SPV_INTEL_subgroups

Name Strings

SPV_INTEL_subgroups

Contact

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

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

Contributors

Ben Ashbaugh, Intel
Biju George, Intel
Michael Kinsner, Intel
Mariusz Merecki, Intel

Notice

Status

Final Draft

Version

Last Modified Date

2018-10-22

Revision

Dependencies

This extension is written against the SPIR-V Specification, Version 1.2 Revision 1.

This extension requires SPIR-V 1.0.

Overview

The goal of this extension is to allow programmers to improve the performance of their applications by taking advantage of the fact that some work items in a work group execute together as a group (a "subgroup"), and that work items in a subgroup can use hardware features that are not available to all work items in a work group. Specifically, this extension is designed to allow work items in a subgroup to share data without the use of local memory and work group barriers, and to utilize specialized hardware to load and store blocks of data from images or buffers.

This extension builds upon "subgroups" functionality that is already in core SPIR-V, so this extension reuses many of the names, concepts, and instructions already described in SPIR-V. The key additions in this extension are:

Intel subgroups adds "shuffle" instructions to allow data interchange between work items within a subgroup without the use of local memory or barriers.
Intel subgroups adds "block read and write" instructions to take advantage of specialized hardware to read or write blocks of data from or to buffers or images.

This extension has a source language counterpart extension for the OpenCL-C kernel language, cl_intel_subgroups, which can be used for online compilation in an OpenCL environment.

Extension Name

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

OpExtension "SPV_INTEL_subgroups"

New Capabilities

This extension introduces new capabilities:

SubgroupShuffleINTEL
SubgroupBufferBlockIOINTEL
SubgroupImageBlockIOINTEL

New Instructions

Instructions added under the SubgroupShuffleINTEL capability:

OpSubgroupShuffleINTEL
OpSubgroupShuffleDownINTEL
OpSubgroupShuffleUpINTEL
OpSubgroupShuffleXorINTEL

Instructions added under the SubgroupBufferBlockIOINTEL capability:

OpSubgroupBlockReadINTEL
OpSubgroupBlockWriteINTEL

Instructions added under the SubgroupImageBlockIOINTEL capability:

OpSubgroupImageBlockReadINTEL
OpSubgroupImageBlockWriteINTEL

Token Number Assignments

SubgroupShuffleINTEL	5568
SubgroupBufferBlockIOINTEL	5569
SubgroupImageBlockIOINTEL	5570
OpSubgroupShuffleINTEL	5571
OpSubgroupShuffleDownINTEL	5572
OpSubgroupShuffleUpINTEL	5573
OpSubgroupShuffleXorINTEL	5574
OpSubgroupBlockReadINTEL	5575
OpSubgroupBlockWriteINTEL	5576
OpSubgroupImageBlockReadINTEL	5577
OpSubgroupImageBlockWriteINTEL	5578

Modifications to the SPIR-V Specification, Version 1.2

Capabilities

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

Capability	Implicitly Declares
5568	SubgroupShuffleINTEL	SPV_INTEL_subgroups
5569	SubgroupBufferBlockIOINTEL	SPV_INTEL_subgroups
5570	SubgroupImageBlockIOINTEL	SPV_INTEL_subgroups

Capability

Implicitly Declares

Enabled by Extension

5568

SubgroupShuffleINTEL

SPV_INTEL_subgroups

5569

SubgroupBufferBlockIOINTEL

SPV_INTEL_subgroups

5570

SubgroupImageBlockIOINTEL

SPV_INTEL_subgroups

Instructions

Modify Section 3.32.21, Group Instructions, adding to the end of the list of instructions:

OpSubgroupShuffleINTEL

Allows data to be arbitrarily transferred between invocations in a subgroup. The data that is returned for this invocation is the value of Data for the invocation identified by InvocationId.

InvocationId need not be the same value for all invocations in the subgroup.

Result Type may be a scalar or vector type.

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

InvocationId must be a 32-bit integer type scalar.

Capability:
SubgroupShuffleINTEL

5571

<id> Result Type

<id> Result

<id> Data

<id> InvocationId

OpSubgroupShuffleDownINTEL

Allows data to be transferred from an invocation in the subgroup with a higher SubgroupLocalInvocationId down to a invocation in the subgroup with a lower SubgroupLocalInvocationId.

There are two data sources to this built-in function: Current and Next. To determine the result of this built-in function, first let the unsigned shuffle index be equivalent to the sum of this invocation’s SubgroupLocalInvocationId plus the specified Delta:

If the shuffle index is less than the SubgroupMaxSize, the result of this built-in function is the value of the Current data source for the invocation with SubgroupLocalInvocationId equal to the shuffle index.

If the shuffle index is greater than or equal to the SubgroupMaxSize but less than twice the SubgroupMaxSize, the result of this built-in function is the value of the Next data source for the invocation with SubgroupLocalInvocationId equal to the shuffle index minus the SubgroupMaxSize.

All other values of the shuffle index are considered to be out-of-range.

Delta need not be the same value for all invocations in the subgroup.

Result Type may be a scalar or vector type.

The type of Current and Next must be the same as Result Type.

Delta must be a 32-bit integer type scalar.

