# US 2019/0377706 A1 · Dynamic simd instruction issue target selection

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

- Number: US 2019/0377706 A1
- Document ID: US20190377706A1
- Status: Published application
- Kind: A1
- Application number: 16/005,790
- Filed: 2018-06-12
- Published: 2019-12-12
- Earliest claimed: 2018-06-12
- Assignees: Arm Limited
- Applicants: Arm Limited
- Inventors: Cedric Denis Robert AIRAUD; Luca NASSI; Damien Robin MARTIN; Xiaoyang SHEN
- Web page: https://patentgrep.com/patent/US20190377706A1

## CPC

- G06F 1/3243: {Power saving in microcontroller unit}
- G06F 15/8015: {One dimensional arrays, e.g. rings, linear arrays, buses}
- G06F 1/3287: by switching off individual functional units in the computer system
- G06F 1/3293: by switching to a less power-consuming processor, e.g. sub-CPU
- G06F 9/30036: {Instructions to perform operations on packed data, e.g. vector, tile or matrix operations}
- G06F 9/3887: {controlled by a single instruction for multiple data lanes [SIMD]}
- G06F 9/3891: {organised in groups of units sharing resources, e.g. clusters}
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- 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

Apparatuses and methods of data processing are disclosed. An apparatus comprises two data processing clusters each having multiple data processing lanes to perform single instruction multiple data (SIMD) processing. Decoded instructions are issued to at least one of the two data processing clusters. A decoded SIMD instruction specifying a vector length which is more than the width of the data processing lanes of the first data processing cluster has a first part issued to the first data processing cluster for execution. An issuance target for a second remaining part of the decoded SIMD instruction is selected in dependence on a dynamic performance condition. When the dynamic performance condition has a first state the issuance target is the first data processing cluster and when the dynamic performance condition has a second state the issuance target is the second data processing cluster. When the issuance target is the first data processing cluster, to schedule the first and second parts of the decoded SIMD instruction in series.

## Claims

### Claim 1

Apparatus comprising:
a first data processing cluster arranged to perform single instruction multiple data (SIMD) processing comprising a first plurality of data processing lanes, wherein the first plurality of data processing lanes has a first width;
a second data processing cluster arranged to perform SIMD processing comprising a second plurality of data processing lanes, wherein the second plurality of data processing lanes has a second width;
issue circuitry to issue decoded instructions to at least one of the first data processing cluster and the second data processing cluster,
wherein the issue circuitry is responsive to receipt of a decoded SIMD instruction specifying a vector length which is more than the first width:
to issue a first part of the decoded SIMD instruction having the first width to the first data processing cluster for execution;
to select an issuance target for a second part of the decoded SIMD instruction having a remainder width of the vector length less the first width in dependence on a dynamic performance condition,
wherein when the dynamic performance condition has a first state the issuance target is the first data processing cluster and when the dynamic performance condition has a second state the issuance target is the second data processing cluster; and
when the issuance target is the first data processing cluster, to schedule the first and second parts of the decoded SIMD instruction in series.

### Claim 2 (depends on 1)

The apparatus as claimed in claim 1, wherein the issue circuitry is responsive to receipt of a decoded SIMD instruction specifying a vector length which is no more than the first width:
to issue the decoded SIMD instruction specifying a vector length which is no more than the first width to the first data processing cluster for execution.

### Claim 3 (depends on 1)

The apparatus as claimed in claim 1, wherein the first data processing cluster belongs to a first power domain and the second data processing cluster belongs to a second power domain.

### Claim 4 (depends on 3)

The apparatus as claimed in claim 3, wherein the dynamic performance condition is a not-fully-powered state of the second power domain.

### Claim 5 (depends on 3)

The apparatus as claimed in claim 3, wherein the first data processing cluster has access to a first register file and a second register file, the second data processing cluster has access to the second register file, and the first register file and the second register file belong to the first power domain.

### Claim 6 (depends on 3)

The apparatus as claimed in claim 3, further comprising power domain control circuitry responsive to a power-restriction signal to power down the second power domain.

