Skip to content

[SM6.10] LinAlg Validation: MatrixMultiply MatrixMultiplyAccumulate #8779

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
Ashley Coleman (V-FEXrt) merged 3 commits into from
Aug 18, 2026
Merged

[SM6.10] LinAlg Validation: MatrixMultiply MatrixMultiplyAccumulate #8779

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9a82afc
[SM6.10] LinAlg Validation: MatrixMultiply MatrixMultiplyAccumulate
V-FEXrt Aug 14, 2026
57ea9ce
address comments
V-FEXrt Aug 15, 2026
a349ec9
address comments
V-FEXrt Aug 15, 2026
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
4 changes: 4 additions & 0 deletions docs/DXIL.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3216,12 +3216,16 @@ INSTR.LINALGMATRIXDIMKVECKMISMATCH %0 vector size '%1' must b
INSTR.LINALGMATRIXDIMVECTORMISMATCH %0 vector size '%1' must match input matrix M dimension '%2'
INSTR.LINALGMATRIXLAYOUTREQSTRIDE %0 with layout '%1' requires stride 0.
INSTR.LINALGMATRIXLOADTHREADREQUIRESBAB Loading matrix with Thread scope requires ByteAddressBuffer.
INSTR.LINALGMATRIXMATRIXKDIMMUSTMATCH K dim of A matrix '%0' must match K dim of B matrix '%1'. %2 != %3.
INSTR.LINALGMATRIXMATRIXRESDIMMUSTMATCH %0 matrix dimension '%1' must match A.MxB.N '%2'.
INSTR.LINALGMATRIXNOTEXACTMATCH %0 matrix '%1' must exactly match %2 matrix '%3'.
INSTR.LINALGMATRIXOUTPUTBIASVECMISMATCH Output vector element type '%0' must match bias vector element type '%1'
INSTR.LINALGMATRIXREQUIRESLAYOUT2 %0 requires layout %1 or %2.
INSTR.LINALGMATRIXREQUIRESRWBAB %0 requires RWByteAddressBuffer.
INSTR.LINALGMATRIXSCOPEMISMATCH %0 matrix scope '%1' does not match expected scope %2.
INSTR.LINALGMATRIXSCOPEMISMATCH2 %0 matrix scope '%1' does not match expected scope %2 or %3.
INSTR.LINALGMATRIXSCOPEMUSTMATCH3 Matrix scope must be the same for all matrices. %0 '%1', %2 '%3', %4 '%5'.
INSTR.LINALGMATRIXSCOPEMUSTMATCH4 Matrix scope must be the same for all matrices. %0 '%1', %2 '%3', %4 '%5', %6 '%7'.
INSTR.LINALGMATRIXSCOPEREQLAYOUT2 %0 matrix with scope '%1' requires layout %2 or %3 for %4.
INSTR.LINALGMATRIXUNSIGNEDFLOATTYPENOTALLOWED Float-like type '%0' must be signed
INSTR.LINALGMATRIXUSEMISMATCH %0 matrix use '%1' does not match expected use %2.
Expand Down
186 changes: 186 additions & 0 deletions lib/DxilValidation/DxilValidation.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1451,12 +1451,198 @@ static void ValidateLinAlgMatrixMultiply(CallInst *CI,
ValidationContext &ValCtx) {
ValidateLinAlgOpReturnMatrix(CI, ValCtx);
ValidateLinAlgOpParameters(CI, ValCtx);
DxilInst_LinAlgMatrixMultiply Op(CI);
std::optional<LinAlgTargetType> RetMat =
GetCheckedLATT(CI->getType(), ValCtx);
if (!RetMat)
return;
std::optional<LinAlgTargetType> AMat =
GetCheckedLATT(Op.get_matrixA()->getType(), ValCtx);
Comment thread
joaosaffran marked this conversation as resolved.
if (!AMat)
return;
std::optional<LinAlgTargetType> BMat =
GetCheckedLATT(Op.get_matrixB()->getType(), ValCtx);
if (!BMat)
return;

// A is an A matrix
if (AMat->Use != DXIL::MatrixUse::A)
ValCtx.EmitInstrFormatError(CI,
ValidationRule::InstrLinAlgMatrixUseMismatch,
{"A", MatrixUseToString(AMat->Use), "A"});

// B is a B matrix
if (BMat->Use != DXIL::MatrixUse::B)
ValCtx.EmitInstrFormatError(CI,
ValidationRule::InstrLinAlgMatrixUseMismatch,
{"B", MatrixUseToString(BMat->Use), "B"});

