Name Strings

SPV_EXT_ocp_microscaling_types

Contact

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

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

Contributors

  • Kevin Petit, Arm Ltd.

  • Vikram Tarikere, Imagination Technologies

  • Dmitry Sidorov, AMD

  • Jeff Bolz, NVIDIA

  • Alan Baker, Google

Notice

Copyright (c) 2024-2026 The Khronos Group Inc.

Status

  • Approved by the SPIR-V Working Group: 2026-05-13

  • Approved by the Khronos Board of Promoters: 2026-06-26

Version

Last Modified Date

2026-05-08

Revision

1

Dependencies

This extension is written against the SPIR-V Specification, Version 1.6 Revision 5.

This extension requires SPIR-V 1.0.

This extension requires SPV_EXT_float8.

Overview

This extension extends the OpTypeFloat instruction to enable the use of the following data types described in the "OCP Microscaling Formats v1.0" specification:

  • FP4 E2M1

  • FP6 E3M2 and E2M3

  • FP8 E8M0

  • MXINT8

It also adds a new OpBitcastExtractEXT instruction to facilitate the loading of values whose bit width is smaller than a byte.

Extension Name

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

OpExtension "SPV_EXT_ocp_microscaling_types"

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 one of the following FP Encoding:

    • Float6E2M3EXT

    • Float6E3M2EXT

    • Float4E2M1EXT

    • Float8UnsignedE8M0EXT

    • MXInt8EXT must only be used with the following instructions:

    • Constant Creation Instructions

    • Memory Instructions

      • Float8UnsignedE8M0EXT and MXInt8EXT encodings can be used with all storage classes

      • Float4E2M1EXT, Float6E2M3EXT, and Float6E3M2EXT encodings can only be used with the Function and Private storage classes

    • Function Instructions

    • Conversion Instructions

    • Composite Instructions

    • Annotation Instructions

    • Debug Instructions

    • Miscellaneous Instructions

    • Extension Instructions

    • Mode-Setting Instructions

    • OpSelect

    • OpPhi

    • OpReturnValue

    • OpLifetimeStart

    • OpLifetimeStop

FP Encoding

Modify Section 3.55, adding these rows to the table of alternative floating point encodings:

FP Encoding Width(s) Enabling Capabilities

4223

Float6E2M3EXT
The floating point type is encoded as an FP6 E2M3 type, as specified in the "OCP Microscaling Formats v1.0" specification.

6

Float6EXT

4224

Float6E3M2EXT
The floating point type is encoded as an FP6 E3M2 type, as specified in the "OCP Microscaling Formats v1.0" specification.

6

Float6EXT

4225

Float4E2M1EXT
The floating point type is encoded as an FP4 E2M1 type, as specified in the "OCP Microscaling Formats v1.0" specification.

4

Float4EXT

4226

Float8UnsignedE8M0EXT
The floating point type is encoded as an unsigned FP8 E8M0 type, as specified in the "OCP Microscaling Formats v1.0" specification.

8

Float8UnsignedE8M0EXT

4227

MXInt8EXT
The floating point type is encoded as an MXINT8 type, as specified in the "OCP Microscaling Formats v1.0" specification.

8

MXInt8EXT

The "OCP Microscaling Formats v1.0" specification can be found at https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf

Capability

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

Capability Implicitly Declares

4228

Float6EXT
Uses OpTypeFloat to specify types with the Float6E2M3EXT or Float6E3M2EXT FP Encoding and values of this type with a few instructions.

4229

Float4EXT
Uses OpTypeFloat to specify types with the Float4E2M1EXT FP Encoding and values of this type with a few instructions.

4230

Float8UnsignedE8M0EXT
Uses OpTypeFloat to specify types with the Float8UnsignedE8M0EXT FP Encoding and values of this type with a few instructions.

4231

MXInt8EXT
Uses OpTypeFloat to specify types with the MXInt8EXT FP Encoding and values of this type with a few instructions.

4232

BitcastExtractEXT
Uses OpBicastExtractEXT instruction.

Instructions

Modify section 3.56.11 "Conversion Instruction", adding the following instruction:

OpBitcastExtractEXT

Extract a bit field from an object and perform a bit pattern-preserving type conversion.

The bits of Base numbered in [Offset, Offset + Bitwidth(ComponentsOf(Result Type)) - 1] (inclusive) become the bits of the result. Results are computed per component.

Result Type must be a scalar or vector of numerical type with the same number of components as Base.

Base must be a scalar or vector of numerical type whose components must have a bit width greater than that of Result Type.

Offset must be a scalar integer type. Offset is the lowest-order bit of the bit field to extract from Base. It is consumed as an unsigned value.

The resulting value is undefined if Offset or the sum of Offset and the bit width of Result Type is greater than the number of bits in Base.

Capability:
BitcastExtractEXT

5

4195

<id>
Result Type

Result <id>

<id>
Base

<id>
Offset

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 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.

2) What conversions do we want to support? Some conversions between these new types do not seem useful. We decided to allow conversions between FP8 encodings in SPV_EXT_float8.

RESOLVED: This specification does not constrain the conversions that are allowed but client API specifications may (and we expect Vulkan will).

3) Does this extension provide any mechanism to calculate scaled values?

RESOLVED: No, not directly, but it provides the necessary support to express code that does it.

4) Should this extension add support for cooperative matrix operations?

RESOLVED: No. While there is consensus that this would be valuable functionality, there is also consensus that there are multiple approaches that could be followed, and that fully resolving these discussions would unduly defer the release of this extension. It was decided to defer such support to a future extension.

Revision History

Rev Date Author Changes

1

2026-05-08

Kévin Petit

Initial revision