### Claim 7 (depends on 6)

The apparatus as claimed in claim 6, wherein the power-restriction signal is generated under control of performance management circuitry.

### Claim 8 (depends on 6)

The apparatus as claimed in claim 6, wherein the power-restriction signal is generated under control of software.

### Claim 9 (depends on 6)

The apparatus as claimed in claim 6, wherein the issue circuitry and the power domain control circuitry are arranged to perform a coordinated power-on procedure for the second domain, wherein in the coordinated power-on procedure the power domain control circuitry causes the dynamic performance condition to have the first state until the second domain is fully powered and when the second domain is fully powered to cause the dynamic performance condition to have the second state.

### Claim 10 (depends on 1)

The apparatus as claimed in claim 1, wherein the first data processing cluster has access to a first register file and a second register file, and the second data processing cluster has access to the second register file.

### Claim 11 (depends on 10)

The apparatus as claimed in claim 10, wherein the first register file and a second register file together form a combined register file, and wherein bits of a less significant half of the combined register file represent the first register file and bits of a more significant half of the combined register file represent the second register file.

### Claim 12 (depends on 1)

The apparatus as claimed in claim 1, wherein the first data processing cluster has access to a first register file and the second data processing cluster has access to a second register file, and the apparatus is responsive to initiation of a switch of the dynamic performance condition from the second state to the first state to copy content of the second register file to storage accessible to the first data processing cluster.

### Claim 13 (depends on 12)

The apparatus as claimed in claim 12, wherein the first data processing cluster is arranged to perform the SIMD processing by reference to a set of architectural registers, and the apparatus comprises register renaming circuitry to remap the set of architectural registers to a first subset of a first set of physical registers,
wherein the apparatus is responsive to the initiation of the switch of the dynamic performance condition from the second state to the first state to copy content of the second register file to a second subset of the first set of physical registers.

### Claim 14 (depends on 12)

The apparatus as claimed in claim 12, wherein the content of the second register file copied has the second width.

### Claim 15

A method of data processing comprising:
performing single instruction multiple data (SIMD) processing in a first data processing cluster comprising a first plurality of data processing lanes, wherein the first plurality of data processing lanes has a first width;
performing SIMD processing in a second data processing cluster comprising a second plurality of data processing lanes, wherein the second plurality of data processing lanes has a second width;
issuing decoded instructions to at least one of the first data processing cluster and the second data processing cluster; and
in response to receipt of a decoded SIMD instruction specifying a vector length which is more than the first width:
issuing a first part of the decoded SIMD instruction having the first width to the first data processing cluster for execution;
selecting an issuance target for a second part of the decoded SIMD instruction having a remainder width of the vector length less the first width in dependence on a dynamic performance condition,
wherein when the dynamic performance condition has a first state the issuance target is the first data processing cluster and when the dynamic performance condition has a second state the issuance target is the second data processing cluster; and
scheduling the first and second parts of the decoded SIMD instruction in series when the issuance target is the first data processing cluster.

### Claim 16

Apparatus comprising:
first means for performing single instruction multiple data (SIMD) processing comprising a first plurality of data processing lanes, wherein the first plurality of data processing lanes has a first width;
second means for performing SIMD processing comprising a second plurality of data processing lanes, wherein the second plurality of data processing lanes has a second width;
means for issuing decoded instructions to at least one of the first means for performing SIMD processing and the second means for performing SIMD processing; and
in response to receipt of a decoded SIMD instruction specifying a vector length which is more than the first width for causing activation of:
means for issuing a first part of the decoded SIMD instruction having the first width to the first means for performing SIMD processing for execution;
means for selecting an issuance target for a second part of the decoded SIMD instruction having a remainder width of the vector length less the first width in dependence on a dynamic performance condition,
wherein when the dynamic performance condition has a first state the issuance target is the first means for performing SIMD processing and when the dynamic performance condition has a second state the issuance target is the second means for performing SIMD processing; and
means for scheduling the first and second parts of the decoded SIMD instruction in series when the issuance target is the first means for performing SIMD processing.
