Add support for building Python Wheel packages

JonathanSalwan commented 2 years ago

This refers to #1144.

Pending tasks

[x] LLVM support (WIP: Linux and MacOS ok)
[x] Bitwuzla support (WIP: Linux and MacOS ok)
[x] ~~Artifacts on Linux with all Python versions~~
[x] ~~Artifacts on Windows with all Python versions~~
[x] ~~Artifacts on macOS with all Python versions~~
[x] ~~Supporting Triton's Interfaces~~

Any help on the following will be much appreciated! :

Building the `.whl` by default

$ python -m build --wheel
* Creating venv isolated environment...
* Installing packages in isolated environment... (setuptools >= 40.8.0, wheel)
* Getting dependencies for wheel...
running egg_info
creating triton_library.egg-info
[...]
* Installing packages in isolated environment... (wheel)
* Building wheel...
running bdist_wheel
running build
running build_ext
-- The C compiler identification is GNU 11.3.0
-- The CXX compiler identification is GNU 11.3.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
[...]
adding 'triton_library-1.0.0.dist-info/WHEEL'
adding 'triton_library-1.0.0.dist-info/top_level.txt'
adding 'triton_library-1.0.0.dist-info/RECORD'
removing build/bdist.linux-x86_64/wheel
Successfully built triton_library-1.0.0-cp39-cp39-linux_x86_64.whl

Building the `.whl` with a specific interface

$ LLVM_INTERFACE=On python -m build --wheel

By default, interfaces are defined as:

Z3_INTERFACE: On
LLVM_INTERFACE: Off
BITWUZLA_INTERFACE: Off
BOOST_INTERFACE: Off

Installing the package with pip

$ pip install . --user
Processing /home/jonathan/Works/Tools/Triton
  Preparing metadata (setup.py) ... done
Building wheels for collected packages: triton-library
  Building wheel for triton-library (setup.py) ... done
  Created wheel for triton-library: filename=triton_library-1.0.0-cp39-cp39-linux_x86_64.whl size=1518902 sha256=aeb378b42f39a7bff36bc2037aca1fd3fd94d37227609ebd719d6f7afde95ccb
  Stored in directory: /tmp/pip-ephem-wheel-cache-xtz2872_/wheels/07/e0/9a/c02c3eac5862a6b34ec58c12793c45a3eb9827de1c9308c2e3
Successfully built triton-library
Installing collected packages: triton-library
Successfully installed triton-library-1.0.0

JonathanSalwan commented 2 years ago

@cnheitman, I think there were a bug around cmake_args on Linux. Commit 450a345fca636402a2c4b5149f78e8434a4e79fb fixes it (before that, cmake_args already defined were not taken into account). However, with this patch there is another issue when calling copy_file. The commit 404484112f6003ef7375b1d6287b431658424a17 fixes it but I do not know if it breaks something else. Is defining CMAKE_LIBRARY_OUTPUT_DIRECTORY important?

cnheitman commented 2 years ago

Hey @JonathanSalwan o/

Nice!

Indeed, there was a bug there '^^. I just built the .whl for both Ubuntu and Windows and it works fine. Replying to your question, no, CMAKE_LIBRARY_OUTPUT_DIRECTORY does not seem necessary, so let's remove it.

An update on the Bitwuzla support. I tried building it on Windows and I ran into some issue (bitwuzla/bitwuzla/issues/30). To summarize, they are updating the Windows build (code and doc) and it'll be ready in a week or two. I'll work on it again once it's ready. (Btw, I have already built Bitwuzla statically on Linux I'll add it to the python.yml when I have the Windows part working so we generate .whl supporting the same libraries on both OSs.)

JonathanSalwan commented 2 years ago

I've cleaned up CI files. There were duplicates of jobs over different files. Now there is only one linux.yml file that supports wide-integer as a matrix and enables LLVM + Bitwuzla. Thus, I've removed the files: llvm.yml, bitwuzla.yml and wide-integer.yml. I've also renamed cmake.yml to vcpkg.yml which is more explicit.

JonathanSalwan commented 2 years ago

Linux and Windows artifacts for different Python versions have been generated with success.

cnheitman commented 2 years ago

Awesome @JonathanSalwan :D I'll add support for Bitwuzla in Linux later today.

JonathanSalwan commented 2 years ago

\o/. I will try to add macOS tomorrow.

