Notice
Copyright (c) 2026 The Khronos Group Inc. Copyright terms at http://www.khronos.org/registry/speccopyright.html
Dependencies
This extension is written against the SPIR-V Specification, Version 1.6 Revision
7, and SPV_KHR_bfloat16, Revision 1.
This extension requires SPIR-V 1.0 and SPV_KHR_bfloat16.
Overview
This extension extends SPV_KHR_bfloat16 to allow use of bfloat16 types with
additional arithmetic floating-point instructions.
Extension Name
To use this extension within a SPIR-V module, the following OpExtension must be present in the module:
OpExtension "SPV_EXT_bfloat16_arithmetic"
Modifications to the SPIR-V Specification, Version 1.6
Validation Rules
Add validation rules to section 2.16.1 Universal Validation Rules:
-
If the BFloat16ArithmeticEXT capability is declared, then variables with a type that is or includes a floating-point type with the BFloat16KHR encoding may be used with the following instructions:
-
-
OpFNegate
-
OpFAdd
-
OpFSub
-
OpFMul
-
OpFDiv
-
OpFRem
-
OpFMod
-
OpVectorTimesScalar
-
-
Supported OpExtInst instructions from the OpenCL Extended Instruction Set
-
Capabilities
Modify Section 3.2.30, "Capability", adding these rows to the Capability table:
| Capability | Implicitly Declares | |
|---|---|---|
6226 |
BFloat16ArithmeticEXT |
BFloat16TypeKHR |
Interactions with Other Extensions
If SPV_INTEL_masked_gather_scatter is supported then OpMaskedGatherINTEL
and OpMaskedScatterINTEL will also accept vectors of BFloat16KHR and vectors of
pointers to BFloat16KHR.