%2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x1x2xbf16>) -> tensor<1x1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x1x2xf32>) -> tensor<1x1x2xf32>
  return %3 : tensor<1x1x2xf32>
}

// CHECK: func @cast_bf16_batch_matmul_v2_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>

  // CHECK-LABEL: test_batch_matmul_broadcast_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  func.return %0: tensor<2x2x2xf32>
}

// CHECK:  [[VAL_3:%.*]] = "tfl.reshape"([[VAL_2]], %{{.*}}) : (tensor<1x1x1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>

  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_1 = "tf.Const"() {value = dense<0.200000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_2 = "tf.Const"() {value = dense<0.300000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>

  %w = arith.constant dense<[[[[-1.0, 1.0]]], [[[1.0, 2.0]]], [[[-2.0, 1.0]]]]> : tensor<3x1x1x2xf32>
  %b = arith.constant dense<[1.0e-2, 2.1473647e1, -2.1473647e2]> : tensor<3xf32>
  %transpose_conv = "tfl.transpose_conv"(%arg1, %w, %0, %b) {
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, fused_activation_function = "NONE"
  } : (tensor<4xi32>, tensor<3x1x1x2xf32>, tensor<1x5x5x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

  %1 = "tfl.add"(%cst_2, %cst_1) {fused_activation_function = "NONE"} : (tensor<2x2x2xf32>, tensor<  2x2xf32>) -> tensor<2x2x2xf32>
  %2 = "tfl.add"(%cst_0, %cst_2) {fused_activation_function = "NONE"} : (tensor<    2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>

  func.return %0, %1, %2 : tensor<2x2xf32>, tensor<2x2x2xf32>, tensor<2x2x2xf32>

  return %0 : tensor<1x1x1x2xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_5(%arg0: tensor<1x1x1x7xf32>) -> tensor<1x1x1x2xf32>

// CHECK-LABEL: avg_pool2d_6
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>

  return %2 : tensor<?x2x2x1xi8>
}

// -----

// CHECK-LABEL: func @convolution_add_add_f32
func.func @convolution_add_add_f32(
    %lhs: tensor<?x3x2x1xf32>, %rhs: tensor<2x1x1x1xf32>,
    %zp_offset: tensor<?x2x2x1xf32>, %bias: tensor<1xf32>
  ) -> tensor<?x2x2x1xf32> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution
  // CHECK-DAG: %[[combined:.*]] = chlo.broadcast_add %[[conv:.*]], %[[zp_offset:.*]]