Flip transpose of (slow) dynamic goi right-hand GEMM operands.
#8917
+192
−0
Conversation
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=80&v=4){kind=small}
If a `fully-connected` or `batch-matrix-multiply` op will use dynamic `goi` packing for its right-hand operand, we add a transpose of the last two dimensions of the operand and flip the `XNN_FLAG_TRANSPOSE_WEIGHTS` flag. If the right-hand operand is generated by a `transpose` op, then the transposition is done therein, potentially skipping the `transpose` altogether if it becomes a no-op. We do this because, in almost all cases, we only have generic (unoptimized) `gio` packing kernels, but we do have optimized 'transpose' and 'goi' kernels. PiperOrigin-RevId: 811769415
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
Flip transpose of (slow) dynamic
goiright-hand GEMM operands.If a
fully-connectedorbatch-matrix-multiplyop will use dynamicgoipacking for its right-hand operand, we add a transpose of the last two dimensions of the operand and flip theXNN_FLAG_TRANSPOSE_WEIGHTSflag.If the right-hand operand is generated by a
transposeop, then the transposition is done therein, potentially skipping thetransposealtogether if it becomes a no-op.We do this because, in almost all cases, we only have generic (unoptimized)
giopacking kernels, but we do have optimized 'transpose' and 'goi' kernels.