%1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<1x1xf32>>>) -> tensor<1x1xf32>
  %2:2 = "tf.Split"(%cst, %arg0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x4xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)
  %3:2 = "tf.Split"(%cst, %2#0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x2xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>)

// -----

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

  %5 = stablehlo.broadcast_in_dim %1, dims = [2] : (tensor<64xf32>) -> tensor<1x1x64xf32>
  %6 = stablehlo.add %4, %5 : tensor<1x1x64xf32>
  %7 = stablehlo.clamp %2, %6, %3 : tensor<1x1x64xf32>
  func.return %7: tensor<1x1x64xf32>
}
// CHECK: %[[CONST_0:.*]] = stablehlo.constant dense<2.000000e+00>
// CHECK: %[[CONST_1:.*]] = stablehlo.constant dense<2.000000e+00>

}

// CHECK-LABEL: @broadcast_batch_matmul_v2_rhs
func.func @broadcast_batch_matmul_v2_rhs(%arg0: tensor<17x17x17xf32>, %arg1: tensor<17x24xf32>) -> tensor<17x17x24xf32> {
  %cst = arith.constant dense<[17, 17, 24]> : tensor<3xi64>
  %0 = "tf.BroadcastTo"(%arg1, %cst) : (tensor<17x24xf32>, tensor<3xi64>) -> tensor<17x17x24xf32>

  %0:2 = "tf.Split"(%cst, %input) : (tensor<i32>, tensor<4x?x4xf32>) -> (tensor<2x?x4xf32>, tensor<2x?x4xf32>)
  func.return %0#0, %0#1 : tensor<2x?x4xf32>, tensor<2x?x4xf32>
}

// -----

// CHECK-LABEL: @split_match_and_split_into_two
func.func @split_match_and_split_into_two(%input: tensor<4x6xf32>) -> (tensor<2x6xf32>, tensor<2x6xf32>) {

// NOFALLBACK-LABEL: @xla_svd
func.func @xla_svd(%arg0: tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>) {
  // NOFALLBACK: XlaSvd
  %s, %u, %v = "tf.XlaSvd"(%arg0) {max_iter = 1, epsilon = 1.0E-09 : f32, precision_config = ""} : (tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>)
  func.return %s, %u, %v : tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>
}

}

// -----

// CHECK-LABEL: listPushBack
func.func @listPushBack(%arg0: tensor<!tf_type.variant<tensor<?x1xf32>>>, %arg1: tensor<16x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>  {
  %0 = "tf.TensorListPushBack"(%arg0, %arg1) : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<16x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  return %3 : tensor<1x2xf32>
}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])

  %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) ({