Rework bli_packm_blk_var1. #707
Merged
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
Separate the dense code (general/hermitian/symmetric/triangular fully-stored part) from the triangular, diagonal-intersecting code. This allows a more consistent usage of round-robin thread scheduling (even for the dense micro-panels) for the latter.
|
@fgvanzee let me know when you've checked it. |
|
@devinamatthews Looks good. Unless there's more you want to do here, I'll squash with a new commit log entry once CI finishes. |
fgvanzee
added a commit
that referenced
this pull request
May 20, 2024
Details: - Factored some of the structure awareness out of the loop in bli_packm_blk_var1(). So instead of having a single loop with conditionals in the body to handle various kinds of structure (and stored/unstored submatrix placement), we now have a conditional branch to handle various structure/storage scenarios with a loop in each section. This change was originally motivated to choose slab or round- robin partitioning (in the context of triangular matrices) based on the structure of the entire block (or panel) being packed rather than each micropanel individually. Previously, the code would attempt to limit rr to the portion of the block that intersects the diagonal and use slab for the remainder. However, that approach was not well-thought out and in many situations this would lead to inferior load balancing when compared to using round-robin for the entire block (or panel). This commit has the added benefit of incurring less overhead during the packing process now that each of the new loops is simpler. - (cherry picked from commit b6735ca) Switch to l3 sup decorator in gemmlike sandbox. (#704) Details: - Modified the gemmlike sandbox to call bli_l3_sup_thread_decorator() rather than a local analogue of that code. This reduces redundant logic and makes it easier for the sandbox to inherit future improvements to the framework's threading code. - Moved addon/gemmd to addon/old/gemmd. This code has fallen out of date and is taking too much effort to maintain. We will very likely reimplement it completely once future changes are made to the framework proper. - (cherry picked from f956b79)
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.
Separate the dense code (general/hermitian/symmetric/triangular fully-stored part) from the triangular, diagonal-intersecting code. This allows a more consistent usage of round-robin thread scheduling (even for the dense micro-panels) for the latter.