Transpose

Applicability

Product	Supported
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	x
Atlas inference products	√
Atlas training products	√

Function Usage

Transposes the tensor without changing its value.

The transpose operator can be used to rearrange the tensor dimensions.

Specifically, it transposes the shape of input tensor x according to the permutation of dimensions (perm) and outputs the result.

Prototype

If the input and output addresses are different:

const aclTensor *Transpose(const aclTensor *x, const aclTensor *y, const aclTensor *perm, aclOpExecutor *executor)

If the input and output addresses are the same:

const aclTensor *Transpose(const aclTensor *x, const aclIntArray *perm, aclOpExecutor *executor)

Parameters

Parameter	Input/Output	Description
x	Input/Output	Input tensor. The input tensor x must be contiguous memory data. The supported data types are FLOAT16, FLOAT, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, BFLOAT16, BOOL. The data format can be ND. NOTE: BFLOAT16 applies to the following product models: Atlas A2 training products/Atlas A2 inference products Atlas A3 training products/Atlas A3 inference products
y	Output	Output tensor after transposing. It has the same data type and format as x.
perm	Input	Permutation of dimensions of input tensor x, which is an integer array. It can be of the aclIntArray* or aclTensor* type. A maximum of 8D is supported. The value range is [0, Dimension count of x – 1]. The data type can be INT32 or INT64. The data format can be ND.
executor	Input	Operator executor, containing the operator computation process.

Returns

Success: The transposed tensor is returned. Failure: nullptr is returned.

Constraints

A maximum of 8D is supported. That is, the maximum dim value of input x and perm is 8.
The input x must have the same dimensions (dim) as perm.

Examples

// Standard format for creating an OpExecutor and checking parameters
auto uniqueExecutor = CREATE_EXECUTOR();
CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);

// Standard format for converting the input self into a contiguous tensor
auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());
CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);

int64_t dims = selfContiguous->GetViewShape().GetDimNum();
int64_t valuePerm[dims] = {0, 2, 1, 3}; // The middle two dimensions of the original four dimensions are transposed, that is, the first axis and the second axis are exchanged.

auto perm = executor->AllocIntArray(valuePerm, dims);
selfContiguous = l0op::Transpose(selfContiguous, perm, uniqueExecutor.get());

Parent topic: Basic Tensor Operation APIs