Operator registration method and apparatus for deep learning framework, device and storage medium

1. An operator registration method for a deep learning framework, the method comprising:

receiving registration information provided by a user for registering operators with the deep learning framework, the registration information comprising: a custom calculation function, the custom calculation function being written in a manner irrelevant to the deep learning framework;

building operator meta-information in the deep learning framework based on the registration information; and

constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework.

2. The method according to claim 1 , wherein the registration information further comprises: custom input information and custom output information, and the step of building operator meta-information in the deep learning framework based on the registration information comprises:

converting the custom input information into standard input information within the deep learning framework;

converting the custom output information into standard output information within the deep learning framework;

processing the custom calculation function by using macros, so as to obtain a calculation function after macro processing; and

building the operator meta-information in the deep learning framework based on the standard input information, the standard output information and the calculation function after macro processing.

3. The method according to claim 2 , wherein the to-be-registered operator comprises to-be-registered operator description information and a to-be-registered operator kernel function, and the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information comprises:

constructing the to-be-registered operator description information based on the standard input information and the standard output information; and

constructing the to-be-registered operator kernel function based on the calculation function after macro processing.

4. The method according to claim 3 , wherein the to-be-registered operator kernel function comprises unified-form input information and a unified function pointer, and the step of constructing the to-be-registered operator kernel function based on the calculation function after macro processing comprises:

determining a structure for replacing the calculation function after macro processing, the structure comprising a static function, the static function having a unified form corresponding to different custom calculation functions;

taking input information of the static function as the unified-form input information; and

taking a function pointer of the static function as the unified function pointer.

5. The method according to claim 4 , wherein the input information of the static function forms an input information list, the input information list comprises an input tensor list, the custom input information comprises custom input information of at least one data type, the structure comprises at least one specialized substructure, different specialized substructures correspond to different data types, the data type comprises tensors, and the step of constructing the to-be-registered operator kernel function based on the calculation function after macro processing further comprises:

obtaining custom input information of the data types corresponding to the specialized substructures by using specialized substructures in the at least one specialized substructure;

forming the input tensor list with the custom input information whose data type are tensors; and

if the data type further comprises non-tensors and the input information list further comprises another data type list, forming the another data type list with the custom input information whose data type are non-tensors.

6. The method according to claim 1 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

7. The method according to claim 2 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

8. The method according to claim 3 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

9. The method according to claim 4 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

10. The method according to claim 5 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

11. An electronic device, comprising:

at least one processor; and

a memory communicatively connected with the at least one processor;

wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform an operator registration method for a deep learning framework, wherein the method comprises:

receiving registration information provided by a user for registering operators with the deep learning framework, the registration information comprising: a custom calculation function, the custom calculation function being written in a manner irrelevant to the deep learning framework;

building operator meta-information in the deep learning framework based on the registration information; and

constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework.

12. The electronic device according to claim 11 , wherein the registration information further comprises: custom input information and custom output information, and the step of building operator meta-information in the deep learning framework based on the registration information comprises:

converting the custom input information into standard input information within the deep learning framework;

converting the custom output information into standard output information within the deep learning framework;

processing the custom calculation function by using macros, so as to obtain a calculation function after macro processing; and

building the operator meta-information in the deep learning framework based on the standard input information, the standard output information and the calculation function after macro processing.

13. The electronic device according to claim 12 , wherein the to-be-registered operator comprises to-be-registered operator description information and a to-be-registered operator kernel function, and the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information comprises:

constructing the to-be-registered operator description information based on the standard input information and the standard output information; and

constructing the to-be-registered operator kernel function based on the calculation function after macro processing.

14. The electronic device according to claim 13 , wherein the to-be-registered operator kernel function comprises unified-form input information and a unified function pointer, and the step of constructing the to-be-registered operator kernel function based on the calculation function after macro processing comprises:

determining a structure for replacing the calculation function after macro processing, the structure comprising a static function, the static function having a unified form corresponding to different custom calculation functions;

taking input information of the static function as the unified-form input information; and

taking a function pointer of the static function as the unified function pointer.

15. The electronic device according to claim 14 , wherein the input information of the static function forms an input information list, the input information list comprises an input tensor list, the custom input information comprises custom input information of at least one data type, the structure comprises at least one specialized substructure, different specialized substructures correspond to different data types, the data type comprises tensors, and the step of constructing the to-be-registered operator kernel function based on the calculation function after macro processing further comprises:

obtaining custom input information of the data types corresponding to the specialized substructures by using specialized substructures in the at least one specialized substructure;

forming the input tensor list with the custom input information whose data type are tensors; and

if the data type further comprises non-tensors and the input information list further comprises another data type list, forming the another data type list with the custom input information whose data type are non-tensors.

16. The electronic device according to claim 11 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

17. The electronic device according to claim 12 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

18. The electronic device according to claim 13 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

19. The electronic device according to claim 14 , wherein the step of constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework comprises:

determining a current constructor corresponding to a current scenario based on the current scenario; and

constructing the to-be-registered operator within the deep learning framework based on the operator meta-information by using the current constructor, and registering the to-be-registered operator in the global operator table within the deep learning framework.

20. A non-transitory computer readable storage medium with computer instructions stored thereon, wherein the computer instructions are used for causing a computer to perform an operator registration method for a deep learning framework, wherein the method comprises:

receiving registration information provided by a user for registering operators with the deep learning framework, the registration information comprising: a custom calculation function, the custom calculation function being written in a manner irrelevant to the deep learning framework;

building operator meta-information in the deep learning framework based on the registration information; and

constructing a to-be-registered operator within the deep learning framework based on the operator meta-information, and registering the to-be-registered operator in a global operator table within the deep learning framework.