triton.experimental.gluon.language.amd.slice

triton.experimental.gluon.language.amd.slice(source, shape, offsets, _semantic=None)

Extract a register only slice of a tensor.

Returns a view of source of the requested shape starting at offsets, keeping the source’s distributed layout. Because the layout is preserved, the slice introduces no cross-thread data movement; the slice extent and offsets must align to the source layout’s CTA tiling.

Parameters:

• source (tensor) – The tensor to slice. Must have a distributed layout.

• shape (List[int]) – Shape of the extracted slice (same rank as source).

• offsets (List[int]) – Per-dimension start offsets into source.