nvcc Error: Boost 1.64.0 & CUDA 9.0

[Alexander Korsunsky <a.korsunsky@…>]

Compilation of files that include Boost fail with the CUDA Toolkit version 9.

The reason for that is that at boost/config/compiler/nvcc.hpp, line 22 and 26 you use the macro __CUDACC_VER__ for version checking, but the symbol has been deprecated in CUDA 9.

Line 64 in crt/common_functions.h of the CUDA toolkit now reads:

#define __CUDACC_VER__ "__CUDACC_VER__ is no longer supported.  Use __CUDACC_VER_MAJOR__, __CUDACC_VER_MINOR__, and __CUDACC_VER_BUILD__ instead."

The fix seems simple enough, instead of checking for the whole version, you could check separately for __CUDACC_VER_MAJOR__ and __CUDACC_VER_MINOR__ .

[a.huebl@…]

migrated comment from github:

ouch, it might be worth opening a CUDA bug report as long as CUDA 9 is still in RC! (What's your exact nvcc version? 9.0.69?)

deprecating and removing this var are two things, this change in the header goes straight to "unannounced removal".

for the boost fix: is {{CUDACC_VER_MAJOR/MINOR/BUILD}} also available since CUDA 7.5? (I think yes.)

[a.huebl@…]

Are you sure you are using the official CUDA 9 RC (9.0.69) and not an older, early-access version? Or already a further update available since last week, 9.0.103?

[a.huebl@…]

The issue appeared in the 2nd RC, CUDA 9.0.103.

the OP and me opened tickets with nvidia in

​https://developer.nvidia.com/nvidia_bug/1971621 ​https://developer.nvidia.com/nvidia_bug/1971623

and are hoping the the define will be "deprecated" first for a release, so not all previous and current boost releases are rendered broken.

[anonymous]

Was this error resolved somehow?

[anonymous]

Is there any update on the issue? CUDA 9.0 is out, and no changes from the RC, it is still not working, NVIDIA as always avoiding dealing with issues. Is there a fix planned for this? I had to go back to an older version in order to be able to compile Caffe...

[sean.happel@…]

Potential patch

[sean.happel@…]

Example test code (fails to compile without patch)

[sean.happel@…]

I've just attached a potential patch, along with a short test program to showcase the bug. This is affecting multiple libraries, including caffe and tensorflow. Fedora 27 has centered on Boost 1.64, so this issue affects that distro's ability to do cuda development with the latest cuda version. I've submitted a bug request there <​https://bugzilla.redhat.com/show_bug.cgi?id=1530828> requesting the patch be applied downstream, but it would be really nice if this could be fixed in a hotfix for boost itself.

This link <​https://devtalk.nvidia.com/default/topic/858355/cuda_version-in-7-5rc-and-taking-different-code-paths-depending-on-nvcc-version-and-or-c-11-suppor/> indicates that the MAJOR/MINOR macros have been available since at least CUDA 7.5, so I think the patch should be a drop-in replacement.

[a.huebl@…]

This issue is fixed for Boost 1.65.1+

Please see: ​https://github.com/boostorg/config/pull/175

Unfortunately, the define to an error in nvcc / the CUDA toolkit was not changed in the final release of CUDA 9.0... so you have to take a newer version of boost or patch your old release with the same diff as linked.