[BUG]: <cuda/atomic> header should be included only in device compilation mode.

[mfbalin]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug (https://github.com/NVIDIA/cuCollections/issues)

Type of Bug

Compile-time Error

Describe the bug

I want to use cuCollections when compiling for old and new CUDA architectures (sm_35 to sm_90). However, in this case, the minimum cuda architecture is sm_35 and importing cuda/atomic header gives an error.

cuCollections/include/cuco/detail/open_addressing/kernels.cuh

Line 22 in 2101cb3

#include <cuda/atomic>

If there is a way, this include should be performed only in device mode and when compiling for suitable cuda architectures. Otherwise, one can't use cuco::static_map in the host code. I want to dispatch to different kernels depending on the compute capability of the device. Ideally, without limiting which architectures I am providing support for (>=35).

How to Reproduce

1. Compile for old and new architectures.
2. Include cuco/static_map.cuh header.

Expected behavior

The host code should be compilable for any CUDA architecture.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[jrhemstad]

Hey @mfbalin this is a known issue and something high on my list of things I'd like to address in CCCL in the coming months. See NVIDIA/cccl#1083 for tracking this effort.

[mfbalin]

Can I hack __CUDA_ARCH_LIST__ to remove old architectures from it to compile a specific file as a workaround?

[mfbalin]

I want to compile the whole project with 35, 50, ..., 90 but only one file including cuco with 70, ..., 90.

[jrhemstad]

I want to compile the whole project with 35, 50, ..., 90 but only one file including cuco with 70, ..., 90.

Are you using CMake? You should be able to specify per-target properties for a specific file like this:

cmake_minimum_required(VERSION 3.18)
project(cuda_arch_example LANGUAGES CUDA)

# Add your library
add_library(my_cuda_library main.cu other.cu)

# Set the default CUDA architecture for the target
set_target_properties(my_cuda_library PROPERTIES CUDA_ARCHITECTURES "35")

# Specify a different CUDA architecture for a specific file
set_source_files_properties(other.cu PROPERTIES CUDA_ARCHITECTURES "72")

[mfbalin]

I want to compile the whole project with 35, 50, ..., 90 but only one file including cuco with 70, ..., 90.

Are you using CMake? You should be able to specify per-target properties for a specific file like this:

cmake_minimum_required(VERSION 3.18)
project(cuda_arch_example LANGUAGES CUDA)

# Add your library
add_library(my_cuda_library main.cu other.cu)

# Set the default CUDA architecture for the target
set_target_properties(my_cuda_library PROPERTIES CUDA_ARCHITECTURES "35")

# Specify a different CUDA architecture for a specific file
set_source_files_properties(other.cu PROPERTIES CUDA_ARCHITECTURES "72")

Yes, I am using CMake. This looks promising, let me try it. Thank you!

[jrhemstad]

@mfbalin turns out you can't use CUDA_ARCHITECTURES on a per-file basis. You'd need to move other.cu to a separate object library.

[mfbalin]

That works too, so long as there is a way to get it working.

[mfbalin]

@jrhemstad Do you know if these separate object libraries can circularly depend on each other?

EDIT: I will figure out a way. Thank you for the pointers!

[PointKernel]

Closing as resolved