Glow 설치법 및 활용


Glow는 하드웨어 가속기를 위한 머신러닝 컴파일러이자 실행 엔진이다.

ubuntu 18.04 기준 설치

소스 다운

git clone git@github.com:pytorch/glow.git  # or: git clone https://github.com/pytorch/glow.git
cd glow

#submodules 다운
git submodule update --init --recursive

# 필요 package들 설치
sudo apt-get install clang clang-8 cmake graphviz libpng-dev \
    libprotobuf-dev llvm-8 llvm-8-dev ninja-build protobuf-compiler wget \
    opencl-headers libgoogle-glog-dev

# update-alternatives to manage the version of clang/clang++:
sudo update-alternatives --install /usr/bin/clang clang \
    /usr/lib/llvm-8/bin/clang 50
sudo update-alternatives --install /usr/bin/clang++ clang++ \
    /usr/lib/llvm-8/bin/clang++ 50

# Glow uses the system default C/C++ compiler (/usr/bin/C++), and so you may also want to switch your default C/C++ compiler to clang:
sudo update-alternatives --config cc
    # Select the option corresponding to /usr/bin/clang ...
sudo update-alternatives --config c++
    # Select the option corresponding to /usr/bin/clang++ ...

아래와 같이 clang compiler를 선택 한다.

스크린샷 2019-07-17 오후 4.20.37

ProtoBuf 설치법

최신 버전은 git clone으로 다운 받고 정해진 버전은 아래와 같이 release tag를 이용해서 다운 받는다.
https://github.com/protocolbuffers/protobuf/releases/tag/v3.6.1

Git 경우

$ sudo apt-get install autoconf automake libtool curl make g++ unzip
$ git clone https://github.com/protocolbuffers/protobuf.git
$ cd protobuf
$ git submodule update --init --recursive
$ ./autogen.sh

$ ./configure
$ make
$ make check
$ sudo make install
$ sudo ldconfig # refresh shared library cache.
./configure --prefix=/usr

Glow 최종 컴파일 및 실행 파일

Debug mode 컴파일

mkdir build_Debug
cd build_Debug
cmake -G Ninja -DCMAKE_BUILD_TYPE=Debug ..
ninja all

Release mode 컴파일
-DCMAKE_BUILD_TYPE=Release ..

실행

아래와 같이 최종적으로 tests에 테스팅 가능한 실행 파일들이 존재한다. 각각 실행해 보면 정상적으로 설치가 되었는지 확인 가능하다.

스크린샷 2019-07-17 오후 5.02.46

bin 디렉터리 안에는 아래와 같이 다양한 실행파일들이 존재한다. --help 옵션을 넣어서 어떠한 인자들을 넣어야 실행 가능한지 알 수 있다.

실행 가능 파일

InstrGen  fr2en             mnist         resnet-runtime   tracing-compare
NodeGen   image-classifier  model-runner  resnet-training
char-rnn  include-bin       png2bin       resnet-verify
cifar10   lenet-loader      ptb           text-translator

테스팅과 실행: Unit tests

ninja test를 통해서 모든 테스트들을 실행 할 수 있다.
Debug 또는 Release로 빌드를 수행한 디렉토리에서 수행한다.
결과는 아래와 같다.

