conda packaging (experimental WIP)

Robadob commented 1 year ago

Very much work in progress

Changes

Add CMake option FLAMEGPU_BUILD_PYTHON_CONDA
- Enables set(CMAKE_SKIP_RPATH TRUE)
Fix set(CMAKE_CUDA_RUNTIME_LIBRARY shared) placement, so that executables and pyflamegpu link against shared cuda libs

todo

[x] Linux conda recipe
- There are still a few warnings present, but it appears to work.
[x] Windows conda recipe
- Probably want to test this on a managed desktop which doesn't have CUDA 12 installed to be sure
- Not sure if we can find a machine with nvidia drivers, but no cuda or visual studio for a true vanilla test (edwin?)
[ ] Conda CI
- --error-overlinking to conda build command, to werror overlinking.
- Solve how to build variants/metapkg
- Solve optional visualisation version?
- Solve py compatible version no (from CMake?)
[ ] Actually create a conda account so we can upload

Robadob commented 1 year ago

Via LD_DEBUG=libs python3 -c "import pyflamegpu" 2> ld.log have tested that this causes pyflamegpu to load libs from conda if present. (Removal of rpath and static from --verbose link command was observed too)

1054040:    find library=libnvrtc.so.12 [0]; searching
   1054040:  search path=/home/rob/miniconda3/envs/py38/bin/../lib      (RPATH from file python3)
   1054040:   trying file=/home/rob/miniconda3/envs/py38/bin/../lib/libnvrtc.so.12
   1054040: 
   1054040: find library=libcuda.so.1 [0]; searching
   1054040:  search path=/home/rob/miniconda3/envs/py38/bin/../lib      (RPATH from file python3)
   1054040:   trying file=/home/rob/miniconda3/envs/py38/bin/../lib/libcuda.so.1
   1054040:  search path=/usr/local/cuda-12.2/lib64:tls/haswell/x86_64:tls/haswell:tls/x86_64:tls:haswell/x86_64:haswell:x86_64:        (LD_LIBRARY_PATH)
   1054040:   trying file=/usr/local/cuda-12.2/lib64/libcuda.so.1
   1054040:   trying file=tls/haswell/x86_64/libcuda.so.1
   1054040:   trying file=tls/haswell/libcuda.so.1
   1054040:   trying file=tls/x86_64/libcuda.so.1
   1054040:   trying file=tls/libcuda.so.1
   1054040:   trying file=haswell/x86_64/libcuda.so.1
   1054040:   trying file=haswell/libcuda.so.1
   1054040:   trying file=x86_64/libcuda.so.1
   1054040:   trying file=libcuda.so.1
   1054040:  search cache=/etc/ld.so.cache
   1054040:   trying file=/lib/x86_64-linux-gnu/libcuda.so.1
   1054040: 
   1054040: find library=libcudart.so.12 [0]; searching
   1054040:  search path=/home/rob/miniconda3/envs/py38/bin/../lib      (RPATH from file python3)
   1054040:   trying file=/home/rob/miniconda3/envs/py38/bin/../lib/libcudart.so.12

Note libcuda.so is always provided by the driver.

Robadob commented 1 year ago

Conda post-build log of first successful package build. https://gist.github.com/Robadob/9b5f33c377a8bb45793058725769d8d5

Robadob commented 1 year ago

Summary of warnings (Linux)

Warning: rpath /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_build_env/lib is outside prefix /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_h_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_p (removing it)

CMake isn't removing all RPaths?

Disabling the CMake rpath command did not change this warning, nor add anything related (assuming i updated git right)

Overlinking (resolved)

