%45 = mhlo.reshape %44 : (tensor<1x64x32x1xf32>) -> tensor<1x64x32xf32>
  %46 = "mhlo.broadcast_in_dim"(%45) <{broadcast_dimensions = dense<[0, 1, 2]> : tensor<3xi64>}> : (tensor<1x64x32xf32>) -> tensor<1x64x32x32xf32>
  %47 = mhlo.multiply %43, %46 : tensor<1x64x32x32xf32>
  %48 = mhlo.add %30, %47 : tensor<1x64x32x32xf32>

// CHECK-LABEL: squeezeToReshape
func.func @squeezeToReshape(%arg0: tensor<1x1x2xf32>) -> tensor<2xf32> {
  %0 = "tfl.squeeze"(%arg0) : (tensor<1x1x2xf32>) -> tensor<2xf32>
  func.return %0 : tensor<2xf32>

  // CHECK-DAG: [[CONST:.*]] = arith.constant dense<2> : tensor<1xi32>
  // CHECK: %[[RESULT:.*]] = "tfl.reshape"(%arg0, %[[CONST:.*]]) : (tensor<1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
  // CHECK: return %[[RESULT]]
}

  %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>)
}

  }

  // CHECK-LABEL: batchmatmulv2
  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>

  func.func @operand_as_shape(%18: tensor<i32>, %39: tensor<1x4x4x32xf32>) -> () {
    %cst_5 = arith.constant dense<512> : tensor<i32>
    %19 = "tf.Pack"(%18, %cst_5) {N = 2 : i64, T = i32, axis = 0 : i64, device = ""} : (tensor<i32>, tensor<i32>) -> tensor<2xi32>
    // CHECK: -> tensor<1x512xf32>
    %40 = "tf.Reshape"(%39, %19) {T = f32, Tshape = i32, device = ""} : (tensor<1x4x4x32xf32>, tensor<2xi32>) -> tensor<?x?xf32>
   func.return
  }

// does not alter the last dimension and it restores the batch dimensions
// collapsed by the FullyConnected op due to `keep_num_dims=false`. For example,
//
//   // %input: tensor<4x16x32xf32>
//   %fc = tfl.fully_connected(%input, %filter, %bias)
//           {keep_num_dims = false, weights_format = "DEFAULT"}
//   %shape = arith.constant dense<[4, 16, 32]> : tensor<3xi32>

}

// -----

func.func @testPReluInvalidBroadcast(%arg0: tensor<15x14x2x14xf32>, %arg1: tensor<1x1x3xf32>) -> tensor<15x14x2x14xf32> {
  // expected-error @+1 {{'tfl.prelu' op operands don't have broadcast-compatible shapes}}
  %0 = "tfl.prelu"(%arg0, %arg1) : (tensor<15x14x2x14xf32>, tensor<1x1x3xf32>) -> tensor<15x14x2x14xf32>
  func.return %0 : tensor<15x14x2x14xf32>
}

// -----

  // CHECK-DAG: [[VAL4:%.+]] = "tf.Const"() <{value = dense<[2, 0, 1, 3]> : tensor<4xi64>}>
  // CHECK-DAG: [[VAL5:%.+]] = "tf.PadV2"(%arg0, [[VAL2]], [[VAL3]])
  // CHECK-SAME: tensor<1x15x2xf32>
  // CHECK-DAG: [[VAL6:%.+]] = "tf.Reshape"([[VAL5]], [[VAL1]])
  // CHECK-DAG: [[VAL7:%.+]] = "tf.Transpose"([[VAL6]], [[VAL4]])
  // CHECK-DAG: [[VAL8:%.+]] = "tf.Reshape"([[VAL7]], [[VAL0]])