Adapter Plug-in APIs - Atlas Data Center Solution V100R020C00 TBE Custom Operator Development Guide(Training)

Atlas Data Center Solution V100R020C00 TBE Custom Operator Development Guide(Training)

Huawei uses machine translation combined with human proofreading to translate this document to different languages in order to help you better understand the content of this document. Note: Even the most advanced machine translation cannot match the quality of professional translators. Huawei shall not bear any responsibility for translation accuracy and it is recommended that you refer to the English document (a link for which has been provided).

Adapter Plug-in APIs

Class OpRegistrationData

Overview

After completing the implementation code of a custom operator, you need to develop an adaptation plug-in to map an operator developed based on a third-party framework to that supported by the Ascend AI processor. Call the REGISTER_CUSTOM_OP macro to implement operator conversion. As shown in the following, the calling starts with REGISTER_CUSTOM_OP. Then, APIs such as FrameworkType, OriginOpType, and ParseParamsFn are linked with a dot (.).

Example:

REGISTER_CUSTOM_OP("OpType")    
   .FrameworkType(TENSORFLOW)     
   .OriginOpType("OriginOpType")    
   .ParseParamsFn(ParseParamFunc);

Constructor and Destructor

Function

OpRegistrationData constructor and destructor.

Syntax

OpRegistrationData(const std::string& om_optype);

~OpRegistrationData()

Parameter Description

Parameter	Input/Output	Description
om_optype	Input	Specifies the operator type supported by models of the Ascend AI Processor.

Return Value

The OpRegistrationData constructor returns an object of the OpRegistrationData type.

Exception Handling

None

Restriction

None

REGISTER_CUSTOM_OP

Function

Registers an operator with a specified name.

Syntax

REGISTER_CUSTOM_OP(name)

Parameter Description

Parameter	Input/Output	Description
name	Input	Operator type name.

FrameworkType

Description

Sets the framework type of the original model.

Prototype

OpRegistrationData &FrameworkType(const domi::FrameworkType& fmk_type)

Parameters

Parameter	Input/Output	Description
fmk_type	Input	Framework type 0: Caffe 3: TensorFlow The enum values of FrameworkType are defined as follows: enum FrameworkType { CAFFE = 0, TENSORFLOW = 3, ANDROID_NN, FRAMEWORK_RESERVED, };

Parameter

Input/Output

Description

fmk_type

Input

Framework type

0: Caffe
3: TensorFlow

The enum values of FrameworkType are defined as follows:

enum FrameworkType

{

CAFFE = 0,

TENSORFLOW = 3,

ANDROID_NN,

FRAMEWORK_RESERVED,

};

OriginOpType

Description

Sets the operator type or list of operator types of the original model.

Prototype

OpRegistrationData &OriginOpType(const std::initializer_list<std::string>& ori_optype_list);

OpRegistrationData &OriginOpType(const std::string& ori_optype)

Parameters

Parameter	Input/Output	Description
ori_optype_list	Input	List of operator types of the original model
ori_optype	Input	Operator type of the original model

ParseParamsFn

Description

Registers the function for parsing operator attributes.

Restrictions

ParseParamsFn will not be available in future releases. User ParseParamsByOperatorFn instead.

Prototype

OpRegistrationData &ParseParamsFn(const ParseParamFunc& parseParamFn)

Parameters

Parameter	Input/Output	Description
parseParamFn	Input	ParseParamFunc for parsing operator attributes. For details, see Callback ParseParamFunc. If the attributes of a TensorFlow operator are consistent with those of an operator of the Ascend AI Processor (both the number and sequence), you can use the AutoMappingFn function to implement automatic mapping.

Parameter

Input/Output

Description

parseParamFn

Input

ParseParamFunc for parsing operator attributes. For details, see Callback ParseParamFunc.

If the attributes of a TensorFlow operator are consistent with those of an operator of the Ascend AI Processor (both the number and sequence), you can use the AutoMappingFn function to implement automatic mapping.

Callback ParseParamFunc

Customizes and implements a function of class FusionParseParamFunc to map attributes of the original model to an offline model adapted to Ascend AI Processor and writes the result to an Operator class.

Status ParseParamFunc(const Message* op_origin, ge::Operator& op_dest)

Table 10-48 Parameter description

Parameter	Input/Output	Description
op_origin	Input	Data structure in protobuf format (from the Caffe prototxt file), which stores the operator information
op_dest	Output	Operator data structure of the offline model adapted to Ascend AI Processor, which stores the operator information For details about class Operator, see Class Operator.

Parameter

Input/Output

Description

op_origin

Input

Data structure in protobuf format (from the Caffe prototxt file), which stores the operator information

op_dest

Output

Operator data structure of the offline model adapted to Ascend AI Processor, which stores the operator information

For details about class Operator, see Class Operator.

ParseParamsByOperatorFn

Description

Registers the function for parsing the attributes of a custom Caffe operator.

Prototype

OpRegistrationData &ParseParamsByOperatorFn(const ParseParamByOpFunc &parse_param_by_op_fn)

Parameters

Parameter	Input/Output	Description
parse_param_by_op_fn	Input	Function for parsing the attributes of a custom Caffe operator. For details, see Callback Function ParseParamByOpFunc.

Callback Function ParseParamByOpFunc

Customizes and implements a function of class ParseParamByOpFunc to map attributes of the original model to an offline model adapted to Ascend AI Processor and writes the result to an Operator class.

Status ParseParamByOpFunc(const ge::Operator & op_origin, ge::Operator& op_dest)

Table 10-49 Parameters

Parameter	Input/Output	Description
op_origin	Input	An object of class Operator defined by the framework, including the custom operator attributes of the parsed Caffe model. For details about class Operator, see Class Operator.
op_dest	Output	Operator data structure of the offline model adapted to Ascend AI Processor, storing the operator information. For details about class Operator, see Class Operator.

