Skip to content

[SM6.10][specs/791] Align LinAlg Matrix Mul with spec #8183

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
V-FEXrt merged 1 commit into from
Feb 20, 2026
+113 −54
Merged

[SM6.10][specs/791] Align LinAlg Matrix Mul with spec #8183

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
4 changes: 2 additions & 2 deletions docs/DXIL.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3076,7 +3076,7 @@ ID Name Description
2147483656 RayQuery_CandidateTriangleObjectPosition returns candidate triangle vertices in object space as <9 x float>
2147483657 RayQuery_CommittedTriangleObjectPosition returns committed triangle vertices in object space as <9 x float>
2147483658 HitObject_TriangleObjectPosition returns triangle vertices in object space as <9 x float>
2147483659 ReservedD0 reserved
2147483659 LinAlgMatrixMultiplyAccumulate Returns the resulting matrix from multiplying A and B and accumulating into C
2147483660 LinAlgFillMatrix fills a matrix with a scalar value
2147483661 LinAlgCopyConvertMatrix Converts and copies the element and use type of the source matrix to the destination matrix with optional transpose
2147483662 LinAlgMatrixLoadFromDescriptor fills a matrix with data from a [RW]ByteAddressBuffer
Expand All @@ -3088,7 +3088,7 @@ ID Name Description
2147483668 LinAlgMatrixStoreToDescriptor stores a matrix to a RWByteAddressBuffer
2147483669 LinAlgMatrixStoreToMemory stores a matrix to groupshared memory
2147483670 LinAlgMatrixQueryAccumulatorLayout returns comptime 0 when accumulator matrix are A layout, 1 when B layout
2147483671 LinAlgMatrixMulOp applies a multiplication op to matrix C using A and B as parameters
2147483671 LinAlgMatrixMultiply Returns the resulting matrix from multiplying A and B
2147483672 LinAlgMatrixAccumulate accumulate A or B matrix into Accumulator matrix following LHS += RHS
2147483673 LinAlgMatVecMul Multiplies a MxK dimension matrix and a K sized input vector
2147483674 LinAlgMatVecMulAdd Multiplies a MxK dimension matrix and a K sized input vector then adds a M sized bias vector
Expand Down
26 changes: 16 additions & 10 deletions include/dxc/DXIL/DxilConstants.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -524,7 +524,6 @@ static const OpCodeTableID TableID = OpCodeTableID::ExperimentalOps;
// Enumeration for ExperimentalOps DXIL operations
enum class OpCode : unsigned {
//
ReservedD0 = 11, // reserved
ReservedD1 = 30, // reserved
ReservedD2 = 31, // reserved
ReservedD3 = 32, // reserved
Expand Down Expand Up @@ -573,8 +572,11 @@ enum class OpCode : unsigned {
14, // fills a matrix with data from a [RW]ByteAddressBuffer
LinAlgMatrixLoadFromMemory =
15, // fills a matrix with data from a groupshared array
LinAlgMatrixMulOp =
23, // applies a multiplication op to matrix C using A and B as parameters
LinAlgMatrixMultiply =
23, // Returns the resulting matrix from multiplying A and B
LinAlgMatrixMultiplyAccumulate =
11, // Returns the resulting matrix from multiplying A and B and
// accumulating into C
LinAlgMatrixOuterProduct = 29, // Outer products an M sized vector and a N
// sized vector producing an MxN matrix
LinAlgMatrixQueryAccumulatorLayout =
Expand Down Expand Up @@ -1263,8 +1265,11 @@ enum class OpCode : unsigned {
EXP_OPCODE(ExperimentalOps,
HitObject_TriangleObjectPosition), // returns triangle vertices in
// object space as <9 x float>
// ReservedD0 = 0x8000000B, 2147483659U, -2147483637
EXP_OPCODE(ExperimentalOps, ReservedD0), // reserved
// LinAlgMatrixMultiplyAccumulate = 0x8000000B, 2147483659U, -2147483637
EXP_OPCODE(ExperimentalOps,
LinAlgMatrixMultiplyAccumulate), // Returns the resulting matrix
// from multiplying A and B and
// accumulating into C
// LinAlgFillMatrix = 0x8000000C, 2147483660U, -2147483636
EXP_OPCODE(ExperimentalOps,
LinAlgFillMatrix), // fills a matrix with a scalar value
Expand Down Expand Up @@ -1316,10 +1321,10 @@ enum class OpCode : unsigned {
LinAlgMatrixQueryAccumulatorLayout), // returns comptime 0 when
// accumulator matrix are A
// layout, 1 when B layout
// LinAlgMatrixMulOp = 0x80000017, 2147483671U, -2147483625
// LinAlgMatrixMultiply = 0x80000017, 2147483671U, -2147483625
EXP_OPCODE(ExperimentalOps,
LinAlgMatrixMulOp), // applies a multiplication op to matrix C
// using A and B as parameters
LinAlgMatrixMultiply), // Returns the resulting matrix from
// multiplying A and B
// LinAlgMatrixAccumulate = 0x80000018, 2147483672U, -2147483624
EXP_OPCODE(ExperimentalOps,
LinAlgMatrixAccumulate), // accumulate A or B matrix into
Expand Down Expand Up @@ -1529,7 +1534,8 @@ enum class OpCodeClass : unsigned {
LinAlgMatrixLength,
LinAlgMatrixLoadFromDescriptor,
LinAlgMatrixLoadFromMemory,
LinAlgMatrixMulOp,
LinAlgMatrixMultiply,
LinAlgMatrixMultiplyAccumulate,
LinAlgMatrixOuterProduct,
LinAlgMatrixQueryAccumulatorLayout,
LinAlgMatrixSetElement,
Expand Down Expand Up @@ -1725,7 +1731,7 @@ enum class OpCodeClass : unsigned {
NodeOutputIsValid,
OutputComplete,

