Skip to content

[WIP][chore](feat) Sync main from branch main-dev - #1920

Open
kingsirvince wants to merge 19 commits into
triton-lang:mainfrom
kingsirvince:0902_main_sync
Open

[WIP][chore](feat) Sync main from branch main-dev#1920
kingsirvince wants to merge 19 commits into
triton-lang:mainfrom
kingsirvince:0902_main_sync

Conversation

@kingsirvince

Copy link
Copy Markdown
Contributor

Sync main from branch main-dev

liwei-link and others added 19 commits September 2, 2026 17:12
* [ssbuffer](fix) insert fixpipe after producer

* [ssbuffer](feat) add merge cube block pass

Signed-off-by: fishofnanqi <1074959344@qq.com>

---------

Signed-off-by: fishofnanqi <1074959344@qq.com>
Co-authored-by: lrf22 <liurf1647@163.com>
Signed-off-by: liuhuan <liuhuan261@huawei.com>
Co-authored-by: shaoyiyang <shaoyiyang1@huawei.com>
Keep the two reserved pure-SIMT scratch pointers null until scratch allocation support is ready. Always append the DTData slot, using nullptr when device printing is disabled, so both launch paths match the kernel ABI.

Signed-off-by: Ritsuka314 <Ritsuka@huawei.com>
…iton-lang#1868)

* [tritonop](refactor) move gather_2d_simd to tutorials and add deprecation warning

* fix precommit
Co-authored-by: shaoyiyang <shaoyiyang1@huawei.com>
Stop passing --enable-lib-call-no-inline=false to bishengir-compile for Ascend950 targets. The explicit override changes vendor code generation and regresses kernels that worked before it was introduced.

Remove the target and compiler-capability helpers that were only used to append this option.

Signed-off-by: zhanwei33 <22679270+zhanwei33@users.noreply.github.com>
Co-authored-by: zhanwei33 <22679270+zhanwei33@users.noreply.github.com>
…ag (triton-lang#1866)

* [ssbuffer](fix) Fall back when outer-scope double-buffer flag budget is exceeded

Inter-core double-buffering in AddMultiBufferOuterScope allocates one
output flag per transfer group via FlagIdManager::acquireId(), which
starts above every existing sync flag (both inter-core transfer syncs
and inter-core memref-dependency syncs). Hardware provides 15 usable
flag IDs (0..14); flag ID 15 is reserved for pipe synchronization and
must not be allocated to cross-core transfers.

The pass now estimates the final max flag ID from the output flags
actually acquired for the transfer groups and, when it would reach 15:
- keeps the module in single-buffer mode (skipping only the
  double-buffer transformation and its load/store crossDeps tagging)
  if the budget pressure comes from the double-buffer itself, or
- falls back with return code 2 (ERRCODE_IGNORED) if the input IR
  already carries a flag ID above 15, which this pass cannot fix.

Previously the pass used a worst-case (maxFlagId + groupCount)
estimate and either silently downgraded or failed on a distinct-flag
count check that miscounted the legitimate pipe flag.

Co-Authored-By: DeLong code

* [UT](feat) add outer-scope flag-budget fallback test

Add Outer-scope-flag-budget-fallback.mlir: input IR uses flags 1..8 across 8 transfer groups and requests inter-core double-buffering (inter_core_buf_count=2). AddMultiBufferOuterScope would push the output flag past MAX_FLAG_ID (14) and must fall back to single-buffer mode, leaving the IR unchanged (no new output flags 9+).

* [ssbuffer](style) fix clang-format on flag-budget LDBG lines
* feat(conv2d): add conv2d op

Add al.conv2d operation for TA.

* fix(conv2d): tag conv2d kernels as mix (cube) kernels

Add Conv2dOp to the existDot walk so it gets mix_mode="mix".

* keep original stride and padding type
Signed-off-by: fishofnanqi <1074959344@qq.com>
* [ssbuffer](feat) extend scenario in splitmatmul

Signed-off-by: fishofnanqi <1074959344@qq.com>

* [ssbuffer](fix) fix operationIsFillZero coredump

Signed-off-by: fishofnanqi <1074959344@qq.com>

---------

Signed-off-by: fishofnanqi <1074959344@qq.com>
… step value (triton-lang#1892)

Signed-off-by: fishofnanqi <1074959344@qq.com>
@github-actions github-actions Bot added compiler Changes to C/C++ compiler backend (lib/, include/) python Changes to Python runtime or bindings ascend-backend Changes to the Ascend NPU backend labels Sep 2, 2026
@kingsirvince kingsirvince changed the title [chore](feat) Sync main from branch main-dev [WIP][chore](feat) Sync main from branch main-dev Sep 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ascend-backend Changes to the Ascend NPU backend compiler Changes to C/C++ compiler backend (lib/, include/) python Changes to Python runtime or bindings

Projects

None yet

Development

Successfully merging this pull request may close these issues.