Parameter

Input/Output

Description

op_origin

Input

An object of class Operator defined by the framework, including the custom operator attributes of the parsed Caffe model. For details about class Operator, see Class Operator.

op_dest

Output

Operator data structure of the offline model adapted to Ascend AI Processor, storing the operator information.

For details about class Operator, see Class Operator.

Restrictions

None

FusionParseParamsFn

Description

Registers a function for parsing the attributes of fused operators.

Prototype

OpRegistrationData &FusionParseParamsFn(const FusionParseParamFunc &fusionParseParamFn)

Parameters

Parameter	Input/Output	Description
fusionParseParamFn	Input	Function for parsing the attributes of fused operators. For details, see Callback FusionParseParamFunc.

Callback FusionParseParamFunc

Customizes and implements a function of class FusionParseParamFunc to map attributes of the original model to the offline model adapted to Ascend AI Processor and fills the result in an Operator class.

Status FusionParseParamFunc(const vector<const google::protobuf::Message *> &v_op_origin, ge::Operator &op_dest)

Table 10-50 Parameters

Parameter	Input/Output	Description
v_op_origin	Input	protobuf data structures (from the prototxt file of the original model) within the scope, which stores the operator information.
op_dest	Output	Data structure of fused operators, which stores the fused operator information. For details, see Class Operator.

Parameter

Input/Output

Description

v_op_origin

Input

protobuf data structures (from the prototxt file of the original model) within the scope, which stores the operator information.

op_dest

Output

Data structure of fused operators, which stores the fused operator information.

For details, see Class Operator.

ParseSubgraphPostFn

Description

Registers the function for mapping between the input and output nodes in the subgraphs of an operator and the input and output nodes of the operator based on the operator type.

Prototype

OpRegistrationData &OpRegistrationData::ParseSubgraphPostFn(const ParseSubgraphFunc &subgraph_post_fn)

Parameters

Parameter	Input/Output	Description
subgraph_post_fn	Input	Function object for mapping between the input and output nodes in the operator subgraphs and the input and output nodes of the operator

Restrictions

None

ImplyType

Function

Sets the operator execution mode.

Syntax

OpRegistrationData& ImplyType(const domi::ImplyType& imply_type)

Parameter Description

Parameter	Input/Output	Description
imply_type	Input	Operator execution mode. enum class ImplyType : unsigned int { BUILDIN = 0,// Built-in operator, which is normally executed by the OME TVM, // Executed after being compiled into a TVM binary file CUSTOM, // The computation logic is customized by the user and executed by the CPU. AI_CPU, // AI CPU custom operator type INVALID = 0xFFFFFFFF, };

Parameter

Input/Output

Description

imply_type

Input

Operator execution mode.

enum class ImplyType : unsigned int

{

BUILDIN = 0,// Built-in operator, which is normally executed by the OME

TVM, // Executed after being compiled into a TVM binary file

CUSTOM, // The computation logic is customized by the user and executed by the CPU.

AI_CPU, // AI CPU custom operator type

INVALID = 0xFFFFFFFF,

};

DelInputWithCond

Function

Deletes an input edge of an operator based on the operator attribute.

Prototype

OpRegistrationData& DelInputWithCond(int inputIdx, const std::string& attrName, bool attrValue)

Parameters

Parameter	Input/Output	Description
inputIdx	Input	Index of the input edge to be deleted
attrName	Input	Attribute name
attrValue	Input	Attribute value

Restrictions

None

DelInputWithOriginalType

Description

Deletes an input edge of an operator based on the operator type.

Prototype

OpRegistrationData &DelInputWithOriginalType(int input_idx, const std::string &ori_type)

Parameters

Parameter	Input/Output	Description
inputIdx	Input	Index of the input edge to be deleted
ori_type	Input	Original operator type

AutoMappingFnDynamic

Function

Implements automatic mapping.

Syntax

Status AutoMappingFnDynamic(const google::protobuf::Message *op_src, ge::Operator &op, std::map<std::string, std::pair<std::string, std::string>> dynamic_name_attr_value, int in_pos = -1, int out_pos = -1)

Parameter Description

Parameter	Input/Output	Description
op_src	Input	Original operator before conversion.
op	Input	Mapped operator.
dynamic_name_attr_value	Input	Actual number of dynamic inputs or outputs. key specifies whether they are inputs or outputs.
in_pos	Input	Port ID for the dynamic input.
out_pos	Input	Port ID for the dynamic output.

AutoMappingSubgraphIndex

Description

Sets the mapping between the input and output of the subgraph and the input and output of the parent node of the main model.

Prototype

Status AutoMappingSubgraphIndex(const ge::Graph &graph,

const std::function<int(int data_index)> &input,

const std::function<int(int netoutput_index)> &output)

Status AutoMappingSubgraphIndex(const ge::Graph &graph,

const std::function<Status(int data_index, int &parent_input_index)> &input,

const std::function<Status(int netoutput_index, int &parent_output_index)> &output)

Parameters

Parameter	Input/Output	Description
graph	Input	Subgraph object
input	Input	Input mapping function
output	Input	Output mapping function

Restrictions

None

Class OpReceiver

Constructor and Destructor

Function

The OpReceiver constructor receives the registration information of a custom operator.

Syntax

OpReceiver(OpRegistrationData& reg_data);

~OpReceiver()

Parameter Description

Parameter	Input/Output	Description
reg_data	Input	Operator information to be registered.

Return Value

Returns an object of class OpReceiver.

Exception Handling

None

Restriction

None

Translation

简体中文

Download

Updated: 2020-11-23

Document ID: EDOC1100155041

Downloads: 10

Average rating:

This Document Applies to these Products