func.func @Int64SliceBeginSize(%arg0: tensor<4x128x32xf32>) -> tensor<1x128x32xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi64>} : () -> tensor<3xi64>
  %1 = "tfl.pseudo_const"() {value = dense<[1, 128, 32]> : tensor<3xi64>} : () -> tensor<3xi64>
  %2 = "tfl.slice"(%arg0, %0, %1) : (tensor<4x128x32xf32>, tensor<3xi64>, tensor<3xi64>) -> tensor<1x128x32xf32>
  func.return %2 : tensor<1x128x32xf32>

// CHECK-NOT: "tfl.batch_matmul"
func.func @Batchmatmul2Fullyconnected(%arg0: tensor<4x128x2xf32>) -> (tensor<4x128x1xf32>) {
  %0 = arith.constant dense<[[1.0], [2.0]]> : tensor<2x1xf32>
  %1 = "tfl.batch_matmul"(%arg0, %0) {adj_x = false, adj_y = false, asymmetric_quantize_inputs = false} : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>
  // CHECK-NEXT: %[[CONST_WEIGHT:.*]] = arith.constant

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

// CHECK:      %[[R1:.*]] = "mhlo.concatenate"(%arg1, %arg3, %arg5, %arg7) <{dimension = 0 : i64}> : (tensor<1x128x72xf32>, tensor<1x128x72xf32>, tensor<1x128x72xf32>, tensor<1x128x72xf32>) -> tensor<4x128x72xf32>
// CHECK:      %[[R2:.*]] = "mhlo.dot_general"(%[[R0]], %[[R1]]) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<
// CHECK-SAME:     lhs_batching_dimensions = [0],

  // CHECK-LABEL: testDilatedConv1DExpandH
  // CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x3xf32>, [[FILTER:%.*]]: tensor<1x5x3x8xf32>)
  // CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() <{value = dense<-3> : tensor<i32>}> : () -> tensor<i32>
  // CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x3xf32>, tensor<i32>) -> tensor<1x1x128x3xf32>

  %0 = "tf.GatherV2"(%cst_0, %arg0, %cst) {batch_dims = 0 : i64, device = ""} : (tensor<128x32xf32>, tensor<6xi64>, tensor<i32>) -> tensor<6x32xf32>
  return %0 : tensor<6x32xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<i32>
// CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() {{.*}} : () -> tensor<128x32xf32>

// CHECK: return [[MUL1]]
}

func.func @batchmatmul2fullyconnected(%arg0: tensor<4x128x2xf32>) -> (tensor<4x128x1xf32>) {
  %0 = "tf.Const"() {value = dense<[[1.0], [2.0]]> : tensor<2x1xf32>} : () -> tensor<2x1xf32>
  %1 = "tf.BatchMatMulV2"(%arg0, %0) : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>

  // CHECK-LABEL: batchmatmul2fullyconnected

  func.return %1 : tensor<128x32xf32>
  // CHECK: %[[FULLY_CONNECTED:.*]] = "tfl.fully_connected"(%arg0, %arg1, %arg2) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<128x64xf32>, tensor<32x64xf32>, tensor<32xf32>) -> tensor<128x32xf32>
  // CHECK: return %[[FULLY_CONNECTED]] : tensor<128x32xf32>
}

  %0 = "tf.CollectivePermute"(%arg0, %source_target_pairs) {
  } : (tensor<128x32xf32>, tensor<3x2xi32>) -> tensor<128x32xf32>

  func.return %0 : tensor<128x32xf32>
}

// -----

// CHECK-LABEL: @cross_replica_sum
func.func @cross_replica_sum(%input: tensor<10xf32>) -> tensor<10xf32> {
  %replica_groups = "tf.Const" () {

// 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>