[SM6.10] Add LinAng Matrix attributes and two new attributed matrix types by hekota · Pull Request #8125 · microsoft/DirectXShaderCompiler

hekota · 2026-02-02T01:25:11Z

Adds new type attribute __LinAlg_Matrix_Attributes that can be attached to __builtin_LinAlg_Matrix built-in matrix handle type to specify the matrix component type, dimensions, use and scope. Includes enabling C11++ -style attributes on types in DXC.

Adds two new AST types that are used to capture the information from the attribute:

AttributedLinAlgMatrixType type is used when the attribute specifies concrete matrix parameters
DependentAttributedLinAlgMatrixType is for cases where the type attributes arguments are dependent on template instantiation

The AttributedLinAlgMatrixType is linked to LinAlgMatrix type alias used in gen_intrin_main.txt for built-in function APIs.

This change also adds the outline of the LinAlg Matrix template class to dx/linalg.h, including the enumeration types used as arguments to the new attribute and as template parameters of Matrix.

Note that the ComponentType enum in dx/linalg.h is defined using values matching the existing internal ComponentType enum that was added in SM 6.9 for cooperative vectors. This is different from the current LinAlg spec. The issue to update the spec is microsoft/hlsl-specs#779.

Fixes #8122

Adds a built-in type `__builtin_LinAlg_Matrix` that will be used in LinAlg Matrix implementation.

…iler into matrix-builtin-handle-type

Fixes microsoft#8122

tools/clang/include/clang/AST/DataRecursiveASTVisitor.h

tools/clang/include/clang/AST/RecursiveASTVisitor.h

tools/clang/include/clang/AST/Type.h

tools/clang/lib/AST/ASTDumper.cpp

tools/clang/lib/Parse/ParseDeclCXX.cpp

tools/clang/lib/Serialization/ASTWriter.cpp

tools/clang/test/SemaHLSL/hlsl/linalg/linalg-matrix-error.hlsl

tex3d · 2026-02-03T00:42:12Z

tools/clang/test/SemaHLSL/hlsl/linalg/linalg-matrix-error.hlsl

+    __builtin_LinAlg_Matrix [[__LinAlg_Matrix_Attributes(ComponentType::I32, 4, 5, MatrixUse::B, MatrixScope::ThreadGroup)]] mat8;
+
+    // expected-error@+1 {{cannot initialize a variable of type '__builtin_LinAlg_Matrix' with an lvalue of type '__builtin_LinAlg_Matrix [[__LinAlg_Matrix_Attributes(ComponentType::I32, 4, 5, MatrixUse::B, MatrixScope::ThreadGroup)]]'}}
+    __builtin_LinAlg_Matrix naked_mat = mat8;


Should it be illegal to declare this type without attributes? Is that the reason for TMP?

We could make it illegal by intercepting every place where a type can be used and adding a special check and error if it is the __builtin_LinAlgMatrix without attributes. Or we can do nothing because the naked typed cannot be used for anything anyway. The __builtin prefix means users should not be using these directly.

tools/clang/lib/Sema/SemaHLSL.cpp

tex3d · 2026-02-03T00:54:09Z

tools/clang/lib/Sema/SemaHLSL.cpp

+
+  // Check that the value is a valid range.
+  int64_t Value = APVal.getLimitedValue();
+  if (Value < 0) {


Should we enforce any maximum?

The spec has maximums for the K dimension but no specified maximum the M or N dimensions:

| Matrix Scope | Scalar element dimensions | | ------------ | ------------------------- | | Thread | [4,128] | | Wave | [4,128] | | ThreadGroup | [1,1024] |

The "scalar element" refers to the dimensions being 4x that number if the data type is an 8-bit type represented in packed data types.

…://github.com/microsoft/DirectXShaderCompiler into linalg-matrix-attributes-and-types

V-FEXrt

LGTM but I did a very rush job on the review. Wanted to mark my general approval before being out the next couple days

github-actions · 2026-02-03T02:31:44Z

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

llvm-beanz

There's a lot of boilerplate here, but it all looks pretty reasonable to me.

llvm-beanz · 2026-02-05T00:20:59Z

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

+#ifdef __hlsl_dx_compiler
+#define SIZE_TYPE int
+#else
+#define SIZE_TYPE uint
+#endif


I don't expect that this header will be portable to clang, so for the purposes of DXC we can simplify this to just:

Suggested change

#ifdef __hlsl_dx_compiler

#define SIZE_TYPE int

#else

#define SIZE_TYPE uint

#endif

#define SIZE_TYPE int

We should probably file an issue to decide what to do about size types generally in HLSL. I don't love that DXC decided size types were signed 32-bit integers, but we may want Clang to match that for at least some language modes (cc @tex3d & @bogner for visibility about Clang and language spec issues).

llvm-beanz · 2026-02-05T00:30:28Z

tools/clang/lib/Sema/SemaHLSL.cpp

+
+  // Check that the value is a valid range.
+  int64_t Value = APVal.getLimitedValue();
+  if (Value < 0) {