// Ret is an Accumulator matrix
if (RetMat->Use != DXIL::MatrixUse::Accumulator)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixUseMismatch,
{"Return", MatrixUseToString(RetMat->Use), "Accumulator"});

// A scope must be Wave or ThreadGroup
if (AMat->Scope != DXIL::MatrixScope::Wave &&
AMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"A", MatrixScopeToString(AMat->Scope), "Wave", "ThreadGroup"});

// B scope must be Wave or ThreadGroup
if (BMat->Scope != DXIL::MatrixScope::Wave &&
BMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"B", MatrixScopeToString(BMat->Scope), "Wave", "ThreadGroup"});

// Ret scope must be Wave or ThreadGroup
if (RetMat->Scope != DXIL::MatrixScope::Wave &&
RetMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"Return", MatrixScopeToString(RetMat->Scope), "Wave", "ThreadGroup"});

// A, B, Ret scope must all be the same
if (AMat->Scope != BMat->Scope || BMat->Scope != RetMat->Scope)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMustMatch3,
{"A", MatrixScopeToString(AMat->Scope), "B",
MatrixScopeToString(BMat->Scope), "Return",
MatrixScopeToString(RetMat->Scope)});

unsigned M = AMat->M;
unsigned AK = AMat->N;
unsigned BK = BMat->M;
unsigned N = BMat->N;

// K dim must match between A and B
if (AK != BK)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixMatrixKDimMustMatch,
{std::to_string(M) + "x" + std::to_string(AK),
std::to_string(BK) + "x" + std::to_string(N), std::to_string(AK),
std::to_string(BK)});

// Return dim must match A.M x B.N
if (RetMat->M != M || RetMat->N != N)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixMatrixResDimMustMatch,
{"Return", std::to_string(RetMat->M) + "x" + std::to_string(RetMat->N),
std::to_string(M) + "x" + std::to_string(N)});
}

static void ValidateLinAlgMatrixMultiplyAccumulate(CallInst *CI,
ValidationContext &ValCtx) {
ValidateLinAlgOpReturnMatrix(CI, ValCtx);
ValidateLinAlgOpParameters(CI, ValCtx);
DxilInst_LinAlgMatrixMultiplyAccumulate Op(CI);
std::optional<LinAlgTargetType> RetMat =
GetCheckedLATT(CI->getType(), ValCtx);
if (!RetMat)
return;
std::optional<LinAlgTargetType> AMat =
GetCheckedLATT(Op.get_matrixA()->getType(), ValCtx);
if (!AMat)
return;
std::optional<LinAlgTargetType> BMat =
GetCheckedLATT(Op.get_matrixB()->getType(), ValCtx);
if (!BMat)
return;
std::optional<LinAlgTargetType> CMat =
GetCheckedLATT(Op.get_matrixC()->getType(), ValCtx);
if (!CMat)
return;

// A is an A matrix
if (AMat->Use != DXIL::MatrixUse::A)
ValCtx.EmitInstrFormatError(CI,
ValidationRule::InstrLinAlgMatrixUseMismatch,
{"A", MatrixUseToString(AMat->Use), "A"});

// B is a B matrix
if (BMat->Use != DXIL::MatrixUse::B)
ValCtx.EmitInstrFormatError(CI,
ValidationRule::InstrLinAlgMatrixUseMismatch,
{"B", MatrixUseToString(BMat->Use), "B"});

// C is an Accumulator matrix
if (CMat->Use != DXIL::MatrixUse::Accumulator)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixUseMismatch,
{"C", MatrixUseToString(CMat->Use), "Accumulator"});

// Ret is an Accumulator matrix
if (RetMat->Use != DXIL::MatrixUse::Accumulator)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixUseMismatch,
{"Return", MatrixUseToString(RetMat->Use), "Accumulator"});

// A scope must be Wave or ThreadGroup
if (AMat->Scope != DXIL::MatrixScope::Wave &&
AMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"A", MatrixScopeToString(AMat->Scope), "Wave", "ThreadGroup"});

// B scope must be Wave or ThreadGroup
if (BMat->Scope != DXIL::MatrixScope::Wave &&
BMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"B", MatrixScopeToString(BMat->Scope), "Wave", "ThreadGroup"});

// C scope must be Wave or ThreadGroup
if (CMat->Scope != DXIL::MatrixScope::Wave &&
CMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"C", MatrixScopeToString(CMat->Scope), "Wave", "ThreadGroup"});

