Arithmetic processing device for deep learning and control method of the arithmetic processing device for deep learning

• US 10,353,705 B2
• Filed: 07/17/2017
• Issued: 07/16/2019
• Est. Priority Date: 08/12/2016
• Status: Active Grant

• Current Assignee: Fujitsu Limited
• Original Assignee: Fujitsu Limited
• Inventors: Kuramoto, Masahiro
• Primary Examiner(s): Yaary, Michael D.
• Application Number:
  US15/651,651
• Publication Number:
  US 20180046458A1
• Time in Patent Office: 729 Days
• Field of Search: None
• CPC Class Codes: G06F 17/16 Matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization matrix transposition G06F7/78
• CPC Class Codes: G06F 9/3001 Arithmetic instructions
• CPC Class Codes: G06F 9/30036 Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
• CPC Class Codes: G06F 9/3012 Organisation of register space, e.g. banked or distributed register file
• CPC Class Codes: G06F 9/30141 Implementation provisions of register files, e.g. ports
• CPC Class Codes: G06F 9/3887 controlled by a single instruction for multiple data lanes [SIMD]
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods

Abstract

A memory 11 stores therein first data and second data each of which has element data that forms a matrix. Arithmetic units 51 to 53 repeat, for each of a first predetermined row of the first data and a second predetermined row of the second data that are stored in the memory 11, by using the element data included in the first predetermined row and the element data included in the second predetermined row, a row portion operation based on the number of columns in the second data and performs, by using results of the row portion operations, an arithmetic operation process that acquires the operation results of the operation that uses the first data and the second data.

Claims

1. An arithmetic processing device comprising:

a processor, a memory and a product-sum calculator, the memory stores first data and second data each of which has element data that forms a matrix;

the product-sum calculator performs a row portion operation by performing, by sequentially shifting a predetermined position from a top by a predetermined number at a time, an operation that arranges, in an order of an arrangement in the first data, element data included in a first predetermined row that has a predetermined number of rows, that acquires first element data that corresponds to the element data arranged from the predetermined position to a number of columns of the second data, that multiplies each of the pieces of the acquired first element data by an associated pieces of the element data included in a second predetermined row that has the predetermined number of rows, and that sums multiplication results, and performs, by using results of the row portion operations, a convolution operation that is performed by shifting, by using the second data as weight data, arrangement positions of the second data in the first data by a predetermined number at a time, and

the processor performs an image recognition by using a result of the convolution operation.

2. The arithmetic processing device according to claim 1, further comprising:

a first operation purpose storage unit;

a second operation purpose storage unit;

a first data control unit that acquires the element data included in the first predetermined row from the first data stored in the memory and that stores the acquired element data in the first operation purpose storage unit; and

a second data control unit that acquires the element data included in the second predetermined row from the second data in the memory and that stores the acquired element data in the second operation purpose storage unit, wherein

the product-sum calculator performs the operation based on the element data stored in the first operation purpose storage unit and the second operation purpose storage unit and instructs, after the end of the row portion operations, the first data control unit and the second data control unit to acquire the element data in another row and store the acquired element data.

3. The arithmetic processing device according to claim 2, wherein

a plurality of the product-sum calculator, the first operation purpose storage units, and the second operation purpose storage units are arranged, and

the arithmetic processing device further comprises a switch that switches connection paths among the product-sum calculator, the first operation purpose storage units, and the second operation purpose storage units.

4. An arithmetic processing device comprising:

a processor, a memory and a product-sum calculator,

the memory stores first data and second data each of which has element data that forms a matrix, the second data being a difference between an output expected value and calculation data that is calculated in a convolution operation that is performed by shifting, by a predetermined number at a time, arrangement positions of weight data arranged in the first data;

the product-sum calculator uses a single row in the first data as a first predetermined row, uses a single row in the second data as a second predetermined row, repeats, for each of the first predetermined row and the second predetermined row, by using the element data included in the first predetermined row and the element data included in the second predetermined row, a row portion operation based on the number of columns of the second data, and performs an operation that calculates a difference between the weight data and a weight expected value by using the first data and the second data, and

the processor performs an image recognition by using a result of the operation.

5. An arithmetic processing device comprising:

a processor, a memory and a product-sum calculator,

the memory stores first data and second data each of which has element data that forms a matrix, the second data being a difference between an output expected value and calculation data that is calculated in a convolution operation that is performed by shifting, by a predetermined number at a time, arrangement positions of weight data arranged in the first data, the first data being a difference between an output expected value and calculation data that is calculated in a convolution operation that is performed by using the second data as weight data and by shifting arrangement positions of the second data in input data by a predetermined number at a time;

the product-sum calculator uses, in a specific layer in one of a plurality of layers arranged in order, a single row in the first data as the first predetermined row, uses, in a specific layer, a single row in the second data as the second predetermined row, repeats, for each of the first predetermined row and the second predetermined row, by using the element data included in the first predetermined row and the element data included in the second predetermined row, a row portion operation based on the number of columns of the second data, and that performs an operation that calculates, for each of the predetermined number by using the first data and the second data, a difference between an output expected value and output data in a layer that is immediately previous to the specific layer, and

the processor performs an image recognition by using a result of the operation.