From: Тимур Отеллович Аблязимов/AI Tools Lab /SRR/Staff Engineer/삼성전자 <t.ablyazimov@samsung.com>
Date: Wed, 12 Dec 2018 17:15:02 +0000 (+0300)
Subject: [nnc] Acl backend system tests for Mobilenet and Inception-v3 operations (#2229)
X-Git-Tag: nncc_backup~1095
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=6270920fd4673af1749d36b3eaf75e1a2cc73b8c;p=platform%2Fcore%2Fml%2Fnnfw.git

[nnc] Acl backend system tests for Mobilenet and Inception-v3 operations (#2229)

The system tests framework for the ACL soft backend has been created. The test cases for the operations used in the Mobilenet and the Inception-v3 operations have been added.

Signed-off-by: Timur Ablyazimov <t.ablyazimov@samsung.com>
---

diff --git a/contrib/nnc/tests/CMakeLists.txt b/contrib/nnc/tests/CMakeLists.txt
index 60bc260..2bf6d7e 100644
--- a/contrib/nnc/tests/CMakeLists.txt
+++ b/contrib/nnc/tests/CMakeLists.txt
@@ -1,3 +1,4 @@
 add_subdirectory(import)
 add_subdirectory(soft_backend)
 add_subdirectory(interpreter)
+add_subdirectory(acl_soft_backend)
diff --git a/contrib/nnc/tests/acl_soft_backend/AclCppOperations.cpp b/contrib/nnc/tests/acl_soft_backend/AclCppOperations.cpp
new file mode 100644
index 0000000..54c2c65
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/AclCppOperations.cpp
@@ -0,0 +1,183 @@
+#include "gtest/gtest.h"
+#include <sstream>
+#include <thread>
+#include <cmath>
+#include <memory>
+#include <H5Cpp.h>
+#include <stdlib.h>
+#include "BuildInfo.h"
+
+using namespace std;
+
+static string netAddr(getenv("ODROID_NET_ADDR") ? getenv("ODROID_NET_ADDR") : "");
+
+static unique_ptr<char[]> readTensorDataFromHdf5File(const string& file_name, vector<int>& shape) {
+  try {
+    H5::H5File h5File(file_name, H5F_ACC_RDONLY);
+    auto tensor_name = h5File.getObjnameByIdx(0);
+    auto dataset = h5File.openDataSet(tensor_name);
+    auto dataspace = dataset.getSpace();
+    auto rank = dataspace.getSimpleExtentNdims();
+
+    if (rank < 2)
+      return nullptr;
+
+    hsize_t dims[rank];
+
+    if (dataspace.getSimpleExtentDims(dims) != rank)
+      return nullptr;
+
+    int size = 1;
+
+    for (int i = 0; i < rank; ++i) {
+      size *= dims[i];
+      shape.push_back(dims[i]);
+    }
+
+    auto result = unique_ptr<char[]>(new char[size * sizeof(float)]);
+    dataset.read(&result[0], H5::PredType::NATIVE_FLOAT);
+    return result;
+  } catch (H5::FileIException&) {
+    return nullptr;
+  }
+}
+
+// TODO: this function was copied from CPPOperations.cpp, move it to a shared place.
