Hi Christian, On 7/24/23 00:14, Christian Kastner wrote:
On 2023-07-22 09:08, Cordell Bloor wrote:To me, this looks like hipGraphMemcpy1D had a broken implementation of host-to-host copies in ROCm 5.2.3. Unfortunately, this function was entirely reworked in ROCm 5.6.0 and the problem section no longer exists at all in the upstream project. If this is a real null pointer dereference bug, it will also affect Bookworm. :(Would you consider this more of a corner case, or rather as something more serious that needs to be fixed in bookworm? If the latter, we could probably fix it in the 12.2 point release.
It's not as bad as I thought. As far as I can tell, the null pointer dereference is unreachable even in 5.2.3-8. It seems that hipGraphLaunch will always return hipErrorOutOfMemory for graphs containing host-to-host copies.
On one hand, I think that HIP Graph will probably see increasing use over time and the lack of a host-to-host copy mechanism may be rather limiting. On the other hand, it is apparently used so infrequently that the upstream project took multiple releases to notice that it was completely broken.
I wouldn't reject a patch if someone did the work to backport this feature into Bookworm, but I don't plan to do it myself.
Sincerely, Cory Bloor