Data Parallel Extensions for Python*

Data Parallel Extensions for Python* extend numerical Python capabilities beyond CPU and allow even higher performance gains on data parallel devices such as GPUs. It consists of three related projects:

dpnp – Data Parallel Extensions for Numpy* - a library that implements a subset of Numpy that can be executed on any data parallel device. The subset is a drop-in replacement of core Numpy functions and numerical data types.
numba-dpex – Data Parallel Extensions for Numba* - extension for Numba compiler that enables programming data parallel devices the same way you program CPU with Numba.
dpctl – Data Parallel Control library that provides utilities for device selection, allocation of data on devices, tensor data structure along with Python Array API standard implementation, and support for creation of user-defined data-parallel extensions.

Learn more

Read more about Data Parallel Extensions for Python

Examples

Examples are located in ./examples. Their names start with the 2-digit number followed by a descriptive name. You can run examples in any order, however, if you are new to Data Parallel Extensions for Python, it is recommended to go in the order examples enumerated.

The following command will run the very first example of using Data Parallel Extensions for Python

python ./examples/01-hello_dpnp.py

Tutorials

Jupyter Notebook-based Getting Started tutorials are located in ./notebooks directory.

To run the tutorial, in the command line prompt type:

jupyter notebook

This will print some information about the notebook server in your terminal, including the URL of the web application (by default, http://localhost:8888):


$ jupyter notebook
[I 08:58:24.417 NotebookApp] Serving notebooks from local directory: /Users/catherine
[I 08:58:24.417 NotebookApp] 0 active kernels
[I 08:58:24.417 NotebookApp] The Jupyter Notebook is running at: http://localhost:8888/
[I 08:58:24.417 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).

It will then open your default web browser to this URL.

When the notebook opens in your browser, you will see the Notebook Dashboard, which will show a list of the notebooks, files, and subdirectories in the directory where the notebook server was started. Navigate to the notebook of your interest and open it in the dashboard.

For more information please refer to Jupyter documentation

Benchmarks

Data Parallel Extensions for Python provide a set of benchmarks illustrating different aspects of implementing the performant code with Data Parallel Extensions for Python. Benchmarks represent some real life numerical problem or some important part (kernel) of real life application. Each application/kernel is implemented in several variants (not necessarily all variants):

Pure Python: Typically the slowest and used just as a reference implementation
numpy: Same application/kernel implemented using NumPy library
dpnp: Modified numpy implementation to run on a specific device. You can use numpy as a baseline while evaluating the dpnp implementation and its performance
numba @njit array-style: application/kernel implemented using NumPy and compiled with Numba. You can use numpy as a baseline when evaluate numba @njit array-style implementat and its performance
numba @njit direct loops (prange): Same application/kernel implemented using Numba compiler using direct loops. Sometimes array-style programming is cumbersome and performance inefficient. Using direct loop programming may lead to more readable and performance code. Thus, while evaluating the performance of direct loop implementation it is useful to compare array-style Numba implementation as a baseline
numba-dpex @dpjit array-style: Modified numba @njit array-style implementation to compile and run on a specific device. You can use vanilla Numba implementation as a baseline while comparing numba-dpex implementation details and performance. You can also compare it against dpnp implementation to see how much extra performance numba-dpex can bring when you compile NumPy code for a given device
numba-dpex @dpjit direct loops (prange): Modified numba @njit direct loop implementation to compile and run on a specific device. You can use vanilla Numba implementation as a baseline while comparing numba-dpex implementation details and performance. You can also compare it against dpnp implementation to see how much extra performance numba-dpex can bring when you compile NumPy code for a given device
numba-dpex @dpjit kernel: Kernel-style programming, which is close to @cuda.jit programming model used in vanilla Numba
numba-mlir: Array-style, direct loops and kernel-style implementations for experimental MLIR-based backend for Numba
cupy: NumPy-like implementation using CuPy to run on CUDA-compatible devices
@cuda.jit: Kernel-style Numba implementation to run on CUDA-compatible devices
Native SYCL: Most applications/kernels also have DPC++ implementation, which can be used to compare performance of above implementations to DPC++ compiled code.

For more details please refer to dpbench documentation.

Demos

There are several demo applications illustrating the power of the Data Parallel Extensions for Python. They are:

Monte Carlo Pi - The Monte Carlo method to estimate the value of $\pi$.
Mandelbrot Set - Visualization of the breathtaking process of diving in the famous Mandelbrot fractal
Game of Life - Visualization of the life evolution using famous Conway's model

For more details please refer to the documentation located in the individual demo directory.

chudur-budur / DPEP