WIP : MI300A Support #170
Open
WIP : MI300A Support #170
+205
−46
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
|
|
Thanks Paul. I will take a look at this shortly. In the mean time, could you check the metadata for commit 8fbf57e? It looks like there are two names on that commit, and it's confusing the license approval mechanism. |
PaulMullowney
force-pushed
the
MI300A
branch
from
8fbf57e
to
54048af
Compare
|
I rebased against the current develop and squashed all commits into 1. |
PaulMullowney
force-pushed
the
MI300A
branch
from
54048af
to
a04594a
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR is a work in progress however I wanted to get feedback sooner rather than later.
For optimal performance, MI300A requires that buffers used in GPU aware MPI be allocated via hipMalloc. To implement this, I added new public APIs: RESERVE_GAM and GET_ALLOCATION_GAM. Any device buffer which is used in GPU aware communication needs to be allocated through these methods.
The buffered and growing allocator classes get new data members/methods that mimic the original implementation. The distinction is that once it is time to perform an allocation deep in REALLOCATE_GROWING_GAM_ALLOCATION in growing_allocator_mod.F90, the code does a hipMalloc for AMD hardware. For Nvidia we use the original ALLOCATE and ACC ENTER DATA CREATE pathway.