Name Strings

SPV_EXT_float8

Contact

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

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

Contributors

  • Kevin Petit, Arm Ltd.

  • Jeff Bolz, NVIDIA

  • Stu Smith, AMD

  • Tobias Hector, AMD

  • Victor Lomuller, Codeplay

  • Alan Baker, Google

  • Alexander Galazin, Imagination

  • Ben Ashbaugh, Intel

  • Vikram Tarikere, Imagination

  • Dmitry Sidorov, Intel

  • Graeme Leese, Broadcom

  • Samuel Huang, Mediatek

Notice

Copyright (c) 2024-2025 The Khronos Group Inc.

Status

  • Complete

  • Approved by the SPIR Working Group: 2025-04-16

  • Approved by the Khronos Board of Promoters: 2025-05-30

Version

Last Modified Date

2025-04-16

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 float8 E4M3 and E5M2 types with cooperative matrices as well as conversions between them and other floating-point or integer types.

Extension Name

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

OpExtension "SPV_EXT_float8"

Modifications to the SPIR-V Specification, Version 1.6

Validation Rules

Add the following bullets to section 2.16.11, "Universal Validation Rules":

  • Objects with a type that is or includes a floating-point type with the Float8E4M3EXT or Float8E5M2EXT FP Encoding must only be used with the following instructions:

    • Constant Creation Instructions

    • Memory Instructions

    • Function Instructions

    • Conversion Instructions

    • Composite Instructions

    • Annotation Instructions

    • Debug Instructions

    • Miscellaneous Instructions

    • Extension Instructions

    • Mode-Setting Instructions

    • OpSelect

    • OpPhi

    • OpReturnValue

    • OpLifetimeStart

    • OpLifetimeStop

    • OpCooperativeMatrixMulAddKHR

FP Encoding

Modify Section 3.51, adding these rows to the table of alternative floating point encodings:

FP Encoding Width(s) Enabling Capabilities

4214

Float8E4M3EXT
The floating point type is encoded as an FP8 E4M3 type, as specified in the "FP8 Formats For Deep Learning" whitepaper (https://arxiv.org/abs/2209.05433).

8

Float8EXT

4215

Float8E5M2EXT
The floating point type is encoded as an FP8 E5M2 type, as specified in the "FP8 Formats For Deep Learning" whitepaper (https://arxiv.org/abs/2209.05433).

8

Float8EXT

Decoration

Modify Section 3.20, "Decoration", adding these rows to the Decoration table:

Decoration Extra Operands Enabling Capabilities

4216

SaturatedToLargestFloat8NormalConversionEXT
Indicates that a conversion to a floating-point type with FP Encoding Float8E4M3EXT or Float8E5M2EXT of a value which is outside of the representable range of Result Type, or infinite, results in the largest representable normal (i.e. non-infinite) value with a matching sign.

This decoration must only be applied to the OpFConvert, OpConvertSToF, or OpConvertUToF instructions with a Result Type that uses an Float8E4M3EXT or Float8E5M2EXT FP Encoding.

Float8EXT

Capability

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

Capability Implicitly Declares

4212

Float8EXT
Uses OpTypeFloat to specify types with the Float8E4M3EXT or Float8E5M2EXT FP Encoding and values of this type with a few instructions.

4213

Float8CooperativeMatrixEXT
Uses cooperative matrix with a Component Type of OpTypeFloat with the Float8E4M3EXT or Float8E5M2EXT encoding.

Float8EXT, CooperativeMatrixKHR

Instructions

Type-Declaration Instructions

Add the following requirement to OpTypeCooperativeMatrixKHR:

If Component Type has a Float8E4M3EXT or Float8E5M2EXT encoding then Float8CooperativeMatrixEXT must be declared.

Conversion Instructions

Modify section 3.42.11, "Conversion Instructions" to add the following text to the description of the OpFConvert, OpConvertSToF, and OpConvertUToF instructions:

When converting to floating-point values with the Float8E4M3EXT or Float8E5M2EXT encoding, out-of-range values and infinity are converted to largest representable finite value with a matching sign if the conversion is decorated with SaturatedToLargestFloat8NormalConversionEXT, otherwise they are converted to NaN if the destination type uses the Float8E4M3EXT FP encoding or infinity with a matching sign if the destination type uses the Float8E5M2EXT FP encoding.

Issues

1) Should this extension add support for OpDot, likely enabled by a separate capability?

RESOLVED: No. OpDot currently only supports operations for which the input and output data types are the same and those are not typically useful with types like FP8 that can only represent a very small set of values. Adding support for mixed-precision dot product operations could be done but none of the vendors participating in the creation of this extension currently think they could accelerate these operations that can otherwise be expressed with existing dot product operations on existing floating point types and the conversion operations introduced in this extension.

Revision History

Rev Date Author Changes

1

2025-04-16

Kévin Petit

Initial revision