[SYCL][CUDA] Alignment issue with zero-sized local arguments

[nmnobre]

Describe the bug

The alignment strategy for local arguments introduced with PR #5113 does not take into account zero-sized local arguments. Since, in such circumstances, the requested alignment is also zero (see sycl/plugins/cuda/pi_cuda.hpp):

const size_t alignment = std::min(max_alignment, size);,

the program receives signal SIGFPE, Arithmetic exception at runtime at the following modulus test:

localOffset % alignment != 0.

This problem is similar to that reported in issue #1467. I confess I understand little about local arguments or why the application I'm using uses zero-sized local arguments. Please do feel free to disregard this issue if this is just incorrect behaviour by the application.

Environment:

• OS: CentOS Linux release 8.5.2111
• Target device and vendor: NVIDIA Quadro GV100
• DPC++ version: ebb1281 and newer

[JackAKirk]

Hi @nmnobre

We hope that #5683 will fix your issue, but do you have your sample code that ran into the problem so that we can be certain that your particular use case is solved by the patch?

Thanks

[AerialMantis]

Related pull request for adding a test for zero-sized local accessors - intel/llvm-test-suite#1001

[AerialMantis]

This issue has been addressed more generically in the SYCL runtime by #6032, with this patch zero-sized local memory arguments are padded to 1 byte so this case is never encountered in the backends. @nmnobre I believe this issue to be resolved so I will close the ticket, however, if you feel an issue persists we can re-open it.