MSVC 2017 build fails with x86 architecture (v2.0.0-rc2)

[taxen]

Using MSVC 2017 Community on Windows 10 and CMake 3.11.3, I am trying to build Windows binary from jumanpp-2.0.0-rc2.tar.xz. However, I have some trouble to ask the developers for help. (On Linux, I got working binary from the same source. Many thanks!)

Procedure

1. Install MSVC 2017 Comunnity version (chose default set for Desktop development with C++) and CMake 3.11.3 (chose Windows win64-x64 ZIP as the installer gave an error).
2. Download jumanpp-2.0.0-rc2.tar.xz and extract into C:\data\jumanpp-2.0.0-rc2 (I used 7-Zip 18.05).
3. As instructed https://github.com/ku-nlp/jumanpp#building-from-a-package , open a terminal and do the following with the result.

C:\data\jumanpp-2.0.0-rc2>mkdir bld

C:\data\jumanpp-2.0.0-rc2>cd bld

C:\data\jumanpp-2.0.0-rc2\bld>cmake ..  -DCMAKE_BUILD_TYPE=Release    -DCMAKE_INSTALL_PREFIX="C:\Program Files\jumanpp"
-- Building for: Visual Studio 15 2017
-- Selecting Windows SDK version 10.0.17134.0 to target Windows 10.0.16299.
-- The C compiler identification is MSVC 19.14.26430.0
-- The CXX compiler identification is MSVC 19.14.26430.0
-- Check for working C compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx86/x86/cl.exe
-- Check for working C compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx86/x86/cl.exe -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working CXX compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx86/x86/cl.exe
-- Check for working CXX compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx86/x86/cl.exe -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for pthread.h
-- Looking for pthread.h - not found
-- Found Threads: TRUE
-- Could NOT find Protobuf (missing: Protobuf_LIBRARIES Protobuf_INCLUDE_DIR)
Not using Protobuf-based components
-- Eigen3... OK

    Static feature generation:
        COMMAND cgtest02_codegen_binary cgtest02 Test02 C:/data/jumanpp-2.0.0-rc2/bld/src/core/codegen/gen
        DEPENDS cgtest02_codegen_binary cg_2_spec.h
        OUTPUT C:/data/jumanpp-2.0.0-rc2/bld/src/core/codegen/gen/cgtest02.cc C:/data/jumanpp-2.0.0-rc2/bld/src/core/codegen/gen/cgtest02.h
        LIBS jpp_core


    Static feature generation:
        COMMAND jpp_jumandic_cg_codegen_binary jpp_jumandic_cg JumandicStatic C:/data/jumanpp-2.0.0-rc2/bld/src/jumandic/gen
        DEPENDS jpp_jumandic_cg_codegen_binary shared/jumandic_spec.cc
        OUTPUT C:/data/jumanpp-2.0.0-rc2/bld/src/jumandic/gen/jpp_jumandic_cg.cc C:/data/jumanpp-2.0.0-rc2/bld/src/jumandic/gen/jpp_jumandic_cg.h
        LIBS jpp_jumandic_spec

C:/data/jumanpp-2.0.0-rc2/bld/src/jumandic/gen
-- Configuring done
-- Generating done
-- Build files have been written to: C:/data/jumanpp-2.0.0-rc2/bld

4. Doubleclick "C:\data\jumanpp-2.0.0-rc2\bld\jumanpp.sln" and launch MSVC 2017.
5. After some messages and project loading is done, choose ALL_BUILD from Solution Explorer and build with solution configuration of Release (changed from Debug) and Win32 (as default).
6. After a while, I get 19 errors and many warnings and the build fails. Attached gz file is the exact output of the build (1210 lines long). jumanpp-build-log.txt.gz

Since Linux build is fine and it actually works pretty well, something on my Windows is not right. Some advice would be really appreciated.

