Restructure pre-packed matrix layouts so generic GEMM code can be agnostic of panel layout #511
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
robertknight
changed the title
These are documented in the papers referenced in other comments in this module, but add comments inline that are more immediately accessible.
Each kernel now manages a temporary tile of an appropriate size, so the generic GEMM outer loops don't need to know the maximum size.
Divide prepacked into depth blocks with a size that matches the depth block size (`kc`) used during computation. This allows for the generic GEMM code to be agnostic of the layout of panels within a block, as it no longer needs to be able to slice panels along the depth dimension. Instead it just uses a depth block index to pick a panel with the pre-determined block size. This in turn gives the GEMM kernel freedom to choose the layout within each panel. This partly undoes #482.
robertknight
force-pushed
the
gemm-depth-block
branch
from
640017b to
6ec1b1b
Compare
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Outdated suggestions cannot be applied.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.
#510 noted that the layout of pre-packed matrices required the generic GEMM code to know about the layout of data within a packed panel. This revises the layout so that is no longer the case. See the last commit for details. After this change it will be possible to use different panel layouts for each kernel. int8 kernels for example will use dot product instructions that require a different block layout than f32 kernels using FMA instructions.
Along the way I also added some comments about where the cache and register blocking sizes come from. These are explained in the referenced paper, but it is useful to have them more immediately accessible.