# US 2024/0256436 A1 · Parallel program control system and method

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

- Number: US 2024/0256436 A1
- Document ID: US20240256436A1
- Status: Published application
- Kind: A1
- Application number: 18/162,491
- Filed: 2023-01-31
- Published: 2024-08-01
- Earliest claimed: 2023-01-31
- Assignees: Arm Limited
- Applicants: Arm Limited
- Inventors: Kévin Petit
- Web page: https://patentgrep.com/patent/US20240256436A1

## CPC

- G06F 12/023: {Free address space management}
- G06F 9/4881: {Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues}
- G06F 12/0207: {with multidimensional access, e.g. row/column, matrix}
- G06F 2212/1016: Performance improvement
- G06F 2212/1044: Space efficiency improvement
- G06F 2212/251: Local memory within processor subsystem
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A system, method and computer program product configured to control a plurality of parallel programs operating in an n-dimensional hierarchical iteration space over an n-dimensional data space, comprising: a processor and a memory configured to accommodate the plurality of parallel programs and the data space; a memory access control decoder configured to decode memory location references to regions of the n-dimensional data space from indices in the plurality of parallel programs; and an execution orchestrator responsive to the memory access control decoder and configured to sequence regions of the n-dimensional hierarchical iteration space of the plurality of parallel programs to honour a data requirement of at least a first of the plurality of parallel programs having a data dependency on at least a second of the plurality of parallel programs.

## Claims

### Claim 1

A system configured to control a plurality of parallel programs operating in an n-dimensional hierarchical iteration space over an n-dimensional data space, comprising:
a processor and a memory configured to accommodate the plurality of parallel programs and the n-dimensional data space;
a memory access control decoder configured to decode memory location references to regions of the n-dimensional data space from indices in the plurality of parallel programs; and
an execution orchestrator responsive to the memory access control decoder and configured to sequence regions of the n-dimensional hierarchical iteration space of the plurality of parallel programs to honour a data requirement of at least a first of the plurality of parallel programs having a data dependency on at least a second of the plurality of parallel programs.

### Claim 2 (depends on 1)

The system according to claim 1, wherein the data requirement comprises a contract to restrict a program's memory access to a specific region.

### Claim 3 (depends on 1)

The system according to claim 1, wherein the memory access control decoder is further configured to decode memory location references of a partially specified tensor to determine dimensions in memory of sufficient size to accommodate the tensor and at least one affine transform of the tensor.

### Claim 4 (depends on 3)

The system according to claim 3, wherein the execution orchestrator is responsive to the memory access control decoder to allocate a memory region of the determined dimensions.

### Claim 5 (depends on 1)

The system according to claim 1, wherein the execution orchestrator sequencing the regions comprises fusing together programs having a data dependency into a single execution entity.

### Claim 6 (depends on 1)

The system according to claim 1, wherein the execution orchestrator sequencing the regions comprises tiling execution of the first and the second programs.

### Claim 7 (depends on 1)

The system according to claim 1, wherein the memory access control decoder is configured to decode a relationship between a memory location in the n-dimensional data space and an index in one of the plurality of parallel programs.

### Claim 8 (depends on 1)

The system according to claim 1, wherein the parallel programs comprise single instruction multiple thread programs.

### Claim 9

A method of operating a controller to control a plurality of parallel programs operating in an n-dimensional hierarchical iteration space over an n-dimensional data space, comprising:
decoding memory location references to regions of the n-dimensional data space from indices in the plurality of parallel programs; and
responsive to the decoding of the memory location references, sequencing regions of the n-dimensional hierarchical iteration space of the plurality of parallel programs to honour a data requirement of at least a first of the plurality of parallel programs having a data dependency on at least a second of the plurality of parallel programs.

### Claim 10 (depends on 9)

The method according to claim 9, wherein honouring the data requirement comprises honouring a contract to restrict a program's memory access to a specific region.

### Claim 11 (depends on 9)

The method according to claim 9, further comprising decoding memory location references of a partially specified tensor to determine dimensions in memory of sufficient size to accommodate the tensor and at least one affine transform of the tensor.

### Claim 12 (depends on 11)

The method according to claim 11, further comprising allocating a memory region of the determined dimensions.

### Claim 13 (depends on 11)

The method according to claim 11, wherein sequencing the regions comprises fusing together programs having a data dependency into a single execution entity.

### Claim 14 (depends on 11)

The method according to claim 11, wherein sequencing the regions comprises tiling execution of the first and the second programs.

### Claim 15 (depends on 11)

The method according to claim 11, further comprising decoding a relationship between a memory location in the n-dimensional data space and an index in one of the plurality of parallel programs.

### Claim 16

A computer program product stored in a non-transitory computer readable medium and comprising computer readable program code to:
decode memory location references to regions of an n-dimensional data space from indices in a plurality of parallel programs; and
sequence regions of an n-dimensional hierarchical iteration space of the plurality of parallel programs to honour a data requirement of at least a first of the plurality of parallel programs having a data dependency on at least a second of the plurality of parallel programs.

### Claim 17 (depends on 16)

The computer program product according to claim 16, wherein honouring the data requirement comprises honouring a contract to restrict a program's memory access to a specific region.

### Claim 18 (depends on 16)

The computer program product according to claim 16, further comprising decoding memory location references of a partially specified tensor to determine dimensions in memory of sufficient size to accommodate the tensor and at least one affine transform of the tensor.

### Claim 19 (depends on 18)

The computer program product according to claim 18, further comprising allocating a memory region of the determined dimensions.

### Claim 20 (depends on 16)

The computer program product according to claim 16, further comprising decoding a relationship between a memory location in the n-dimensional data space and an index in one of the plurality of parallel programs.
