amypad / cuvec Goto Github PK

View Code? Open in Web Editor NEW

79.0 4.0 2.0 730 KB

Unifying Python/C++/CUDA memory: Python buffered array ↔️ `std::vector` ↔️ CUDA managed memory

Home Page: https://amypad.github.io/CuVec/

License: Other

CMake 13.90% Python 42.66% Cuda 37.77% SWIG 3.38% Makefile 2.30%

c cuda python array vector buffer cpython cpython-extensions cpython-api cpp

cuvec's People

Contributors

Stargazers

Watchers

Forkers

geogubd wangshuxihe00

cuvec's Issues

Installation fails in WSL environment

Unsuccessful attempt to install this package on a WSL2 Ubuntu system as a (AmyPET) dependency. Some required libraries are missing, perhaps?

System details:

(amypet_dev) user@host:/.../AmyPET$ uname -r
5.10.60.1-microsoft-standard-WSL2
(amypet_dev) user@host:/.../AmyPET$ python -V
Python 3.7.13
(amypet_dev) user@host:/.../AmyPET$ pip -V
pip 22.0.4 from /home/user/venvs/amypet_dev/lib/python3.7/site-packages/pip (python 3.7)

pip error trace:

(amypet_dev) user@host:/.../AmyPET$ pip install cuvec
Collecting cuvec
  Using cached cuvec-2.11.1.tar.gz (37 kB)
  Installing build dependencies ... done
  Getting requirements to build wheel ... done
  Preparing metadata (pyproject.toml) ... done
Building wheels for collected packages: cuvec
  Building wheel for cuvec (pyproject.toml) ... error
  error: subprocess-exited-with-error

  × Building wheel for cuvec (pyproject.toml) did not run successfully.
  │ exit code: 1
  ╰─> [83 lines of output]
      WARNING:cuvec.setup:Import or CUDA device detection error:
      NVML Shared Library Not Found
      Not searching for unused variables given on the command line.
      -- The C compiler identification is GNU 7.5.0
      -- Detecting C compiler ABI info
      -- Detecting C compiler ABI info - done
      -- Check for working C compiler: /usr/bin/cc - skipped
      -- Detecting C compile features
      -- Detecting C compile features - done
      -- The CXX compiler identification is GNU 7.5.0
      -- Detecting CXX compiler ABI info
      -- Detecting CXX compiler ABI info - done
      -- Check for working CXX compiler: /usr/bin/c++ - skipped
      -- Detecting CXX compile features
      -- Detecting CXX compile features - done
      -- Configuring done
      -- Generating done
      -- Build files have been written to: /tmp/pip-install-ty00gks4/cuvec_d337dca4039c4731a10fcf56a873fa80/_cmake_test_compile/build
      -- The C compiler identification is GNU 7.5.0
      -- The CXX compiler identification is GNU 7.5.0
      -- Detecting C compiler ABI info
      -- Detecting C compiler ABI info - done
      -- Check for working C compiler: /usr/bin/cc - skipped
      -- Detecting C compile features
      -- Detecting C compile features - done
      -- Detecting CXX compiler ABI info
      -- Detecting CXX compiler ABI info - done
      -- Check for working CXX compiler: /usr/bin/c++ - skipped
      -- Detecting CXX compile features
      -- Detecting CXX compile features - done
      CMake Error at /tmp/pip-build-env-4csnmbpo/overlay/lib/python3.7/site-packages/cmake/data/share/cmake-3.22/Modules/FindPackageHandleStandardArgs.cmake:230 (message):
        Could NOT find Python3 (missing: Python3_INCLUDE_DIRS Python3_LIBRARIES
        Development Development.Module Development.Embed) (found version "3.7.13")
      Call Stack (most recent call first):
        /tmp/pip-build-env-4csnmbpo/overlay/lib/python3.7/site-packages/cmake/data/share/cmake-3.22/Modules/FindPackageHandleStandardArgs.cmake:594 (_FPHSA_FAILURE_MESSAGE)
        /tmp/pip-build-env-4csnmbpo/overlay/lib/python3.7/site-packages/cmake/data/share/cmake-3.22/Modules/FindPython/Support.cmake:3166 (find_package_handle_standard_args)
        /tmp/pip-build-env-4csnmbpo/overlay/lib/python3.7/site-packages/cmake/data/share/cmake-3.22/Modules/FindPython3.cmake:490 (include)
        CMakeLists.txt:12 (find_package)
...
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for cuvec
Failed to build cuvec
ERROR: Could not build wheels for cuvec, which is required to install pyproject.toml-based projects