// Ret scope must be Wave or ThreadGroup
if (RetMat->Scope != DXIL::MatrixScope::Wave &&
RetMat->Scope != DXIL::MatrixScope::ThreadGroup)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMismatch2,
{"Return", MatrixScopeToString(RetMat->Scope), "Wave", "ThreadGroup"});

// A, B, C, Ret scope must all be the same
if (AMat->Scope != BMat->Scope || BMat->Scope != CMat->Scope ||
CMat->Scope != RetMat->Scope)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixScopeMustMatch4,
{"A", MatrixScopeToString(AMat->Scope), "B",
MatrixScopeToString(BMat->Scope), "C",
MatrixScopeToString(CMat->Scope), "Return",
MatrixScopeToString(RetMat->Scope)});

unsigned M = AMat->M;
unsigned AK = AMat->N;
unsigned BK = BMat->M;
unsigned N = BMat->N;

// K dim must match between A and B
if (AK != BK)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixMatrixKDimMustMatch,
{std::to_string(M) + "x" + std::to_string(AK),
std::to_string(BK) + "x" + std::to_string(N), std::to_string(AK),
std::to_string(BK)});

// C dim must match A.M x B.N
if (CMat->M != M || CMat->N != N)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixMatrixResDimMustMatch,
{"C", std::to_string(CMat->M) + "x" + std::to_string(CMat->N),
std::to_string(M) + "x" + std::to_string(N)});

// Return dim must match A.M x B.N
if (RetMat->M != M || RetMat->N != N)
ValCtx.EmitInstrFormatError(
CI, ValidationRule::InstrLinAlgMatrixMatrixResDimMustMatch,
{"Return", std::to_string(RetMat->M) + "x" + std::to_string(RetMat->N),
std::to_string(M) + "x" + std::to_string(N)});
}

static void ValidateLinAlgMatrixOuterProduct(CallInst *CI,
Expand Down
26 changes: 17 additions & 9 deletions tools/clang/test/CodeGenDXIL/hlsl/linalg/builtins/matrixmatrixmultiply/nominal.hlsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,14 +6,22 @@
void main() {
// CHECK-LABEL: define void @main()

// CHECK: call %dx.types.LinAlgMatrixC4M5N4U1S2 @dx.op.linAlgMatrixMultiply.mC4M5N4U1S2.mC4M5N4U1S2.mC4M5N4U1S2(i32 -2147483625,
// CHECK-SAME: %dx.types.LinAlgMatrixC4M5N4U1S2 %{{.*}}, %dx.types.LinAlgMatrixC4M5N4U1S2 %{{.*}}) ; LinAlgMatrixMultiply(matrixA,matrixB)
// Matrix<I32, 8, 4, A, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 8, 4, 0, 2)]] matA;
// Matrix<I32, 4, 8, B, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 4, 8, 1, 2)]] matB;
// Matrix<I32, 8, 8, Acc, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 8, 8, 2, 2)]] matC;

// CHECK2: call void @"dx.hl.op..void (i32, %dx.types.LinAlgMatrixC4M5N4U1S2*, %dx.types.LinAlgMatrixC4M5N4U1S2,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M5N4U1S2)"(i32 412, %dx.types.LinAlgMatrixC4M5N4U1S2* %mat2,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M5N4U1S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M5N4U1S2 %{{[0-9]+}})
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 5, 4, 1, 2)]] mat1;
__builtin_LinAlg_FillMatrix(mat1, 1);
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 5, 4, 1, 2)]] mat2;
__builtin_LinAlg_MatrixMatrixMultiply(mat2, mat1, mat1);
__builtin_LinAlg_FillMatrix(matA, 1);
__builtin_LinAlg_FillMatrix(matB, 2);

// CHECK: call %dx.types.LinAlgMatrixC4M8N8U2S2 @dx.op.linAlgMatrixMultiply.mC4M8N8U2S2.mC4M8N4U0S2.mC4M4N8U1S2(i32 -2147483625,
// CHECK-SAME: %dx.types.LinAlgMatrixC4M8N4U0S2 %{{.*}}, %dx.types.LinAlgMatrixC4M4N8U1S2 %{{.*}}) ; LinAlgMatrixMultiply(matrixA,matrixB)

