<tensor<640xf32>>>, tensor<!tf_type.resource<tensor<640xf32>>>, tensor<!tf_type.resource<tensor<512x640xf32>>>, tensor<!tf_type.resource<tensor<640x640xf32>>>, tensor<!tf_type.resource<tensor<4096xf32>>>, tensor<!tf_type.resource<tensor<640x4096xf32>>>, tensor<f32>, tensor<!tf_type.resource<tensor<4096xf32>>>, tensor<!tf_type.resource<tensor<640xf32>>>, tensor<!tf_type.resource<tensor<512xf32>>>, tensor<!tf_type.resource<tensor<640xf32>>>, tensor<!tf_type.resource<tensor<640xf32>>>, tensor<!tf_t...

  %4 = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %5 = "tf.Range"(%4, %3, %2) : (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<4096xi32>
  %6 = "tf.Equal"(%1, %5) : (tensor<*xi32>, tensor<4096xi32>) -> tensor<*xi1>
  %7 = "tf.Cast"(%6) : (tensor<*xi1>) -> tensor<*xf32>
  %8 = "tf.BatchMatMulV2"(%7, %arg0) {adj_x = false, adj_y = false} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  %0:4 = "tfl.basic_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4) {fused_activation_function = "RELU", cell_clip = 1.0 : f32, proj_clip = 2.0 : f32} : (tensor<1x384xf32>, tensor<1x96xf32>, tensor<384x480xf32>, tensor<384xf32>, tensor<1x96xf32>) -> (tensor<1x96xf32>, tensor<1x96xf32>, tensor<1x480xf32>, tensor<1x384xf32>)
  func.return %0#0 : tensor<1x96xf32>

// CHECK:             mhlo.return %[[VAL_5]] : tensor<f32>
// CHECK:           }) : (tensor<20x6xf32>, tensor<4xi32>, tensor<4x6xf32>) -> tensor<20x6xf32>
// CHECK:           return %[[VAL_3]] : tensor<20x6xf32>
// CHECK:         }
func.func @convert_scatter_update(%arg0: tensor<20x6xf32>, %arg1: tensor<4xi32>, %arg2: tensor<4x6xf32>) -> tensor<20x6xf32> {
  %0 = "mhlo.scatter"(%arg0, %arg1, %arg2) ({
  ^bb0(%arg3: tensor<f32>, %arg4: tensor<f32>):

^bb0(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>):
  %cst = arith.constant dense<1.0> : tensor<40xf32>
  %0:2 = "tfl.fully_connected"(%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<40x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> (tensor<40x40xf32>, tensor<40x40xf32>)
  func.return %0 : tensor<40x40xf32>

^bb0(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>):
  %cst = arith.constant dense<1.0> : tensor<40xf32>
  %0:2 = "tfl.fully_connected"(%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<40x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> (tensor<40x40xf32>, tensor<40x40xf32>)
  func.return %0 : tensor<40x40xf32>

  // return %[[RES1]], %[[RES2]], %[[RES3]]
  func.return %0#0, %0#1, %0#2 : tensor<4x6xf32>, tensor<4x6xf32>, tensor<4x6xf32>
}

// -----

// CHECK-LABEL: func @unpack_dynamic
func.func @unpack_dynamic(%arg0: tensor<?x?x2xf32>) -> (tensor<?x?xf32>, tensor<?x?xf32>) {

  %0 = "tfl.unidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6,...

  // CHECK-SAME:      : (tensor<40x37xf32>) -> tensor<40x37xf32>
  // CHECK-NEXT:    %[[cst:.*]] = "tfl.pseudo_const"() <{value = dense<
  // CHECK-SAME:      1.000000e+00> : tensor<40xf32>}> : () -> tensor<40xf32>
  // CHECK-NEXT:    %[[fc:.*]]:2 = "tfl.fully_connected"(%[[stat]], %arg1,
  // CHECK-NEXT:    %[[stat1:.*]] = "quantfork.stats"(%[[fc]]#0)
  // CHECK-SAME:    <{axis = 1 : i64,

  // CHECK-NEXT:  %11 = mhlo.multiply %8, %10 : tensor<1x300x300x40xf32>
  // CHECK-NEXT:  %12 = mhlo.convert %arg2 : tensor<40xf32>
  // CHECK-NEXT:  %13 = "mhlo.broadcast_in_dim"(%12) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<40xf32>) -> tensor<1x300x300x40xf32>
  // CHECK-NEXT:  %14 = mhlo.add %11, %13 : tensor<1x300x300x40xf32>
  // CHECK-NEXT:  %15 = mhlo.constant dense<0.000000e+00> : tensor<f32>