Introduce concat kernel (#4887)

Introduce concat kernel from tflite
Use concat kernel in cpu backend

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

3 files changed, 37 insertions, 16 deletions

diff --git a/runtimes/neurun/backend/cpu/kernel/CMakeLists.txt b/runtimes/neurun/backend/cpu/kernel/CMakeLists.txt
index 6617d627e..1f35ce975 100644
--- a/runtimes/neurun/backend/cpu/kernel/CMakeLists.txt
+++ b/runtimes/neurun/backend/cpu/kernel/CMakeLists.txt
@@ -7,7 +7,7 @@ target_include_directories(${LIB_NEURUN_KERNEL_CPU} PUBLIC ${CMAKE_SOURCE_DIR}/e
 
 target_link_libraries(${LIB_NEURUN_KERNEL_CPU} nnapi-header)
 target_link_libraries(${LIB_NEURUN_KERNEL_CPU} tensorflow-lite)
-target_link_libraries(${LIB_NEURUN_KERNEL_CPU} nnfw_lib_misc)
+target_link_libraries(${LIB_NEURUN_KERNEL_CPU} nnfw_lib_misc nnfw_lib_cker)
 
 set_target_properties(${LIB_NEURUN_KERNEL_CPU} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 set_target_properties(${LIB_NEURUN_KERNEL_CPU} PROPERTIES OUTPUT_NAME kernel_cpu)
diff --git a/runtimes/neurun/backend/cpu/kernel/ConcatLayer.cc b/runtimes/neurun/backend/cpu/kernel/ConcatLayer.cc
index dbe15fca6..16fe4eb61 100644
--- a/runtimes/neurun/backend/cpu/kernel/ConcatLayer.cc
+++ b/runtimes/neurun/backend/cpu/kernel/ConcatLayer.cc
@@ -17,9 +17,10 @@
 
 #include "ConcatLayer.h"
 
-#include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h"
 #include "OperationUtils.h"
 
+#include <cker/operation/Concatenation.h>
+
 namespace neurun
 {
 namespace backend
@@ -40,18 +41,18 @@ bool ConcatLayer::concatenationFloat32()
 {
   uint32_t num_inputs = _inputShapes.size();
 
-  tflite::ConcatenationParams op_params;
+  nnfw::cker::ConcatenationParams op_params;
   op_params.axis = _axis;
   op_params.inputs_count = num_inputs;
 
-  std::vector<::tflite::RuntimeShape *> inputDimsPtr;
-  std::vector<::tflite::RuntimeShape> inputDims;
+  std::vector<nnfw::cker::Shape *> inputDimsPtr;
+  std::vector<nnfw::cker::Shape> inputDims;
   inputDimsPtr.reserve(num_inputs);
   inputDims.reserve(num_inputs);
 
   for (uint32_t i = 0; i < num_inputs; i++)
   {
-    inputDims.push_back(convertShapeToTFLiteShape(_inputShapes[i]));
+    inputDims.push_back(convertShapeToCkerShape(_inputShapes[i]));
     inputDimsPtr.push_back(&inputDims[i]);
   }
 
@@ -62,9 +63,8 @@ bool ConcatLayer::concatenationFloat32()
     inputFloatPtrs.emplace_back(reinterpret_cast<const float *>(ptr));
   }
 
-  ::tflite::optimized_ops::Concatenation<float>(
-      op_params, inputDimsPtr.data(), inputFloatPtrs.data(),
-      convertShapeToTFLiteShape(_outputShape), _outputData.f);
+  nnfw::cker::Concatenation<float>(op_params, inputDimsPtr.data(), inputFloatPtrs.data(),
+                                   convertShapeToCkerShape(_outputShape), _outputData.f);
   return true;
 }
 bool ConcatLayer::concatenationQuant8()
@@ -79,7 +79,7 @@ bool ConcatLayer::concatenationQuant8()
     input_scales[i] = _inputShapes[i].scale;
   }
 
-  tflite::ConcatenationParams op_params;
+  nnfw::cker::ConcatenationParams op_params;
   op_params.axis = _axis;
   op_params.inputs_count = num_inputs;
   op_params.input_zeropoint = input_zeropoints.data();
@@ -87,19 +87,18 @@ bool ConcatLayer::concatenationQuant8()
   op_params.output_zeropoint = _outputShape.offset;
   op_params.output_scale = _outputShape.scale;
 
-  std::vector<::tflite::RuntimeShape *> inputDimsPtr;
-  std::vector<::tflite::RuntimeShape> inputDims;
+  std::vector<nnfw::cker::Shape *> inputDimsPtr;
+  std::vector<nnfw::cker::Shape> inputDims;
   inputDimsPtr.reserve(num_inputs);
   inputDims.reserve(num_inputs);
   for (uint32_t i = 0; i < num_inputs; i++)
   {
-    inputDims.push_back(convertShapeToTFLiteShape(_inputShapes[i]));
+    inputDims.push_back(convertShapeToCkerShape(_inputShapes[i]));
     inputDimsPtr.push_back(&inputDims[i]);
   }
 
-  ::tflite::optimized_ops::Concatenation<uint8_t>(
-      op_params, inputDimsPtr.data(), _inputDataPtrs.data(),
-      convertShapeToTFLiteShape(_outputShape), _outputData.u8);
+  nnfw::cker::Concatenation<uint8_t>(op_params, inputDimsPtr.data(), _inputDataPtrs.data(),
+                                     convertShapeToCkerShape(_outputShape), _outputData.u8);
   return true;
 }
 
diff --git a/runtimes/neurun/backend/cpu/kernel/OperationUtils.h b/runtimes/neurun/backend/cpu/kernel/OperationUtils.h
index 0b0649ecd..f38cec072 100644
--- a/runtimes/neurun/backend/cpu/kernel/OperationUtils.h
+++ b/runtimes/neurun/backend/cpu/kernel/OperationUtils.h
@@ -23,6 +23,8 @@
 #include <limits>
 #include <vector>
 
+#include <cker/Shape.h>
+
 #include "tensorflow/contrib/lite/c/builtin_op_data.h"
 #include "tensorflow/contrib/lite/kernels/internal/types.h"
 #include "tensorflow/contrib/lite/kernels/internal/tensor.h"
@@ -107,6 +109,26 @@ inline ::tflite::RuntimeShape convertShapeToTFLiteShape(const Shape &shape)
   return ::tflite::GetTensorShape(raw_shape);
 }
 
+inline nnfw::cker::Shape convertShapeToCkerShape(const Shape &shape)
+{
+  std::vector<int32_t> raw_shape;
+  raw_shape.resize(4);
+
+  for (uint32_t i = 0; i < 4; ++i)
+  {
+    if (i >= shape.dimensions.size())
+    {
+      raw_shape[i] = 1;
+    }
+    else
+    {
+      raw_shape[i] = shape.dimensions[i];
+    }
+  }
+
+  return nnfw::cker::GetShape(raw_shape);
+}
+
 inline TfLiteFusedActivation convertFusedActivation(FuseCode act)
 {
   if (act == ANEURALNETWORKS_FUSED_NONE)