfl.logistic"(%11) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %13 = tfl.mul %arg4, %12 {fused_activation_function = "NONE"} : tensor<4x2xf32>
    %14 = "tfl.relu"(%10#1) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %15 = "tfl.logistic"(%10#0) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %16 = tfl.mul %15, %14 {fused_activation_function = "NONE"} : tensor<4x2xf32>