Add TSCATTER by frank-deng · Pull Request #980 · hw-native-sys/PTOAS

frank-deng · 2026-07-23T03:36:18Z

No description provided.

Zhendong404 · 2026-07-27T07:18:06Z

  let extraClassDeclaration = [{
    bool hasIndexForm() { return static_cast<bool>(getIndexes()); }
    bool hasMaskForm() { return static_cast<bool>(getMaskPatternAttr()); }
+    bool hasAxis() { return static_cast<bool>(getAxis()); }


TileOp IR文档需要同步修改

Zhendong404 · 2026-07-27T07:58:32Z

    PTODpsType:$src,
    PTODpsType:$dst,
    Optional<PTODpsType>:$indexes,
+    Optional<Index>:$axis,


axis应该和maskPattern作为attribute，不要作为operand，因为这是一个编译期静态信息

Zhendong404 · 2026-07-27T08:29:09Z

    %dst_mask = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>
-    pto.tscatter ins(%src_mask, {maskPattern = #pto.mask_pattern<P0101>} : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=64, v_row=1, v_col=64, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst_mask : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    %axis = arith.constant 0 : index
+    pto.tscatter ins(%src_mask, {maskPattern = #pto.mask_pattern<P0101>} : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=64, v_row=1, v_col=64, blayout=row_major, slayout=none_box, fractal=512, pad=0>, %axis : index) outs(%dst_mask : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>)


前面改了tscatter的verifier，要求dst的valid_col大于src的valid_col，这个例子是反过来的

Zhendong404 · 2026-07-27T08:31:05Z

    if (srcValid.size() != 2 || dstValid.size() != 2)
      return emitOpError("expects src and dst to have rank-2 valid_shape");

+    auto axis = getAxis();


verifier需要拒绝同时提供index和axis

Zhendong404 · 2026-07-27T08:32:50Z


+def tscatter(src, dst, *, indexes=None, axis=None, mask_pattern=None):
+    """``pto.tscatter`` tile scatter wrapper."""
+    _pto.tscatter(


DSL也要加校验，报错尽量提前。indexes和axis不能同时提供，axis的取值范围需要约束

Zhendong404 · 2026-07-27T08:36:45Z

-    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=64, v_row=1, v_col=64, blayout=row_major, slayout=none_box, fractal=512, pad=0>
-    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>
-    pto.tscatter ins(%src, {maskPattern = #pto.mask_pattern<P0101>} : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=64, v_row=1, v_col=64, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f16, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>


新增axis参数需要lit用例覆盖，确保ExpandTileOp和EmitC路径能正常工作，特别是EmitC路径下indexes form和axis form两套接口

Zhendong404 · 2026-07-27T08:42:23Z

+@tilelib.tile_template(
+    op="pto.tscatter",
+    target="a5",
+    name="template_tscatter_mask_p0101_col",
+    dtypes=_SCATTER_MASK_DTYPES,
+    iteration_axis="none",
+    op_engine="vector",
+    op_class="other",
+    layouts=("row_major",),
+    loop_depth=2,
+    is_post_update=False,
+    tags=("scatter", "mask", "p0101"),
+    constraints=(_axis_is_col,),
+    id=9,
+)
+def template_tscatter_mask_p0101_col(src: pto.Tile, dst: pto.Tile, axis: pto.i32):
+    _template_tscatter_mask_col(src, dst, 1, 2)
+
+
+@tilelib.tile_template(
+    op="pto.tscatter",
+    target="a5",
+    name="template_tscatter_mask_p1010_row",
+    dtypes=_SCATTER_MASK_DTYPES,
+    iteration_axis="none",
+    op_engine="vector",
+    op_class="other",
+    layouts=("row_major",),
+    loop_depth=2,
+    is_post_update=False,
+    tags=("scatter", "mask", "p1010"),
+    constraints=(_axis_is_row,),
+    id=3,
+)
+def template_tscatter_mask_p1010_row(src: pto.Tile, dst: pto.Tile, axis: pto.i32):
+    _template_tscatter_mask_row(src, dst, (False,))
+
+
+@tilelib.tile_template(
+    op="pto.tscatter",
+    target="a5",
+    name="template_tscatter_mask_p1010_col",
+    dtypes=_SCATTER_MASK_DTYPES,
+    iteration_axis="none",
+    op_engine="vector",
+    op_class="other",
+    layouts=("row_major",),
+    loop_depth=2,
+    is_post_update=False,
+    tags=("scatter", "mask", "p1010"),
+    constraints=(_axis_is_col,),
+    id=10,
+)
+def template_tscatter_mask_p1010_col(src: pto.Tile, dst: pto.Tile, axis: pto.i32):
+    _template_tscatter_mask_col(src, dst, 0, 2)


