[locomotiv] execution for elementwise add (#6079)

* [locomotiv] execution for elementwise add

This will introduce execution for EltwiseAdd node

Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* validate shape

* remove using namespace

* check if same domain not tensor

diff --git a/compiler/locomotiv/src/Node.lst b/compiler/locomotiv/src/Node.lst
index 3ec1915..2290a88 100644 (file)
--- a/compiler/locomotiv/src/Node.lst
+++ b/compiler/locomotiv/src/Node.lst
@@ -11,6 +11,7 @@ NODE(ConstGen)
 NODE(Conv2D)
 NODE(DepthwiseConv2D)
 NODE(DepthwiseFilterEncode)
+NODE(EltwiseAdd)
 NODE(FeatureDecode)
 NODE(FeatureEncode)
 NODE(FilterEncode)

diff --git a/compiler/locomotiv/src/Node/EltwiseAdd.cpp b/compiler/locomotiv/src/Node/EltwiseAdd.cpp
new file mode 100644 (file)
index 0000000..6af1982
--- /dev/null
+++ b/compiler/locomotiv/src/Node/EltwiseAdd.cpp
@@ -0,0 +1,80 @@
+/*
+ * 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 "NodeExecution.h"
+
+#include "NodeDataImpl.h"
+#include "NodeDomain.h"
+#include "Validation.h"
+
+#include <nncc/core/ADT/tensor/Shape.h>
+#include <nncc/core/ADT/tensor/Buffer.h>
+#include <nncc/core/ADT/tensor/IndexEnumerator.h>
+#include <nncc/core/ADT/tensor/LexicalLayout.h>
+
+using nncc::core::ADT::tensor::IndexEnumerator;
+using nncc::core::ADT::tensor::LexicalLayout;
+using nncc::core::ADT::tensor::make_buffer;
+
+#include <cassert>
+#include <stdexcept>
+
+namespace locomotiv
+{
+
+void NodeExecution::execute(loco::EltwiseAdd *eltwise_add)
+{
+  auto lhs_data = annot_data(eltwise_add->lhs());
+  auto rhs_data = annot_data(eltwise_add->rhs());
+
+  validate(lhs_data && rhs_data, "Input not ready");
+  validate(annot_domain(eltwise_add->lhs()) == annot_domain(eltwise_add->rhs()),
+           "Wrong input domain");
+  validate(lhs_data->dtype() == rhs_data->dtype(), "Wrong input type");
+  validate(*lhs_data->shape() == *rhs_data->shape(), "Wrong input shape");
+
+  std::unique_ptr<NodeData> eltwise_add_data = nullptr;
+
+  switch (lhs_data->dtype())
+  {
+    case loco::DataType::FLOAT32:
+    {
+      auto lhs_bufptr = lhs_data->as_f32_bufptr();
+      auto rhs_bufptr = rhs_data->as_f32_bufptr();
+      auto eltwise_add_bufptr = make_buffer<float, LexicalLayout>(*lhs_data->shape());
+
+      auto *shape = lhs_data->shape();
+
+      for (IndexEnumerator e{*shape}; e.valid(); e.advance())
+      {
+        const auto &index = e.current();
+        eltwise_add_bufptr.at(index) = lhs_bufptr->at(index) + rhs_bufptr->at(index);
+      }
+
+      eltwise_add_data = make_data(eltwise_add_bufptr);
+      break;
+    }
+    default:
+      throw std::runtime_error("NYI for this DataType");
+  }
+
+  assert(eltwise_add_data != nullptr);
+  erase_annot_data(eltwise_add);
+  annot_data(eltwise_add, std::move(eltwise_add_data));
+  annot_domain(eltwise_add, annot_domain(eltwise_add->lhs()));
+}
+
+} // namespace locomotiv

