Python bindings to Kokkos Views for data interop
Project description
pykokkos-base
This package contains the minimal set of bindings for Kokkos interoperability with Python:
Kokkos::initialize(...)
Kokkos::finalize()
Kokkos::View<...>
Kokkos::DynRankView<...>
By importing this package in Python, you can pass the supported Kokkos Views and DynRankViews from C++ to Python and vice-versa. Furthermore, in Python, these bindings provide interoperability with numpy and cupy arrays:
import kokkos
import numpy as np
view = kokkos.view([2, 2], dtype=kokkos.double, space=kokkos.CudaUVMSpace,
layout=Kokkos.LayoutRight, trait=kokkos.RandomAccess,
dynamic=False)
arr = np.array(view, copy=False)
This package depends on a pre-existing installation of Kokkos
Writing Kokkos in Python
In order to write native Kokkos in Python, see pykokkos.
Installation
You can install this package via CMake or Python's setup.py
. There are two important cmake options:
ENABLE_LAYOUTS
ENABLE_MEMORY_TRAITS
By default, CMake will enable these options if the Kokkos installation was not built with CUDA support. If Kokkos was built with CUDA support, these options will be disabled by default due to unreasonable compilation times (> 1 hour).
Configuring Options via CMake
cmake -DENABLE_LAYOUTS=ON -DENABLE_MEMORY_TRAITS=OFF /path/to/source`
Configuring Options via setup.py
There are three ways to configure the options:
- Via the Python argparse options
--enable-<option>
and--disable-<option>
- Setting the
PYKOKKOS_BASE_SETUP_ARGS
environment variable to the CMake options - Passing in the CMake options after a
--
All three lines below are equivalent:
python setup.py install --enable-layouts --disable-memory-traits
PYKOKKOS_BASE_SETUP_ARGS="-DENABLE_LAYOUTS=ON -DENABLE_MEMORY_TRAITS=OFF" python setup.py install
python setup.py install -- -DENABLE_LAYOUTS=ON -DENABLE_MEMORY_TRAITS=OFF
Configuring Options via pip
Pip does not handle build options well. Thus, it is recommended to use the PYKOKKOS_BASE_SETUP_ARGS
environment variable noted above. However, using the --install-option
for pip is possible but
each "space" must have it's own --install-option
, e.g. all of the following are equivalent:
All three lines below are equivalent:
pip install pykokkos-base --install-option=--enable-layouts --install-option=--disable-memory-traits
pip install pykokkos-base --install-option=-- --install-option=-DENABLE_LAYOUTS=ON --install-option=-DENABLE_MEMORY_TRAITS=OFF
pip install pykokkos-base --install-option={--enable-layouts,--disable-memory-traits}
pip install pykokkos-base --install-option={--,-DENABLE_LAYOUTS=ON,-DENABLE_MEMORY_TRAITS=OFF}
pip install pykokkos-base
will build against the latest release in the PyPi repository. In order to pip install from this repository, usepip install --user -e .
Example
Overview
This example is designed to emulate a work-flow where the user has code using Kokkos in C++ and writes python bindings to those functions. A python script is used as the "main"
:
ex-numpy.py
imports the kokkos bindings- Calls a routine in the "users" python bindings to a C++ function which returns a
Kokkos::View
- This view is then converted to a numpy array in python and printed via the numpy capabilities.
Files
- ex-generate.cpp
- This is the python bindings to the user code
- user.cpp
- This is the implementation of the user's code which returns a
Kokkos::View<double**, Kokkos::HostSpace>
- This is the implementation of the user's code which returns a
- ex-numpy.py
- This is the "main"
ex-numpy.py
import argparse
import gc
import numpy as np
#
# The python bindings for generate_view are in ex-generate.cpp
# The declaration and definition of generate_view are in user.hpp and user.cpp
# The generate_view function will return a Kokkos::View and will be converted
# to a numpy array
from ex_generate import generate_view
#
# Importing this module is necessary to call kokkos init/finalize and
# import the python bindings to Kokkos::View which generate_view will
# return
#
import kokkos
def main(args):
# get the kokkos view
view = generate_view(args.ndim)
# verify the type id
print("Kokkos View : {}".format(type(view).__name__))
# wrap the buffer protocal as numpy array without copying the data
arr = np.array(view, copy=False)
# verify type id
print("Numpy Array : {} (shape={})".format(type(arr).__name__, arr.shape))
# demonstrate the data is the same as what was printed by generate_view
for i in range(arr.shape[0]):
print(" view({}) = {}".format(i, arr[i]))
if __name__ == "__main__":
kokkos.initialize()
parser = argparse.ArgumentParser()
parser.add_argument("-n", "--ndim", default=10,
help="X dimension", type=int)
args = parser.parse_args()
main(args)
# make sure all views are garbage collected
gc.collect()
kokkos.finalize()
Build and Run
mkdir build
cd build
cmake -DBUILD_EXAMPLES=ON ..
make
python ./ex-numpy.py
Expected Output
view(0) = 0 0
view(1) = 0 1
view(2) = 2 0
view(3) = 0 3
view(4) = 4 0
view(5) = 0 5
view(6) = 6 0
view(7) = 0 7
view(8) = 8 0
view(9) = 0 9
Sum of view: 45
extent(0): 10
stride(0): 2
Kokkos View : KokkosView_HostSpace_double_2
Numpy Array : ndarray (shape=(10, 2))
view(0) = [0. 0.]
view(1) = [0. 1.]
view(2) = [2. 0.]
view(3) = [0. 3.]
view(4) = [4. 0.]
view(5) = [0. 5.]
view(6) = [6. 0.]
view(7) = [0. 7.]
view(8) = [8. 0.]
view(9) = [0. 9.]
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
File details
Details for the file pykokkos-base-0.0.2.tar.gz
.
File metadata
- Download URL: pykokkos-base-0.0.2.tar.gz
- Upload date:
- Size: 601.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/3.1.1 pkginfo/1.7.0 requests/2.25.1 setuptools/56.2.0 requests-toolbelt/0.9.1 tqdm/4.60.0 CPython/3.7.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 489251b5d4674f5eede24288242c2d061299a76439aa6c4653ea416821169581 |
|
MD5 | 8eac58a5a47fc1bf21dfd653789dacd0 |
|
BLAKE2b-256 | e91b6fb7ba78b02cb5d00010b52f7a0abf0cad6444b4053077b0aea6de9b4e71 |