+bool areFloatsNear(float a, float b, int32_t ulp, float eps) {
+  assert(ulp < (1 << 23) && "this algorithm is not applicable for such large diffs");
+  assert(eps >= 0 && "epsilon should be positive number");
+  if (fabs(a - b) <= eps)
+    return true;
+  // since this point need to dind difference between numbers
+  // in terms of ULP
+  int32_t ai;
+  int32_t bi;
+  memcpy(&ai, &a, sizeof(float));
+  memcpy(&bi, &b, sizeof(float));
+  // compare mantissa of numbers
+  if (ai > bi)
+    return ai - bi <= ulp;
+  return bi - ai <= ulp;
+}
+
+static void compareHdf5Files(const string& file_name1, const string& file_name2) {
+  vector<int> shape1;
+  auto tensor1 = readTensorDataFromHdf5File(file_name1, shape1);
+  float* tensorData1 = reinterpret_cast<float*>(&tensor1[0]);
+  ASSERT_NE(tensorData1, nullptr);
+  vector<int> shape2;
+  auto tensor2 = readTensorDataFromHdf5File(file_name2, shape2);
+  float* tensorData2 = reinterpret_cast<float*>(&tensor2[0]);
+  ASSERT_NE(tensorData2, nullptr);
+  ASSERT_EQ(shape1.size(), shape2.size());
+  int size = 1;
+
+  for (int i = 0; i < shape1.size(); ++i) {
+    ASSERT_EQ(shape1[i], shape2[i]);
+    size *= shape1[i];
+  }
+
+  for (int i = 0; i < size; ++i) {
+    ASSERT_TRUE(areFloatsNear(tensorData1[i], tensorData2[i], 32, 1e-6));
+  }
+}
+
+static string genTmpDirName() {
+  string result("/tmp/nnc_test_");
+  stringstream ss;
+  ss << this_thread::get_id();
+  result += ss.str();
+
+  return result;
+}
+
+static bool runOnOdroid(const string& remote_cmd) {
+  string cmd = "ssh " + netAddr + " \"" + remote_cmd + "\"";
+  return system(cmd.c_str()) == 0;
+}
+
+static bool copyToOdroid(const string& src, const string& dst) {
+  string cmd("scp -q " + src + " " + netAddr + ":" + dst);
+  return system(cmd.c_str()) == 0;
+}
+
+static bool copyFromOdroid(const string& src, const string& dst) {
+  string cmd("scp -q " + netAddr + ":" + src + " " + dst);
+  return system(cmd.c_str()) == 0;
+}
+
+static void runAclSystemTest(const string& name) {
+  // Ensure the Odroid device net address was set.
+  ASSERT_TRUE(!netAddr.empty());
+
+  // The name of the temporary directory which is generated on the remote device.
+  string dir_name = genTmpDirName();
+
+  // Insure there is no such the directory on the remote device.
+  ASSERT_TRUE(runOnOdroid("rm -rf " + dir_name));
+
+  // Create the temporary directory on the remote device.
+  ASSERT_TRUE(runOnOdroid("mkdir " + dir_name));
+
+  // Copy the executable artifact file to the remote device.
+  ASSERT_TRUE(copyToOdroid(binDir + "/" + name + "/nnc_test", dir_name));
+
+  // Copy the artifact parameter file to the remote device.
+  ASSERT_TRUE(copyToOdroid(binDir + "/" + name + "/AclArtifact.par", dir_name));
+
+  // Copy the model input HDF5 file to the remote device.
+  ASSERT_TRUE(copyToOdroid(binDir + "/" + name + "/in_" + name + "_caffe.hdf5",
+                           dir_name + "/in.hdf5"));
+
+  // Switch to the artifact directory on the remote device and run the artifact.
+  ASSERT_TRUE(runOnOdroid("cd " + dir_name + "; ./nnc_test"));
+
+  // Copy the resulting file from the remote device to the host.
+  ASSERT_TRUE(copyFromOdroid(dir_name + "/out.hdf5", binDir + "/" + name));
+
+  // Remove the temporary test case directory from the remote device.
+  ASSERT_TRUE(runOnOdroid("rm -rf " + dir_name));
+
+  // Compare the resulting HDF5 file with the reference one.
