[Flang][OpenMP] Fix one part of the descriptor type performance regression noted in LC-1761#2413
Open
agozillon wants to merge 1 commit into
Open
[Flang][OpenMP] Fix one part of the descriptor type performance regression noted in LC-1761#2413agozillon wants to merge 1 commit into
agozillon wants to merge 1 commit into
Conversation
Collaborator
Collaborator
|
killing npsdb, will restart in 2 hours |
Collaborator
|
!PSDB |
Collaborator
…ssion noted in LC-1761 This PR tries to fix the second performance regression noted in LC-1761 by removing the OMP_MAP_OBJ_AND_PTR flag that was put in place to work-around an unintended early write back issue in another PR and then addressing that breakage a different way, by removing the application of MEMBER_OF forced on pointers by the declare mapper infrastructure that does not respect MEMBER_OF being omitted from an entry, it will ALWAYS apply it without this PR (which conceptually makes sense except that pointers can belong to structures, and they can have data, that isn't part of the structure, but is part of the overall mapping). These changes take us through a path in the runtime that avoids forced re-maps caused by MAP_OBJ_AND_PTR. NOTE: This PR only applies to Fortran, as I've attempted to restrict the declare mapper infrastrcuture to only apply this tweak for fortran via an option only passed through the Flang path, it should default to off for Clang. There's a larger rework of this for attach maps in progress by Intel so this PR is likely a stop-gap for the more permanent solutiuon. Although, they intend to fix it upstream in a similar manner, just with a lot more functionality (discussed the solution a bit in e-mail with Abhinav at Intel, we agree this fix is the way to go for the attach map style of pointer mapping).
agozillon
force-pushed
the
amd/dev/agozillo/performance-regression-fix-1
branch
from
63726cd to
d22a671
Compare
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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 tries to fix the second performance regression noted in LC-1761 by removing the OMP_MAP_OBJ_AND_PTR flag that was put in place to work-around an unintended early write back issue in another PR and then addressing that breakage a different way, by removing the application of MEMBER_OF forced on pointers by the declare mapper infrastructure that does not respect MEMBER_OF being omitted from an entry, it will ALWAYS apply it without this PR (which conceptually makes sense except that pointers can belong to structures, and they can have data, that isn't part of the structure, but is part of the overall mapping). These changes take us through a path in the runtime that avoids forced re-maps caused by MAP_OBJ_AND_PTR.
NOTE: This PR only applies to Fortran, as I've attempted to restrict the declare mapper infrastructure to only apply this tweak for fortran via an option only passed through the Flang path, it should default to off for Clang.
There's a larger rework of this for attach maps in progress by Intel so this PR is likely a stop-gap for the more permanent solution. Although, they intend to fix it upstream in a similar manner, just with a lot more functionality (discussed the solution a bit in e-mail with Abhinav at Intel, we agree this fix is the way to go for the attach map style of pointer mapping).