Name Strings

SPV_KHR_bfloat16

Contact

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

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

Contributors

  • Tobias Hector, AMD

  • Stu Smith, AMD

  • Jeff Bolz, Nvidia

  • Kévin Petit, Arm

  • David Neto, Google

  • Graeme Leese, Broadcom

  • Ruihao Zhang, Qualcomm

  • Mark Sheppard, Imagination

  • Ben Ashbaugh, Intel

  • Dmitry Sidorov, Intel

  • Victor Mustya, Intel

  • Alan Baker, Google

Notice

Copyright (c) 2023-2025 The Khronos Group Inc. Copyright terms at http://www.khronos.org/registry/speccopyright.html

Status

  • Approved by the SPIR-V Working Group: 2025-01-29

  • Approved by the Khronos Board of Promoters: 2025-03-14

Version

Last Modified Date

2025-01-29

Revision

1

Dependencies

This extension is written against the SPIR-V Specification, Version 1.6 Revision 4.

This extension requires SPIR-V 1.0.

This extension interacts with SPV_KHR_cooperative_matrix.

Overview

This extension extends the OpTypeFloat instruction to enable the use of bfloat16 types with cooperative matrices and dot products, and enabling conversions between other types.

Extension Name

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

OpExtension "SPV_KHR_bfloat16"

Modifications to the SPIR-V Specification, Version 1.6

FP Encoding

Add a new enum:

FP Encoding Width(s) Enabling Capabilities

0

BFloat16KHR
The floating point type is encoded as the bfloat16 type, which is equivalent to the IEEE754 binary32 floating point encoding but with 16 fewer bits in its significand. This is encoded per chapter 3 of the IEEE 754-2008 standard with the following encoding parameters:

- bias is 127
- sign bit is 1
- w (exponent) is 8
- t (significand) is 7
- k (width) is 16

The additional precision parameters p and emax are calculated as p = t+1 and emax = bias. The radix (or base) b is 2.

The accuracy of floating-point operations on BFloat16KHR values is the same as operations on IEEE754 binary32 floating-point values, with the result rounded using round-to-nearest-even or round-to-zero rounding. It is implementation-defined whether intermediate values are rounded.

16

BFloat16TypeKHR

Type-Declaration Instructions

Add the following requirement to OpTypeCooperativeMatrixKHR:

If Component Type has a *BFloat16KHR* encoding then *BFloat16CooperativeMatrixKHR* must be declared.

Conversion Instructions

Add the following paragraph to OpFConvert, OpConvertFToU, OpConvertFToS, OpConvertSToF, and OpConvertUToF:

Conversions to or from floating-point values with the `BFloat16KHR` encoding first convert the source value to IEEE754 binary32, and then from IEEE754 binary32 to the target format.
When converting from `BFloat16KHR` to IEEE754 binary32, the additional significand bits are padded with 0s, producing the exact same value.
When converting from IEEE754 binary32 to `BFloat16KHR`, the rounding mode is defined by the client API.

Arithmetic Instructions

Add the following requirement to OpDot:

If the Result Type (and the component types of Vector 1 and Vector 2) has a *BFloat16KHR* encoding then *BFloat16DotProductKHR* must be declared.

Capabilities

Modify Section 3.31, Capability, adding this row to the Capability table:

Capability Implicitly Declares

5116

BFloat16TypeKHR
Uses OpTypeFloat to specify types with the BFloat16KHR floating point encoding and the use of conversion functions for those types.

5117

BFloat16DotProductKHR
Uses vectors with a Component Type of OpTypeFloat with the BFloat16KHR encoding with OpDot.

BFloat16TypeKHR

5118

BFloat16CooperativeMatrixKHR
Uses cooperative matrices with a Component Type of OpTypeFloat with the BFloat16KHR encoding.

BFloat16TypeKHR, CooperativeMatrixKHR

Validation Rules

Add the following bullets to section 2.16.1, Universal Validation Rules:

Issues

What is the origin of the naming for bfloat16?

The type was originally named "brain float" as it was developed by Google Brain, an artificial intelligence group at Google, and it is a 16-bit type. The name is generally shortened to bfloat16.

Revision History

Rev Date Author Changes

1

2025-01-29

TobiasH

Initial revision