ViewCopy

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: copies the src tensor as the view described by src_size, src_stride, and src_storage_offset to the dst tensor as the view described by dst_size, dst_stride, and dst_storage_offset, and returns dst.

• Computation formula:

  $$\text{dst}\left[\text{dst\_storage\_offset}+\sum _{i=0}^{n-1}{\text{dst\_index}}_i\cdot {\text{dst\_stride}}_i\right]=\text{src}\left[\text{src\_storage\_offset}+\sum _{i=0}^{n-1}{\text{src\_index}}_i\cdot {\text{src\_stride}}_i\right]$$

  where $n$ is the number of view dimensions; ${\text{dst\_index}}_i$ and ${\text{src\_index}}_i$ are the indices of the dst and src views in the $i$-th dimension respectively, with the value ranges $0\le {\text{dst\_index}}_i<{\text{dst\_size}}_i$ and $0\le {\text{src\_index}}_i<{\text{src\_size}}_i$. Since dst_size and src_size have the same shape, the indices of corresponding dimensions are element-wise equal; $\text{dst\_storage\_offset}$ and $\text{src\_storage\_offset}$ are the view start offsets; ${\text{dst\_stride}}_i$ and ${\text{src\_stride}}_i$ are the strides of the $i$-th dimension. The content in square brackets is the element subscript in the tensor storage area.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
dst	Input	Destination tensor to be written to. Its view is described by dst_size, dst_stride, and dst_storage_offset.	BFLOAT16, FLOAT16, FLOAT, UINT8, INT8, UINT16, INT16, UINT32, INT32, UINT64, INT64, BOOL, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN	ND
dst_size	Input	View shape of the dst tensor, a 1D tensor, whose elements are non-negative and whose number of elements is the same as that of dst_stride.	INT32, INT64	ND
dst_stride	Input	View stride of the dst tensor, a 1D tensor, whose elements are non-negative and whose number of elements is the same as that of dst_size.	INT32, INT64	ND
dst_storage_offset	Input	View start offset of the dst tensor, a 1D tensor, whose elements are non-negative and do not exceed the element range of the dst tensor.	INT32, INT64	ND
src	Input	Source tensor to be read. Its view is described by src_size, src_stride, and src_storage_offset.	BFLOAT16, FLOAT16, FLOAT, UINT8, INT8, UINT16, INT16, UINT32, INT32, UINT64, INT64, BOOL, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN	ND
src_size	Input	View shape of the src tensor, a 1D tensor, whose elements are non-negative and whose number of elements is the same as that of src_stride.	INT32, INT64	ND
src_stride	Input	View stride of the src tensor, a 1D tensor, whose elements are non-negative and whose number of elements is the same as that of src_size.	INT32, INT64	ND
src_storage_offset	Input	View start offset of the src tensor, a 1D tensor, whose elements are non-negative and do not exceed the element range of the src tensor.	INT32, INT64	ND
dst	Output	Copy result, which shares the same memory as the input dst (in-place write).	BFLOAT16, FLOAT16, FLOAT, UINT8, INT8, UINT16, INT16, UINT32, INT32, UINT64, INT64, BOOL, HIFLOAT8, FLOAT8_E5M2, FLOAT8_E4M3FN	ND

Constraint Description

• dst_size and src_size have the same shape.
• The number of dimensions of dst_size is the same as that of dst_stride; the number of dimensions of src_size is the same as that of src_stride.
• The elements of dst_size and src_size are non-negative.
• The elements of dst_stride and src_stride are non-negative.
• dst_storage_offset and src_storage_offset are non-negative and do not exceed the element range of the corresponding tensors.
• For each axis i, dst_storage_offset + (dst_size[i] - 1) * dst_stride[i] does not exceed the number of elements of the dst tensor. The same applies to src.
• For the dtype, type conversion, and determinism constraints at the aclnn API layer, see aclnnInplaceCopy.

Invocation Description

Invocation Method	Sample Code	Description
aclnn invocation	test_aclnn_view_copy	Invokes the ViewCopy operator through the aclnnInplaceCopy API.
Graph-mode invocation	-	Invokes the ViewCopy operator by building a graph through the operator IR.