CALCULATION PROCESSING APPARATUS AND METHOD

• US 20100223219A1
• Filed: 06/11/2008
• Published: 09/02/2010
• Est. Priority Date: 06/15/2007
• Status: Active Grant

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Ito, Yoshinori, Yamamoto, Takahisa, Kato, Masami
• Granted Patent: US 7,937,346 B2
• Time in Patent Office: Days
• US Class Current: 706/27
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A calculation processing apparatus for executing network calculations defined by hierarchically connecting a plurality of logical processing nodes that apply calculation processing to input data, sequentially designates a processing node which is to execute calculation processing based on sequence information that specifies an execution order of calculations of predetermined processing units to be executed by the plurality of processing nodes, so as to implement the network calculations, and executes the calculation processing of the designated processing node in the processing unit to obtain a calculation result. The calculation apparatus allocates partial areas of a memory to the plurality of processing nodes as ring buffers, and writes the calculation result in the memory while circulating a write destination of data to have a memory area corresponding to the amount of the calculation result of the processing unit as a unit.

Claims

1. A calculation processing apparatus hierarchically connecting a plurality of logical processing nodes that apply calculation processing to input data, said apparatus comprising:

designation means for sequentially designating a processing node which is to execute calculation processing from the plurality of processing nodes based on sequence information that specifies an execution order of calculations of predetermined processing units to be executed by the plurality of processing nodes;

execution means for executing the calculation processing of the processing node designated by said designation means unit in the processing unit to obtain a calculation result; and

memory control means for allocating ring buffers to the plurality of processing nodes respectively, and writing the calculation result of the processing node obtained by said execution means in the corresponding ring buffer while circulating a write destination of data in units of the amount corresponding to the calculation result of the processing unit.

2. The apparatus according to claim 1, wherein said memory control means determines allocations of the ring buffers to the processing nodes based on a logical connection state of the plurality of processing nodes.

3. The apparatus according to claim 2, wherein said memory control means determines the circular number of each ring buffer allocated to the processing node based on a size of data to be referred to when a processing node connected to a subsequent stage of that processing node executes calculations of the processing unit.

4. The apparatus according to any one of claims 1 to 3, wherein the input data is two-dimensional image data,

said execution means acquires a calculation result for one line in a horizontal direction by executing calculation processing to have one line of the image data in the horizontal direction as a unit, and

said memory control means circulates the write destination of data to have a memory area for one line in the horizontal direction as a unit.

5. The apparatus according to claim 2, wherein said designation means changes the sequence information in accordance with a change of calculations to be executed, and

said memory control means changes the allocations of the ring buffers to the processing nodes based on a connection structure of the processing nodes in the changed calculations.

6. The apparatus according to any one of claims 1 to 5, further comprising:

holding means for holding circulate positions indicating the write destinations of the calculation results in the ring buffers in association with the plurality of processing nodes.

7. The apparatus according to any one of claims 1 to 6, wherein said memory control means unit sets the circular numbers of the ring buffers for respective processing nodes.

8. The apparatus according to any one of claims 1 to 6, wherein said memory control means sets the circular numbers of the ring buffers for respective layers to which the processing nodes belong.

9. The apparatus according to any one of claims 1 to 8, wherein the ring buffer is a partial area of a memory having a continuous address space, and is shared by allocating different subspaces of the continuous address space to the plurality of processing nodes.

10. The apparatus according to any one of claims 1 to 9, wherein the calculation processing is calculation processing of a convolutional neural network.

11. A method of controlling a calculation processing apparatus hierarchically connecting a plurality of logical processing nodes that apply calculation processing to input data, said method comprising:

a designation step of sequentially designating a processing node which is to execute calculation processing from the plurality of processing nodes based on sequence information that specifies an execution order of calculations of predetermined processing units to be executed by the plurality of processing nodes;

an execution step of executing the calculation processing of the processing node designated in the designation step in the processing unit to obtain a calculation result; and

a memory control step of allocating ring buffers to the plurality of processing nodes respectively, and writing the calculation result of the processing node obtained in the execution step in the corresponding ring buffer while circulating a write destination of data in units of the amount corresponding to the calculation result of the processing unit as a unit.

12. The method according to claim 11, wherein in the memory control step, allocations of the ring buffers to the processing nodes are determined based on a logical connection state of the plurality of processing nodes.

13. The method according to claim 12, wherein in the memory control step, the circular number of each ring buffer allocated to the processing node is determined based on a size of data to be referred to when a processing node connected to a subsequent stage of that processing node executes calculations of the processing unit.

14. The method according to any one of claims 11 to 13, wherein the input data is two-dimensional image data,

in the execution step, a calculation result for one line in a horizontal direction is acquired by executing calculation processing to have one line of the image data in the horizontal direction as a unit, and

in the memory control step, the write destination of data is circulated to have a memory area for one line in the horizontal direction as a unit.

15. The method according to claim 12, wherein in the designation step, the sequence information is changed in accordance with a change of calculations to be executed, and

in the memory control step, the allocations of the ring buffers to the processing nodes are changed based on a connection structure of the processing nodes in the changed calculations.

16. The method according to any one of claims 11 to 15, further comprising:

a holding step of holding circulate positions indicating the write destinations of the calculation results in the ring buffers in association with the plurality of processing nodes.

17. The method according to any one of claims 11 to 16, wherein in the memory control step, the circular numbers of the ring buffers are set for respective processing nodes.

18. The method according to any one of claims 11 to 16, wherein in the memory control step, the circular numbers of the ring buffers are set for respective layers to which the processing nodes belong.

19. The method according to any one of claims 11 to 18, wherein the ring buffer is a partial area of a memory having a continuous address space, and is shared by allocating different subspaces of the continuous address space to the plurality of processing nodes.

20. The method according to any one of claims 11 to 19, wherein the calculation processing is calculation processing of a convolutional neural network.

22. A computer-readable storage medium storing a computer program for making a computer execute a control method according to any one of claims 11 to 20.

21. (canceled)