%8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

// 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) ({

//CHECK-NEXT:}

func.func @iota() -> tensor<3x4xf32> {
 %0 = "vhlo.iota_v1" () <{iota_dimension = #vhlo.integer_v1<0 : i64>}> : () -> tensor<3x4xf32>
 return %0 : tensor<3x4xf32>
}

//CHECK:func.func private @iota() -> tensor<3x4xf32> {
//CHECK-NEXT: %0 = "vhlo.iota_v1"() <{iota_dimension = #vhlo.integer_v1<0 : i64>}> : () -> tensor<3x4xf32>
//CHECK-NEXT: return %0 : tensor<3x4xf32>
//CHECK-NEXT:}

  %0 = "mhlo.reshape"(%arg0) : (tensor<1x1x512xf32>) -> tensor<1x512xf32>
  %1 = "mhlo.dot"(%0, %arg1) : (tensor<1x512xf32>, tensor<512x13x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x13xf32>
  %2 = "mhlo.reshape"(%1) : (tensor<1x13xf32>) -> tensor<1x1x13xf32>
  func.return %2 : tensor<1x1x13xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

// CHECK:          return %[[VALUES]], %[[INDICES]] : tensor<1x4xf32>, tensor<1x4xi32>
// CHECK:        }
func.func @convert_approx_top_k_custom_call(%arg0: tensor<1x4xf32>, %arg1: tensor<1x4xi32>, %arg2: tensor<f32>, %arg3: tensor<i32>) -> (tensor<1x4xf32>, tensor<1x4xi32>) {
  %0:2 = mhlo.custom_call @ApproxTopK(%arg0, %arg1, %arg2, %arg3) {

  // CHECK: return %arg0
  func.return %0 : tensor<2x4xf32>
}

// CHECK-LABEL: func @testBroadcastToNoOp
func.func @testBroadcastToNoOp(%arg0: tensor<2x4xf32>, %arg1: tensor<2xi32>) -> tensor<2x4xf32> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<2x4xf32>, tensor<2xi32>) -> tensor<2x4xf32>

  // CHECK: return %arg0
  func.return %0 : tensor<2x4xf32>
}

// dimensions into a single dimension. For example,
//
//   %shape = arith.constant dense<[1, 128, 64]> : tensor<3xi32>
//   %reshape = tfl.reshape(%input, %shape) // %input: tensor<128x64xf32>
//   %fc = tfl.fully_connected(%reshape, %filter, %bias)
//           {keep_num_dims = false, weights_format = "DEFAULT"}
//
// can be canonicalized to
//

    // CHECK-NOT: tf.TPUPartitionedOutputV2
    %partitioned_output:2 = "tf.TPUPartitionedOutputV2"(%computation) {_XlaSharding = "\08\03\1A\02\01\02\22\02\00\01", partition_dims = [1, 2]} : (tensor<3x4xf32>) -> (tensor<3x2xf32>, tensor<3x2xf32>)
    // CHECK: "tf.AssignVariableOp"(%arg0, %[[PARALLEL_EXECUTE_OUTPUT]]#0)