jemin@jemin:~/development/glow/build_Debug$ ninja test
[0/1] Running tests...
Test project /home/jemin/development/glow/build_Debug
      Start  1: BackendCorrectnessTest
 1/35 Test  #1: BackendCorrectnessTest ..............   Passed   11.51 sec
      Start  2: BackendTest
 2/35 Test  #2: BackendTest .........................   Passed    1.35 sec
      Start  3: BasicIRTest
 3/35 Test  #3: BasicIRTest .........................   Passed    0.14 sec
      Start  4: Caffe2ImporterTest
 4/35 Test  #4: Caffe2ImporterTest ..................   Passed    1.93 sec
      Start  5: DeviceManagerTest
 5/35 Test  #5: DeviceManagerTest ...................   Passed    0.39 sec
      Start  6: ThreadPoolExecutorTest
 6/35 Test  #6: ThreadPoolExecutorTest ..............   Passed    1.43 sec
      Start  7: Float16Test
 7/35 Test  #7: Float16Test .........................   Passed    0.00 sec
      Start  8: GemmTest
 8/35 Test  #8: GemmTest ............................   Passed    0.09 sec
      Start  9: GlowOnnxifiManagerTest
 9/35 Test  #9: GlowOnnxifiManagerTest ..............   Passed    0.09 sec
      Start 10: GradCheckTest
10/35 Test #10: GradCheckTest .......................   Passed    3.63 sec
      Start 11: GraphGradTest
11/35 Test #11: GraphGradTest .......................   Passed    0.07 sec
      Start 12: GraphOptzTest
12/35 Test #12: GraphOptzTest .......................   Passed    0.11 sec
      Start 13: GraphSchedulerTest
13/35 Test #13: GraphSchedulerTest ..................   Passed    0.01 sec
      Start 14: GraphTest
14/35 Test #14: GraphTest ...........................   Passed    0.45 sec
      Start 15: HostManagerTest
15/35 Test #15: HostManagerTest .....................   Passed    3.94 sec
      Start 16: HyphenTest
16/35 Test #16: HyphenTest ..........................   Passed    0.83 sec
      Start 17: IROptTest
17/35 Test #17: IROptTest ...........................   Passed    0.01 sec
      Start 18: ImageTest
18/35 Test #18: ImageTest ...........................   Passed    0.24 sec
      Start 19: LLVMIRGenTest
19/35 Test #19: LLVMIRGenTest .......................   Passed    0.04 sec
      Start 20: MLTest
20/35 Test #20: MLTest ..............................   Passed   17.52 sec
      Start 21: MemoryAllocatorTest
21/35 Test #21: MemoryAllocatorTest .................   Passed    0.14 sec
      Start 22: OnnxImporterTest
22/35 Test #22: OnnxImporterTest ....................   Passed    0.30 sec
      Start 23: OnnxExporterTest
23/35 Test #23: OnnxExporterTest ....................   Passed    0.06 sec
      Start 24: OperatorGradTest
24/35 Test #24: OperatorGradTest ....................   Passed    0.04 sec
      Start 25: OperatorTest
25/35 Test #25: OperatorTest ........................   Passed    6.80 sec
      Start 26: PartitionerTest
26/35 Test #26: PartitionerTest .....................   Passed    0.11 sec
      Start 27: RecommendationSystemTest
27/35 Test #27: RecommendationSystemTest ............   Passed  226.65 sec
      Start 28: ProvisionerTest
28/35 Test #28: ProvisionerTest .....................   Passed    0.54 sec
      Start 29: QuantizationTest
29/35 Test #29: QuantizationTest ....................   Passed    3.96 sec
      Start 30: TensorsTest
30/35 Test #30: TensorsTest .........................   Passed    0.17 sec
      Start 31: TensorPoolTest
31/35 Test #31: TensorPoolTest ......................   Passed    0.01 sec
      Start 32: ThreadPoolTest
32/35 Test #32: ThreadPoolTest ......................   Passed    0.01 sec
      Start 33: TraceEventsTest
33/35 Test #33: TraceEventsTest .....................   Passed    6.74 sec
      Start 34: TypeAToTypeBFunctionConverterTest
34/35 Test #34: TypeAToTypeBFunctionConverterTest ...   Passed    0.05 sec
      Start 35: UtilsTest
35/35 Test #35: UtilsTest ...........................   Passed    0.01 sec

100% tests passed, 0 tests failed out of 35

Label Time Summary:
EXPENSIVE    = 226.65 sec*proc (1 test)

Total Test time (real) = 289.40 sec

C++ API examples

example 디렉터리에 각종 예제가 존재 한다.

