tilelang.ascend.language.reduce_op¶

Ascend dialect of the reduce operators.

Same surface as tilelang.language.reduce_op, plus the SimdVF direct path: a SimdVF region has no thread domain, so the common implementation’s fragment round-trip (alloc_fragment + copy in + reduce + copy out) for shared-to-shared reductions cannot lower there. Inside SimdVF the reduce is emitted directly on the UB regions and AscendSimdVFLowerParallel lowers it to the SIMD reduction intrinsics.

The thin reduce_* wrappers are redeclared because the common ones call the common reduce by module-level binding and would bypass this dialect’s dispatch.

Functions¶

`reduce`(buffer, out, reduce_type, dim, clear[, batch, ...])	Common `tilelang.language.reduce_op.reduce()` plus the SimdVF path.
`reduce_max`(buffer, out[, dim, clear, batch, annotations])	See `tilelang.language.reduce_op.reduce_max()`.
`reduce_min`(buffer, out[, dim, clear, batch, annotations])	See `tilelang.language.reduce_op.reduce_min()`.
`reduce_sum`(buffer, out[, dim, clear, batch, annotations])	See `tilelang.language.reduce_op.reduce_sum()`.
`reduce_abssum`(buffer, out[, dim, batch, annotations])	See `tilelang.language.reduce_op.reduce_abssum()`.
`reduce_absmax`(buffer, out[, dim, clear, batch, ...])	See `tilelang.language.reduce_op.reduce_absmax()`.
`reduce_bitand`(buffer, out[, dim, clear, batch, ...])	See `tilelang.language.reduce_op.reduce_bitand()`.
`reduce_bitor`(buffer, out[, dim, clear, batch, annotations])	See `tilelang.language.reduce_op.reduce_bitor()`.
`reduce_bitxor`(buffer, out[, dim, clear, batch, ...])	See `tilelang.language.reduce_op.reduce_bitxor()`.

Module Contents¶

tilelang.ascend.language.reduce_op.reduce(buffer, out, reduce_type, dim, clear, batch=1, annotations=None)¶

Common tilelang.language.reduce_op.reduce() plus the SimdVF path.

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
reduce_type (str)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_max(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_max().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_min(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_min().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_sum(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_sum().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_abssum(buffer, out, dim=-1, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_abssum().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_absmax(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_absmax().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_bitand(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_bitand().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_bitor(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_bitor().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None

tilelang.ascend.language.reduce_op.reduce_bitxor(buffer, out, dim=-1, clear=True, batch=1, annotations=None)¶

See tilelang.language.reduce_op.reduce_bitxor().

Parameters:

buffer (tvm.tirx.Buffer)
out (tvm.tirx.Buffer)
dim (int)
clear (bool)
batch (int)
annotations (dict | None)

Return type:

None