+  compareHdf5Files(binDir + "/" + name + "/ref.hdf5", binDir + "/" + name + "/out.hdf5");
+}
+
+TEST(acl_cpp_operations_test, convolution) {
+  runAclSystemTest("convolution");
+}
+
+TEST(acl_cpp_operations_test, depthwise_convolution) {
+  runAclSystemTest("depthwise_convolution");
+}
+
+TEST(acl_cpp_operations_test, convolution_with_bias) {
+  runAclSystemTest("convolution_with_bias");
+}
+
+TEST(acl_cpp_operations_test, scale) {
+  runAclSystemTest("scale");
+}
+
+TEST(acl_cpp_operations_test, relu) {
+  runAclSystemTest("relu");
+}
+
+TEST(acl_cpp_operations_test, pooling_max) {
+  runAclSystemTest("pooling_max");
+}
+
+TEST(acl_cpp_operations_test, pooling_avg) {
+  runAclSystemTest("pooling_avg");
+}
+
+TEST(acl_cpp_operations_test, concatenate) {
+  runAclSystemTest("concatenate");
+}
+
+TEST(acl_cpp_operations_test, reshape) {
+  runAclSystemTest("reshape");
+}
+
+TEST(acl_cpp_operations_test, fully_connected) {
+  runAclSystemTest("fully_connected");
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/BuildInfo.h.in b/contrib/nnc/tests/acl_soft_backend/BuildInfo.h.in
new file mode 100644
index 0000000..c229a4f
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/BuildInfo.h.in
@@ -0,0 +1,6 @@
+#ifndef _NNC_BUILD_INFO_H_IN_H_
+#define _NNC_BUILD_INFO_H_IN_H_
+
+static std::string binDir = "${CMAKE_CURRENT_BINARY_DIR}";
+
+#endif //_NNC_BUILD_INFO_H_IN_H_
diff --git a/contrib/nnc/tests/acl_soft_backend/CMakeLists.txt b/contrib/nnc/tests/acl_soft_backend/CMakeLists.txt
new file mode 100644
index 0000000..da31ad5
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/CMakeLists.txt
@@ -0,0 +1,133 @@
+function(acl_warn MESSAGE)
+  message(WARNING "The ACL backend system tests will not be built: ${MESSAGE}")
+endfunction(acl_warn)
+
+if(NOT TARGET caffegen)
+  acl_warn("the CAFFEGEN was not built.")
+  return()
+endif(NOT TARGET caffegen)
+
+if(NOT TARGET nni)
+  acl_warn("the NNKIT was not built.")
+  return()
+endif(NOT TARGET nni)
+
+# Path to the folder where the Odroid root folder is either mounted or copied.
+if(NOT DEFINED ENV{ODROID_MIRROR_DIR})
+  acl_warn("the ODROID_MIRROR_DIR environment variable was not defined.\n\
+    As a cross-build is done, this variable should point to a directory\n\
+    which is either mounted to (with e.g. SSHFS) or contains a copy of the target\n\
+    (e.g. Odroid XU4) device file system.")
+  return()
+endif()
+
+# Path to the ACL library root on the Odroid device.
+if(NOT DEFINED ENV{ODROID_ACL_DIR})
+  acl_warn("the ODROID_ACL_DIR environment variable was not defined.\n\
+    As a cross-build is done, this variable should contain the path to the root directory\n\
+    of the Arm Compute Library on the target (e.g. Odroid XU4) device.")
+  return()
+endif()
+
+# Path to the HDF5 library on the Odroid.
+# It must most likely be: /usr/lib/arm-linux-gnueabihf/hdf5/serial
+if(NOT DEFINED ENV{ODROID_H5_DIR})
+  acl_warn("the ODROID_H5_DIR environment variable was not defined.\n\
+    As a cross-build is done, this variable should contain the path to the root directory\n\
+    of the HDF5 library on the target (e.g. Odroid XU4) device.\n\
+    Often it would be the: /usr/lib/arm-linux-gnueabihf/hdf5/serial directory.")
+  return()
+endif()
+
+find_package(HDF5 COMPONENTS CXX REQUIRED)
+
+# Provide the test suite with the information where to locate executalbes to run etc.
+configure_file(BuildInfo.h.in BuildInfo.h)
+
+add_nncc_test(nnc_acl_soft_backend_system_test AclCppOperations.cpp)
+target_include_directories(nnc_acl_soft_backend_system_test PRIVATE ${CMAKE_CURRENT_BINARY_DIR}
+    ${HDF5_INCLUDE_DIRS})
+target_link_libraries(nnc_acl_soft_backend_system_test ${HDF5_CXX_LIBRARIES})
+
+file(GLOB MODELS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} "models/*.prototxt")
+
+# Loop over all the existing system test models.