NumOpClasses = 224, // exclusive last value of enumeration
NumOpClasses = 225, // exclusive last value of enumeration
};
// OPCODECLASS-ENUM:END

Expand Down
46 changes: 40 additions & 6 deletions include/dxc/DXIL/DxilInstructions.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10500,6 +10500,41 @@ struct DxilInst_HitObject_TriangleObjectPosition {
void set_hitObject(llvm::Value *val) { Instr->setOperand(1, val); }
};

/// This instruction Returns the resulting matrix from multiplying A and B and
/// accumulating into C
struct DxilInst_LinAlgMatrixMultiplyAccumulate {
llvm::Instruction *Instr;
// Construction and identification
DxilInst_LinAlgMatrixMultiplyAccumulate(llvm::Instruction *pInstr)
: Instr(pInstr) {}
operator bool() const {
return hlsl::OP::IsDxilOpFuncCallInst(
Instr, hlsl::OP::OpCode::LinAlgMatrixMultiplyAccumulate);
}
// Validation support
bool isAllowed() const { return true; }
bool isArgumentListValid() const {
if (4 != llvm::dyn_cast<llvm::CallInst>(Instr)->getNumArgOperands())
return false;
return true;
}
// Metadata
bool requiresUniformInputs() const { return false; }
// Operand indexes
enum OperandIdx {
arg_matrixA = 1,
arg_matrixB = 2,
arg_matrixC = 3,
};
// Accessors
llvm::Value *get_matrixA() const { return Instr->getOperand(1); }
void set_matrixA(llvm::Value *val) { Instr->setOperand(1, val); }
llvm::Value *get_matrixB() const { return Instr->getOperand(2); }
void set_matrixB(llvm::Value *val) { Instr->setOperand(2, val); }
llvm::Value *get_matrixC() const { return Instr->getOperand(3); }
void set_matrixC(llvm::Value *val) { Instr->setOperand(3, val); }
};

/// This instruction fills a matrix with a scalar value
struct DxilInst_LinAlgFillMatrix {
llvm::Instruction *Instr;
Expand Down Expand Up @@ -10859,15 +10894,14 @@ struct DxilInst_LinAlgMatrixQueryAccumulatorLayout {
bool requiresUniformInputs() const { return false; }
};