diff --git a/compiler/locomotiv/src/Node/EltwiseAdd.test.cpp b/compiler/locomotiv/src/Node/EltwiseAdd.test.cpp
new file mode 100644 (file)
index 0000000..2899dcc
--- /dev/null
+++ b/compiler/locomotiv/src/Node/EltwiseAdd.test.cpp
@@ -0,0 +1,121 @@
+/*
+ * 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 "NodeExecution.h"
+
+#include "locomotiv/NodeData.h"
+#include "NodeDataImpl.h"
+#include "NodeDomain.h"
+
+#include <nncc/core/ADT/tensor/Shape.h>
+#include <nncc/core/ADT/tensor/Buffer.h>
+#include <nncc/core/ADT/tensor/LexicalLayout.h>
+#include <nncc/core/ADT/tensor/Index.h>
+#include <nncc/core/ADT/tensor/IndexEnumerator.h>
+
+#include <gtest/gtest.h>
+
+using nncc::core::ADT::tensor::Shape;
+using nncc::core::ADT::tensor::LexicalLayout;
+using nncc::core::ADT::tensor::make_buffer;
+using nncc::core::ADT::tensor::IndexEnumerator;
+
+/*
+test case generated from the following:
+
+x = tf.constant([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18],
+                shape=[1, 3, 3, 2], dtype=tf.float32)
+y = tf.constant([-1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16, -17, -18],
+                shape=[1, 3, 3, 2], dtype=tf.float32)
+out = tf.math.add(x, y)
+
+with tf.Session() as sess:
+    print(sess.run(out))
+*/
+TEST(NodeExecution_EltwiseAdd, f32)
+{
+  float x_val[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18};
+  float y_val[] = {-1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16, -17, -18};
+  float out_val[] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+
+  // make EltwiseAdd(Pull, Pull)
+  auto g = loco::make_graph();
+  Shape input_shape{1, 3, 3, 2}; // NHWC
+
+  auto inp_lhs = g->nodes()->create<loco::Pull>();
+  {
+    inp_lhs->dtype(loco::DataType::FLOAT32);
+    inp_lhs->shape({1, 3, 3, 2});
+  }
+
+  auto inp_rhs = g->nodes()->create<loco::Pull>();
+  {
+    inp_rhs->dtype(loco::DataType::FLOAT32);
+    inp_rhs->shape({1, 3, 3, 2});
+  }
+
+  auto eltwise_add = g->nodes()->create<loco::EltwiseAdd>();
+  {
+    eltwise_add->lhs(inp_lhs);
+    eltwise_add->rhs(inp_rhs);
+  }
+
+  // Make and assign data to two pull nodes
+  auto inp_lhs_buf = make_buffer<float, LexicalLayout>(input_shape);
+  {
+    int n = 0;
+    for (IndexEnumerator e{inp_lhs_buf.shape()}; e.valid(); e.advance())
+    {
+      inp_lhs_buf.at(e.current()) = x_val[n++];
+    }
+  }
+
+  auto inp_rhs_buf = make_buffer<float, LexicalLayout>(input_shape);
+  {
+    int n = 0;
+    for (IndexEnumerator e{inp_rhs_buf.shape()}; e.valid(); e.advance())
+    {
+      inp_rhs_buf.at(e.current()) = y_val[n++];
+    }
+  }
+
+  auto inp_lhs_data = locomotiv::make_data(inp_lhs_buf);
+  locomotiv::annot_data(inp_lhs, std::move(inp_lhs_data));
+  locomotiv::annot_domain(inp_lhs, loco::Domain::Tensor);
+
+  auto inp_rhs_data = locomotiv::make_data(inp_rhs_buf);
+  locomotiv::annot_data(inp_rhs, std::move(inp_rhs_data));
+  locomotiv::annot_domain(inp_rhs, loco::Domain::Tensor);
+
+  // run the network
+  locomotiv::NodeExecution::get().run(eltwise_add);
+
+  // get result
+  auto eltwise_add_data = locomotiv::annot_data(eltwise_add);
+
+  // comparing the result
+  ASSERT_NE(eltwise_add_data, nullptr);
+  ASSERT_EQ(eltwise_add_data->dtype(), loco::DataType::FLOAT32);
+  ASSERT_EQ(*(eltwise_add_data->shape()), Shape({1, 3, 3, 2}));
+
+  uint32_t n = 0;
+  for (IndexEnumerator e{*(eltwise_add_data->shape())}; e.valid(); e.advance())
+  {
+    ASSERT_FLOAT_EQ(eltwise_add_data->as_f32_bufptr()->at(e.current()), out_val[n++]);
+  }
+
+  ASSERT_EQ(locomotiv::annot_domain(eltwise_add), loco::Domain::Tensor);
+}