1 files changed, 150 insertions, 0 deletions
diff --git a/compiler/moco-tf/src/Canonicalization/RsqrtCanonicalizer.cpp b/compiler/moco-tf/src/Canonicalization/RsqrtCanonicalizer.cpp
new file mode 100644
index 000000000..c31dbf6d6
--- /dev/null
+++ b/compiler/moco-tf/src/Canonicalization/RsqrtCanonicalizer.cpp
@@ -0,0 +1,150 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "RsqrtCanonicalizer.h"
+
+#include <moco/IR/TFDialect.h>
+#include <moco/Support/TFShapeInferenceHelper.h>
+
+#include <moco/Log.h>
+
+#include <loco/Service/TypeInference.h>
+
+#include <stdex/Memory.h>
+#include <oops/UserExn.h>
+
+namespace
+{
+
+template <typename T>
+bool prepare_const_gen(loco::ConstGen *const_node, const loco::TensorShape &tensorshape, T value);
+
+template <>
+bool prepare_const_gen<float>(loco::ConstGen *const_node, const loco::TensorShape &tensorshape,
+                              float value)
+{
+  LOGGER(l);
+
+  uint32_t const_num_elements = 1;
+
+  auto dtype = loco::DataType::FLOAT32;
+  const_node->dtype(dtype);
+
+  auto rank = tensorshape.rank();
+  const_node->rank(rank);
+  for (uint32_t r = 0; r < rank; ++r)
+  {
+    if (tensorshape.dim(r).known())
+      const_node->dim(r) = tensorshape.dim(r);
+    else
+      return false;
+
+    assert(tensorshape.dim(r).value() > 0);
+
+    const_num_elements *= tensorshape.dim(r).value();
+  }
+
+  INFO(l) << "prepare_const_gen : Elements = " << const_num_elements;
+
+  const_node->size<loco::DataType::FLOAT32>(const_num_elements);
+  for (uint32_t i = 0; i < const_num_elements; ++i)
+  {
+    const_node->at<loco::DataType::FLOAT32>(i) = value;
+  }
+
+  return true;
+}
+
+bool canonicalize_rsqrt(loco::Graph *graph, moco::TFRsqrt *node)
+{
+  /**
+   * @note This will replace TFRsqrt node with Canonical EltwiseSqrt + EltwiseRealDiv
+   *
+   *       Before
+   *                 A --- TFRsqrt -- C
+   *       After
+   *                    +- TFRsqrt --
+   *                    |
+   *                    |   ConstGen --+
+   *                    |               \
+   *                 A -+- EltwiseSqrt -- EltwiseDiv -- C
+   *
+   *       Where
+   *                 A : features of TFRsqrt
+   *                 C : a node that uses TFSqrt as an input
+   *                 TFRsqrt is disconnected from C
+   *                 TFRsqrt is converted to 1 / EltwiseSqrt
+   */
+
+  auto nodeshape = moco::node_shape(node);
+  if (nodeshape.domain() == loco::Domain::Unknown)
+  {
+    // We need this shape information
+    assert(false); // this shouldn't happen, let's add an alarm
+    return false;
+  }
+  auto tensorshape = nodeshape.as<loco::TensorShape>();
+
+  if (!loco::dtype_known(node))
+  {
+    // We need type of this node
+    return false;
+  }
+
+  auto sqrt_node = graph->nodes()->create<loco::EltwiseSqrt>();
+  auto eltdiv_node = graph->nodes()->create<loco::EltwiseDiv>();
+  auto const_node = graph->nodes()->create<loco::ConstGen>();
+
+  auto dtype = loco::dtype_get(node);
+
+  switch (dtype)
+  {
+    case loco::DataType::FLOAT32:
+      if (!prepare_const_gen<float>(const_node, tensorshape, 1.0f))
+        throw oops::UserExn("Cannot handle unknown shape", node->name());
+      break;
+
+    default:
+      throw oops::UserExn("Unsupported data type", node->name());
+  }
+
+  auto node_A = node->x();
+
+  // update connections
+  sqrt_node->input(node_A);
+  eltdiv_node->lhs(const_node);
+  eltdiv_node->rhs(sqrt_node);
+
+  // replace node
+  replace(node).with(eltdiv_node);
+
+  return true;
+}
+
+} // namespace
+
+namespace moco
+{
+namespace tf
+{
+
+bool RsqrtCanonicalizer::transform(TFRsqrt *node) const
+{
+  return canonicalize_rsqrt(node->graph(), node);
+}
+
+} // namespace tf
+} // namespace moco