The spec has maximums for the K dimension but no specified maximum the M or N dimensions:

| Matrix Scope | Scalar element dimensions | | ------------ | ------------------------- | | Thread | [4,128] | | Wave | [4,128] | | ThreadGroup | [1,1024] |

The "scalar element" refers to the dimensions being 4x that number if the data type is an 8-bit type represented in packed data types.

tools/clang/lib/Sema/SemaHLSL.cpp

…ectly in error reporting

…iler into linalg-matrix-attributes-and-types

pow2clk

I just implemented almost this exact change, so I had many thoughts. Nothing is really blocking, but some things you may want to consider

pow2clk · 2026-02-03T20:12:29Z

tools/clang/lib/AST/TypePrinter.cpp

+void TypePrinter::printAttributedLinAlgMatrixBefore(
+    const AttributedLinAlgMatrixType *T, raw_ostream &OS) {
+  printBefore(T->getWrappedType(), OS);
+}


Slightly confusing to put the before after after 😄

pow2clk · 2026-02-04T00:06:08Z

tools/clang/include/clang/Basic/DiagnosticSemaKinds.td

+def err_hlsl_linalg_matrix_attribute_on_invalid_type
+    : Error<"matrix attributes can only be applied to %0">;
+def err_hlsl_linalg_attributed_matrix_required
+    : Error<"argument type must be linear algebra matrix with attributes">;


The attributes are required for the linalg matrix type aren't they? Infrequent as I think this will be produced, I think it might be clearer to just say "linear algebra matrix" or "linear algebra matrix type"

pow2clk · 2026-02-04T00:24:07Z

tools/clang/lib/Sema/SemaHLSL.cpp

+                        // indexer object used to implement .mips[1].
+  AR_TOBJ_STRING,       // Represents a string
+  AR_TOBJ_DEPENDENT,    // Dependent type for template.
+  AR_TOBJ_LINALG_MATRIX // LinAlg Matric type


This is the change I'm least sure about. It provides a way to enforce argument matching, but I was looking at this as just another TOBJ_OBJECT. As it stands, there are a few other switches that will need LINALG_MATRIX entries if this is to persist. Pretty much wherever TOBJ_OBJECT shows up, this will be needed too.

pow2clk · 2026-02-05T00:02:29Z

include/dxc/dxcapi.internal.h

-  LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS = 54,
+  LICOMPTYPE_BUILTIN_TRIANGLE_POSITIONS = 55,

 #ifdef ENABLE_SPIRV_CODEGEN


At some point, I think it would be simpler to follow the example of #7231 and remove these ifdefs. I don't know if this is that moment though. Maybe should be done as a separate change, but as you shuffled around a few numbers above, maybe you have thoughts.

pow2clk · 2026-02-05T00:06:52Z

tools/clang/include/clang/AST/DataRecursiveASTVisitor.h

 DEF_TRAVERSE_TYPELOC(EnumType, {})
+// HLSL Change Start
+DEF_TRAVERSE_TYPELOC(AttributedLinAlgMatrixType, {})
+DEF_TRAVERSE_TYPELOC(DependentAttributedLinAlgMatrixType, {})


Can these be empty? I traversed each of the expressions and types similar to how ExtVector and DependentExtVector above do.

pow2clk · 2026-02-05T00:19:15Z

tools/clang/include/clang/AST/Type.h

+  hlsl::DXIL::MatrixUse getUse() const { return Use; }
+  hlsl::DXIL::MatrixScope getScope() const { return Scope; }
+
+  void appendMangledAttributes(llvm::raw_ostream &OS) const;


I don't see any other types with this implemented as part of the type class. They put this in the mangling file. It's not that important, but I'm curious why it was placed here. Is it because the enums are converted from integers when they are stored here? Since they come in as integers and even the mangling treats them as such, I just kept them as integers.

pow2clk · 2026-02-05T00:24:07Z

tools/clang/include/clang/Serialization/ASTBitCodes.h

-      TYPE_ADJUSTED              = 42
+      TYPE_ADJUSTED = 42,
+      /// \brief An AttributedLinAlgMatrixType record.
+      TYPE_ATTRIBUTED_LINALG_MATRIX = 43


I don't see this being used anywhere. Is it necessary?

pow2clk · 2026-02-05T00:28:34Z

tools/clang/lib/AST/ASTDumper.cpp

+                T->getComponentType())
+         << ", " << T->getRows() << ", " << T->getCols() << ", "
+         << hlsl::ConvertLinAlgMatrixUseToString(T->getUse()) << ", "
+         << hlsl::ConvertLinAlgMatrixScopeToString(T->getScope()) << ")]]";


If this is ever to be read, representing use and scope as strings is going to make that trickier, requiring a header somehow to define the values

pow2clk · 2026-02-05T00:34:19Z

tools/clang/lib/AST/TypePrinter.cpp

+     << hlsl::ConvertLinAlgMatrixComponentTypeToString(T->getComponentType())
+     << ", " << T->getRows() << ", " << T->getCols() << ", "
+     << hlsl::ConvertLinAlgMatrixUseToString(T->getUse()) << ", "
+     << hlsl::ConvertLinAlgMatrixScopeToString(T->getScope()) << ")]]";