> [During the post-build phase, the shared libraries in the newly created package are examined for linkages which are not provided by the package's requirements or a predefined list of system libraries.](https://docs.conda.io/projects/conda-build/en/stable/resources/define-metadata.html#whitelisting-shared-libraries) ``` WARNING (pyflamegpu,lib/python3.11/site-packages/pyflamegpu/_pyflamegpu.so): $RPATH/libnvrtc.so.12 not found in packages, sysroot(s) nor the missing_dso_whitelist. .. is this binary repackaging? WARNING (pyflamegpu,lib/python3.11/site-packages/pyflamegpu/_pyflamegpu.so): $RPATH/libcuda.so.1 not found in packages, sysroot(s) nor the missing_dso_whitelist. .. is this binary repackaging? WARNING (pyflamegpu,lib/python3.11/site-packages/pyflamegpu/_pyflamegpu.so): $RPATH/libcudart.so.12 not found in packages, sysroot(s) nor the missing_dso_whitelist. .. is this binary repackaging? ``` I've resolved this by moving cuda dependencies with required `.so` from `build` to `host`. ```yml requirements: build: ... - libcurand-dev # [linux] Build as no .so export (we don't use host curand) - swig ==4.0.2 # Have to manually direct swig to find this - git - cmake host: - cuda-cudart-dev # [linux] Host as require .so export - cuda-nvrtc-dev # [linux] Host as require .so export ``` However, I'm concerned this will make an implicit requirement on the `-dev` variant of the packages rather than the none-dev versions (which contain the same `.so`). Yet, placing the none-dev versions in `run` instead does not resolve the warning. ```yml build: ... - libcurand-dev # [linux] Build as no .so export (we don't use host curand) - cuda-cudart-dev # [linux] Host as require .so export - cuda-nvrtc-dev # [linux] Host as require .so export host: ... run: ... - cuda-cudart # [linux] - cuda-nvrtc # [linux] ``` Additionally, `libcuda.so` provided by the driver is ignored via ```yml build: missing_dso_whitelist: - $RPATH/libcuda.so.1 # Ignore as this comes from CUDA driver ```

Possible we can force an error for these warnings with --error-overlinking

WARNING (pyflamegpu): interpreter (Python) package conda-forge::python-3.11.6-hab00c5b_0_cpython in requirements/run but it is not used (i.e. it is overdepending or perhaps statically linked? If that is what you want then add it to `build/ignore_run_exports`)
WARNING (pyflamegpu): interpreted library (Python) package conda-forge::astpretty-2.1.0-pyhd8ed1ab_0 in requirements/run but it is not used (i.e. it is overdepending or perhaps statically linked? If that is what you want then add it to `build/ignore_run_exports`)

No idea how to resolve these, we know them to be required but for whatever reason it thinks otherwise.

The suggested build/ignore_run_exports seems to be documented(1, 2), that it instead prevents dependencies of the named requirements from having their dependencies required too?? Which may not be what we want.

build:
  ignore_run_exports:
    - python # Not clear why conda thinks this isn't used
    - astpretty # Not clear why conda thinks this isn't used

