cuda-racecheck ignores error-exitcode option

Hi all,

I’m trying to integrate cuda-memcheck into our unit testing pipeline. Normally, running cuda-memcheck with

--error-exitcode 1

results in an exit code which is 0 if no errors were found, and 1 otherwise. This makes verifying the memory check very easy.

This works for all cuda-memcheck tools (including memcheck itself), except the racecheck tool. This tool seems to always exit with return code 0 no matter how many hazards (errors) were found.

Has anyone bumped up against this issue before? As a workaround, I’m looking to parse the text output of the racecheck tool and look for strings indicating that errors were found. However it’d be really nice to have consistent functionality across all tools.

From the description, this sounds like a bug in cuda-memcheck. Assuming you are using CUDA 7.0 (so this is not an old bug that may laready be fixed), I would suggest filing a bug report using the form linked from the CUDA registered developer website.

I’ve filed a bug internally with NVIDIA. You’re welcome to file your own of course on the developer portal. I don’t have any further details at this time except to say that I’ve witnessed and confirmed the reported behavior (in CUDA 7.5RC).

Glad to know I’m not the only one! I’ll go ahead and submit a bug report as well. Hopefully this way I’ll be notified if they do indeed fix it soon.