I again wonder if printing string names of enums that aren't externally available is going to make this less computer readable, though more human readable.

pow2clk · 2026-02-05T00:36:39Z

tools/clang/include/clang/Basic/Attr.td

 }

+def HLSLLinAlgMatrixAttributes : TypeAttr {
+  let Spellings = [CXX11<"", "__LinAlgMatrix_Attributes", 2015>];


If you were to put hlsl in the now empty string, the attribute would require a "hlsl:" prefix, which seems in keeping with the langopt inclusion below as well as how resources were done in modern LLVM

Suggested change

let Spellings = [CXX11<"", "__LinAlgMatrix_Attributes", 2015>];

let Spellings = [CXX11<"hlsl", "__LinAlgMatrix_Attributes", 2015>];

hekota and others added 9 commits January 22, 2026 12:54

[SM6.10] Add built-in type for LinAlg Matrix handle

5390e54

Adds a built-in type `__builtin_LinAlg_Matrix` that will be used in LinAlg Matrix implementation.

Fallback to default debug infor handling

be693b2

Fix type name string

ed855c6

chore: autopublish 2026-01-29T18:27:59Z

9b6f7a7

Merge branch 'main' of https://github.com/microsoft/DirectXShaderComp…

deaac93

…iler into matrix-builtin-handle-type

Move tests back under linalg/builtins

716c63a

Update __buitin_LinAlg_Matrix codegen

8e956a8

Merge branch 'main' of https://github.com/microsoft/DirectXShaderComp…

2e32605

…iler into matrix-builtin-handle-type

Add LinAng Matrix attributes and two new attributed matrix types

3065524

Fixes microsoft#8122

github-project-automation bot added this to HLSL Roadmap Feb 2, 2026

github-project-automation bot moved this to New in HLSL Roadmap Feb 2, 2026

hekota marked this pull request as ready for review February 2, 2026 18:14

hekota requested a review from llvm-beanz February 2, 2026 18:15

hekota changed the title ~~Add LinAng Matrix attributes and two new attributed matrix types~~ [SM6.10] Add LinAng Matrix attributes and two new attributed matrix types Feb 2, 2026

hekota linked an issue Feb 2, 2026 that may be closed by this pull request

Implement attributed matrix handle in the AST #8122

Open

tex3d reviewed Feb 2, 2026

View reviewed changes

tools/clang/include/clang/AST/DataRecursiveASTVisitor.h Show resolved Hide resolved

tex3d reviewed Feb 2, 2026

View reviewed changes

tools/clang/include/clang/AST/RecursiveASTVisitor.h Show resolved Hide resolved

hekota mentioned this pull request Feb 3, 2026

[SM6.10] Add built-in type for LinAlg Matrix handle #8090

Merged

Code review feedback

e9a6bb6

tex3d reviewed Feb 3, 2026

View reviewed changes

hekota added 2 commits February 2, 2026 17:09

Rename the type to __builtin_LinAlgMatrix

eea0b2b

Merge branch 'user/hekota/pr8090-matrix-builtin-handle-type' of https…

94a389f

…://github.com/microsoft/DirectXShaderCompiler into linalg-matrix-attributes-and-types

V-FEXrt reviewed Feb 3, 2026

View reviewed changes

Rename attributes to __LinAlgMatrix_Attributes

c0f4761

hekota and others added 5 commits February 2, 2026 19:04

code review feedback - add missing attributes, remove stale comments

cd120b0

Do not break spirv C++11-style attributes

7f61a48

Few more changes (alignment and type visiting)

f63ceb0

one more rename

57a71b0

chore: autopublish 2026-02-04T18:29:11Z

afe3f11

llvm-beanz reviewed Feb 5, 2026

View reviewed changes

hekota added 2 commits February 4, 2026 17:05

code review feedback - simplify SIZE_TYPE definition, use numbers dir…

d4b899d

…ectly in error reporting

Merge branch 'main' of https://github.com/microsoft/DirectXShaderComp…

9bce569

…iler into linalg-matrix-attributes-and-types

hekota changed the base branch from user/hekota/pr8090-matrix-builtin-handle-type to main February 5, 2026 01:09

pow2clk reviewed Feb 5, 2026

View reviewed changes

	let Spellings = [CXX11<"", "__LinAlgMatrix_Attributes", 2015>];
	let Spellings = [CXX11<"hlsl", "__LinAlgMatrix_Attributes", 2015>];

Conversation

hekota commented Feb 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hekota Feb 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

V-FEXrt left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot commented Feb 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvm-beanz left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

pow2clk left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

hekota commented Feb 2, 2026 •

edited

Loading

hekota Feb 3, 2026 •

edited

Loading

github-actions bot commented Feb 3, 2026 •

edited

Loading