/// This instruction applies a multiplication op to matrix C using A and B as
/// parameters
struct DxilInst_LinAlgMatrixMulOp {
/// This instruction Returns the resulting matrix from multiplying A and B
struct DxilInst_LinAlgMatrixMultiply {
llvm::Instruction *Instr;
// Construction and identification
DxilInst_LinAlgMatrixMulOp(llvm::Instruction *pInstr) : Instr(pInstr) {}
DxilInst_LinAlgMatrixMultiply(llvm::Instruction *pInstr) : Instr(pInstr) {}
operator bool() const {
return hlsl::OP::IsDxilOpFuncCallInst(Instr,
hlsl::OP::OpCode::LinAlgMatrixMulOp);
return hlsl::OP::IsDxilOpFuncCallInst(
Instr, hlsl::OP::OpCode::LinAlgMatrixMultiply);
}
// Validation support
bool isAllowed() const { return true; }
Expand Down
61 changes: 34 additions & 27 deletions lib/DXIL/DxilOperations.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2823,16 +2823,15 @@ static const OP::OpCodeProperty ExperimentalOps_OpCodeProps[] = {
{{0x2}},
{{0x0}}}, // Overloads: f

{OC::ReservedD0,
"ReservedD0",
OCC::Reserved,
"reserved",
Attribute::None,
0,
{},
{}}, // Overloads: v

// Linear Algebra Operations
{OC::LinAlgMatrixMultiplyAccumulate,
"LinAlgMatrixMultiplyAccumulate",
OCC::LinAlgMatrixMultiplyAccumulate,
"linAlgMatrixMultiplyAccumulate",
Attribute::None,
4,
{{0x200}, {0x200}, {0x200}, {0x200}},
{{0x0}, {0x0}, {0x0}, {0x0}}}, // Overloads: o,o,o,o
{OC::LinAlgFillMatrix,
"LinAlgFillMatrix",
OCC::LinAlgFillMatrix,
Expand Down Expand Up @@ -2921,10 +2920,10 @@ static const OP::OpCodeProperty ExperimentalOps_OpCodeProps[] = {
0,
{},
{}}, // Overloads: v
{OC::LinAlgMatrixMulOp,
"LinAlgMatrixMulOp",
OCC::LinAlgMatrixMulOp,
"linAlgMatrixMulOp",
{OC::LinAlgMatrixMultiply,
"LinAlgMatrixMultiply",
OCC::LinAlgMatrixMultiply,
"linAlgMatrixMultiply",
Attribute::None,
3,
{{0x200}, {0x200}, {0x200}},
Expand Down Expand Up @@ -3950,15 +3949,16 @@ void OP::GetMinShaderModelAndMask(OpCode C, bool bWithTranslation,
minor = 10;
return;
}
// Instructions: LinAlgFillMatrix=2147483660,
// LinAlgCopyConvertMatrix=2147483661, LinAlgMatrixLoadFromMemory=2147483663,
// LinAlgMatrixLength=2147483664, LinAlgMatrixGetCoordinate=2147483665,
// LinAlgMatrixGetElement=2147483666, LinAlgMatrixSetElement=2147483667,
// Instructions: LinAlgMatrixMultiplyAccumulate=2147483659,
// LinAlgFillMatrix=2147483660, LinAlgCopyConvertMatrix=2147483661,
// LinAlgMatrixLoadFromMemory=2147483663, LinAlgMatrixLength=2147483664,
// LinAlgMatrixGetCoordinate=2147483665, LinAlgMatrixGetElement=2147483666,
// LinAlgMatrixSetElement=2147483667,
// LinAlgMatrixStoreToDescriptor=2147483668,
// LinAlgMatrixStoreToMemory=2147483669, LinAlgMatrixMulOp=2147483671,
// LinAlgMatrixStoreToMemory=2147483669, LinAlgMatrixMultiply=2147483671,
// LinAlgMatrixAccumulate=2147483672,
// LinAlgMatrixAccumulateToMemory=2147483676
if ((2147483660 <= op && op <= 2147483661) ||
if ((2147483659 <= op && op <= 2147483661) ||
(2147483663 <= op && op <= 2147483669) ||
(2147483671 <= op && op <= 2147483672) || op == 2147483676) {
major = 6;
Expand Down Expand Up @@ -6557,13 +6557,14 @@ Function *OP::GetOpFunc(OpCode opCode, Type *pOverloadType) {
A(pHit);
break;

