%cst: tensor<f32>) -> tensor<1x64x56x56xf32> {
  %tpos = stablehlo.transpose %input, dims = [0, 3, 1, 2]
      : (tensor<1x114x114x64xf32>) -> tensor<1x64x114x114xf32>
  %ret = "stablehlo.reduce_window"(%tpos, %cst)
      <{window_dimensions = array<i64: 1, 1, 3, 3>, window_strides = array<i64: 1, 1, 2, 2>}> ({
  ^bb0(%arg0: tensor<f32>, %arg1: tensor<f32>):
    %max = stablehlo.maximum %arg0, %arg1 : tensor<f32>

  %13 = "tfl.square"(%6) : (tensor<f32>) -> tensor<f32>

  func.return %7, %8, %9, %10, %11, %12, %13 : tensor<f32>, tensor<f32>, tensor<f32>, tensor<f32>, tensor<f32>, tensor<f32>, tensor<f32>
}

// CHECK-LABEL: @max_with_neg_f32_max_val
// CHECK-SAME: (%[[ARG0:.+]]: tensor<f32>)
func.func @max_with_neg_f32_max_val(%arg0 : tensor<f32>) -> (tensor<f32>, tensor<f32>) {

// CHECK-NEXT: [[res:%.+]] = "quantfork.dcast"([[conv]]) : (tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>) -> tensor<*xf32>

    func @outside_compilation() -> tensor<f32> {
      %0 = "tf_device.cluster"() ( {
        %1 = "tf.Const"() {_xla_outside_compilation = "0", value = dense<1.0> : tensor<f32>} : () -> (tensor<f32>)
        %2 = "tf.Identity"(%1) {_xla_outside_compilation = "0"} : (tensor<f32>) -> (tensor<f32>)
        %3 = "tf.AddV2"(%1, %2) : (tensor<f32>, tensor<f32>) -> (tensor<f32>)
        tf_device.return %3 : tensor<f32>

  %0:3 = "tf.CaseRegion"(%index) ({
      %mul = "tf.Mul"(%arg0, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
      "tf.Yield"(%arg1, %mul, %arg2) : (tensor<f32>, tensor<f32>, tensor<!tf_type.resource>) -> ()
    }, {
      %sub = "tf.Sub"(%arg0, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
      "tf.Yield"(%arg1, %sub, %arg2) : (tensor<f32>, tensor<f32>, tensor<!tf_type.resource>) -> ()
    }, {

    // CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() <{value = dense<3> : tensor<i32>}> : () -> tensor<i32>
    // CHECK-NEXT: %[[UQ:.*]] = "tf.UniformQuantize"(%arg0, %cst, %cst_0) <{quantization_axis = -1 : i64, quantization_max_val = 127 : i64, quantization_min_val = -128 : i64}> : (tensor<1xf32>, tensor<f32>, tensor<i32>) -> tensor<1x!tf_type.qint8>

// CHECK: return %[[CONV2D]] : tensor<1x3x3x2x!quant.uniform<i8:f32, 4.000000e+00>>

// -----

func.func @convolution_upstream_srq_non_const_filter(%arg0: tensor<1x3x3x4x!quant.uniform<i8:f32, 1.000000e+0:-100>>, %arg1: tensor<3x3x4x2x!quant.uniform<i8:f32:3, {2.000000e+2, 3.000000e+3}>>) -> tensor<1x3x3x2x!quant.uniform<i8:f32, 4.000000e+0>> {

}

func.func @branch0(%arg0: tensor<f32>, %arg1: tensor<f32>) -> tensor<f32> {
  %0 = "tf.Add" (%arg0, %arg1) {device = "/device:CPU:0"}  : (tensor<f32>, tensor<f32>) -> tensor<f32>
  func.return %0 : tensor<f32>
}

func.func @branch1(%arg0: tensor<f32>, %arg1: tensor<f32>) -> tensor<f32> {
  %0 = "tf.Add" (%arg0, %arg1) {device = "/device:CPU:0"}  : (tensor<f32>, tensor<f32>) -> tensor<f32>

@testUnidirectionalSequenceLstmWithIntermediates(%arg0: tensor<? x ? x ? x f32>, %arg1: tensor<? x ? x f32>, %arg2: tensor<? x ? x f32>, %arg3: tensor<? x ? x f32>, %arg4: tensor<? x ? x f32>, %arg5: tensor<? x ? x f32>, %arg6: tensor<? x ? x f32>, %arg7: tensor<? x ? x f32>, %arg8: tensor<? x ? x f32>, %arg9: tensor<? x f32>, %arg10: tensor<? x f32>, %arg11: tensor<? x f32>, %arg12: tensor<? x f32>, %arg13: tensor<? x f32>, %arg14: tensor<? x f32>, %arg15: tensor<? x f32>, %arg16: tensor<? x ? x f32>, %arg17:...

  %4 = "tfl.reshape"(%3, %cst) : (tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>, tensor<2xi32>) -> tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>
  %5 = "tfl.softmax"(%4) {beta = 1.000000e+00 : f32} : (tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>) -> tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>