From: 오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 <hseok82.oh@samsung.com>
Date: Mon, 17 Jun 2019 10:15:56 +0000 (+0900)
Subject: Remove axis restriction on concat elimination (#5414)
X-Git-Tag: submit/tizen/20190809.050447~670
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=06f98b60da87ff6dd24538ac60ccbbaf7c096378;p=platform%2Fcore%2Fml%2Fnnfw.git

Remove axis restriction on concat elimination (#5414)

* Remove axis restriction on concat elimination

Remove axis restruction that is allowed channel axis only on concat elimination

Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>

* Remove old comment
---

diff --git a/runtimes/neurun/backend/acl_common/TemplTensorBuilder.h b/runtimes/neurun/backend/acl_common/TemplTensorBuilder.h
index decc5ae..8d96aad 100644
--- a/runtimes/neurun/backend/acl_common/TemplTensorBuilder.h
+++ b/runtimes/neurun/backend/acl_common/TemplTensorBuilder.h
@@ -221,17 +221,7 @@ void TemplTensorBuilder<T_ITensor, T_Tensor, T_SubTensor, T_Object>::prepare(voi
       assert(asDataType(info.type().type()) == parent_tensor->info()->data_type());
       // TODO Change to set data_layout for each front-end
       auto shape = asTensorShape(info.shape(), _layout, _apply_dim_correction_map[current]);
-
-      // Only support axis: 3 (channel)
-      ::arm_compute::Coordinates coordinates;
-      coordinates.set_num_dimensions(4);
-      assert(info.offset()[0] == 0);
-      assert(info.offset()[1] == 0);
-      assert(info.offset()[2] == 0);
-      // TODO Change to set data_layout for each front-end
-      auto channel_index = get_data_layout_dimension_index(
-          parent_tensor->info()->data_layout(), ::arm_compute::DataLayoutDimension::CHANNEL);
-      coordinates[channel_index] = info.offset()[3];
+      ::arm_compute::Coordinates coordinates = asTensorCoordinate(info.offset(), _layout);
       auto tensor = std::make_shared<T_SubTensor>(parent_tensor.get(), shape, coordinates, true);
       _subtensors[current] = tensor;
       stack.pop();
diff --git a/runtimes/neurun/core/src/compiler/SubTensorAnalyzer.cc b/runtimes/neurun/core/src/compiler/SubTensorAnalyzer.cc
index 5859adc..efbd558 100644
--- a/runtimes/neurun/core/src/compiler/SubTensorAnalyzer.cc
+++ b/runtimes/neurun/core/src/compiler/SubTensorAnalyzer.cc
@@ -21,6 +21,7 @@
 #include "cpp14/memory.h"
 #include "model/OperandIndexSequence.h"
 #include "util/logging.h"
+#include "util/Coordinates.h"
 
 namespace neurun
 {
@@ -42,25 +43,28 @@ void SubTensorAnalyzer::visit(const model::operation::ConcatNode &node)
     return;
   }
 
-  // NOTE This implementation assumes concat over feature depth
-  // TODO Remove this assumption
-  int32_t axis = _ctx.at(axis_index).asScalar<int32_t>();
-  if (axis != 3)
-  {
-    VERBOSE(SUBTENSOR) << "Cannot handle axis is not channel" << std::endl;
-    return;
-  }
+  int32_t axis_raw = _ctx.at(axis_index).asScalar<int32_t>();
 
   auto &output_index = node.getOutputs().at(0);
   auto &inputs = node.getInputs();
 
   int32_t axis_point = 0;
+  const auto rank = _ctx.at(output_index).shape().rank();
+  uint32_t axis = axis_raw < 0 ? (axis_raw + rank) : axis_raw;
+  assert(rank > axis);
+
   for (auto &input_index : inputs)
   {
     auto input_shape = _ctx.at(input_index).shape();
-    std::vector<int32_t> offset = {0, 0, 0, 0};
-    offset[axis] = axis_point;
-    neurun::util::Coordinates coordinate_info({offset[0], offset[1], offset[2], offset[3]});
+    assert(rank == input_shape.rank());
+
+    neurun::util::Coordinates coordinate_info{};
+    for (uint32_t i = 0; i < rank; i++)
+    {
+      coordinate_info.set(i, 0);
+    }
+    coordinate_info.set(axis, axis_point);
+
     std::unique_ptr<graph::operand::ParentInfo> parentInfo =
         nnfw::cpp14::make_unique<graph::operand::ParentInfo>(output_index, coordinate_info);