%cst0 = mhlo.constant dense<[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]> : tensor<2x3xf32>
  %cst1 = mhlo.constant dense<[[[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], [[7.0, 8.0, 9.0], [10.0, 11.0, 12.0]]]]> : tensor<1x2x2x3xf32>
  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<[1, 3]> : tensor<2xi64>}> : (tensor<2x3xf32>) -> tensor<1x2x2x3xf32>
  %1 = mhlo.multiply %0, %cst1 : tensor<1x2x2x3xf32>

// CHECK-DAG: %[[input_18:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-8.000000e-01, 1.600000e+00]> : tensor<2xf32>}> : (tensor<1x4xf32>) -> tensor<1x4xf32>
// CHECK-DAG: %[[input_19:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-2.000000e+00, 4.000000e+00]> : tensor<2xf32>}> : (tensor<1x2xf32>) -> tensor<1x2xf32>

    %2:2 = "tf.RecvTPUEmbeddingActivations"() {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D"} : () -> (tensor<2x2xf32>, tensor<4x4xf32>)
    "tf.SendTPUEmbeddingGradients"(%2#0, %2#1) {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D", operandSegmentSizes = array<i32: 2, 0>} : (tensor<2x2xf32>, tensor<4x4xf32>) -> ()
    tf_device.return

    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
    return %2 : tensor<1x3xf32>
  }
// CHECK: func.func private @quantize_dot_general_with_bias_same_shape_fn(%[[ARG_0:.+]]: tensor<1x2xf32>) -> tensor<1x3xf32> attributes {tf._original_func_name = "main_0"}

    module attributes {} {
      func.func @main(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> attributes {} {
        %0 = stablehlo.constant dense<1.0> : tensor<4x3xf32>
        %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
        return %1 : tensor<1x3xf32>
      }
    }
  )mlir");
  ASSERT_TRUE(module_op);

    %9 = stablehlo.convert %2 : (tensor<2x3xi32>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

  %5 = "tf.Relu6"(%y) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

  func.return %3 : tensor<2x1xf32>
}
}

// CHECK:   func @simpleTest(%[[VAL_0:.*]]: tensor<1xf32>, %[[VAL_1:.*]]: tensor<1xf32>, %[[VAL_2:.*]]: tensor<1xf32>, %[[VAL_3:.*]]: tensor<1xf32>) -> tensor<2x1xf32> {

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

  %2 = mhlo.constant dense<6.400000e+01> : tensor<64xf32>
  %3 = mhlo.constant dense<3.200000e+01> : tensor<64xf32>
  %4 = mhlo.constant dense<5.000000e-01> : tensor<64xf32>
  %5 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<64xf32>
  %6 = mhlo.add %5, %4 : tensor<64xf32>
  %7 = mhlo.multiply %6, %3 : tensor<64xf32>
  %8 = mhlo.divide %7, %2 : tensor<64xf32>
  %9 = mhlo.floor %8 : tensor<64xf32>