这些版本能合并吗？好像没必要分出这么多版本，TileOp的属性是可以在函数里面获取的

Zhendong404 · 2026-07-27T08:49:00Z

index 模式的索引 dtype 规则三层不一致
C++ verifier（verifyIndexedForm）：indexes 只能是 i16/i32 族，且有位宽规则——data 4B → idx 4B，data 2B → idx 2B，data 1B → idx 2B；
DSL 文档："indexes must have i32 or ui32 dtype"；
模板 dtype 矩阵：NUMERIC_DTYPES × {i32, ui32}

Zhendong404 · 2026-07-27T08:52:00Z


 from .metadata import ScalarSpec, VectorSpec, ViewSpec

+_MASK_PATTERN_TOKENS = {"p0101", "p1010", "p0001", "p0010", "p0100", "p1000", "p1111"}


这个是TScatter专用的constraint，似乎没有必要加到公共constraint系统里，否则以后任何模板只要 tag 里碰巧含 "p0101" 之类就会被这个规则过滤。更干净的做法是把 "context attr 必须命中 template tags" 做成模板声明式的元数据（比如 require_tag_from=("mask_pattern",)），或者干脆用普通 constraint 函数表达。同理 _no_mask_pattern(mask_pattern=True) 这个"默认 True、收到字符串就为 False"的写法能工作但非常隐晦，依赖"约束参数名与 context attr 名巧合相同"这一隐式契约，建议至少加注释。

Zhendong404 · 2026-07-27T08:53:43Z

+def _make_index_inputs(name, src_dtype, idx_dtype, src_shape, idx_shape):
+    np_dt = npy_dtype(src_dtype)
+    np_idx_dt = npy_dtype(idx_dtype)
+    rng = np.random.RandomState(hash(name) & 0xFFFFFFFF)


建议改成zlib.crc32(name.encode()) 之类稳定种子

Zhendong404 · 2026-07-28T07:05:28Z

+def _static_valid_dim(tile, index):
+    static_valid_shape = getattr(tile, "_template_static_valid_shape", None)
+    if static_valid_shape is None:
+        static_valid_shape = getattr(tile, "static_valid_shape", None)
+    if static_valid_shape is not None and static_valid_shape[index] is not None:
+        return static_valid_shape[index]
+    return tile.valid_shape[index]
+
+
+def _template_tscatter_mask_row(src: pto.Tile, dst: pto.Tile, interleave_args):
+    dtype = dst.dtype
+    valid_rows = _static_valid_dim(src, 0)
+    valid_cols = _static_valid_dim(src, 1)


valid_row/valid_col可以通过src.valid_shape直接拿，不用自己实现接口

frank-deng · 2026-07-29T01:27:06Z

tscatter.zip

frank-deng force-pushed the tscatter branch 8 times, most recently from 7e63f3d to e8965d1 Compare July 27, 2026 05:59

Zhendong404 requested changes Jul 27, 2026

View reviewed changes

frank-deng force-pushed the tscatter branch 2 times, most recently from 09c9fa3 to e20ab52 Compare July 27, 2026 14:27

Zhendong404 requested changes Jul 28, 2026

View reviewed changes

frank-deng added 2 commits July 29, 2026 09:33

vscatter add non-32b type support

8a203fc

Add TSCATTER

75412af

frank-deng force-pushed the tscatter branch from e20ab52 to 75412af Compare July 29, 2026 01:39

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add TSCATTER - #980

Add TSCATTER#980
frank-deng wants to merge 2 commits into
hw-native-sys:mainfrom
frank-deng:tscatter

frank-deng commented Jul 23, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 27, 2026

Uh oh!

Zhendong404 Jul 28, 2026

Uh oh!

frank-deng commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants


		from .metadata import ScalarSpec, VectorSpec, ViewSpec

		_MASK_PATTERN_TOKENS = {"p0101", "p1010", "p0001", "p0010", "p0100", "p1000", "p1111"}

Uh oh!

Conversation

frank-deng commented Jul 23, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

frank-deng commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants