Description
Problem Description
Hi,
As indicated at #104 , I have seen that the current implementation of interprocess events is suboptimal. In particular, the fact that it uses a stream callback to decrement the internal flags causes the CPU to not be able to queue further GPU work ahead.
I believe that the implementation could be made better by using monotonically increasing counters and stream memory wait operations. Using such monotonically increasing counters would remove the need for decreasing the counters with the CPU side as the CPU could just check if the counter value is greater than desired.
This is quite important to fix this as interprocess events are quite used for the multi process multi gpu applications.
Best regards,
Epliz
Operating System
Ubuntu 22.04
CPU
Intel xeon
GPU
MI300x
ROCm Version
ROCm 6.3.0
ROCm Component
No response
Steps to Reproduce
No response
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support
No response
Additional Information
No response