[SM6.10] Add BFloat16 Support

[Ashley Coleman (V-FEXrt)]

Fixes #8722

microsoft/hlsl-specs#907 added BFloat16 as an allowed ComponentType in the LinAlg API. This PR adds support for it to the implementation.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Copilot]

Pull request overview

Adds BFloat16 support to the Shader Model 6.10 LinAlg API and DXIL validation.

Changes:

• Defines and exposes the BFloat16 component type.
• Adds packed two-per-uint type traits.
• Allows BFloat16 matrices in validation and serialization.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
include/dxc/DXIL/DxilConstants.h	Defines the DXIL component value.
include/dxc/DxilContainer/RDAT_LibraryTypes.inl	Adds RDAT enum serialization.
lib/DxilValidation/DxilValidation.cpp	Permits BFloat16 LinAlg matrices.
lib/DxilValidation/DxilValidationUtils.cpp	Formats the component name.
tools/clang/lib/Headers/hlsl/dx/linalg.h	Exposes BFloat16 and its packing traits.
tools/clang/lib/Sema/SemaHLSL.cpp	Adds Sema string conversion.
tools/clang/test/LitDXILValidation/LinAlgMatrix/linalgmatrix-illegal-component-type.ll	Verifies validator acceptance.

Suppressed comments (1)

tools/clang/lib/Headers/hlsl/dx/linalg.h:143

• Marking BFloat16 as non-native invalidates the Matrix class comment at lines 241-242, which says every non-native scalar is an 8-bit type with four elements per scalar. Update that comment to describe packed component types generically (or explicitly include the two-BFloat16 case) so it matches this specialization.

  static const bool IsNativeScalar = false;
  static const uint ElementsPerScalar = 2;

[Damyan Pepper (damyanp)]

We didn't ship linalg.h for SM 6.9 did we?

[Ashley Coleman (V-FEXrt)]

Nope, but we shipped the type in DxilConstants.h in SM6.9!

I updated this here because the header says this type needs to exactly match the DxilConstants.h type

[Ashley Coleman (V-FEXrt)]

We could update the type in DxilConstants.h to say "new for 6.10" which would be fine by me but not wholly accurate (I don't think the new types could actually be used for anything even though they are published in 6.9)

[Damyan Pepper (damyanp)]

🤷

[Damyan Pepper (damyanp)]

LGTM

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated no new comments.

Suppressed comments (1)

include/dxc/DXIL/DxilConstants.h:203

• This adds a user-visible Shader Model 6.10 feature, so it needs an entry under Experimental Shader Model 6.10 in docs/ReleaseNotes.md per the release-note policy in CONTRIBUTING.md. Please add the release note (or point to the related PR that will provide the shared coverage).

  BFloat16 = 23,

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Suppressed comments (2)

docs/ReleaseNotes.md:45

• BFloat16 is gated to experimental SM 6.10, but this entry is under Upcoming Release, which makes it appear to ship in the next non-preview release. Move it under Upcoming Preview Release → Experimental Shader Model 6.10 (and match the surrounding past-tense, period-terminated style).

#### LinAlg Matrix
- Adds `BFloat16` to the ComponentType enum in DxilConstants and the linalg
  header [#8722](https://github.com/microsoft/DirectXShaderCompiler/issues/8722)

tools/clang/lib/Headers/hlsl/dx/linalg.h:143

• This specialization makes a non-native component contain two values per scalar, so the Matrix comment at lines 241–242 (“we have an 8-bit type” and “4 elements”) is now false for BFloat16. Please generalize that comment to describe packed/non-native component types and use ElementsPerScalar rather than asserting the 8-bit layout.

template <> struct ComponentTypeTraits<ComponentType::BFloat16> {
  using Type = uint;
  static const bool IsNativeScalar = false;
  static const uint ElementsPerScalar = 2;

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.