// CHECK2: call void @"dx.hl.op..void (i32, %dx.types.LinAlgMatrixC4M8N8U2S2*, %dx.types.LinAlgMatrixC4M8N4U0S2,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M4N8U1S2)"(i32 412, %dx.types.LinAlgMatrixC4M8N8U2S2* %matC,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M8N4U0S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M4N8U1S2 %{{[0-9]+}})

__builtin_LinAlg_MatrixMatrixMultiply(matC, matA, matB);
}
38 changes: 23 additions & 15 deletions ...s/clang/test/CodeGenDXIL/hlsl/linalg/builtins/matrixmatrixmultiplyaccumulate/nominal.hlsl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,20 +12,28 @@ void main() {
// CHECK: ; LinAlgFillMatrix(value)
// CHECK: ; LinAlgFillMatrix(value)

// CHECK: call %dx.types.LinAlgMatrixC4M5N3U1S2
// CHECK-SAME: @dx.op.linAlgMatrixMultiplyAccumulate.mC4M5N3U1S2.mC4M5N4U1S2.mC4M4N3U1S2.mC4M5N3U1S2
// CHECK-SAME: (i32 -2147483637, %dx.types.LinAlgMatrixC4M5N4U1S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M4N3U1S2 %{{[0-9]+}},
// CHECK-SAME: %dx.types.LinAlgMatrixC4M5N3U1S2 %{{[0-9]+}}) ; LinAlgMatrixMultiplyAccumulate(matrixA,matrixB,matrixC)
// Matrix<I32, 8, 4, A, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 8, 4, 0, 2)]] matA;
// Matrix<I32, 4, 8, B, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 4, 8, 1, 2)]] matB;
// Matrix<I32, 8, 8, Acc, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 8, 8, 2, 2)]] matC;
// Matrix<I32, 8, 8, Acc, ThreadGroup>
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 8, 8, 2, 2)]] matR;

// CHECK2: call void @"dx.hl.op..void (i32, %dx.types.LinAlgMatrixC4M5N3U1S2*, %dx.types.LinAlgMatrixC4M5N4U1S2,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M4N3U1S2, %dx.types.LinAlgMatrixC4M5N3U1S2)"(i32 413,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M5N3U1S2* %{{.*}}, %dx.types.LinAlgMatrixC4M5N4U1S2 %{{[0-9]+}},
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M4N3U1S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M5N3U1S2 %{{[0-9]+}})
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 5, 4, 1, 2)]] mat1;
__builtin_LinAlg_FillMatrix(mat1, 1);
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 4, 3, 1, 2)]] mat2;
__builtin_LinAlg_FillMatrix(mat2, 2);
__builtin_LinAlgMatrix [[__LinAlgMatrix_Attributes(4, 5, 3, 1, 2)]] mat3;
__builtin_LinAlg_FillMatrix(mat3, 3);
__builtin_LinAlg_MatrixMatrixMultiplyAccumulate(mat3, mat1, mat2, mat3);
__builtin_LinAlg_FillMatrix(matA, 1);
__builtin_LinAlg_FillMatrix(matB, 2);
__builtin_LinAlg_FillMatrix(matC, 3);

// CHECK: call %dx.types.LinAlgMatrixC4M8N8U2S2
// CHECK-SAME: @dx.op.linAlgMatrixMultiplyAccumulate.mC4M8N8U2S2.mC4M8N4U0S2.mC4M4N8U1S2.mC4M8N8U2S2
// CHECK-SAME: (i32 -2147483637, %dx.types.LinAlgMatrixC4M8N4U0S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M4N8U1S2 %{{[0-9]+}},
// CHECK-SAME: %dx.types.LinAlgMatrixC4M8N8U2S2 %{{[0-9]+}}) ; LinAlgMatrixMultiplyAccumulate(matrixA,matrixB,matrixC)

// CHECK2: call void @"dx.hl.op..void (i32, %dx.types.LinAlgMatrixC4M8N8U2S2*, %dx.types.LinAlgMatrixC4M8N4U0S2,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M4N8U1S2, %dx.types.LinAlgMatrixC4M8N8U2S2)"(i32 413,
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M8N8U2S2* %{{.*}}, %dx.types.LinAlgMatrixC4M8N4U0S2 %{{[0-9]+}},
// CHECK2-SAME: %dx.types.LinAlgMatrixC4M4N8U1S2 %{{[0-9]+}}, %dx.types.LinAlgMatrixC4M8N8U2S2 %{{[0-9]+}})

__builtin_LinAlg_MatrixMatrixMultiplyAccumulate(matR, matA, matB, matC);
}
Loading
Loading