Skip to content

[Doc][SPIR-V] Move FP4 E2M1 to SPV_EXT_ocp_microscaling_types #22528

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
uditagarwal97 merged 1 commit into from
Jul 6, 2026
+34 −144
Merged

[Doc][SPIR-V] Move FP4 E2M1 to SPV_EXT_ocp_microscaling_types #22528

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
174 changes: 32 additions & 142 deletions sycl/doc/design/spirv-extensions/SPV_INTEL_float4.asciidoc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,12 +1,11 @@
:extension_name: SPV_INTEL_float4

:hf4_capability_name: Float4E2M1INTEL
:hf4_capability_token: 6212
:hf4_matrix_capability_name: Float4E2M1CooperativeMatrixINTEL
:hf4_matrix_capability_token: 6213
:hf4_encoding: 6214
:hf4_encoding_name: Float4E2M1EXT

:khr_matrix_capability_name: CooperativeMatrixKHR
:float4ext_capability_name: Float4EXT

:joint_matrix_url: https://https://github.com/intel/llvm/tree/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_joint_matrix.asciidoc
:fp_conv_url: https://github.com/intel/llvm/tree/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_fp_conversions.asciidoc
Expand Down Expand Up @@ -55,8 +54,8 @@ please let us know!

[width="40%",cols="25,25"]
|========================================
| Last Modified Date | 2025-10-24
| Revision | 2
| Last Modified Date | 2026-07-02
| Revision | 3
|========================================

== Dependencies
Expand All @@ -74,172 +73,62 @@ This extension interacts with {fp8_url}[*SPV_EXT_float8*] extension.

This extension interacts with {fp_conv_url}[*SPV_INTEL_fp_conversions*] extension.

This extension depends on the *SPV_EXT_ocp_microscaling_types* extension, which
provides the *Float4E2M1EXT* encoding that this extension uses as a
cooperative-matrix component type.

This extension requires SPIR-V 1.0.

Overview
--------

This extension extends the *OpTypeFloat* instruction to enable the definition of `FP4E2M1`
floating-point format that has one sign bit, two exponent bits and one mantissa bits.

The `FP4E2M1` special values are defined by the table below.

[options="header"]
[width="80%"]
[cols="1,2"]
|====
| ^| `FP4E2M1`
| Exponent Bias | 1
| Max normal
| S.11.1 = 6.0 (1.5 * 2^2^)

| Min normal
| S.01.0 = 1.0 (1.0 * 2^0^)
The `FP4E2M1` floating-point type, which has one sign bit, two exponent bits and
one mantissa bit, is defined by the multi-vendor *SPV_EXT_ocp_microscaling_types*
extension as the *Float4E2M1EXT* encoding.

| Max subnormal
| S.00.1 = 0.5 (0.5 * 2^0^)

| Min subnormal
| S.00.1 = 0.5 (0.5 * 2^0^)

| Infinity | N/A
| NaN | N/A

|====
This extension enables that type to be used as an *OpTypeCooperativeMatrixKHR*
component type, for which the microscaling extension has no analog.

== Modifications to the SPIR-V Specification, Version 1.6

Binary Form
~~~~~~~~~~~

FP Encoding
~~~~~~~~~~~

Add a new enum:

--
[cols="^2,14,2,4",options="header",width = "100%"]
|====
2+^.^| FP Encoding | Width(s) | Enabling Capabilities
| {hf4_encoding} | *Float4E2M1INTEL* +
The floating point type is encoded as a 4-bit float type.
This is encoded with the following encoding parameters: +

- _bias_ is 1
+
- _sign bit_ is 1
+
- _w_ (exponent) is 2
+
- _t_ (significand) is 1
+
- _k_ (width) is 4
| 4 | *Float4E2M1INTEL*

|===
--
The FP4 (E2M1) type is represented with the *{hf4_encoding_name}* floating-point
encoding defined by the *SPV_EXT_ocp_microscaling_types* extension.
This extension adds the capability to use that type as an *OpTypeCooperativeMatrixKHR*
component type, for which the microscaling extension has no analog.

=== Capabilities

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

--
[options="header"]
|====
2+^| Capability ^| Implicitly Declares
| {hf4_capability_token} | *{hf4_capability_name}* +
Uses *Float4E2M1INTEL* floating-point encoding. +
|
| {hf4_matrix_capability_token} | *{hf4_matrix_capability_name}* | *{khr_matrix_capability_name}*
2+^| Capability ^| Implicitly Declares
| {hf4_matrix_capability_token} | *{hf4_matrix_capability_name}* +
Uses a cooperative matrix with a _Component Type_ of *OpTypeFloat* with the
*{hf4_encoding_name}* encoding. | *{float4ext_capability_name}*, *{khr_matrix_capability_name}*
|====
--

=== Memory Layout

Add to Section 2.18.1. Memory Layout, FPE2M1 4 layout:

Scalar floating point variables with a `Width` of 4 can only be declared in the `Private` or `Function` storage classes.
In other storage classes, they must be included in an `OpTypeVector` with an even `Component Count`, where the first component in every pair is in bits 0-3 of the corresponding byte, and the second component is in bits 4-7.

