Convolution calculation apparatus and method

Patent metadata

Field	Value
Publication number	US-11907329-B2
Application number	US-202117328006-A
Country	US
Kind code	B2
Filing date	May 24, 2021
Priority date	Sep 18, 2020
Publication date	Feb 20, 2024
Grant date	Feb 20, 2024

Abstract

A convolution calculation apparatus applied for convolution calculation of a convolution layer includes a decompression circuit, a data combination circuit and a calculation circuit. The decompression circuit decompresses compressed weighting data of a convolution kernel of the convolution layer to generate decompressed weighting data. The data combination circuit combines the decompressed weighting data and non-compressed data of the convolution kernel to restore a data order of weighting data of the convolution kernel. The calculation circuit performs calculation according to the weighting data of the convolution kernel and input data of the convolution layer. Since the compressed weighting data of the convolution kernel is transmitted to the convolution calculation apparatus in advance, the compressed weighting data is first decompressed and then convolution calculation is performed accordingly, hence reducing the storage amount and transmission bandwidth used by the convolution kernel in an electronic apparatus.

First claim

What is claimed is:

1. A convolution calculation apparatus, applied for convolution calculation of a convolution layer, the apparatus comprising:
a decompression circuit, decompressing compressed weighting data of a convolution kernel of the convolution layer to generate decompressed weighting data;
a data combination circuit, combining the decompressed weighting data and non-compressed data of the convolution kernel to restore a data order of weighting data of the convolution kernel;
a calculation circuit, performing calculation according to the weighting data of the convolution kernel and input data of the convolution layer; and
a conversion circuit, coupled between the data combination circuit and the calculation circuit, performing format conversion on the weighting data of the convolution kernel, and outputting the converted weighting data to the calculation circuit.

2. A convolution calculation apparatus, applied for convolution calcalution of a convolution layer, the apparatus comprising:
a decompression circuit, decompressing compressed weighting data of a convolution kernel of the convolution layer to generate decompressed weighting data;
a data combination circuit, combining the decompressed weighting data and non-compressed data of the convolution kernel to restore a data order of weighting data of the convolution kernel; and
a calculation circuit, performing calculation according to the weighting data of the convolution kernel and input data of the convolution layer,
wherein the weighting data of the convolution kernel comprises first-part data and second-part data, the compressed weighting data corresponds to the first-part data and the non-compressed weighting data corresponds to the second-part data.

3. The convolution calculation apparatus according to claim 2 , wherein the weighting data of the convolution kernel consists of a plurality of 8-bit binary values, and the first-part data of the weighting data consists of at least 2-bit binary values in each of the 8-bit binary values.

4. The convolution calculation apparatus according to claim 1 , wherein the compressed weighting data is compressed by using a Huffman compression format, and the decompression circuit decompresses the compressed weighting data according to a Huffman code table.

5. The convolution calculation apparatus according to claim 1 , wherein the compressed weighting data is stored in a first block of a memory, the non-compressed weighting data is stored in a second block of the memory, and the first block is different from the second block.

6. The convolution calculation apparatus according to claim 1 , wherein the compressed weighting data is compressed by a device outside an electronic apparatus provided with the convolution calculation apparatus.

7. A convolution calculation apparatus, applied for convolution calculation of a convolution layer, the apparatus comprising:
a decompression circuit, decompressing compressed weighting data of a convolution kernel of the convolution layer to generate decompressed weighting data;
a calculation circuit, performing calculation according to the decompressed weighting data and input data of the convolution layer; and
a conversion circuit, coupled between the decompression circuit and the calculation circuit, performing format conversion on the decompressed weighting data, and outputting the converted weighting data to the calculation circuit;
wherein, the compressed weighting data and the input data are respectively stored in different blocks of a memory.

8. The convolution calculation apparatus according to claim 7 , wherein the compressed weighting data is compressed by using a Huffman compression format, and the decompression circuit decompressed the compressed weighting data according to a Huffman code table.

9. The convolution calculation apparatus according to claim 7 , wherein the compressed weighting data is compressed by a device outside an electronic apparatus provided with the convolution calculation apparatus.

Assignees

• Xiamen Sigmastar Tech Ltd
• Sigmastar Technology Ltd

Inventors

• Bao Fabo
• Liu Donghao
• Zhu Wei
• Zheng Chengwei

Classifications

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06F17/153 Primary

  Multidimensional correlation or convolution · CPC title

• G06F9/545

  where tasks reside in different layers, e.g. user- and kernel-space · CPC title

• G06N3/045 Primary

  Combinations of networks · CPC title