Skip to content

[solvers] Virtual GetSolution()#1426

Closed
averinevg wants to merge 4 commits into
developfrom
virtual_get_solution
Closed

[solvers] Virtual GetSolution()#1426
averinevg wants to merge 4 commits into
developfrom
virtual_get_solution

Conversation

@averinevg
Copy link
Copy Markdown
Contributor

No description provided.

@codecov

This comment was marked as off-topic.

@averinevg averinevg closed this Feb 16, 2022
@averinevg averinevg deleted the virtual_get_solution branch February 18, 2022 14:32
assistant-librarian Bot pushed a commit that referenced this pull request Oct 1, 2025
[MIOpen] Implement a couple of improvements for bwd and wrw
 CK kernel selection (#1426)

This PR includes a couple of changes to improve the selection of CK kernels in MIOpen. These include

1. _Add -1 to the list of splitk values to iterate over for wrw._

Prior to this change, splitk values could be the power of 2 values in
the range 1 to 128, but CK has a special splitk autodeduce value of -1
that in some cases can provide better performance. This PR adds -1 to
the set of splitk CK values, which improves the overall performance of
CK solvers.

4. _Get the workspace size from CK for grp wrw, grp bwd, and 3D grp
wrw._

MIOpen was setting the CK workspace size to zero if some instances did
not require a workspace for a particular shape, but this precluded some
CK instances that might require a workspace from even being considered.
And in some cases, the CK solver performance was reduced because the CK
instance that needed the workspace might actually have been the fastest
instance for that solver.  This PR changes MIOpen to get the workspace size from CK for the grp
wrw, grp bwd, and 3D grp wrw solvers, and no longer excludes kernels
that require a non-zero workspace.

## Test Plan

Run a large assortment of fwd, bwd, and wrw shapes using *all* solvers and ensure there is an
improvement in the CK-only solvers when compared to all solvers for wrw and no regressions for fwd and bwd.

## Test Result

- [x] Performance test results on Fremont shapes indicate CK-solver
performance has improved for wrw
- [x] CI passes
- [x] Run a large set of of the Tuna Gold shapes
- [x] Also need to wait for CK commit
29446da1d57170a8bda47a452113ef7e44363a04 to be staged/promoted
- [x] Investigate hanging kernel issue on bwd path

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
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.

1 participant