The "__CUDA_ARCH_LIST__" macro dosen't work on Jetson Xavier NX with nvcc 11.4

[Lurvelly]

Hi All,

I am trying to use cuCollections library on my Jetson Xavier NX with customer key, but get the error that said "build/_deps/cuco-src/include/cuco/static_map.cuh(157): error: static assertion failed with "A key type larger than 8B is supported for only sm_70 and up."

As all known, the Jetson Xavier NX is Volta and support sm_72, so I double checked my CMakelist.txt, I am sure I have added these " --generate-code arch=compute_72,code=sm_72" in the CMAKE_CUDA_FLAGS.

Then I found that the "CUDA_ARCH_LIST" macro dosen't work, that's why there is no define "CUCO_HAS_INDEPENDENT_THREADS".

cuCollections/include/cuco/detail/__config

Line 22 in 2dbcfc7

#define CUCO_CUDA_MINIMUM_ARCH _NV_FIRST_ARG(__CUDA_ARCH_LIST__)

cuCollections/include/cuco/detail/__config

Lines 32 to 34 in 2dbcfc7

	#if (CUCO_CUDA_MINIMUM_ARCH >= 700)
	#define CUCO_HAS_INDEPENDENT_THREADS
	#endif

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_May__4_00:02:26_PDT_2022
Cuda compilation tools, release 11.4, V11.4.239
Build cuda_11.4.r11.4/compiler.31294910_0

[PointKernel]

Thanks for reporting the issue.

Related to #245

IIRC, __CUDA_ARCH_LIST__ is supported by nvcc 11.5 or plus. @Lurvelly Is it possible to install 11.5 on your Jetson?

cc @sleeepyjack

[Lurvelly]

Hi, @PointKernel

Thanks for your reply.

I will try to check if my Jetson can upgrade the nvcc than 11.5. If not, I would like to use the #define CUCO_CUDA_MINIMUM_ARCH 720 to replace current macro at

cuCollections/include/cuco/detail/__config

Line 22 in 2dbcfc7

#define CUCO_CUDA_MINIMUM_ARCH _NV_FIRST_ARG(__CUDA_ARCH_LIST__)

It seems to be the simplest way to solve my issue, isn't it? If I did, will there be another issue in the cuCollections ?

[PointKernel]

I would like to use the #define CUCO_CUDA_MINIMUM_ARCH 720 to replace current macro

Yeah, that should work just fine. (Somehow I remember a valid CUDA arch has 2 digits only, so it would be either 720 or 72, yours to discover!)

will there be another issue in the cuCollections?

Hopefully not, do let us know if you have any further issues. 😄

[Lurvelly]

Yeah, that should work just fine. (Somehow I remember a valid CUDA arch has 2 digits only, so it would be either 720 or 72, yours to discover!)

I have double checked the CUDA arch on my computer, it should be 720.

Hopefully not, do let us know if you have any further issues.

Ok, I will test the cuCollections on my Jetson, hope it works fine. And if I meet another issue about cuCollections, I will let you know asap.

Thanks again for your reply.

Best Regards

Louis Li