Capability:
SubgroupShuffleINTEL

5572

<id> Result Type

<id> Result

<id> Current

<id> Next

<id> Delta

OpSubgroupShuffleUpINTEL

Allows data to be transferred from an invocation in the subgroup with a lower SubgroupLocalInvocationId up to an invocation in the subgroup with a higher SubgroupLocalInvocationId.

There are two data sources to this built-in function: Previous and Current. To determine the result of this built-in function, first let the signed shuffle index be equivalent to this invocation’s SubgroupLocalInvocationId minus the specified Delta:

If the shuffle index is greater than or equal to zero and less than the SubgroupMaxSize, the result of this built-in function is the value of the Current data source for the invocation with SubgroupLocalInvocationId equal to the shuffle index.

If the shuffle index is less than zero but greater than or equal to the negative SubgroupMaxSize, the result of this built-in function is the value of the Previous data source for the invocation with SubgroupLocalInvocationId equal to the shuffle index plus the SubgroupMaxSize.

All other values of the shuffle index are considered to be out-of-range.

Delta need not be the same value for all invocations in the subgroup.

Result Type may be a scalar or vector type.

The type of Previous and Current must be the same as Result Type.

Delta must be a 32-bit integer type scalar.

Capability:
SubgroupShuffleINTEL

5573

<id> Result Type

<id> Result

<id> Previous

<id> Current

<id> Delta

OpSubgroupShuffleXorINTEL

Allows data to be transferred between invocations in a subgroup as a function of the invocation_s SubgroupLocalInvocationId. The data that is returned for this invocation is the value of Data for the invocation with SubgroupLocalInvocationId equal to this invocation’s SubgroupLocalInvocationId XOR_d with the specified Value. If the result of the XOR is greater than SubgroupMaxSize then it is considered out-of-range.

Value need not be the same for all invocations in the subgroup.

Result Type may be a scalar or vector type.

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

Value must be a 32-bit integer type scalar.

Capability:
SubgroupShuffleINTEL

5574

<id> Result Type

<id> Result

<id> Data

<id> Value

OpSubgroupBlockReadINTEL

Reads one or more components of Result data for each invocation in the subgroup from the specified Ptr as a block operation.

The data is read strided, so the first value read is:

Ptr[ SubgroupLocalInvocationId ]

and the second value read is:

Ptr[ SubgroupLocalInvocationId + SubgroupMaxSize ]

etc.

Result Type may be a scalar or vector type, and its component type must be equal to the type pointed to by Ptr.

The type of Ptr must be a pointer type, and must point to a scalar type.

Capability:
SubgroupBufferBlockIOINTEL

5575

<id> Result Type

<id> Result

<id> Ptr

OpSubgroupBlockWriteINTEL

Writes one or more components of Data for each invocation in the subgroup from the specified Ptr as a block operation.

The data is written strided, so the first value is written to:

Ptr[ SubgroupLocalInvocationId ]

and the second value written is:

Ptr[ SubgroupLocalInvocationId + SubgroupMaxSize ]

etc.

The type of Ptr must be a pointer type, and must point to a scalar type.

The component type of Data must be equal to the type pointed to by Ptr.

Capability:
SubgroupBufferBlockIOINTEL

5576

<id> Ptr

<id> Data

OpSubgroupImageBlockReadINTEL

Reads one or more components of Result data for each invocation in the subgroup from the specified Image at the specified Coordinate as a block operation. Note that the Coordinate is a byte coordinate, not a texel coordinate. Also note that the image data is read without format conversion, so each invocation may read multiple image elements.

The data is read row-by-row, so the first value read is from the row specified by the y-component of the provided Coordinate, the second value is read from the row specified by the y-component of the provided Coordinate plus one, etc.

Result Type may be a scalar or vector type.

Image must be an object whose type is OpTypeImage with a Sampled operand of 0 or 2. If the Sampled operand is 2, then some dimensions require a capability.

Coordinate is an integer scalar or vector. The x-component is a byte coordinate into rows of the image and remaining coordinates are non-normalized texel coordinates.

Capability:
SubgroupImageBlockIOINTEL

5577

<id> Result Type

<id> Result

<id> Image

<id> Coordinate

OpSubgroupImageBlockWriteINTEL

Writes one or more components of Data for each invocation in the subgroup to the specified Image at the specified Coordinate as a block operation. Note that the Coordinate is a byte coordinate, not a texel coordinate. Also note that the image data is read without format conversion, so each invocation may write multiple image elements.

The data is written row-by-row, so the first value is written to the row specified by the y-component of the provided Coordinate, the second value is written to the row specified by the y-component of the provided Coordinate plus one, etc.

Image must be an object whose type is OpTypeImage with a Sampled operand of 0 or 2. If the Sampled operand is 2, then some dimensions require a capability.

Coordinate is an integer scalar or vector. The x-component is a byte coordinate into rows of the image and remaining coordinates are non-normalized texel coordinates.

Result Type may be a scalar or vector type.

Capability:
SubgroupImageBlockIOINTEL

5578

<id> Image

<id> Coordinate

<id> Data

Validation Rules

None.

Issues

None.

Revision History

Rev	Date	Author	Changes
1	2017-09-29	Ben Ashbaugh	Initial revision
2	2018-10-22	Ben Ashbaugh	Minor formatting updates.