[TIR, CUDA] Add pass to replace global to shared memory copy with cp.async #11658
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
Member
|
Amazing work!! |
vinx13
approved these changes
Jun 9, 2022
|
|
||
| Stmt VisitStmt_(const BufferStoreNode* store) { | ||
| if (in_async && (store->buffer.scope() == "shared" || store->buffer.scope() == "shared.dyn")) { | ||
| if (auto* load = store->value.as<BufferLoadNode>()) { |
Member
There was a problem hiding this comment.
I'm wondering how we should handle the case that the value is not BufferLoad? For padding case maybe this can rely on the intrin provide predicated read, not sure about more complicated case. But this PR is good, no action needed for now
Member
Author
There was a problem hiding this comment.
Interesting. Indeed, currently it only supports the fixed pattern shared[...] = global[...]. But I think we can add more patterns as they come up, as long as we can extract the src pointer and the offset.
Kathryn-cat
pushed a commit
to Kathryn-cat/tvm
that referenced
this pull request
Jun 10, 2022
…async (apache#11658) * [TIR, CUDA] Add pass to replace global to shared memory copy with cp.async * add missing doc * black * missing src * clang format * clang format * check against nested async scope
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.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 pass looks for global to shared memory copy enclosed in the new
tir::attr::async_scopescope, and replace that with PTXcp.asyncintrinsics added in #11368.This pass is disabled by default, since
cp.asyncis only supported by NV gpus with sm >= 80. For now, the attrtir::attr::async_scopeand the proper synchronization need to be inserted manually in the input TIR. But I have a working branch https://github.com/apache/tvm/compare/main...masahi:inject-async-copy?expand=1 that automatically inserts such async regions and synchronizations as part of the software pipeline transform.@vinx13 @junrushao1994 @tqchen @csullivan