Repository navigation
[rocjitsu] Distribute strided VM policy lookups across a larger cache - #13809
Merged
Merged
Conversation
Use 128 policy snapshots and offset successive page blocks to reduce repeated replacement on strided memory accesses while preserving short contiguous runs across block boundaries. Share the index calculation with both private-memory hint lookups and preserve range, generation and mutation-token validation. Cover contiguous pages, tiled rows and large strides, plus distinct private policies and stale-hint rejection after mutation. On a Threadripper PRO 9995WX with fixed physical CPU affinity and no CPU reservation: - gfx1100 FP32 Tensile: median warm GEMM time falls 31.9%, from 3.314 to 2.258 seconds across three process repeats. Whole-process CPU cycles fall 38.5% and L2 data-request misses fall 36.9%. - Across 37 numerical ML smoke cells, summed process time falls 4.5%; the equal-cell geometric mean falls 2.5%. This combines screening runs with repeated comparisons and is not a production workload estimate. - The unchanged hipRAND MT19937 initialization case is effectively neutral: 141.4 to 142.2 seconds across two process repeats. Four fresh comparisons do not reproduce the initially flagged Gluon FP8 or integrated Tensile GSU slowdowns. Small differences remain uncertain on the shared host. Validation: 496 focused tests and all 37 numerical smoke cells pass. Completed two local peanut-review rounds.
❌ PR Check — Action Required
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🚫 Please fix the failed policies before requesting reviews. The following policy checks failed:
The |
atgutier
approved these changes
Oct 20, 2026
This was referenced Oct 21, 2026
Contributor
TheRock Submodule Bump ActivityNewest first
|
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
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.
Use 128 policy snapshots and offset successive page blocks to reduce repeated replacement on strided memory accesses while preserving short contiguous runs across block boundaries. Share the index calculation with both private-memory hint lookups and preserve range, generation and mutation-token validation.
Cover contiguous pages, tiled rows and large strides, plus distinct private policies and stale-hint rejection after mutation.
On a Threadripper PRO 9995WX with fixed physical CPU affinity and no CPU reservation:
Four fresh comparisons do not reproduce the initially flagged Gluon FP8 or integrated Tensile GSU slowdowns. Small differences remain uncertain on the shared host.
Validation: 496 focused tests and all 37 numerical smoke cells pass. Completed two local peanut-review rounds.