aclopInferShape

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product / Atlas A3 inference product	√
Atlas A2 training product / Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	√
Atlas inference product	√
Atlas training product	√

Function Usage

Infers the output shape of an operator based on the input shape and other necessary arguments.

Consider the following possibilities:

If the accurate output shape of the operator can be inferred, the accurate output shape is returned.
If the accurate output shape of the operator cannot be inferred, but the range of the output shape can be obtained, the dynamic dimension in the tensor description is set to –1 in the output parameter outputDesc. In this scenario, you can call aclGetTensorDescDimRange to obtain the size range of the specified dimension from the tensor description.
(Reserved) If the accurate output shape and shape range cannot be obtained, the runtime dimension in the tensor description is set to –2 in the output parameter outputDesc.

Prototype

      
           aclError aclopInferShape(const char *opType,
int numInputs,
aclTensorDesc *inputDesc[],
aclDataBuffer *inputs[],
int numOutputs,
aclTensorDesc *outputDesc[],
aclopAttr *attr)

Parameters

Parameter	Input/Output	Description
opType	Input	Pointer to the operator type name.
numInputs	Input	Number of input tensors.
inputDesc	Input	Pointer array of the operator input tensor description. For details about the type definition, see aclTensorDesc. Call aclCreateTensorDesc to create data of the aclTensorDesc type in advance. The array length must be consistent with numInputs.
inputs	Input	Pointer array of the operator input tensor. For details about the type definition, see aclDataBuffer. Call aclCreateDataBuffer to create data of the aclDataBuffer type in advance. In Ascend EP form, allocate host memory for storing the input tensor data of the operator. In Ascend RC form, allocate device memory for storing the input tensor data of the operator. In Ctrl CPU open form, allocate device memory for storing the input tensor data of the operator.
numOutputs	Input	Number of output tensors.
outputDesc	Output	Pointer array of the operator output tensor description. For details about the type definition, see aclTensorDesc. Call aclCreateTensorDesc to create data of the aclTensorDesc type in advance. The array length must be consistent with numOutputs.
attr	Input	Operator attribute. For details about the type definition, see aclopAttr.

Returns

0 on success; else, failure. For details, see aclError.

Restrictions

Before aclopInferShape is called to infer the output shape of a dynamic-input or dynamic-output operator, aclSetTensorDescName must be called to set the names of all input and output tensors. For details about the input and output names defined in the operator IR prototypes, see .

For a required input, optional input, or required output, use the exact name defined in the operator IR prototypes.
For a dynamic input or dynamic output, use the exact name defined in the operator IR prototypes with a sequence number suffix. Dynamic inputs/outputs are indexed starting at 0.

For example, if an operator has two inputs (the first one is a required input x, the second one is a dynamic input y, including two inputs) and one required output z, the code for calling the aclSetTensorDescName API is as follows:

aclSetTensorDescName(inputTensorDesc[0], "x");
aclSetTensorDescName(inputTensorDesc[1], "y0");
aclSetTensorDescName(inputTensorDesc[2], "y1");
aclSetTensorDescName(outputTensorDesc[0], "z");

Parent topic: Single-Operator Model Execution