+foreach(MODEL IN ITEMS ${MODELS})
+  # Set the model-related variables used inside the loop.
+  set(MODEL_FILE ${CMAKE_CURRENT_SOURCE_DIR}/${MODEL})
+  get_filename_component(MODEL_NAME ${MODEL} NAME_WE)
+  set(MODEL_DIR ${CMAKE_CURRENT_BINARY_DIR}/${MODEL_NAME})
+  set(COPIED_MODEL_FILE ${MODEL_DIR}/${MODEL_NAME}.prototxt)
+  set(INITIALIZED_MODEL_FILE ${MODEL_DIR}/${MODEL_NAME}.prototxt.weights)
+  set(ENCODED_MODEL_FILE ${MODEL_DIR}/${MODEL_NAME}.caffemodel)
+  set(INPUT_FILE ${MODEL_DIR}/in_${MODEL_NAME}_caffe.hdf5)
+  set(REFERENCE_OUTPUT ${MODEL_DIR}/ref.hdf5)
+  set(MODEL_DIR_TARGET nnc_acl_soft_backend_${MODEL_NAME}_dir)
+  set(COPIED_MODEL_TARGET nnc_acl_soft_backend_${MODEL_NAME}_prototxt)
+  set(INITIALIZED_MODEL_TARGET nnc_acl_soft_backend_${MODEL_NAME}_weights)
+  set(ENCODED_MODEL_TARGET nnc_acl_soft_backend_${MODEL_NAME}_caffemodel)
+  set(INPUT_TARGET nnc_acl_soft_backend_${MODEL_NAME}_input)
+  set(REFERENCE_OUTPUT_TARGET nnc_acl_soft_backend_${MODEL_NAME}_reference_output)
+  set(MAIN_FILE_TARGET nnc_acl_soft_backend_${MODEL_NAME}_main)
+  set(ARTIFACT_TARGET_CMAKE nnc_acl_soft_backend_${MODEL_NAME}_cmake)
+  set(ARTIFACT_TARGET_TOOLCHAIN nnc_acl_soft_backend_${MODEL_NAME}_toolchain)
+  set(ARTIFACT_TARGET nnc_acl_soft_backend_${MODEL_NAME}_artifact)
+  set(ARTIFACT_TARGET_MAKE nnc_acl_soft_backend_${MODEL_NAME}_artifact_make)
+  set(ARTIFACT_TARGET_BINARY nnc_acl_soft_backend_${MODEL_NAME}_artifact_binary)
+
+  # Create a directrory for handling the model in the binary directory.
+  add_custom_target(${MODEL_DIR_TARGET} ALL ${CMAKE_COMMAND} -E make_directory ${MODEL_DIR})
+
+  # Copy the model prototxt to the model build directory.
+  add_custom_target(${COPIED_MODEL_TARGET} ALL ${CMAKE_COMMAND} -E copy ${MODEL_FILE} ${COPIED_MODEL_FILE})
+
+  # Copy the artifact main and project files into the model subfolder inside the binary directory.
+  add_custom_target(${MAIN_FILE_TARGET} ALL ${CMAKE_COMMAND}
+      -E copy ${CMAKE_CURRENT_SOURCE_DIR}/artifact_cmake/main.cpp ${MODEL_DIR}/main.cpp)
+
+  # Copy the artifact CMakeLists.txt to the artifact build directory.
+  add_custom_target(${ARTIFACT_TARGET_CMAKE} ALL ${CMAKE_COMMAND}
+      -E copy ${CMAKE_CURRENT_SOURCE_DIR}/artifact_cmake/CMakeLists.txt ${MODEL_DIR}/CMakeLists.txt)
+
+  # Copy the artifact toolchain file to the artifact build directory.
