[AMD] Improve Scheduling for Async BF16 GEMM#802
Open
raikonenfnu wants to merge 40 commits intoshared/triton-gfx950-launchfrom
Open
[AMD] Improve Scheduling for Async BF16 GEMM#802raikonenfnu wants to merge 40 commits intoshared/triton-gfx950-launchfrom
raikonenfnu wants to merge 40 commits intoshared/triton-gfx950-launchfrom
Conversation
4-stage FA experiment Cluster assignment
…anonicalize can fold it
…it barrier from Membar
I'll submit a PR upstream later.
This reverts commit 718ee20.
Initial support over already arranged ops.
…ction based on the loop. This is not meant as a permanent solution just to make this branch useable for other workloads
Computation part interleaves mfma and ds_read Placed extra conditional barrier to overlap computation part and buffer_load part. Dot slicing by plognjen at https://github.com/plognjen/triton/tree/slice_dot_scaled requires vmcnt fix to achieve full performance.
Fix incorrect condition to choose enable transforms. Fix missing tokens to the local_load
Only enable for 256x256x256 tilesize
…950 (triton-lang#6744)" This reverts commit 4ecc496.
… BufferLoadToLocal to avoid implicit barrier from Membar" This reverts commit 012793a.
* [AMD] Generalize PingPong to have different type of Load/Store Ops This main motivation behind this commit is to add support for PingPong with AsyncOps. In order to accomplish that we made these changes: - Fork "determineDotMemoryOps" to "determineDotAsyncMemoryOps" that handles async memory ops. - Refactor validation and pruning of memory ops to "pruneDotMemoryOps" S.T we can have clean interface for it's async memory ops counterpart "pruneAsyncDotMemoryOps". - Plumb "useBlockPingpong" into StreamPipeliner S.T it can adjust AsyncWait stage/cluster to hoist first AsyncWait and allow set AsyncWait towards the end of the loop to make it easier for 4 PP cluster to move it before the 3rd dot-slice / 2 s_barrier before localLoads this is to ensure no race conditions. - Add check to enable handling of dotSOps (dot scaled) VS dotOps (dot) Signed-off-by: Stanley Winata <stanley.winata@amd.com> Co-authored-by: Alexander Weinrauch <alexander.weinrauch@amd.com>
Overlapping buffer_load and local_load+dot
…t on GFX950 (triton-lang#6744)"" This reverts commit f6065b9.
…ton-lang#6844) This commit improves how we create the mfma-like layout for optimizing global store by using linear layout composition. Along the way fixes a few implemenation issues. --------- Co-authored-by: Yi Qian <yi.qian@amd.com>
avoid wrongly enabled.
Requirement to enable the transform : mxfp4, 128x128x512 tile size, async_copy, num_stages=2, num_warps=8
- Use single AsyncWait (1030 -> 1070) - Move local load before global load to hide latency (1076) - Move slice local load(3) to the cluster before dot(3) (1080.5) - Update clusterBarrier to schedBarrier + s_barrier + schedBarrier (1086) Signed-off-by: Stanley Winata <stanley.winata@amd.com>
raikonenfnu
force-pushed
the
raikonenfnu/BetterAsyncBF16Schedule
branch
from
02d3120 to
8c3abdc
Compare
jungpark-mlir
requested changes
May 21, 2025
jungpark-mlir
left a comment
jungpark-mlir
left a comment
There was a problem hiding this comment.
[WIP]
Nothing wrong with this code but this requires async_copy slicing to achieve its full performance.
Holding until the related issue is addressed. We'll be working on the planB.
antiagainst
force-pushed
the
shared/triton-gfx950-launch
branch
from
77c00fa to
a259f0a
Compare
7 tasks
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 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.
The core Triton is a small number of people, and we receive many PRs (thank
you!). To help us review your code more quickly, if you are a new
contributor (less than 3 PRs merged) we ask that you complete the following
tasks and include the filled-out checklist in your PR description.
Complete the following tasks before sending your PR, and replace
[ ]with[x]to indicate you have done them.I am not making a trivial change, such as fixing a typo in a comment.
I have written a PR description following these
rules.
I have run
pre-commit run --from-ref origin/main --to-ref HEAD.Select one of the following.
/testforlittests/unittestfor C++ tests/python/testfor end-to-end testsFILL THIS IN.Select one of the following.
littests.littests I have added follow these best practices,including the "tests should be minimal" section. (Usually running Python code
and using the instructions it generates is not minimal.)