// CHECK:  %[[ADD2:.*]] = "tf.Add"(%[[MUL2]], %[[SUB]])
// CHECK:  %[[BATCHNORM1_a:[^,]+]], {{.*}} = "tf.FusedBatchNormV3"(%[[ADD2]], %[[CONSTANT1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
// CHECK:  %[[BATCHNORM1_b:[^,]+]], {{.*}} = "tf.FusedBatchNormV3"(%[[BATCHNORM1_a]], %[[CONSTANT1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
// CHECK:  "tf.FusedBatchNormV3"(%[[BATCHNORM1_b]], %[[ARG1]], %[[ARG2]], %[[ARG3]], %[[ARG4]])
}

// the checks on intermediate values, ideally we should find equivalent
// checks that guarantees the resultant ops are valid.
// The extra conditions are the broadcasting conditions.
//
// The pattern lower FusedBatchNormV3 to arithmetic ops.
// Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//
// Let multiplier = scale / sqrt(variance + epsilon),
// to compute

func.func @fusedBatchNormV3_training_dynamic_unsupported1(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<?xf32>, %arg2: tensor<?xf32>, %arg3: tensor<?xf32>, %arg4: tensor<?xf32>) -> (tensor<?x?x?x?xf32>) {
  // CHECK: tf.FusedBatchNormV3

      self.assertTrue(self._contains_op(output_graphdef, 'XlaCallModule'))
    else:
      self.assertTrue(self._contains_quantized_function_call(output_graphdef))
    self.assertFalse(self._contains_op(output_graphdef, 'FusedBatchNormV3'))

  @parameterized.named_parameters(
      ('to_tf_with_int32_input_type', dtypes.int32, quant_opts_pb2.TF),
      ('to_xla_with_int32_input_type', dtypes.int32, quant_opts_pb2.XLA),

func.func @testFusedBatchNormToBatchNormV3(%arg0: tensor<8x8x8x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> (tensor<8x8x8x8xf32>) {
  // CHECK: "tf.FusedBatchNormV3"
  // CHECK-SAME: device = "/job:localhost/replica:0/task:0/device:GPU:0"

        support generators and Sequences.
*   `tf.lite`
    *   Legalization for `NMS` ops in TFLite.
    *   add `narrow_range` and `axis` to `quantize_v2` and `dequantize` ops.
    *   Added support for `FusedBatchNormV3` in converter.
    *   Add an `errno`-like field to `NNAPI` delegate for detecting `NNAPI`
        errors for fallback behaviour.
    *   Refactors `NNAPI` Delegate to support detailed reason why an operation

    // TF_LayoutSensitiveInterface:
    StringRef GetOptimalLayout(const RuntimeDevices& devices);
    LogicalResult UpdateDataFormat(StringRef data_format);
  }];
}

def TF_FusedBatchNormV3Op : TF_Op<"FusedBatchNormV3", [Pure, TF_FoldOperandsTransposeInterface, TF_LayoutSensitiveInterface]> {
  let summary = "Batch normalization.";

  let description = [{
Note that the size of 4D Tensors are defined by either "NHWC" or "NCHW".