"illegal instruction" Error when using gsl::span as Kernel Parameter about gsl-lite

[mbeutel]

Thank you for the report. This was indeed a regression and should be fixed by 035f9d3.

We have compile-only CI for CUDA, but unfortunately this bug surfaced only at runtime. Although I don't think any CI vendor offers CUDA hardware in an OSS plan, we should at least add some proper runtime tests to the test suite.

---

(The bug was not specific to any CUDA version, but since you mentioned CUDA 9.2 I'll point out that we currently don't test (and hence, do not support) anything older than CUDA 11.0. Is there a pressing need to support versions older than 11.0?

Also, please don't include the <gsl/gsl> header, which is for MS-GSL compatibility only; use <gsl/gsl-lite.hpp> instead.)

[mbeutel]

Thanks. We used to have CI for CUDA 10.x, and I suppose it wouldn't be too difficult to reinstate that.

[pauleonix]

We have a server for developement and small simulations which uses compute_35 architecture GPUs, so as far as I understand, we wont be able to update much further than CUDA 11, as their architecture is already deprecated in 11. That server was still on CUDA 7 or 8 a few weeks ago, so 9 is already good progress for us ;)
The main reason, that we don't have 11 yet is that we chose the easy way of using the package manager. Debian Buster has a backport of CUDA 11, but it doesn't seem complete (compute-sanitizer didn't work for me when I tried it). I hope Debian Bullseye will get a release date soon.
On the cluster where I mainly work CUDA 11 was added as a module just this month...

So I would say that supporting the most recent 2 major CUDA versions might be more reasonable given how slow adaption is in some areas.