+  add_custom_target(${ARTIFACT_TARGET_TOOLCHAIN} ALL ${CMAKE_COMMAND}
+      -E copy ${CMAKE_CURRENT_SOURCE_DIR}/artifact_cmake/odroid.cmake ${MODEL_DIR}/odroid.cmake)
+
+  # Initialize the model with weights.
+  add_custom_target(${INITIALIZED_MODEL_TARGET} ALL cat ${COPIED_MODEL_FILE} | GLOG_minloglevel=2
+      $<TARGET_FILE:caffegen> init > ${INITIALIZED_MODEL_FILE}
+      DEPENDS ${COPIED_MODEL_TARGET})
+
+  # Encode the model.
+  add_custom_target(${ENCODED_MODEL_TARGET} ALL cat ${INITIALIZED_MODEL_FILE} | GLOG_minloglevel=2
+      $<TARGET_FILE:caffegen> encode > ${ENCODED_MODEL_FILE}
+      DEPENDS ${INITIALIZED_MODEL_TARGET})
+
+  add_custom_target(${INPUT_TARGET} ALL $<TARGET_FILE:tensor_gen> data ${MODEL_NAME} 1 3 4 8
+      WORKING_DIRECTORY ${MODEL_DIR})
+
+    message("INPUT_FILE = ${INPUT_FILE}")
+
+  # Generate the 'reference' output with NNKIT.
+  add_custom_target(${REFERENCE_OUTPUT_TARGET} ALL $<TARGET_FILE:nni>
+      --backend $<TARGET_FILE:nnkit_caffe_backend> --backend-arg ${INITIALIZED_MODEL_FILE}
+      --pre $<TARGET_FILE:nnkit_HDF5_import_action> --pre-arg ${INPUT_FILE}
+      --post $<TARGET_FILE:nnkit_HDF5_export_action> --post-arg ${REFERENCE_OUTPUT}
+      DEPENDS $<TARGET_FILE:nni> ${INPUT_TARGET} ${INITIALIZED_MODEL_TARGET})
+
+  # Generate an artifact from the model.
+  add_custom_target(${ARTIFACT_TARGET} ALL $<TARGET_FILE:nnc>
+      --caffe -m ${ENCODED_MODEL_FILE} -o AclArtifact -d ${MODEL_DIR} --target=arm-gpu-c++
+      DEPENDS $<TARGET_FILE:nnc> ${ENCODED_MODEL_FILE})
+
+  # Generate a Makefile for the artifact cross-building.
+  add_custom_target(${ARTIFACT_TARGET_MAKE} ALL cmake .
+      -DCMAKE_TOOLCHAIN_FILE=odroid.cmake
+      DEPENDS ${ARTIFACT_TARGET_CMAKE} ${ARTIFACT_TARGET_TOOLCHAIN} ${ARTIFACT_TARGET}
+      WORKING_DIRECTORY ${MODEL_DIR})
+
+  # Cross-build the artifact with the generated Makefile.
