# US 2021/0241082 A1 · Method for accelerating operations and accelerator apparatus

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0241082 A1
- Document ID: US20210241082A1
- Status: Published application
- Kind: A1
- Application number: 17/048,048
- Filed: 2019-04-09
- Published: 2021-08-05
- Earliest claimed: 2018-04-19
- Assignees: AImotive Kft.
- Applicants: Aimotive Kft.
- Inventors: Gyula NAGY; Márton FEHÉR; István Herman WEIMANN
- Web page: https://patentgrep.com/patent/US20210241082A1

## CPC

- G06N 3/063: using electronic means
- G06F 12/0207: {with multidimensional access, e.g. row/column, matrix}
- G06F 12/0813: with a network or matrix configuration
- G06F 12/0846: Cache with multiple tag or data arrays being simultaneously accessible
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06F 2212/1016: Performance improvement
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

A method for accessing and processing data by an accelerator apparatus includes retrieving at least a part of input data to be processed by the accelerator apparatus, segmenting the input data into a plurality of adjacent input tiles, the input tiles having a pre-determined size, storing at least one of the plurality of adjacent input tiles in a data buffer of the accelerator apparatus, accessing data elements of the at least one of the plurality of adjacent input tiles stored in the data buffer by positioning an input window over the data elements stored in the data buffer, to generate a plurality of input areas, wherein the input window is adjustable according to a set of parameters, and sequentially processing at least one of the plurality of input areas to at least partially process the at least one of the plurality of adjacent input tiles by the accelerator apparatus.

## Claims

### Claim 1

A method for accessing and processing data by an accelerator apparatus, comprising:
retrieving at least a part of input data to be processed by the accelerator apparatus;
segmenting the input data into a plurality of adjacent input tiles, the input tiles having a pre-determined size;
storing at least one of the plurality of adjacent input tiles in a data buffer of the accelerator apparatus;
accessing data elements of the at least one of the plurality of adjacent input tiles stored in the data buffer by positioning an input window over the data elements stored in the data buffer, to generate a plurality of input areas, wherein the input window is adjustable according to a set of parameters; and
sequentially processing at least one of the plurality of input areas to at least partially process the at least one of the plurality of adjacent input tiles by the accelerator apparatus.

### Claim 2 (depends on 1)

The method of claim 1, further comprising sequentially storing another one of the plurality of adjacent input tiles in the data buffer, and accessing data of the other one of the plurality of adjacent input tiles using the input window.

### Claim 3 (depends on 1)

The method of claim 1, wherein the set of parameters includes a stride value and a dilation value, wherein the dilation value defines a distance between two successive positions of the input window.

### Claim 4 (depends on 3)

The method of claim 3, further comprising determining a processing mode of the accelerator apparatus, wherein, in a first processing mode, the stride value defines a spacing between data elements sampled by the input window, and wherein, in a second processing mode, the stride value defines a number of data elements generated by a single data element of the input window.

### Claim 5 (depends on 1)

The method of claim 1, wherein the set of parameters further includes a mapping to a coefficient matrix stored in a coefficient buffer, wherein an input area is processed by applying at least one element of the coefficient matrix to the input area according to the mapping.

### Claim 6 (depends on 5)

The method of claim 5, further comprising retrieving the at least one element of the coefficient matrix in parallel with generating the input area, and applying the at least one element of the coefficient matrix on the input area using at least one processing element.

### Claim 7 (depends on 1)

The method of claim 1, wherein accessing the data elements via the input window is hard-wired in a circuitry implementing the data buffer.

### Claim 8 (depends on 1)

The method of claim 1, wherein the data buffer is a two-dimensional data buffer and each of the plurality of input areas represents a two-dimensional area.

### Claim 9 (depends on 1)

The method of claim 1, wherein the accelerator apparatus implements a convolution operation on the input data and wherein the input data represent data elements of a layer of a neural network.

### Claim 10 (depends on 1)

The method of claim 1, wherein the input data include sensor data.

### Claim 11 (depends on 1)

An accelerator apparatus for processing input data, comprising:
an interface configured to couple the apparatus to an interconnect; and
at least one processing module configured to process input data, wherein the accelerator apparatus is configured to retrieve at least some of the input data via the interface and process the input data using the at least one processing module according to claim 1.

### Claim 12 (depends on 11)

The accelerator apparatus of claim 11, wherein the apparatus is configured to retrieve input data via the interface, the input data representing at least a part of an input layer of a neural network, wherein the at least one processing module processes at least a part of a layer of the neural network.

### Claim 13 (depends on 11)

The accelerator apparatus of claim 11, further comprising a cache module configured to store at least a portion of data processed by the at least one processing module, wherein data stored in the cache module represent at least a part of a next layer of the neural network.

### Claim 14 (depends on 11)

The accelerator apparatus of claim 11, wherein the accelerator apparatus is a convolutional neural network accelerator.

### Claim 15 (depends on 11)

A system comprising:
a processing host;
at least one memory;
an accelerator apparatus according to claim 11; and
an interconnect coupling the processing host, the at least one memory, and the accelerator apparatus, wherein the accelerator apparatus is configured to retrieve input data from the at least one memory via the interconnect.