MAC-OS (10.14.6 Mojave)

다운

git clone git@github.com:pytorch/glow.git  # or: git clone https://github.com/pytorch/glow.git
cd glow
git submodule update --init --recursive

필수 페키지 설치

brew install cmake graphviz libpng ninja protobuf wget glog autopep8
brew install llvm@7

링크

ln -s "/usr/local/opt/llvm@7/bin/clang-format" "/usr/local/bin/clang-format"
ln -s "/usr/local/opt/llvm@7/bin/clang-tidy" "/usr/local/bin/clang-tidy"

Mac OS 10.14 (Mojava)의 경우 ninja all시 header 파일 몇개가 없어서 빌드가 실패 한다.
아래와 같이 CLT을 실행해서 추가적으로 설치 한다.

open /Library/Developer/CommandLineTools/Packages/macOS_SDK_headers_for_macOS_10.14.pkg

설치 명령어

mkdir build_Debug
cd ./build_Debug
cmake -G Ninja -DCMAKE_BUILD_TYPE=Debug -DLLVM_DIR=/usr/local/opt/llvm@7/lib/cmake/llvm ..

#결과
-- Found glog with new-style glog target.
-- Found LLVM 7.1.0
-- Using LLVMConfig.cmake in: /usr/local/opt/llvm@7/lib/cmake/llvm
Adding CPU backend.
Adding Interpreter backend.
--
-- ******** Summary ********
--   CMake version         : 3.13.4
--   CMake command         : /usr/local/Cellar/cmake/3.13.4/bin/cmake
--   System                : Darwin
--   C++ compiler          : /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/c++
--   C++ compiler version  : 10.0.1.10010046
--   CXX flags             :  -Wall -Wnon-virtual-dtor -fno-exceptions -fno-rtti -Wnon-virtual-dtor
--   Build type            : Debug
--   Compile definitions   : GIT_SHA1="cd686e48";GIT_DATE="2019-07-26";WITH_PNG;GLOW_WITH_LLVMIRCODEGEN=1;GLOW_WITH_CPU=1;GOOGLE_PROTOBUF_NO_RTTI;ONNX_NAMESPACE=glow_onnx
--   CMAKE_PREFIX_PATH     :
--   CMAKE_INSTALL_PREFIX  : /usr/local
--   CMAKE_MODULE_PATH     : /Users/jeminlee/development/glow/cmake/modules
--
--   ONNX version          : 1.5.0
--   ONNX NAMESPACE        : glow_onnx
--   ONNX_BUILD_TESTS      : OFF
--   ONNX_BUILD_BENCHMARKS : OFF
--   ONNX_USE_LITE_PROTO   : OFF
--   ONNXIFI_DUMMY_BACKEND : OFF
--   ONNXIFI_ENABLE_EXT    : OFF
--
--   Protobuf compiler     : /usr/local/bin/protoc
--   Protobuf includes     : /usr/local/include
--   Protobuf libraries    : /usr/local/lib/libprotobuf.dylib
--   BUILD_ONNX_PYTHON     : OFF
-- Failed to find LLVM FileCheck
-- git Version: v1.5.0
-- Version: 1.5.0
-- Performing Test HAVE_THREAD_SAFETY_ATTRIBUTES -- failed to compile
-- Performing Test HAVE_STD_REGEX -- success
-- Performing Test HAVE_GNU_POSIX_REGEX -- failed to compile
-- Performing Test HAVE_POSIX_REGEX -- success
-- Performing Test HAVE_STEADY_CLOCK -- success
Skipping adding test en2gr_cpu_test because it requires a models directory. Configure with -DGLOW_MODELS_DIR.
Skipping adding test en2gr_quantization_test because it requires a models directory. Configure with -DGLOW_MODELS_DIR.
Skipping adding test en2gr_cpu_partition_test because it requires a models directory. Configure with -DGLOW_MODELS_DIR.
Skipping adding test en2gr_cpu_config_test because it requires a models directory. Configure with -DGLOW_MODELS_DIR.
Skipping adding test resnet_runtime_test because it requires a models directory. Configure with -DGLOW_MODELS_DIR.
-- Configuring done
-- Generating done
-- Build files have been written to: /Users/jeminlee/development/glow/build_Debug
$ ninja all

