SliceLastDim

English | 简体中文

Product Support Status

Product	Supported
Ascend 950PR&950DT products	√
Atlas A3 products	√
Atlas A2 products	√
Atlas 200I/500 A2 inference products	√
Atlas inference products	×
Atlas training products	√

Function Description

• Operator function: slices the last dimension of the input tensor, with the slice range specified by start and end and the step by stride.

• Computation example:

  • Input x shape = (2, 3, 10), start = 0, end = 10, stride = 2
  • Output y shape = (2, 3, 5)
  • y[i, j, k] = x[i, j, start + k * stride] = x[i, j, 2 * k]

• Computation formula: The shape of the output y is the same as that of the input x, except that only the last dimension becomes:

  y.shape[-1] = ceil((end - start) / stride)
  y[..., k] = x[..., start + k * stride]
  

  start and end support negative indexing (relative to the length of the last dimension); start is clamped to [0, lastDim], and end is clamped to [0, lastDim].

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	Input tensor.	INT8, INT16, INT32, INT64, FLOAT16, FLOAT, BFLOAT16	ND
y	Output	Output tensor, with the same data type and format as x.	Same as x	ND
start	Attribute	Start index of the last-dimension slice; negative indexing is supported.	INT	-
end	Attribute	End index of the last-dimension slice; negative indexing is supported.	INT	-
stride	Optional Attribute	Slice step, which defaults to 1 and must be >= 1.	INT	-

Constraint Description

• The rank of the input tensor x is in the range [1, 8].
• stride must be >= 1; when the output is non-empty, stride must additionally satisfy stride < 2^40, otherwise it is rejected during the Host Tiling stage.
• start and end support negative indexing; a negative value is converted to a positive index by adding the length of the last dimension; start is clamped to [0, lastDim], and end is clamped to [0, lastDim].
• Empty-axis input is supported: when a non-last dimension of x is 0, or the slice result is empty (e.g. end <= start), the output is an empty tensor of the corresponding shape: the Host-side Tiling returns via a short-circuit path, and the Kernel directly skips the copy.
• Row-stride upper bound: the row stride of the continuous copy, i.e. (input last-dimension length - output last-dimension length) x the number of bytes of the data type, must be less than 2^40 bytes.
• Dynamic shape support: dynamic Shape (dimensions marked as -1) and dynamic-compile-static are supported: a non-last -1 dimension is propagated to the output as-is, and when the last dimension is -1, the output last dimension is inferred as -1. An unknown-rank input (shape marked as -2) is passed through with an unknown-rank output, and the final shape is derived at runtime from the actual input.

Exception Validation

Violation	Validation Stage	Error Behavior
The rank (number of dimensions) of x is not in [1, 8]	Graph-compilation inference / Host Tiling	Returns GRAPH_FAILED, presenting EZ0012 and recording the shape-dim validation log
stride < 1, or stride >= 2^40 when the output is non-empty	Graph-compilation inference / Host Tiling	Returns GRAPH_FAILED, presenting EZ0026 and recording the stride validation log
The data type of x is not in the supported set (INT8, INT16, INT32, INT64, FLOAT16, FLOAT, BFLOAT16)	Graph-compilation inference / Host Tiling	Returns GRAPH_FAILED, presenting EZ0020 and recording the dtype validation log

Note: Error codes are wrapped by graph compilation or the running framework upon validation failure; different integration paths may only return GRAPH_FAILED while keeping the detailed logs. Some invalid inputs may be intercepted by framework-level checks before reaching operator validation (e.g. E80012 on some paths for rank out-of-range), in which case the framework-level code takes precedence.

Invocation Description

Invocation Method	Invocation Sample	Description
Graph-mode invocation	test_geir_slice_last_dim	Invokes the SliceLastDim operator by building a graph through the operator IR.