+  add_custom_target(${ARTIFACT_TARGET_BINARY} ALL make
+      DEPENDS ${ARTIFACT_TARGET_MAKE}
+      WORKING_DIRECTORY ${MODEL_DIR})
+endforeach(MODEL)
diff --git a/contrib/nnc/tests/acl_soft_backend/artifact_cmake/CMakeLists.txt b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/CMakeLists.txt
new file mode 100644
index 0000000..ceecded
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/CMakeLists.txt
@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.5)
+project(nnc_test)
+
+set(CMAKE_CXX_STANDARD 11)
+
+set(ODROID_MIRROR_DIR $ENV{ODROID_MIRROR_DIR})
+set(ODROID_ACL_DIR $ENV{ODROID_ACL_DIR})
+set(ODROID_ACL_INC_DIR ${ODROID_ACL_DIR}/include)
+set(ODROID_ACL_BUILD_DIR ${ODROID_ACL_DIR}/build)
+
+find_library(OPEN_CL OpenCL /usr/lib/arm-linux-gnueabihf)
+find_library(ARM_COMPUTE arm_compute PATHS ${ODROID_ACL_BUILD_DIR})
+find_library(ARM_COMPUTE_CORE arm_compute_core PATHS ${ODROID_ACL_BUILD_DIR})
+find_package(HDF5 COMPONENTS CXX REQUIRED)
+
+add_executable(nnc_test main.cpp AclArtifact.cpp)
+
+target_include_directories(nnc_test PRIVATE . ${ODROID_MIRROR_DIR}${ODROID_ACL_INC_DIR}
+    ${ODROID_MIRROR_DIR}${ODROID_ACL_DIR} ${ODROID_MIRROR_DIR}${HDF5_INCLUDE_DIRS})
+target_link_libraries(nnc_test ${ARM_COMPUTE} ${ARM_COMPUTE_CORE} ${OPEN_CL} ${HDF5_CXX_LIBRARIES})
+target_compile_definitions(nnc_test PRIVATE ARM_COMPUTE_CL)
diff --git a/contrib/nnc/tests/acl_soft_backend/artifact_cmake/main.cpp b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/main.cpp
new file mode 100644
index 0000000..03adfa2
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/main.cpp
@@ -0,0 +1,110 @@
+#include "AclArtifact.h"
+#include <iostream>
+#include <memory>
+#include <H5Cpp.h>
+
+using namespace std;
+using namespace arm_compute;
+
+static unique_ptr<char[]> getTensorData(CLTensor& tensor) {
+  auto buf = unique_ptr<char[]>(new char[tensor.info()->total_size()]);
+  tensor.map();
+  Window window;
+  window.use_tensor_dimensions(tensor.info()->tensor_shape());
+  Iterator i(&tensor, window);
+  char* ptr = &buf[0];
+
+  execute_window_loop(window, [&i, &ptr](const Coordinates&) {
+    memcpy(ptr, i.ptr(), sizeof(float));
+    ptr += sizeof(float);
+  }, i);
+
+  tensor.unmap();
+  return buf;
+}
+
+static void readTensor(CLTensor& tensor, H5::DataSet& dataset) {
+  auto buf = unique_ptr<char[]>(new char[tensor.info()->total_size()]);
+  dataset.read(&buf[0], H5::PredType::NATIVE_FLOAT);
+  tensor.map();
+  Window window;
+  window.use_tensor_dimensions(tensor.info()->tensor_shape());
+  Iterator i(&tensor, window);
+  char* ptr = &buf[0];
+
+  execute_window_loop(window, [&i, &ptr](const Coordinates&) {
+    memcpy(i.ptr(), ptr, sizeof(float));
+    ptr += sizeof(float);
+  }, i);
+
+  tensor.unmap();
+}
+
+static bool readTensorFromHDF5File(CLTensor& tensor, const string& file_name) {
+  // Read from the .hdf5 file
+  try {
+    H5::H5File h5File(file_name, H5F_ACC_RDONLY);
+    auto tensor_name = h5File.getObjnameByIdx(0);
+    auto dataset = h5File.openDataSet(tensor_name);
+    auto dataspace = dataset.getSpace();
+    auto rank = dataspace.getSimpleExtentNdims();
+
+    if (rank < 2)
+      return false;
+
+    hsize_t dims[rank];
+
+    if (dataspace.getSimpleExtentDims(dims) != rank)
+      return false;
+
+    TensorShape shape;
+    shape.set_num_dimensions(rank - 1);
+
+    for (int i = 1; i < rank; ++i)
+      shape[rank - i - 1] = dims[i];
+
+    readTensor(tensor, dataset);
+  } catch (H5::FileIException&) {
+    return false;
+  }
+
+  return true;
+}
+
+static void writeTensorToHDF5File(CLTensor& tensor, const string& tensor_name,
+                                  const string& file_name) {