[1/300] Running gen_proto.py on onnx/onnx.in.proto
Processing /Users/jeminlee/development/glow/thirdparty/onnx/onnx/onnx.in.proto
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx_glow_onnx-ml.proto
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx_glow_onnx-ml.proto3
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx-ml.pb.h
generating /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx_pb.py
[5/300] Running gen_proto.py on onnx/onnx-operators.in.proto
Processing /Users/jeminlee/development/glow/thirdparty/onnx/onnx/onnx-operators.in.proto
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx-operators_glow_onnx-ml.proto
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx-operators_glow_onnx-ml.proto3
Writing /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx-operators-ml.pb.h
generating /Users/jeminlee/development/glow/build_test/lib/Importer/build_onnx/onnx/onnx_operators_pb.py
[42/300] InstrGen: Generating instructions.
Writing instr descriptors to:
    /Users/jeminlee/development/glow/build_test/glow/AutoGenInstr.h
    /Users/jeminlee/development/glow/build_test/glow/AutoGenInstr.cpp
    /Users/jeminlee/development/glow/build_test/glow/AutoGenInstr.def
    /Users/jeminlee/development/glow/build_test/glow/AutoGenIRBuilder.h
    /Users/jeminlee/development/glow/build_test/glow/AutoGenIRBuilder.cpp
    /Users/jeminlee/development/glow/build_test/glow/AutoGenIRGen.h
[45/300] NodeGen: Generating nodes.
Writing node descriptors to:
    /Users/jeminlee/development/glow/build_test/glow/AutoGenNodes.h
    /Users/jeminlee/development/glow/build_test/glow/AutoGenNodes.cpp
    /Users/jeminlee/development/glow/build_test/glow/AutoGenNodes.def
[300/300] Linking CXX executable tests/RuntimeBench

문제해결 (Troubleshooting)

git pull 후 tests/googletest/googletest/CMakeLists.txt:133 (cxx_library) 관련 에러

Glow는 Thirdparty library들이 많이 엮여 있다. 따라서 관련 라이브러리들도 모두 업데이트 해야한다.

해결 방법

git submodule sync
Synchronizing submodule url for 'tests/OutputCheck'
Synchronizing submodule url for 'tests/googlebenchmark'
Synchronizing submodule url for 'tests/googletest'
Synchronizing submodule url for 'thirdparty/foxi'
Synchronizing submodule url for 'thirdparty/fp16'
Synchronizing submodule url for 'thirdparty/onnx'
Synchronizing submodule url for 'thirdparty/pybind11'

git submodule update --init --recursive
Submodule path 'tests/googletest': checked out '703bd9caab50b139428cea1aaff9974ebee5742e'
Submodule path 'thirdparty/foxi': checked out '97fe555430a857581b9b826ecd955e4f0a3653f0'

llvm 경로 문제

보통 자동 설치하면 llvm 경로는 알아서 잡힌다.

경로가 잡히지 않으면 아래와 같이 DLLVM_DIR 옵션을 준다.

cmake -G Ninja ../glow \
    -DCMAKE_BUILD_TYPE=Debug \
    -DLLVM_DIR=/usr/local/opt/llvm@7/lib/cmake/llvm


'Data Science > Embedded Deep learning' 카테고리의 다른 글

Glow 설치법  (0) 2019.11.19
Coral Dev Board (Google Edge TPU) 설정 및 사용후기  (1) 2019.08.13
VTA on FPGA Board  (2) 2019.04.02
PYNQ: Python productivity on ZYNQ  (0) 2019.04.02
TVM 설치 방법  (0) 2019.04.02
TensorRT이용한 Xavier DLA (NVDLA) 실행  (3) 2019.02.08

+ Recent posts