ParallelConcat

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: concatenates N same-shaped tensors whose first dimension has a size of 1 along the first dimension. Each tensor in the dynamic input values has the shape [1, d_1..d_k], the output output_data has the shape [N, d_1..d_k], and the content of row i is the flattened content of the i-th input. Data is moved bit by bit, and arbitrary bit patterns such as NaN/Inf/±0 remain unchanged.

• Computation formula:

  $$\text{output\_data}[i,d_1,\dots ,d_k]={\text{values}}_i[0,d_1,\dots ,d_k],i=0,1,\dots ,N-1$$

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
values	Input	Dynamic input, which is a list of tensors to be concatenated. There are N tensors in total, and each tensor has the shape [1, d_1..d_k], which is values_i in the formula. The size of the first dimension must be 1, and all tensors must have the same shape and data type.	FLOAT, FLOAT16, BFLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, BOOL	ND
output_data	Output	Concatenation result tensor, whose shape is [N, d_1..d_k], which is output_data in the formula. Its data type is the same as that of values.	FLOAT, FLOAT16, BFLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, BOOL	ND
shape	Attribute	Explicit shape of the output tensor, which is [N, d_1..d_k]. Required attribute.	LIST_INT	-
N	Attribute	Number of tensors in the dynamic input values. N≥1 and N==shape[0]. Required attribute.	INT	-

Constraint Description

• The size of the first dimension of each input tensor must be 1, and the rank range is [1,8] (a scalar input with rank 0 is invalid).
• All input tensors must have the same shape and data type, and attr shape[1:] must be consistent with input shape[1:].
• The shape of the output output_data must be equal to attr shape.
• The empty tensor scenario is supported (when any d_m is 0, the output shape is [N, 0]).

Invocation Description

Invocation Method	Invocation Sample	Description
GE graph mode	test_geir_parallel_concat	Invokes the ParallelConcat operator by building a graph through the operator IR.