Skip to content

Conversation

awlauria
Copy link
Contributor

No description provided.

@awlauria
Copy link
Contributor Author

Opened as an alternative to #8574

@awlauria awlauria force-pushed the atomics_core_2 branch 2 times, most recently from c0754f3 to 573352b Compare March 12, 2021 21:05
Signed-off-by: Austen Lauria <awlauria@us.ibm.com>
Choose the read-modify-write path over the atomics
path when the address isn't aligned.

Signed-off-by: Austen Lauria <awlauria@us.ibm.com>
It's a common function between rdma and ucx, put it
somewhere both can access it.

Signed-off-by: Austen Lauria <awlauria@us.ibm.com>
Signed-off-by: Austen Lauria <awlauria@us.ibm.com>
Copy link
Member

@hjelmn hjelmn left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@devreal
Copy link
Contributor

devreal commented Mar 14, 2021

I'm not sure I understand how this deals with misaligned atomics at all levels. We might still end up doing misaligned CPU atomics in mca_btl_base_am_atomic_64, no?

@hjelmn
Copy link
Member

hjelmn commented Mar 14, 2021

The BTL atomics should never be used with unaligned values. If we enforce that requirement this is sufficient since we have no other path that would attempt it.

@awlauria
Copy link
Contributor Author

@devreal the test where I found this, c_reqops.c, passes with this change as well as #8574.

@awlauria awlauria merged commit 5cbc19e into open-mpi:master Mar 15, 2021
@awlauria awlauria deleted the atomics_core_2 branch March 15, 2021 02:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants