Name Strings

SPV_KHR_subgroup_rotate

Contact

To report problems with this extension, please open a new issue at:

https://github.com/KhronosGroup/SPIRV-Headers

Contributors

  • Kevin Petit, Arm Ltd.

  • Ruihao Zhang, Qualcomm

  • Faith Ekstrand, Collabora

  • Graeme Leese, Broadcom

  • Alan Baker, Google

  • Caio Oliveira, Intel

  • Jeff Bolz, NVIDIA

Notice

Copyright (c) 2022 The Khronos Group Inc. Copyright terms at http://www.khronos.org/registry/speccopyright.html

Status

  • Approved by the SPIR-V Working Group: 2022-03-02

  • Approved by the Khronos Board of Promoters: 2022-04-15

Version

Last Modified Date

2022-03-02

Revision

1

Dependencies

This extension is written against the SPIR-V Specification Version 1.5 Revision 5.

This extension requires SPIR-V 1.3.

Overview

This extension adds a new instruction that enables rotating values across invocations within a subgroup. Taking the example of a subgroup of size 16, a rotation by an amount of 2 would, when executed by the invocation identified by id 0, return the value from the invocation identified by the id 2. The same rotation instruction, when executed by the invocation identified by id 14, would return the value from the invocation identified by id 0.

A rotation by an amount of N rotates values "down" N invocations within the subgroup.

A rotation by an amount of (SubgroupSize - N) rotates values "up" N invocations within the subgroup.

Extension Name

To use this extension within a SPIR-V module, the following OpExtension must be present in the module:

OpExtension "SPV_KHR_subgroup_rotate"

Modifications to the SPIR-V Specification, Version 1.5

Capabilities

Modify Section 3.31, "Capability", adding these rows to the Capability table:

Capability Implicitly declares

6026

GroupNonUniformRotateKHR
Use OpGroupNonUniformRotateKHR instruction

GroupNonUniform

Instructions

Add the new instruction:

OpGroupNonUniformRotateKHR

Return the Value of the invocation whose id within the group is calculated as follows:

LocalId = SubgroupLocalInvocationId if Execution is Subgroup or LocalInvocationId if Execution is Workgroup
RotationGroupSize = ClusterSize when ClusterSize is present, otherwise
RotationGroupSize = SubgroupMaxSize if the Kernel capability is declared and SubgroupSize if not.
Invocation ID = ( (LocalId + Delta) & (RotationGroupSize - 1) ) + (LocalId & ~(RotationGroupSize - 1))

Result Type must be a scalar or vector of floating-point type, integer type, or Boolean type.

Execution is a Scope. It must be either Workgroup or Subgroup.

The type of Value must be the same as Result Type.

Delta must be a scalar of integer type, whose Signedness operand is 0.
Delta must be dynamically uniform within Execution.

Delta is treated as unsigned and the resulting value is undefined if the selected lane is inactive.

ClusterSize is the size of cluster to use. ClusterSize must be a scalar of integer type, whose Signedness operand is 0. ClusterSize must come from a constant instruction. Behavior is undefined unless ClusterSize is at least 1 and a power of 2. If ClusterSize is greater than the declared SubGroupSize, executing this instruction results in undefined behavior.

Capability:
GroupNonUniformRotateKHR

6 + variable

4431

<id>
Result Type

Result <id>

Scope <id>
Execution

<id>
Value

<id>
Delta

Optional <id> ClusterSize

Issues

None.

Revision History

Rev Date Author Changes

1

2022-03-02

Kevin Petit

Initial revision