[Blackwell] Refactor/slightly generalize warp specialization #6597
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
…ry for scales This ensures scales produced by TMA are elligible for transfer to tensor memory in later lowering. git-pr-chain: csullivan/support_desc_load_tmem_copy
… for tl.dot_scaled This enables automatic warp specialization for block scaled workloads. git-pr-chain: csullivan/support_block_scales_in_warp_spec
…n_warp_spec' into mogball/fmha
Mogball
changed the title
manman-ren
reviewed
Apr 30, 2025
| passes.ttgpuir.add_warp_specialize(pm, opt.num_stages) | ||
| passes.ttgpuir.add_pipeline(pm, opt.num_stages, dump_enabled) | ||
| passes.ttgpuir.add_combine_tensor_select_and_if(pm) | ||
| nvidia.passes.ttnvgpuir.add_promote_lhs_to_tmem(pm) |
There was a problem hiding this comment.
I needed this same change for getting FA to run with WarpSpec on Blackwell.
| @@ -100,7 +100,8 @@ DenseMap<Operation *, int> deserializeLatencies(Operation *op); | |||
| Value createScalarAlloc(ImplicitLocOpBuilder &rewriter, Type type, | |||
| unsigned numBuffers); | |||
| // Create an allocation and init the mbarriers. | |||
| Value createBarrierAlloc(scf::ForOp forOp, int numBarriers); | |||
| Value createBarrierAlloc(scf::ForOp forOp, int numBarriers, | |||
| int arriveCount = 1); | |||
There was a problem hiding this comment.
Is this part of refactoring? Or is it addressing a separate issue?
There was a problem hiding this comment.
This is part of the refactor. Load groups can have multiple consumers
| @@ -297,6 +298,8 @@ mlir::triton::getDefinitionAndDistance(scf::ForOp forOp, Value value) { | |||
| return {nullptr, 0}; | |||
| ++distance; | |||
| value = forOp.getYieldedValues()[arg.getArgNumber() - 1]; | |||
| if (!seen.insert(value).second) | |||
| return {nullptr, 0}; | |||
| } | |||
There was a problem hiding this comment.
This also doesn't feel like refactoring :]
There was a problem hiding this comment.
Some of the refactoring exposed a bug :P
Mogball
changed the title
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.
This is basically a rewrite of LoadMMASpecialization to cleanly separate the warp specialization of loads and MMAs into discrete steps. This allows warp specialization of loads and MMAs separate from each other, and supports an arbitrary number of load groups and MMAs. This still places all loads and MMAs in the same partitions.
In addition, this PR separates the actual partition assignment from the multibuffering and loop lowering step, like SWP. This should make it easier to tweak partitioning strategies.