psath

Thanks for the suggestion, I'll see if I can pin down what's corrupting that state.

In practice the offset is zero when the device id is getting corrupted. Will be tracing back the corruption after some meetings today. However that's an extremely useful correction to...

I'm still poking at this as I have time. I've worked around the assertion failure by explicitly doing a copy command group. However, my debug build of the hipSYCL runtime...

That's... extremely disconcerting and off-topic. Have we met? While I'm not exactly trying to anonymize my profile (it has papers associated with it after all), this is not an appropriate...

All signs point to it being some issue with host-associated buffers, as long as I don't initialize the buffer with a host pointer (neither stack nor heap), and instead just...

> The address sanitizer may have trouble with GPU builds as in that case you have external devices randomly writing to memory and lots of things going on at kernel...

> Thanks for the update! If there are no host pointers, there will be no writeback and the destruction of the buffer will be much simpler. > > I've just...

> I wonder if it causes an issue if we submit operations for the buffer when the buffer is already in the destructor. But I have just checked, the actual...

> Can you please check if the issue persists with PR #395? Rebuilding and running now. When I tested with the attached MWE last night (pre #395), without optimization flags...

Yep #395 resolves it in the MWE. The USE_HOST_PTR version with and without the flush passed all 1000 tests (and reliably copied the data back to the host) Will try...