CUDA Core Compute Libraries (CCCL) for Python

13.2.27 · active · verified Sat Apr 11

The `nvidia-cuda-cccl` Python package provides Pythonic interfaces to the NVIDIA CUDA Core Compute Libraries (CCCL), specifically CUB and Thrust. CCCL itself unifies essential CUDA C++ libraries (Thrust, CUB, libcudacxx) to offer building blocks for writing safe and efficient CUDA C++ code. This Python binding enables developers to leverage GPU-accelerated parallel algorithms and cooperative primitives directly from Python, easing the implementation of custom algorithms without needing to drop down to C++. The library is currently in 'experimental' status, meaning its API and feature set can evolve rapidly.

Warnings

breaking The `cuda-cccl` Python package is in 'experimental' status, meaning its API and feature set can change rapidly. Users should anticipate potential breaking changes between minor versions.
Fix: Always refer to the latest documentation and release notes before upgrading. Pin exact versions in production environments.
breaking CCCL is not forward compatible with the CUDA Toolkit. An older version of CCCL will not be compatible with a newer CUDA Toolkit. However, a newer CCCL can generally be integrated with an older CUDA Toolkit.
Fix: Ensure that your CCCL version is the same or newer than the version included with your CUDA Toolkit installation. Always use the latest compatible CCCL version for your CUDA Toolkit.
breaking For C++ users, CCCL 3.0 (bundled with CUDA Toolkit 13.0) introduced significant breaking changes, including dropping support for C++11/14, CUDA Toolkit versions prior to 12.0, and older host compilers (GCC < 7, Clang < 14, MSVC < 2019). It also removed support for ICC and CUDA Dynamic Parallelism v1.
Fix: Update your C++ standard to C++17 or newer, ensure your CUDA Toolkit is 12.0+, and use compatible host compilers. Consult the CCCL 2.x to 3.0 Migration Guide for detailed steps.
breaking Starting with CUDA Toolkit 13.0 (which includes CCCL 3.0), the on-disk header location for CCCL in the CUDA Toolkit installation moved to `${CTK_ROOT}/include/cccl/`. Also, the `nvidia-cuda-cccl` PyPI package will no longer use `cuXX` suffixes (e.g., `nvidia-cuda-cccl-cu12`) to allow upstream libraries to select the CUDA version.
Fix: Update build systems to reflect the new header paths. For Python installations, prefer `pip install cuda-cccl[cuXX]` as specified, or `pip install cuda-cccl` for the general version when CUDA 13.0+ is in use, allowing the upstream to manage the CUDA version selection.
gotcha When compiling CUDA C++ code with `nvcc`, it automatically adds CCCL headers to your include path. However, if compiling with other compilers, you must manually update your build system's include search path to point to the CCCL headers.
Fix: For non-nvcc compilers, explicitly add the path to CCCL headers (e.g., `/usr/local/cuda/include` or `${CTK_ROOT}/include/cccl/`) to your build system's include paths (`-I`). Avoid `-isystem` and use `-I` to prevent collisions with implicitly included headers.

Install

pip install cuda-cccl[cu13] # For CUDA 13.x Pip (CUDA 13.x)
pip install cuda-cccl[cu12] # For CUDA 12.x Pip (CUDA 12.x)
conda install -c conda-forge cccl-python Conda (conda-forge)

Imports

compute
```
import cuda.compute
```
Provides device-level parallel algorithms like reduce, scan, and sort.
coop
```
import cuda.coop
```
Provides block and warp-level cooperative primitives for custom CUDA kernels.

Quickstart

This example demonstrates how to perform a parallel reduction sum on a CuPy array using `cuda.compute.reduce`. It showcases a basic use case of the Pythonic interface to GPU-accelerated algorithms.

import cuda.compute
import cupy as cp

def reduce_sum_example():
    # Initialize CuPy array on GPU
    arr = cp.arange(1000, dtype=cp.int32)
    
    # Perform a reduction sum using cuda.compute
    result = cuda.compute.reduce(arr, op='sum')
    
    print(f"Original array sum (CuPy): {arr.sum()}")
    print(f"Reduced sum (cuda.compute): {result}")
    assert result == arr.sum()

if __name__ == "__main__":
    reduce_sum_example()

view raw JSON →