[eiennohito]

I am using the same procedure (the only difference is that I was using mingw-msys64 cmake and not the Windows installer) and it works (with AppVeyor CI as well). I'll try to dig into it.

A question: does this script work for you? (A build out of MSVS GUI, but still using MSVC toolkit).

mkdir build-dir
cd build-dir
cmake -G "Visual Studio 15 2017 Win64" ..
cmake --build . --config Release
ctest -C Release --output-on-failure

[taxen]

Wow! I did exactly you suggested me to do and voila! Only ctest output is enough, I think, but in case you need I attach the whole log file (6272 line long).

C:\data\jumanpp-2.0.0-rc2\build-dir>ctest -C Release --output-on-failure
Test project C:/data/jumanpp-2.0.0-rc2/build-dir
      Start  1: jpp_util_test
 1/10 Test  #1: jpp_util_test ....................   Passed    0.07 sec
      Start  2: testing_infra_test
 2/10 Test  #2: testing_infra_test ...............   Passed    0.05 sec
      Start  3: jpp_core_tests
 3/10 Test  #3: jpp_core_tests ...................   Passed    0.27 sec
      Start  4: jpp_core_analysis_tests
 4/10 Test  #4: jpp_core_analysis_tests ..........   Passed    0.72 sec
      Start  5: jpp_codegen_tests
 5/10 Test  #5: jpp_codegen_tests ................   Passed    0.09 sec
      Start  6: jpp_core_spec_tests
 6/10 Test  #6: jpp_core_spec_tests ..............   Passed    0.06 sec
      Start  7: jpp_core_train_tests
 7/10 Test  #7: jpp_core_train_tests .............   Passed    0.20 sec
      Start  8: jpp_jumandic_tests
 8/10 Test  #8: jpp_jumandic_tests ...............   Passed    0.33 sec
      Start  9: jpp_bug_tests
 9/10 Test  #9: jpp_bug_tests ....................   Passed    0.14 sec
      Start 10: jpp_rnn_tests
10/10 Test #10: jpp_rnn_tests ....................   Passed    0.22 sec

100% tests passed, 0 tests failed out of 10

Total Test time (real) =   2.24 sec

jumanpp-build-log2.txt.gz

Some further questions:

1. After the build is done, I manually grab "{where source is located}\jumanpp-2.0.0-rc2\build-dir\src\jumandic\Release\jumanpp_v2.exe" and "{where source is located}\jumanpp-2.0.0-rc2\model\jumandic.jppmdl" and put them into "C:\Program Files\jumanpp" and "C:\Program Files\jumanpp\libexec\jumanpp" respectively (as config file "{where source is located}\jumanpp-2.0.0-rc2\model\jumandic.conf.in" suggested to put the dic file like that). Is there a command to install after build?

2. I also copy the config file and put it in the binary directory, rename it to jumandic.conf and try to run, but fails (tried with original file name as well). I set a model option in command line right now. How can I direct the binary to use the config file automatically without -c option? Using -c option to specify the config file fails because I put it under C:\Program Files\jumanpp which include a space. Escaping a space by \ does not seem to work for me. Strange escape is not even needed in command line when double quoted.

C:\Program Files\jumanpp>jumanpp_v2.exe -c jumandic.conf sample.txt
failed to load model from disk: InvalidState: could not access file: "C:/Program errcode=123
backtrace:
    jumanpp::util::MappedFile::open at c:\data\jumanpp-2.0.0-rc2\src\util\mmap_impl_win32.h:66
    jumanpp::core::model::FilesystemModel::open at c:\data\jumanpp-2.0.0-rc2\src\core\impl\model_io.cc:111
    jumanpp::core::JumanppEnv::loadModel at c:\data\jumanpp-2.0.0-rc2\src\core\env.cc:11
    jumanpp::jumandic::JumanppExec::init at c:\data\jumanpp-2.0.0-rc2\src\jumandic\shared\jumandic_env.cc:26

Here is my content of jumandoc.conf

--model="C:/Program Files/jumanpp/libexec/jumanpp/jumandic.jppmdl"
--rnn-nce-bias=5.62844432562
--rnn-unk-constant=-3.4748115191
--rnn-unk-length=-2.92994951022
--feature-weight-perceptron=1
--feature-weight-rnn=0.0176

3. Is there any way to control output character code? Above terminal is mintty. Windows default terminal won't display the output correctly.

