aicpu_ext_info_handle

The APIs described in this section are reserved and may be changed or deprecated in the future. They do not need your attention.

Table 1 API list

API Definition

Description

AicpuExtInfoHandler(const std::string &nodeName, const uint32_t inputNum, const uint32_t outputNum, const ge::UnknowShapeOpType unknownType)

Indicates the constructor for managing AI CPU extended arguments.

nodeName_(nodeName)

Indicates the operator name.

inputNum_(inputNum)

Indicates the number of operator inputs.

outputNum_(outputNum)

Indicates the number of operator outputs.

unknownType_(unknownType)

Indicates the type of an operator.

AicpuExtInfoHandler()

Indicates the constructor for managing AI CPU extended arguments.

GenTfExtBuffer(const FVector<const aclTensor *> &inputs, const FVector<aclTensor *> &outputs, std::string &taskExtInfo)

Generates the extended arguments of the TensorFlow framework.

GenCCExtBuffer( const FVector<const aclTensor *> &inputs, const FVector<aclTensor *> &outputs, std::string &taskExtInfo)

Generates the extended arguments of the CANN framework.

Parse(const std::string &extInfo, uint8_t *hostAddr)

Parses the extended arguments.

UpdateInputShape(const uint32_t inputIndex, const gert::Shape &inputShape)

Updates the input shape in the extended arguments.

UpdateOutputShape(const uint32_t outputIndex, const gert::Shape &outputShape)

Updates the output shape in the extended arguments.

GetOutputShapeAndType(const uint32_t outputIndex, gert::Shape &shape, ge::DataType &dataType)

Obtains the shape and type in the extended arguments.

CopyH2D(const rtStream_t stream, const aclOpExecutor *executor, const uint64_t deviceExtMemSize, uint64_t &deviceCacheOffset)

Copies extended arguments from the host to the device.

CopyOutputShapeD2H()

Updates the output shape of the operator on the device to the host.

GetExtInfoDeviceBuffer(const aclOpExecutor *executor, const uint64_t deviceExtMemSize, uint64_t &deviceCacheOffset)

Obtains the address of memory reserved on the device for extended arguments.

GenerateKernelId()

Generates a kernel ID.

UpdateOutputShapeFromExtInfo(const FVector<aclTensor *> &outputs, aclrtStream stream)

Obtains the output shape from the device and updates the information to the host.

UpdateInputAndOutputShape(const FVector<const aclTensor *> &inputs, const FVector<aclTensor *> &outputs, aclrtStream stream, const aclOpExecutor *executor, const uint64_t deviceExtMemSize, uint64_t &deviceCacheOffset)

Updates the actual input and output shape to the device before the second phase is executed.

SetSpace(void *space)

Sets the task space to be used.

ParseExtShapeType(AicpuExtInfo &aicpuExtInfo)

Parses the shape and type in the extended arguments.

ParseExtInputShape(AicpuExtInfo &aicpuExtInfo)

Parses the input shape in the extended arguments.

ParseExtOutputShape(AicpuExtInfo &aicpuExtInfo)

Parses the output shape in the extended arguments.

AppendExtOpName(std::string &taskExtInfo)

Appends the operator name to extended arguments.

AppendExtShapeType(std::string &taskExtInfo)

Appends the shape and type to extended arguments.

AppendExtBitMap(std::string &taskExtInfo)

Appends bitmap information to extended arguments.

AppendExtInfoShape(const FVector<const aclTensor *> &tensors, const aicpu::FWKAdapter::FWKTaskExtInfoType type, std::string &taskExtInfo, bool isTf = false)

Appends shape information to extended arguments.

AppendSessionInfo(std::string &taskExtInfo)

Appends session information to extended arguments.

UpdateShape(const gert::Shape &shape, AicpuShapeAndType *const shapeAndType)

Updates shape information.

GetShapeAndType(const AicpuShapeAndType &shapeAndType, gert::Shape &shape, ge::DataType &dataType)

Obtains the shape and type in the extended arguments.