Carthesian Product error with nvc++

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
,
1

Hi, I think I found a bug with the cartesian product header from this tutorial and parallel STL algorithms with nvc++.

When I use the par_unseq policy on for_each to iterate on a virtual (1D vector) 3D tensor with the cartesian product, it seems that I encounter data corruption. I don’t encounter this error with a sequential policy using nvc++ or with the par_unseq policy using gcc and tbb.

I made the following reproducer.

Let me know if you need more details about this potential bug.

edit :
Configuration :

  • HPC SDK 24.1, 24.3, 24.5
  • CUDA 12.3, 12.4
  • Linux x86_64
1 Like
2

Hi - I have a colleague who is going to look at this and get back to you on it. It will probably be next week before he’s able to do so, but I just wanted to give you a quick update that we’re going to check out the issue and investigate! Thank you for making us aware. We’ll update you when we know more.

3

Perfect, thanks for your help.

4

We’ve confirmed that this is an error in cartesian product where it can return an off-by-one value when offloading. In this case the “y” values returned are 2, 2, 3, 3, 4, 4 when they should be 2, 2, 2, 3, 3, 3.

We’ve added an issue report, NVBugs #4656068, and will have engineering investigate.

Thanks for the report!

5

Hi, I just applied the latest bug fix from the tutorial, and it works now.

I updated the reproducer.

Thanks for your help.

1 Like
6

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.