// CHECK-NEXT:    "vhlo.return_v1"(%1) : (tensor<f32>) -> ()
  // CHECK-NEXT: }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  %0 = stablehlo.reduce(%arg0 init: %arg1) applies stablehlo.add across dimensions = [1, 2] : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

// -----

// CHECK-LABEL: op_with_region_mixed_tfl_shlo_tfl

    %9:4 = "tf.CustomAggregator"(%6) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<1x64xf32>) -> (tensor<1x64xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    return %9#0 : tensor<1x64xf32>
  }

}

func.func @QDQsFollowedByTranspose(tensor<1x2xf32>) -> (tensor<2x1xf32>) {
^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %0 = "tfl.unidirectional_sequence_lstm"(%arg0,

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>

// CHECK-LABEL: testBasicLstm
func.func @testBasicLstm(%arg0: tensor<1x384xf32>, %arg1: tensor<1x96xf32>, %arg2: tensor<384x480xf32>, %arg3: tensor<384xf32>, %arg4: tensor<1x96xf32>) -> (tensor<1x96xf32>, tensor<1x96xf32>) {

    "vhlo.return_v1" (%421) : (tensor<1xf32>) -> ()
   }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

//CHECK:func.func private @reduce(%arg0: tensor<1x16x16x320xf32>, %arg1: tensor<f32>) -> tensor<1x320xf32> {
//CHECK-NEXT:  %0 = "vhlo.reduce_v1"(%arg0, %arg1) <{dimensions = #vhlo.tensor_v1<dense<[1, 2]> : tensor<2xi64>>}> ({ 

// CHECK-LABEL: reduce_window_max_activation_transpose
// CHECK-SAME: (%[[ARG:.+]]: tensor<16x8xf32>) -> tensor<4x8xf32>
func.func @reduce_window_max_activation_transpose_rank2(%arg0: tensor<16x8xf32>) -> tensor<4x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({

  } : (tensor<9x10x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<9x10xf32>
  %0 = "tf.MatMul"(%input, %filter_new) {
  } : (tensor<8x9xf32>, tensor<9x10xf32>) -> tensor<8x10xf32>
  return %0 : tensor<8x10xf32>
})mlir";
  constexpr absl::string_view kProgram = R"mlir(
func.func @main(%input: tensor<8x9xf32>, %filter: tensor<9x10xi8>) -> tensor<8x10xf32> {
  %filter_scale = "tf.Const"() { value = dense<0.0235> : tensor<f32> } : ()

// CHECK-LABEL: testFoldEnsureShapeOp
func.func @testFoldEnsureShapeOp(%arg0: tensor<10x20xf32>) -> (tensor<10x20xf32>, tensor<10x20xf32>, tensor<20x10xf32>) {
  %0 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<10x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  %1 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<?x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  // Failing case which should not be folded.