+  const TensorShape& orig_shape = tensor.info()->tensor_shape();
+  const TensorShape& transposed_shape = orig_shape;
+  int rank = transposed_shape.num_dimensions();
+  hsize_t dims[rank + 1];
+  dims[0] = 1;
+
+  for (int i = 0; i < rank; ++i)
+    dims[rank - i] = transposed_shape[i];
+
+  // Write to the .hdf5 file
+  H5::H5File h5File(file_name, H5F_ACC_TRUNC);
+  H5::DataSpace dataspace(rank + 1, dims);
+  auto dataset = h5File.createDataSet(tensor_name, H5::PredType::IEEE_F32BE, dataspace);
+  dataset.write(&getTensorData(tensor)[0], H5::PredType::NATIVE_FLOAT);
+}
+
+int main(int argc, char* argv[]) {
+  CLScheduler::get().default_init();
+
+  if (!CLScheduler::get().is_initialised()) {
+    cout << "Failed to initialise the ACL scheduler" << endl;
+    return 1;
+  }
+
+  AclArtifact artifact;
+  CLTensor& artifact_in = artifact.getInput();
+  readTensorFromHDF5File(artifact_in, "in.hdf5");
+
+  artifact.Inference();
+
+  CLTensor& artifact_out = artifact.getOutput();
+  writeTensorToHDF5File(artifact_out, "out", "out.hdf5");
+
+  return 0;
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/artifact_cmake/odroid.cmake b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/odroid.cmake
new file mode 100644
index 0000000..92faea0
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/artifact_cmake/odroid.cmake
@@ -0,0 +1,17 @@
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR arm)
+
+set(CMAKE_C_COMPILER /usr/bin/arm-linux-gnueabihf-gcc)
+set(CMAKE_CXX_COMPILER /usr/bin/arm-linux-gnueabihf-g++)
+
+set(ODROID_MIRROR_DIR $ENV{ODROID_MIRROR_DIR})
+set(ODROID_H5_DIR $ENV{ODROID_H5_DIR})
+
+set(CMAKE_FIND_ROOT_PATH ${ODROID_MIRROR_DIR}${ODROID_H5_DIR})
+set(CMAKE_SYSROOT ${ODROID_MIRROR_DIR})
+set(CMAKE_PREFIX_PATH ${ODROID_MIRROR_DIR})
+
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
diff --git a/contrib/nnc/tests/acl_soft_backend/models/concatenate.prototxt b/contrib/nnc/tests/acl_soft_backend/models/concatenate.prototxt
new file mode 100644
index 0000000..78d0f8f
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/concatenate.prototxt
@@ -0,0 +1,23 @@
+name: "CONCANET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "inception_a1_output"
+  type: "Concat"
+  bottom: "data"
+  bottom: "data"
+  bottom: "data"
+  top: "inception_a1_output"
+  phase: TEST
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/convolution.prototxt b/contrib/nnc/tests/acl_soft_backend/models/convolution.prototxt
new file mode 100644
index 0000000..176940c
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/convolution.prototxt
@@ -0,0 +1,34 @@
+name: "CONVONET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "conv1"
+  type: "Convolution"
+  bottom: "data"
+  top: "conv1"
+  param {
+    lr_mult: 1
+    decay_mult: 1
+  }
+  convolution_param {
+    num_output: 3
+    bias_term: false
+    pad: 1
+    kernel_size: 3
+    stride: 2
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/convolution_with_bias.prototxt b/contrib/nnc/tests/acl_soft_backend/models/convolution_with_bias.prototxt
new file mode 100644
index 0000000..804bac1
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/convolution_with_bias.prototxt
@@ -0,0 +1,38 @@
+name: "CONVONET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "conv1"
+  type: "Convolution"
+  bottom: "data"
+  top: "conv1"
+  param {
+    lr_mult: 1
+    decay_mult: 1
+  }
+  convolution_param {
+    num_output: 3
+    bias_term: true
+    pad: 1
+    kernel_size: 3
+    stride: 2
+    weight_filler {
+      type: "msra"
+    }
+    bias_filler {
+      type: "constant"
+      value: 3.14
+    }
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/depthwise_convolution.prototxt b/contrib/nnc/tests/acl_soft_backend/models/depthwise_convolution.prototxt