//
case OpCode::ReservedD0:
A(pV);
// Linear Algebra Operations
case OpCode::LinAlgMatrixMultiplyAccumulate:
EXT(0);
A(pI32);
EXT(1);
EXT(2);
EXT(3);
break;

// Linear Algebra Operations
case OpCode::LinAlgFillMatrix:
EXT(0);
A(pI32);
Expand Down Expand Up @@ -6637,7 +6638,7 @@ Function *OP::GetOpFunc(OpCode opCode, Type *pOverloadType) {
A(pI32);
A(pI32);
break;
case OpCode::LinAlgMatrixMulOp:
case OpCode::LinAlgMatrixMultiply:
EXT(0);
A(pI32);
EXT(1);
Expand Down Expand Up @@ -7013,7 +7014,6 @@ llvm::Type *OP::GetOverloadType(OpCode opCode, llvm::Function *F) {
case OpCode::GetGroupWaveIndex:
case OpCode::GetGroupWaveCount:
case OpCode::ClusterID:
case OpCode::ReservedD0:
case OpCode::LinAlgMatrixQueryAccumulatorLayout:
case OpCode::ReservedD1:
case OpCode::ReservedD2:
Expand Down Expand Up @@ -7070,13 +7070,20 @@ llvm::Type *OP::GetOverloadType(OpCode opCode, llvm::Function *F) {
return llvm::StructType::get(Ctx,
{FT->getParamType(1), FT->getParamType(2)});

case OpCode::LinAlgMatrixMultiplyAccumulate:
if (FT->getNumParams() < 4)
return nullptr;
return llvm::StructType::get(Ctx,
{FT->getReturnType(), FT->getParamType(1),
FT->getParamType(2), FT->getParamType(3)});

case OpCode::LinAlgMatrixSetElement:
if (FT->getNumParams() < 4)
return nullptr;
return llvm::StructType::get(
Ctx, {FT->getReturnType(), FT->getParamType(1), FT->getParamType(3)});