Package Verification (no longer present?)

``` Package verification results: ----------------------------- /tmp/rob/tmp9cb176fx/pyflamegpu-2.0.0rc1-hbd1bc14_0.tar.bz2: C1141 Found python file "lib/python3.11/site-packages/pyflamegpu/codegen/codegen.py" without a corresponding pyc file ``` Conda is supposed to automatically compile all `.py`, i've only seen guides for exempting files. This may have self resolved, however I can't even see the "Package verification results" title anymore so unclear.

ClobberWarning: This transaction has incompatible packages due to a shared path.
  packages: conda-forge/linux-64::cuda-nsight-12.0.78-ha770c72_0, conda-forge/linux-64::cuda-cuobjdump-12.0.76-h59595ed_0, conda-forge/linux-64::cuda-cuxxfilt-12.0.76-h59595ed_0, conda-forge/linux-64::cuda-nvprune-12.0.76-h59595ed_0, conda-forge/linux-64::cuda-sanitizer-api-12.0.90-h59595ed_0, conda-forge/linux-64::cuda-nvprof-12.0.90-h59595ed_0, conda-forge/linux-64::cuda-nvvp-12.0.90-h59595ed_0
  path: 'LICENSE'
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-cuobjdump-12.0.76-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-cuxxfilt-12.0.76-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-nvprune-12.0.76-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-sanitizer-api-12.0.90-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-nvprof-12.0.90-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE
ClobberWarning: Conda was asked to clobber an existing path.
  source path: /home/rob/miniconda3/envs/py311/pkgs/cuda-nvvp-12.0.90-h59595ed_0/LICENSE
  target path: /home/rob/miniconda3/envs/py311/conda-bld/pyflamegpu_1698661724842/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placeh/LICENSE

Not too sure what clobber is, some kind of path rewrite? This would suggest it's trying to package the license of various cuda packages into our package though which seems wrong. Unable to find a resolution via googling, it currently seems harmless (clobber can cause install failures in bad environments, but that doesn't appear to be the case here)

ptheywood commented 1 year ago

However, I'm concerned this will make an implicit requirement on the -dev variant of the packages rather than the none-dev versions (which contain the same .so).

-dev might be able to go in build and have the non-dev versions in host / run?

Not sure about the rest of the things to resolve, might just have to be more trial and error.

Robadob commented 1 year ago

The currently built linux conda package appears to work, however it's a bit awkward to install it with dependencies.

Create directory test_channel/noarch and place the package's .tar.gz inside that
conda index test_channel (requires conda-build)
conda install -c file:///home/rob/test_channel pyflamegpu (abs path may vary ~ does not work)

Robadob commented 1 year ago

Windows log from first successful conda packaging

https://gist.github.com/Robadob/9e3b7a94991762f76a22d864d727d89a

Robadob commented 1 year ago

Summary of warnings (Windows)

Missing dlls, both CUDA should probably be provided by driver/CUDA, so will add to DSO whitelist. The VC++ ones appear similar to the python warnings we had, if they're vc++ redist they are required despite whatever conda thinks.

WARNING (pyflamegpu,Lib/site-packages/pyflamegpu/_pyflamegpu.pyd): $RPATH/nvrtc64_120_0.dll not found in packages, sysroot(s) nor the missing_dso_whitelist.
.. is this binary repackaging?
   INFO (pyflamegpu,Lib/site-packages/pyflamegpu/_pyflamegpu.pyd): Needed DSO C:/Windows/System32/nvcuda.dll found in $SYSROOT
WARNING (pyflamegpu,Lib/site-packages/pyflamegpu/_pyflamegpu.pyd): $RPATH/cudart64_12.dll not found in packages, sysroot(s) nor the missing_dso_whitelist.
.. is this binary repackaging?
WARNING (pyflamegpu): dso library package conda-forge::vc14_runtime-14.36.32532-hdcecf7f_17 in requirements/run but it is not used (i.e. it is overdepending or perhaps statically linked? If that is what you want then add it to `build/ignore_run_exports`)
WARNING (pyflamegpu): dso library package conda-forge::ucrt-10.0.22621.0-h57928b3_0 in requirements/run but it is not used (i.e. it is overdepending or perhaps statically linked? If that is what you want then add it to `build/ignore_run_exports`)

Robadob commented 1 year ago

Windows vis build now seems to work.

Linux is stuck on CMake finding OpenGL. OpenGL is not a conda package as it's more of a system lib.

This guide suggests that core dependency tree (CDT) for mesa should be used, - {{ cdt('mesa-libgl-devel') }} # [linux], this does appear to compile, however this is very old so we would rather use lbglvnd as mesa may fail to run.

This conda recipe includes it's cdt libglvnd-opengl. However when we attempt that (on Waimu) it attempts to install the COS6 cdt version which does not exist for any libglvnd package.

There isn't much documentation about the COS6/7 feature, conda forge states that - sysroot_linux-64 2.17 # [linux64] will cause COS7 to be used (as also seen in the above example's recipe), however that does not make a difference in our case. Unclear if this is a conda-forge specific build feature.

More research and testing required.

Robadob commented 1 year ago

trying to get cos7 to work

cdt cos7 seems alot like a conda-forge feature.

- sysroot_{{ target_platform }} 2.17 # [linux] or - sysroot_linux-64 2.17 # [linux64]

https://conda-forge.org/docs/maintainer/knowledge_base.html#using-centos-7

Doesn't make a difference, maybe conda-forge specific.

# [cdt_name=='cos7']

https://github.com/conda-forge/cdt-builds/issues/41

Doesn't make a difference, maybe conda-forge specific.

c_stdlib:
  - sysroot                     # [linux]
c_stdlib_version:
  - 2.17                        # [linux]

- {{ stdlib('c') }}

https://github.com/conda/conda-build/blob/main/docs/source/resources/compiler-tools.rst#expressing-the-relation-between-compiler-and-its-standard-library

Doesn't work, adding pkg {{ stdlib('c') }} or {{ stdlib('cxx') }} causes an early RegEx failure when conda build is called.

Robadob commented 1 year ago

Shall see how far I can work through this list for publishing on conda-forge this week.

https://conda-forge.org/docs/maintainer/adding_pkgs.html

Robadob commented 1 year ago

https://github.com/conda-forge/staged-recipes/pull/24483

Working on this here.

FLAMEGPU / FLAMEGPU2