[Unity][Dlight] Handle Epilogue Broadcasting #15252
Merged
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
This PR improves the Decode-GEMV scheduling by further analyzing its epilogue pattern. The existing behavior assumes that the outcome of cross-thread reduction stays in register files local to each thread, which is further used to calculate the epilogue in the same thread. This strategy means the cross-thread reduction outcome is stored only on thread 0, while the other threads cannot participate in subsequent computation (i.e. epilogue). Related: apache#15192. When the epilogue is relatively lightweight, i.e. elementwise add, casting on scalars, this strategy is optimal. However, once the outcome needs to be broadcasted to compute over a non-trivial region, for example, act as a normalizer of `np.mean`, it would become much slower because only one thread in a thread block is effectively used. In this case, we will need to broadcast the cross-thread reduction outcome in shared memory, making it visible to other threads, and then bind the compute region to all threads in the threadblock.
Collaborator
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.
Generated by tvm-bot |
Member
Author
|
CC @MasterJH5574 on cross-thread reduction, @Hzfengsy and @spectrometerHBH for e2e performance |
Hzfengsy
approved these changes
Jul 6, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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Outdated suggestions cannot be applied.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.
This PR improves the Decode-GEMV scheduling by further analyzing its epilogue pattern.
The existing behavior assumes that the outcome of cross-thread reduction stays in register files local to each thread, which is further used to calculate the epilogue in the same thread.
This strategy means the cross-thread reduction outcome is stored only on thread 0, while the other threads cannot participate in subsequent computation (i.e. epilogue). Related: #15192.
When the epilogue is relatively lightweight, i.e. elementwise add, casting on scalars, this strategy is optimal. However, once the outcome needs to be broadcasted to compute over a non-trivial region, for example, act as a normalizer of
np.mean, it would become much slower because only one thread in a thread block is effectively used.In this case, we will need to broadcast the cross-thread reduction outcome in shared memory, making it visible to other threads, and then bind the compute region to all threads in the threadblock.