Thank you!

[eiennohito]

Nice to know that the build have passed. That means that at least MSBuild files generated by CMake are sane. I wonder that is the problem with the VS Solution.

1. We want to build an installer for Windows. #81. Installing from the CMake source build... is a bit advanced thing on Windows.
2. Thats a known bug :P. Please don't use spaces in the path names for the time being. I will fix it together with #83.
3. I do not plan to add any character code convertion functionality in the binary itself. You can use (or write) a wrapper script if you want.

[taxen]

Thank you for clearing those things out. Totally understandable especially about output code 3. This issue is now solved, I think.

[eiennohito]

I have looked over your build log once more. It seems that by default CMake tries to generate build files for 32-bit. I haven't tested that configuration.

It probably should work, but a lot of Juman++ internal optimizations assume that we are using a 64-bit platform and have a sane number of 64-bit registers, which is obviously false on x86.

[taxen]

I once tried cmake -G "Visual Studio 15 2017" .. but not cmake -G "Visual Studio 15 2017 Win64" .. So... Win64 makes the difference, I see.

[DoumanAsh]

Visual Studio 15 2017 is always 32 bit project and you need append Win64 for 64-bit builds

It should be possible to add warning to CMake to generate at least warning in such cases

[eiennohito]

Win32 build should not fail. It is a bug. It should be significantly slower (my guess is ~10x) than Win64 version though.

[DoumanAsh]

Oh ok, I can check it out on my windows laptop since I have toolchain

[eiennohito]

Tests still fail, so I reopen this issue.

Still, I don't know if this is a good idea to support 32 bit architecture at all. The problem is: Juman++ memory maps analysis model and it can be large: 300m. When embedding this thing into other applications you can run out of virtual memory very fast on 32 bits.

Also, hashing algorithm I use assumes that we have a lot of 64-bit integer registers. That will cause slowdown on x32 (I don't know how much though). Hashing is the cornerstone of the analysis speed by the way.

[DoumanAsh]

@eiennohito Yeah, sorry, I suspect yesterday I tested it incorrectly.

Looking at tests:

TEST_CASE("murmur hash works with strings") {
  auto string = "this is a test string";
  auto ptr = reinterpret_cast<const jumanpp::u8*>(string);
  auto hv = murmurhash3_memory(ptr, ptr + sizeof(string), 10);
  CHECK(hv == 0x2e62c68031e4659dULL);
}

Even tests themself assume 64 bit integers, but internally murmur_hash uses size_t so you can easily lose accuracy

[eiennohito]

That's a bug, it should use uint64_t. Otherwise models will not be portable.

[DoumanAsh]

There are actually lots of warnings about implicit casts between size_t and uint64_t so to be sure I'd suggest to go through all of them... but it is lots to take on, so might take a while

[eiennohito]

At least MurmurHash internal hashing representation should use fixed 64 bit integers.

But sized integer hygiene is... not good, yep.

[DoumanAsh]

@eiennohito Just letting you know I for now give up with murmurhash3_memory tests, I just do not see where narrowing could happen that starts failing test. I doubt it is critical issue since x64 is recommended to be used, but well maybe someone else will try to tackle it. Or me at later time.

[eiennohito]

Thanks for digging. I'm busy with a paper and upcoming PhD thesis right now, so I can't guarantee when I look into it, but I will try.

[eiennohito]

If your project is sorta largish, I strongly recommend you to use Juman++ via gRPC. You will be protected from crashes and other weird things that way.

https://github.com/eiennohito/jumanpp-grpc