JonathanSalwan commented 2 years ago

Artifacts for macOS are ok

JonathanSalwan commented 2 years ago

LLVM static support for Linux and macOS is ok. For Windows we have to hack here and there.

n0thing2speak commented 1 year ago

m1 chip,macOS 12.4, python3.10.5 arm64 version,it can be compiled successfully, but the following error occurs when using it

Python 3.10.5 | packaged by conda-forge | (main, Jun 14 2022, 07:07:06) [Clang 13.0.1 ] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import triton
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
SystemError: initialization of triton did not return an extension module

cnheitman commented 1 year ago

How did you build it exactly? What interfaces did you turn on?

n0thing2speak commented 1 year ago

I think this might be some problems for the python installed in miniconda? I compiled and installed it successfully with the system's python, and I don't know why.

Here is the cmds:

git clone https://github.com/JonathanSalwan/Triton
cd Triton
mkdir build ; cd build
cmake .. -DCMAKE_INSTALL_PREFIX="$PREFIX" \
  -DCMAKE_C_COMPILER=gcc-12 \
  -DCMAKE_CXX_COMPILER=g++-12 \
  -DBOOST_ROOT=$(brew --prefix boost) \
  -DPYTHON_LIBRARY=/Users/$USER/mambaforge/lib/libpython3.10.dylib \
  -DPYTHON_INCLUDE_DIR=/Users/$USER/mambaforge/include/python3.10
make -j8 
cd ..
python -m build --wheel
cd dist 
pip install triton_library-1.0.0rc1-cp310-cp310-macosx_12_0_arm64.whl --user

This command will have the above problem in python installed by conda, but works well in python with Xcode version

cnheitman commented 1 year ago

First, you don't need to run cmake by your own for building the whl package, python -m build --wheel take care of that. So, you only need to type:

git clone https://github.com/JonathanSalwan/Triton
cd Triton
python -m build --wheel
cd dist 
pip install triton_library-1.0.0rc1-cp310-cp310-macosx_12_0_arm64.whl --user

I think the problem might be related to the python library that is being linked to triton. Can you send me the output of python -m build --wheel (at some point it runs cmake which prints information about which python lib is using).

n0thing2speak commented 1 year ago

In fact, when I don't run CMake, I get a bunch of errors. Sorry, I uninstalled the miniconda environment, and now the python installed from brew runs very well, do you need the arm64 wheel? I uploaded it belo triton_library-1.0.0rc1-cp39-cp39-macosx_12_0_arm64.whl.zip w

cnheitman commented 1 year ago

Hey! An update on this issue.

Linux: I tweaked and fixed the building process. I did some testing and it's working fine.
macOS: I fixed the packaging process. Now it adds to the final whl package all the libs it needs to run on its own. I haven't tested it yet but it should be working fine.
Windows: I have a working building process which includes Z3 and Bitwuzla. I also managed to add to the final whl all the libs it needs to run on its own. There are two pending issues to solve. First, I have this script working locally. I haven't added it to the CI yet (building Bitwuzla is complex and adding it to the CI is not straightforward). Second, I'm still working on adding LLVM support. I have disabled the CI for Windows for now. I tested the package I have so far (Z3 + Bitwuzla) and it's working fine.

Btw, I'll do a PR with these changes once the CI on my fork finish running.

cnheitman commented 1 year ago

Hey o/ Another update.

Windows: I managed to compile LLVM and build the whl package successfully. Now, the windows version of the package have support for Z3, Bitwuzla and LLVM. For now, I will build the package locally (that is, I will not add it to CI). The reason for this is: 1) Bitwuzla is complex to build on Windows, and 2) the only way to add LLVM support was to building it from scratch, which requires a lot of cpu time and disk. That being say, the package is already on PyPI and working for Python 3.8, 3.9 and 3.10 (same for the Linux one).
macOS: The package is ready but I still need to check they are working properly. Once I do that I'll push them to PyPI.

JonathanSalwan commented 1 year ago

You rockz @cnheitman !

cnheitman commented 1 year ago

Last update on this issue.

The packages for both Linux and Windows are working and on PyPI for Python 3.8, 3.9 and 3.10 with support for Z3, Bitwuzla and LLVM. It can be installed by running pip install triton-libray.
The packages for Linux are generated using Github Actions. The packages for Windows are generated locally by me until I find a way to automatize it (complex setup).
The packages for mac are still pending (they need some tweaking). I'll open a new issue for it.

JonathanSalwan / Triton