Open
Description
In an hacky attempt tried to enable the device and system memory scope support for AMD MI210 GPU reverting the commit
9947a81 to the 08b19b2 but it didn't work out for 2 GPU system. The same code I run at level_zero 2GPU failed with AMD GPU at random iterations due to inconsistencies.
I was wondering is there an easy way to have consistent device and system wide atomic memory_scope ?