new file mode 100644
index 0000000..b2b1ce5
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/depthwise_convolution.prototxt
@@ -0,0 +1,35 @@
+name: "DEPTHCONVNET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "conv2_1/dw"
+  type: "Convolution"
+  bottom: "data"
+  top: "conv2_1/dw"
+  param {
+    lr_mult: 1
+    decay_mult: 1
+  }
+  convolution_param {
+    num_output: 3
+    bias_term: false
+    pad: 1
+    kernel_size: 3
+    group: 3
+    stride: 1
+    weight_filler {
+      type: "msra"
+    }
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/fully_connected.prototxt b/contrib/nnc/tests/acl_soft_backend/models/fully_connected.prototxt
new file mode 100644
index 0000000..beca801
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/fully_connected.prototxt
@@ -0,0 +1,29 @@
+name: "FULLYCONNET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "fc8"
+  type: "InnerProduct"
+  param { lr_mult: 1 decay_mult: 1 }
+  inner_product_param {
+    num_output: 10
+    weight_filler {
+      type: "gaussian"
+      std: 0.01
+    }
+    bias_term: false
+  }
+  bottom: "data"
+  top: "fc8"
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/pooling_avg.prototxt b/contrib/nnc/tests/acl_soft_backend/models/pooling_avg.prototxt
new file mode 100644
index 0000000..ad93a6e
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/pooling_avg.prototxt
@@ -0,0 +1,26 @@
+name: "POOLINGAVGNET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "inception_a1_pool"
+  type: "Pooling"
+  bottom: "data"
+  top: "inception_a1_pool"
+  pooling_param {
+    pool: AVE
+    kernel_size: 3
+    stride: 1
+    pad: 1
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/pooling_max.prototxt b/contrib/nnc/tests/acl_soft_backend/models/pooling_max.prototxt
new file mode 100644
index 0000000..7ac060a
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/pooling_max.prototxt
@@ -0,0 +1,26 @@
+name: "POOLINGMAXNET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "inception_a1_pool"
+  type: "Pooling"
+  bottom: "data"
+  top: "inception_a1_pool"
+  pooling_param {
+    pool: MAX
+    kernel_size: 3
+    stride: 1
+    pad: 1
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/relu.prototxt b/contrib/nnc/tests/acl_soft_backend/models/relu.prototxt
new file mode 100644
index 0000000..5f4932f
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/relu.prototxt
@@ -0,0 +1,20 @@
+name: "RELUNET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "relu1"
+  type: "ReLU"
+  bottom: "data"
+  top: "relu1"
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/reshape.prototxt b/contrib/nnc/tests/acl_soft_backend/models/reshape.prototxt
new file mode 100644
index 0000000..1f48e39
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/reshape.prototxt
@@ -0,0 +1,26 @@
+name: "RESHAPENET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "reshape"
+  type: "Reshape"
+  bottom: "data"
+  top: "output"
+  reshape_param {
+    shape {
+      dim: 1
+      dim: 96
+    }
+  }
+}
diff --git a/contrib/nnc/tests/acl_soft_backend/models/scale.prototxt b/contrib/nnc/tests/acl_soft_backend/models/scale.prototxt
new file mode 100644
index 0000000..b113056
--- /dev/null
+++ b/contrib/nnc/tests/acl_soft_backend/models/scale.prototxt
@@ -0,0 +1,30 @@
+name: "SCALENET"
+layer {
+  name: "input"
+  type: "Input"
+  top: "data"
+  input_param {
+    shape {
+      dim: 1
+      dim: 3
+      dim: 4
+      dim: 8
+    }
+  }
+}
+layer {
+  name: "scale1"
+  type: "Scale"
+  bottom: "data"
+  top: "scale1"
+  param {
+    lr_mult: 1
+    decay_mult: 0
+  }
+  scale_param {
+    filler {
+      value: 2.71
+    }
+    bias_term: false
+  }
+}