TensorFlow 연산 융합

def get_implements_signature():
  implements_signature = [
    # 'name' will be used as a name for the operation.
    'name: "my_custom_fused_op"',
    # attr "tfl_fusable_op" is required to be set with true value.
    'attr {key: "tfl_fusable_op" value { b: true } }',
    # Example attribute "example_option" that the op accepts.
    'attr {key: "example_option" value { i: %d } }' % 10
  ]
  return ' '.join(implements_signature)

@tf.function(experimental_implements=get_implements_signature())
def my_custom_fused_op(input_1, input_2):
  # An empty function that represents pre/post processing example that
  # is not represented as part of the Tensorflow graph.
  output_1 = tf.constant(0.0, dtype=tf.float32, name='first_output')
  output_2 = tf.constant(0.0, dtype=tf.float32, name='second_output')
  return output_1, output_2

class TestModel(tf.Module):
  def __init__(self):
    super(TestModel, self).__init__()
    self.conv_1 = tf.keras.layers.Conv2D(filters=1, kernel_size=(3, 3))
    self.conv_2 = tf.keras.layers.Conv2D(filters=1, kernel_size=(3, 3))

  @tf.function(input_signature=[
      tf.TensorSpec(shape=[1, 28, 28, 3], dtype=tf.float32),
      tf.TensorSpec(shape=[1, 28, 28, 3], dtype=tf.float32),
  ])
  def simple_eval(self, input_a, input_b):
    return my_custom_fused_op(self.conv_1(input_a), self.conv_2(input_b))

  TfLiteRegistration reg;
  // This name must match the name specified in the implements signature.
  static constexpr char kOpName[] = "my_custom_fused_op";
  reg.custom_name = kOpName;
  reg.prepare = [](TfLiteContext* context, TfLiteNode* node) -> TfLiteStatus {
    // Add your code.
    return kTfLiteOk;
  };
  reg.invoke = [](TfLiteContext* context, TfLiteNode* node) -> TfLiteStatus {
    // Add your coder.
    return kTfLiteOk;
  };
  reg.builtin_code = kTfLiteCustom;
  resolver->AddCustom(kOpName, &reg);

  @tf.function(
        experimental_implements="embedding_lookup")
    def EmbFprop(embs, ids_vec):
      """Embedding forward prop.

      Effectively, it computes:
        num = size of ids_vec
        rets = zeros([num, embedding dim])
        for i in range(num):
          rets[i, :] = embs[ids_vec[i], :]
        return rets

      Args:
        embs: The embedding matrix.
        ids_vec: A vector of int32 embedding ids.

      Returns:
        The result of embedding lookups. A matrix of shape
        [num ids in ids_vec, embedding dims].
      """
      num = tf.shape(ids_vec)[0]
      rets = inplace_ops.empty([num] + emb_shape_suf, py_utils.FPropDtype(p))

      def EmbFpropLoop(i, embs, ids_vec, rets):
        # row_id = ids_vec[i]
        row_id = tf.gather(ids_vec, i)
        # row = embs[row_id]
        row = tf.reshape(tf.gather(embs, row_id), [1] + emb_shape_suf)
        # rets[i] = row
        rets = inplace_ops.alias_inplace_update(rets, [i], row)
        return embs, ids_vec, rets

      _, _, rets = functional_ops.For(
          start=0,
          limit=num,
          delta=1,
          inputs=[embs, ids_vec, rets],
          body=EmbFpropLoop,
          rewrite_with_while=compiled)
      if len(weight_shape) > 2:
        rets = tf.reshape(rets, [num, symbolic.ToStatic(p.embedding_dim)])
      return rets

void PrepareCompositeFunctionsPass::ConvertTFImplements(FuncOp func,
                                                        StringAttr attr) {
  if (attr.getValue() == "embedding_lookup") {
    func.eraseBody();
    func.addEntryBlock();
    // Convert the composite embedding_lookup function body to a
    // TFLite fused embedding_lookup op.
    ConvertEmbeddedLookupFunc convert_embedded_lookup(func);
    if (failed(convert_embedded_lookup.VerifySignature())) {
      return signalPassFailure();
    }
    convert_embedded_lookup.RewriteFunc();
  } else if (attr.getValue() == mlir::TFL::kKerasLstm) {
     func.eraseBody();
     func.addEntryBlock();
     OpBuilder builder(func.getBody());
     if (failed(ConvertKerasLSTMLayer(func, &builder))) {
       return signalPassFailure();
     }
  } else if (.....) /* Other fusions can plug in here */
}

void RewriteFunc() {
    Value lookup = func_.getArgument(1);
    Value value = func_.getArgument(0);
    auto output_type = func_.getType().getResult(0);

    OpBuilder builder(func_.getBody());
    auto op = builder.create<mlir::TFL::EmbeddingLookupOp>(
        func_.getLoc(), output_type, lookup, value);

    builder.create<mlir::ReturnOp>(func_.getLoc(), op.getResult());
  }