CUDA CCCL integration

jwm · October 8, 2025, 8:17pm

I recently tried to upgrade my CUDA version from 12.8.0 to 13.1 (then dropped back to 12.9.1). Apparently I’ve been hit by the upgrade to the CCCL. My immediate symptom is that <thrust/xxx> is no longer found.

At the same time, I downgraded CMake to 3.26.5, since a package was available for RHEL9. If this is fixed in a later version, I can upgrade.

As usual, CUDA code compiled by nvcc builds just fine, but C++ code built by the native compiler (in my case g++ 13.3.1) can’t find the thrust headers.

It seems that I need to add CCCL::CCCL as a link-library to my target, but I haven’t come across any examples of how to just use it. I don’t understand why this isn’t part of

The CCCL has an example project that talks about including CCCL as an external project; I don’t need to do that, since it is part of the CUDA Toolkit in my nvidia base image.

Can anyone (and I’m hoping @robert.maynard will read this) give me the “proper” way to add CCCL to a project that uses CUDA?

LegalizeAdulthood · October 8, 2025, 9:41pm

Did you read through these instructions?

github.com/NVIDIA/cccl

lib/cmake/thrust/README.md

main

# Using Thrust with CMake

Thrust provides configuration files that simplify using Thrust
from other CMake projects. Requirements:

- Thrust >= 1.9.10
- CMake >= 3.15

See the [Fixing Legacy FindThrust.cmake](#fixing-legacy-findthrustcmake)
section for solutions that work on older Thrust versions.

## Quick Start: Use CCCL Example

The CCCL repository contains a [stand-alone example](https://github.com/NVIDIA/cccl/tree/main/examples/thrust_flexible_device_system) project the demonstrates how to use Thrust with any device system (CUDA, OpenMP, TBB, or CPP) from CMake using the CCCL package.
The choice of device system is flexible and determined at configure-time via the `CCCL_THRUST_DEVICE_SYSTEM` CMake variable, which initializes the default `CCCL::Thrust` target's configuration.

If the default `CCCL::Thrust` target is not desired and your project would like more control over the Thrust configuration, set the CMake variable `CCCL_ENABLE_DEFAULT_THRUST_TARGET=OFF` and use the `thrust_create_target` function manually, as described in the following document.

Note that `find_package(CCCL)` internally calls `find_package(Thrust)`. These may be used interchangeably if you're only interested in the Thrust component of CCCL.

This file has been truncated. show original

jwm · October 8, 2025, 9:46pm

I had not found those instructions, but I stumbled into most of it. The

thrust_create_target(Thrust)

part I did not do (although I found it in the Thrust cmake file; I wasn’t sure what to do with it). That was the ticket. Thanks!

robert.maynard · October 9, 2025, 1:41pm

Also see the CCCL migration guide: CCCL 2.x ‐ CCCL 3.0 migration guide — CUDA Core Compute Libraries

jwm · October 9, 2025, 3:45pm

I did see that, but it was missing the critical README buried down in cmake/lib/thrust (I poked around that tree and that’s the only README down there).