case OpCode::LinAlgMatrixMulOp:
case OpCode::LinAlgMatrixMultiply:
case OpCode::LinAlgMatrixAccumulate:
case OpCode::LinAlgMatVecMul:
case OpCode::LinAlgMatrixOuterProduct:
Expand Down
4 changes: 2 additions & 2 deletions lib/HLSL/HLOperationLower.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7680,9 +7680,9 @@ constexpr IntrinsicLower gLowerTable[] = {
{IntrinsicOp::IOP___builtin_LinAlg_MatrixAccumulate, EmptyLower,
DXIL::OpCode::LinAlgMatrixAccumulate},
{IntrinsicOp::IOP___builtin_LinAlg_MatrixMatrixMultiply, EmptyLower,
DXIL::OpCode::LinAlgMatrixMulOp},
DXIL::OpCode::LinAlgMatrixMultiply},
{IntrinsicOp::IOP___builtin_LinAlg_MatrixMatrixMultiplyAccumulate,
EmptyLower, DXIL::OpCode::LinAlgMatrixMulOp},
EmptyLower, DXIL::OpCode::LinAlgMatrixMultiplyAccumulate},
{IntrinsicOp::IOP___builtin_LinAlg_MatrixQueryAccumulatorLayout, EmptyLower,
DXIL::OpCode::LinAlgMatrixQueryAccumulatorLayout},
{IntrinsicOp::IOP___builtin_LinAlg_MatrixAccumulateToDescriptor, EmptyLower,
Expand Down
2 changes: 1 addition & 1 deletion utils/hct/gen_intrin_main.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -411,7 +411,7 @@ void [[min_sm=6.10]] __builtin_LinAlg_MatrixStoreToDescriptor(in LinAlgMatrix ma
void [[min_sm=6.10]] __builtin_LinAlg_MatrixStoreToMemory(in LinAlgMatrix matrix, in int GroupSharedMem, in uint offset, in uint stride, in uint layout);
uint [[min_sm=6.10]] __builtin_LinAlg_MatrixQueryAccumulatorLayout();
void [[min_sm=6.10]] __builtin_LinAlg_MatrixMatrixMultiply(out LinAlgMatrix matrixC, in LinAlgMatrix matrixA, in LinAlgMatrix matrixB);
void [[min_sm=6.10]] __builtin_LinAlg_MatrixMatrixMultiplyAccumulate(out LinAlgMatrix matrixC, in LinAlgMatrix matrixA, in LinAlgMatrix matrixB);
void [[min_sm=6.10]] __builtin_LinAlg_MatrixMatrixMultiplyAccumulate(out LinAlgMatrix matrixR, in LinAlgMatrix matrixA, in LinAlgMatrix matrixB, in LinAlgMatrix matrixC);
void [[min_sm=6.10]] __builtin_LinAlg_MatrixAccumulate(out LinAlgMatrix matrixC, in LinAlgMatrix matrixLHS, in LinAlgMatrix matrixRHS);
void [[min_sm=6.10]] __builtin_LinAlg_MatrixVectorMultiply(out numeric<> ret, in LinAlgMatrix mat, in numeric<> input, in uint input_interp);
void [[min_sm=6.10]] __builtin_LinAlg_MatrixVectorMultiplyAdd(out numeric<> ret, in LinAlgMatrix mat, in numeric<> input, in uint input_interp, in numeric<> bias, in uint bias_interp);
Expand Down
24 changes: 18 additions & 6 deletions utils/hct/hctdb.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1174,8 +1174,8 @@ def populate_categories_and_models_ExperimentalOps(self):
+ "LinAlgMatrixGetCoordinate,LinAlgMatrixGetElement,"
+ "LinAlgMatrixSetElement,LinAlgMatrixStoreToDescriptor,"
+ "LinAlgMatrixLoadFromMemory,LinAlgMatrixStoreToMemory,"
+ "LinAlgMatrixAccumulateToMemory,LinAlgMatrixMulOp,"
+ "LinAlgMatrixAccumulate"
+ "LinAlgMatrixAccumulateToMemory,LinAlgMatrixMultiply,"
+ "LinAlgMatrixMultiplyAccumulate,LinAlgMatrixAccumulate"
):
i.category = "Linear Algebra Operations"
i.shader_model = experimental_sm
Expand Down Expand Up @@ -6341,7 +6341,19 @@ def populate_ExperimentalOps(self):
)

# Linear Algebra Ops
op_table.reserve_dxil_op_range("ReservedD", 1)
add_dxil_op(
"LinAlgMatrixMultiplyAccumulate",
"LinAlgMatrixMultiplyAccumulate",
"Returns the resulting matrix from multiplying A and B and accumulating into C",
"o,o,o,o",
"",
[
db_dxil_param(0, "$x0", "", "resulting matrix"),
Copy link
Collaborator

@spall spall Feb 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is skipping over 1 intentional?

Copy link
Collaborator Author

@V-FEXrt V-FEXrt Feb 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yep! 1 is the const int opcode which is inserted automatically

db_dxil_param(2, "$x1", "matrixA", "A matrix"),
db_dxil_param(3, "$x2", "matrixB", "B matrix"),
db_dxil_param(4, "$x3", "matrixC", "C matrix"),
],
)

add_dxil_op(
"LinAlgFillMatrix",
Expand Down Expand Up @@ -6530,9 +6542,9 @@ def populate_ExperimentalOps(self):
)

add_dxil_op(
"LinAlgMatrixMulOp",
"LinAlgMatrixMulOp",
"applies a multiplication op to matrix C using A and B as parameters",
"LinAlgMatrixMultiply",
"LinAlgMatrixMultiply",
"Returns the resulting matrix from multiplying A and B",
Copy link
Collaborator

@spall spall Feb 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this one doesn't need db_dxil_params?

Copy link
Collaborator Author

@V-FEXrt V-FEXrt Feb 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It has them just below the diff!

The params of this op didn't change, just the name as description

Copy link
Collaborator

@spall spall Feb 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

haha im just blind!

"o,o,o",
"",
[
Expand Down