=== Instructions

==== 3.42.11. Conversion Instructions

* Add the following paragraphs to *OpFConvert*:
+
When converting to floating-point values with the *Float4E2M1INTEL* encoding, out-of-range
values and infinity and are converted to largest representable finite value with a matching sign.
Conversion from NaNs is implementation-defined. +
+

==== 3.49.6. Type-Declaration Instructions

Add the following requirement to *OpTypeCooperativeMatrixKHR*:

If _Component Type_ has a *Float4E2M1INTEL* encoding then *{hf4_matrix_capability_name}* must be declared.
If _Component Type_ has a *{hf4_encoding_name}* encoding then *{hf4_matrix_capability_name}* must be declared.

Validation Rules
~~~~~~~~~~~~~~~~

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

* Variables with a type that is or includes a floating-point type with the *Float4E2M1INTEL* encoding must only be used with the following instructions:
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_miscellaneous_instructions[Miscellaneous Instructions] :
*** OpUndef
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_constant_creation_instructions[Constant Creation Instructions] :
*** OpConstant
*** OpConstantNull
*** OpConstantOp
*** OpConstantComposite
*** OpConstantCompositeContinuedINTEL
*** OpCooperativeMatrixConstructCheckedINTEL
*** OpSpecConstant
*** OpSpecConstantOp
*** OpSpecConstantComposite
*** OpSpecConstantCompositeContinuedINTEL
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_arithmetic_instructions[Arithmetic Instructions] :
*** OpCooperativeMatrixMulAddKHR
*** OpCooperativeMatrixMulAddScaledINTEL
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_composite_instructions[Composite Instructions] :
*** OpVectorExtractDynamic
*** OpVectorInsertDynamic
*** OpVectorShuffle
*** OpCompositeConstruct
*** OpCompositeExtract
*** OpCompositeInsert
*** OpCopyObject
*** OpCopyLogical
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_memory_instructions[Memory Instructions] :
*** OpPtrEqual
*** OpPtrNotEqual
*** OpPtrDiff
*** OpCooperativeMatrixLoadKHR
*** OpCooperativeMatrixStoreKHR
*** OpCooperativeMatrixLoadCheckedINTEL
*** OpCooperativeMatrixStoreCheckedINTEL
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_function_instructions[Function Instructions] :
*** OpFunction
*** OpFunctionParameter
*** OpFunctionCall
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_conversion_instructions[Conversion Instructions] :
*** OpConvertSToF
*** OpFConvert
*** OpConvertPtrToU
*** OpConvertUToPtr
*** OpPtrCastToGeneric
*** OpGenericCastToPtr
*** OpGenericCastToPtrExplicit
*** OpBitcast
*** OpClampConvertFToFINTEL
*** OpBiasedRoundFToFINTEL
*** OpClampBiasedRoundFToFINTEL
*** OpBiasedRoundFToSINTEL
** https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_control_flow_instructions[Control-Flow Instructions] :
*** OpReturnValue
*** OpSelect
*** OpPhi
*** OpLifetimeStart
*** OpLifetimeStop
The *SPV_EXT_ocp_microscaling_types* extension restricts the instructions that a
*{hf4_encoding_name}*-encoded object may be used with, and does not permit the
cooperative-matrix instructions.
This extension additionally permits a cooperative matrix whose _Component Type_
has the *{hf4_encoding_name}* encoding to be used with the following instructions:

* OpCooperativeMatrixMulAddKHR
* OpCooperativeMatrixMulAddScaledINTEL

=== Issues

Expand All @@ -255,4 +144,5 @@ Revision History
|Rev|Date|Author|Changes
|1|2024-06-15|Dmitry Sidorov|Initial revision
|2|2025-10-24|Dmitry Sidorov|Prepare to publish
|3|2026-07-02|Viktoria Maksimova|Represent the FP4E2M1 type with the SPV_EXT_ocp_microscaling_types Float4E2M1EXT encoding; narrow this extension to enabling that type as a cooperative-matrix component type
|========================================
4 changes: 2 additions & 2 deletions sycl/doc/design/spirv-extensions/mini_float_conversions_env.asciidoc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
:fp4_extension_name: SPV_INTEL_float4
:fp4_encoding_name: Float4E2M1INTEL
:fp4_extension_name: SPV_EXT_ocp_microscaling_types
:fp4_encoding_name: Float4E2M1EXT

= Mini-float Types and Conversions Environment Specification

Expand Down
Loading