Make matmulsm mergeable (Fixes #1407) #1417
Merged
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Outdated suggestions cannot be applied.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.
As promised in #1407 I spent some time on making MATMULSM mergeable.
For most protocols, the new implementation will merge as much communication as possible while respecting the batch size.
In order for the allocator to be able to determine the memory addresses accessed by a MATMUSLM, I needed to add some temporary fields to the instruction that contain the used rows and columns. This is somewhat hacky, but it's the cleanest solution I found.
If those fields are not present, the allocator will fall back to the previous logic where the instructions is made to be dependent of all previous memory writes.
Further, the Hemi protocol's implementation will correctly calculate the matrix products, but will do so sequentially as that would require to somehow merge the communication of multiple matrix preps.