build: distribute wheels

Use cibuildwheel setup + options (e.g. https://github.com/benfred/implicit/blob/main/pyproject.toml) to precompile C++/CUDA?

support arrayviews

Improve support for https://github.com/xnd-project/arrayviews

separate cpu/cuda modules

current

cuvec.cuvec: at compile time, use cudaMallocManaged/cudaFree if available, otherwise use malloc/free
- at runtime, people with NVIDIA GPUs are forced to allocate memory on their device

proposed

cuvec.cuvec_cpu: uses malloc/free
cuvec.cuvec_cuda: if available, use cudaMallocManaged/cudaFree, otherwise ImportError
cuvec.cuvec: point to one of the above

pybind11: add native buffer protocol

pybind11::buffer_protocol() and .def_buffer([](NDCuVec &m) -> py::buffer_info { ... }) would simplify the pybind11 wrapping, vis. https://pybind11.readthedocs.io/en/stable/advanced/pycpp/numpy.html#buffer-protocol

add `__cuda_array_interface__`

This would allow easily adding to https://github.com/xnd-project/arrayviews (#23):

CPU arrays (already possible)
CUDA arrays, thereby supporting e.g.: cupy, numba, pyarrow, xnd, cudf, ...

Other references:

https://github.com/rapidsai

`asPyCuVec` type checks

Can check self.dtype for inconsistency with requested type T (safety/error catching)

Improve NumPy subclass

continue using an ndarray subclass
- https://numpy.org/doc/stable/reference/arrays.classes.html
- https://numpy.org/doc/stable/user/basics.subclassing.html
or use array containers
- https://numpy.org/doc/stable/user/basics.dispatch.html
define __array_function__?
define __array_ufunc__?

SWIG support

SWIG provides std_vector.i. CuVec could provide similar (e.g. cuvec.i) to support any language (not just Python) supported by SWIG

build: conda-forge

potential alternative to #27:

conda-forge (CUDA and non-CUDA) recipe for the parallelproj projectors (libparallelproj) [...]: https://github.com/conda-forge/parallelproj-feedstock

The main things I learned/had to reali[s]e where:

All your dependencies need to be on conda-forge

For handling cuda in the meta.yaml, there is # [cuda_compiler_version != "None"] "decorator" to describe cuda/non-cuda specific things

at runtime you need the virtual __cuda conda-forge package as dependency which helps the package manager to figure out which package version needs to be / can be installed on systems with and without cuda https://github.com/conda-forge/parallelproj-feedstock/blob/af55c92209af12cfad72f03078c24d0f0412045b/recipe/meta.yaml#L47

for the builds there are no physical GPUs available meaning that all tests requiring GPUs need to be skipped (this might change in the future)

In the build scripts, you can test the ${cuda_compiler_version} to trigger CUDA / non-cuda specific things https://github.com/conda-forge/parallelproj-feedstock/blob/af55c92209af12cfad72f03078c24d0f0412045b/recipe/build-parallelproj.sh#L7

I think the recipe for libparallelproj is a good starting point. Note that libparallelproj in the meta.yaml I build two packages (libparallelproj with CUDA and non-CUDA version, and also parallelproj which is a minimal python interface to the libs).

The conda-forge team on gitter.im also helped me a lot in the beginning.

— @gschramm

amypad / cuvec Goto Github PK

cuvec's People

Contributors

Stargazers

Watchers

Forkers

cuvec's Issues

current

proposed

Recommend Projects

Recommend Topics

Recommend Org

Jobs