# US 2004/0064670 A1 · Data addressing

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

- Number: US 2004/0064670 A1
- Document ID: US20040064670A1
- Status: Published application
- Kind: A1
- Application number: 10/432,203
- Filed: 2003-10-29
- Published: 2004-04-01
- Earliest claimed: 2000-11-21
- Assignees: Aspex Technology Limited
- Inventors: John Lancaster; Martin Whitaker
- Web page: https://patentgrep.com/patent/US20040064670A1

## CPC

- G06F 9/30043: {LOAD or STORE instructions; Clear instruction}
- G06F 9/345: of multiple operands or results {(addressing multiple banks G06F12/06)}
- G06F 13/28: using burst mode transfer, e.g. direct memory access {DMA}, cycle steal (G06F13/32 takes precedence)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method of carrying out a data fetch operation for a data-parallel processor such as a SIMD processor is described. The operation is specifically involving the use of a plurality of non-sequential data addresses. The method comprises constructing a linear address vector from the non-sequential addresses, and using the address vector in a block fetch command to a data store.

## Claims

### Claim 1

A method of fetching data for processing on a data-parallel processor, the method comprising:
generating a plurality of non-linear data addresses for accessing data stored in a data store;
storing each of the non-linear addresses sequentially in a linear array to form an address vector;
reading out the address vector from the linear array to the data store;
retrieving from the data store the data stored at the plurality of non-linear addresses specified by the address vector; and
providing the read data to the data-parallel processor.

### Claim 2 (depends on 1)

A method according to claim 1, wherein the storing step comprises storing the plurality of non-linear addresses in a data store having a size that corresponds to the number of processing elements in the data-parallel processor.

### Claim 3 (depends on 1, 2)

A method according to claim 1 or 2, further comprising creating a pipeline by carrying out the non-linear address storing step and the retrieving data steps concurrently.

### Claim 4

A method according to any preceding claim, wherein the providing step comprises providing the read data to a plurality of processing elements of the data-parallel processor.

### Claim 5

A method according to any preceding claim, wherein the generating step is carried out by a plurality of data processing elements of the data-parallel processor.

### Claim 6

A method according to any preceding claim, wherein the storing step comprises storing each of the non-linear addresses in a store which is tightly coupled to the structure of the data-parallel processor and sequentially indexing each of the non-linear addresses to form an address vector.

### Claim 7 (depends on 6)

A method according to claim 6, wherein the storing step comprises storing each of the non-linear addresses in a portion of a data cache of the data-parallel processor.

### Claim 8

A method according to any preceding claim wherein the generating step comprises generating each of the plurality of addresses concurrently.

### Claim 9

A method of carrying out a data fetch operation, for a data-parallel processor, involving the use of a plurality of non-sequential data addresses, the method comprising:
constructing a linear address vector from the non-sequential addresses; and
using the address vector in a block fetch command to a data store.

### Claim 10

A combination of a data-parallel processor and an apparatus for fetching data for processing on the data-parallel processor, the combination comprising:
an address generator for generating a plurality of non-linear data addresses for accessing data stored in a main data store;
a temporary data store comprising a linear array, the store being arranged to store each of the non-linear addresses sequentially in the linear array to form an address vector;
a data retrieval controller arranged to read the address vector, to use the same to retrieve from the main data store the data stored at the plurality of non-linear addresses specified by the address vector, and to supply the read data to the data-parallel processor.

### Claim 11 (depends on 10)

A combination according to claim 10, wherein the address generator comprises the data-parallel processor.

### Claim 12 (depends on 10, 11)

A combination according to claim 10 or 11, wherein the address generator further comprises a data transferring controller for storing each of the generated addresses in the temporary store.

### Claim 13 (depends on 10, 11)

A combination according to claim 10 or 11, wherein the temporary data store comprises a conventional address memory.

### Claim 14 (depends on 10)

A combination according to claim 10, wherein the temporary data store comprises a data memory tightly coupled to the data-parallel processor.

### Claim 15 (depends on 10, 11, 12, 13, 14)

A combination according to any of claims 10 to 14, wherein the data retrieval controller comprises a direct memory access (DMA) controller arranged to read block of data from the main data store having sequentially stored data.

### Claim 16

An apparatus for use with a data-parallel processor for carrying out a data fetch operation involving the use of a plurality of non-sequential data addresses, the apparatus comprising:
an address vector generation means for converting the non-sequential data addresses into a linear address vector; and
a data fetch controller arranged to use the address vector in a block fetch command to a data store.
