Skip to content

[Sync](chore) sync triton-ascend main into 36f3bae - #50

Merged
hongziqi merged 10 commits into
TecJesh:3.7-masterfrom
hongziqi:3.7-master-sync-main-0828-36f3bae
Aug 29, 2026
Merged

[Sync](chore) sync triton-ascend main into 36f3bae#50
hongziqi merged 10 commits into
TecJesh:3.7-masterfrom
hongziqi:3.7-master-sync-main-0828-36f3bae

Conversation

@hongziqi

@hongziqi hongziqi commented Aug 29, 2026

Copy link
Copy Markdown
Collaborator

[Sync] 回合 triton-ascend/main (36f3bae) 至 3.7-master

概述

upstream/main 最新代码(commit 36f3baeaa1)回合至 3.7-master-sync-main-0827-872fa1e,同步 main 自 872fa1e 以来 8 个上游提交,主要涉及 ControlFlowOpt 指针控制流解耦、SplitIfByBlockIdPass、SSBuffer 修复及 Msdebug 调试增强。


上游合入提交(8 个)

Commit 说明
120b97f6f6 [Docs] add 'tracking' issue label for long-lived work (triton-lang#1828)
de756c69eb [ssbuffer] fix multi-yield op transfer (triton-lang#1744)
12827daf87 [unordered_lock] fix unordered lock is able to run on A5 (triton-lang#1636)
14ba391adb [Msdebug] stabilize DWAF line rewriting and load/store stepping (triton-lang#1754)
daa0844144 [ssbuffer] Add SplitIfByBlockIdPass to pipeline (triton-lang#1735)
a9238f59d3 [ControlFlowOpt] restore pointer control-flow decoupling (triton-lang#1771)
2bafba14ac [ControlFlowOpt] stabilize pointer control-flow lowering
36f3baeaa1 [SplitIfByBlockId] skip parallel_nsa_fwd_kernel in SplitIfByBlockIdPass

冲突文件及解决策略

merge 过程中产生 2 个冲突文件,均已手动解决:

冲突文件 解决要点
third_party/ascend/lib/TritonToLinalg/BlockPtrAnalysis.cpp 取交集:保留 3.7 侧已有实现,合入 main 新增头文件依赖及 BlockPtrAnalysis.h 中新增接口(kScalarPointerCarrierAttrneedsLegacyBlockDataLoopRewritegetMarkedMakeRangeCarrierSlots 等)
third_party/ascend/lib/TritonToLinalg/TritonToLinalgPass.cpp 逻辑追随 main:采用 main 分支 ControlFlowOpt / 指针控制流解耦相关 pass 编排与 lowering 逻辑

其余文件自动合并,包括 compiler.pyLoadStoreConverter.cppTritonOpConverter.cpp 等。

image

hongziqi and others added 9 commits August 28, 2026 16:23
…ang#1828)

* [Docs](feat) add 'tracking' issue label for long-lived work

* update stale close time to 7
…g#1636)

* sync_block_lock_unordered is able to run on A5

* [syncblock](fix) resolve code pre-commit issue

---------

Co-authored-by: yangkaixin <yangkaixin1@huawei.com>
Co-authored-by: liufengsheng-beep <liufengsheng2@h-partners.com>
* [ssbuffer](feat) Add SplitIfByBlockIdPass to pipeline

Signed-off-by: chenwuyang <chenwuyang1@huawei.com>

* [ssbuffer](fix) allow matmul placeholder for split if

---------

Signed-off-by: chenwuyang <chenwuyang1@huawei.com>
Co-authored-by: chenwuyang <chenwuyang1@huawei.com>
…n-lang#1771)

Reintroduce BlockPtr, TensorPtr, scalar-pointer, and scope control-flow decoupling with the corresponding T2L and T2U handoff updates.

Preserve conversion failure atomicity, scalar carrier materialization, marked range carriers, and descriptor boundary metadata across SCF regions.

Add the reviewed MLIR and end-to-end pytest coverage for pointer control-flow boundaries.
Keep stable scalar pointer bases and select complete-address carriers for opaque scf.if loop backedges.

Preserve rank-N sign-extended offset structure, constrain legacy BlockData address-use discovery, and fold supported scan pointer stores before use analysis.

Keep discrete vector masks attached to extracted memory accesses and include the sparse decode model1 DynamicCV fallback from PR triton-lang#1791.

Add focused MLIR and end-to-end regressions for the repaired paths.
@hongziqi hongziqi closed this Aug 29, 2026
@hongziqi hongziqi reopened this Aug 29, 2026
@hongziqi
hongziqi merged commit 450d44a into TecJesh:3.7-master Aug 29, 2026
10 of 15 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants