fix to enable new wave creation for MI200. - #222
Conversation
|
Pre-commit issues. Please address them first. |
|
Please also add a commit message explaining what the change is doing. When merging for rocgdb we use "rebase + merge", which discards the PR description. So the PR description is only helpful for reviewers. |
aktemur
left a comment
There was a problem hiding this comment.
I don't understand well what problem we are addressing with this PR. Can you explain the problem?
lancesix
left a comment
There was a problem hiding this comment.
Some minor comments.
Note that the commit description missing, it should discuss what this fixes, why do we have an issue and so on, as well as describing the testing done to ensure this is compatible with all the GPUs we maintain compatibility for today.
b8c071a to
fba5ab1
Compare
that fails on MI200 for new waves creation subtest, the fix scales the hog kernel to occupy a significant portion of the GPU's concurrent wave execution capacity, regardless of GPU size.
fba5ab1 to
9f43f0f
Compare
|
Is this specific to MI200? I don't recall seeing this for MI210. I still don't think the problem description is clear here. What exactly are the failures and the failure mode? |
Motivation
step-schedlock-spurious-waves.expfails on MI200 for new waves creation subtest.Technical Details
Scale the hog kernel to occupy a significant portion of the GPU's
concurrent wave execution capacity, regardless of GPU size.
Test Plan
Run gdb.rocm step-schedlock-spurious-waves.exp on MI200.
Test Result
The step-schedlock-spurious-waves.exp should pass completely on MI200.