Repository navigation
[BUG]: nvcc -arch=all fails with cub/cub.cuh included due to CCCL_PP_SPLICE_WITH issue #4967
Description
Is this a duplicate?
- I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct
Type of Bug
Compile-time Error
Component
General CCCL
Describe the bug
When compiling with nvcc with <cub/cub.cuh> or <thrust/thrust.cuh> included and -arch=all, the compile will error with " error: macro "_CCCL_PP_SPLICE_WITH_IMPL1" passed 3 arguments, but takes just 2"
How to Reproduce
- Simple reproduction program:
#include <cub/cub.cuh> int main() { return 0; } - nvcc -E -arch=all repro_splice_bug.cu > repro_splice_bug.cu
- Note the error spam
/opt/cuda-12.9/bin/../targets/x86_64-linux/include/cub/util_macro.cuh:50:1: error: macro "_CCCL_PP_SPLICE_WITH_IMPL1" passed 3 arguments, but takes just 2
50 | CUB_NAMESPACE_BEGIN
| ^~~~~
- Review preprocessed code to confirm it ran out of tokens building the arch string:
namespace cub { inline namespace CUB_200802_SM_500_520_530_600_610_620_700_720_750_800_860_870_890_900_1000_1010_1030__CCCL_PP_SPLICE_WITH_IMPL1 {
Expected behavior
The preprocessor should be able to handle all known architectures and future architectures as well
Reproduction link
No response
Operating System
Slackware -current Linux, but looks to be universal.
nvidia-smi output
No response
NVCC version
nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Tue_May_27_02:21:03_PDT_2025
Cuda compilation tools, release 12.9, V12.9.86
Build cuda_12.9.r12.9/compiler.36037853_0
Metadata
Metadata
Assignees
Labels
Type
Projects
- Status
Activity
bernhardmgruber commentedon Jun 12, 2025
I am sorry I didn't expect we had so many CUDA arches already :D Thanks for the report!
Support more arguments to CCCL_PP_SPLICE_WITH
JKtheSlacker commentedon Jun 12, 2025
Just for posterity, I missed that _CCCL_PP_SPLICE_WITH_IMPL20 was skipped, which appears to have been the actual problem. The proposed fix handles that as well as more future architectures.
Support more arguments to CCCL_PP_SPLICE_WITH (#4972)
Support more arguments to CCCL_PP_SPLICE_WITH (NVIDIA#4972)
Support more arguments to CCCL_PP_SPLICE_WITH (NVIDIA#4972)
Support more arguments to CCCL_PP_SPLICE_WITH (#4972) (#4993)
Support more arguments to CCCL_PP_SPLICE_WITH (#4972) (#4995)
runtimes/cuda: downgrade cuda and cudnn
runtimes/cuda: downgrade cuda and cudnn (#269)
cudaPackages_12_9.cuda_cccl: patch to fix missing _CCCL_PP_SPLICE_WIT…
Work around CCCL 2.8.x arch token overflow (NVIDIA/cccl#4967)
Fix CUDA 13.0 compatibility + harden arch support matrix (#138)
Test the build on CUDA 13.4.1 and 12.9.2 (#560)