Implementation of the CoopVec Inference and Training builtin intrinisics

include/dxc/HLSL/HLOperations.h

@@ -433,6 +433,56 @@ const unsigned kNodeHandleToResCastOpIdx = 1;
 const unsigned kAnnotateNodeHandleNodePropIdx = 2;
 const unsigned kAnnotateNodeRecordHandleNodeRecordPropIdx = 2;
 
+// Linear Algebra Operations
+
+// MatVecMul
+const unsigned kMatVecMulOutputVectorIdx = 1;
+const unsigned kMatVecMulIsOutputUnsignedIdx = 2;
+const unsigned kMatVecMulInputVectorIdx = 3;
+const unsigned kMatVecMulIsInputUnsignedIdx = 4;
+const unsigned kMatVecMulInputInterpretationIdx = 5;
+const unsigned kMatVecMulMatrixBufferIdx = 6;
+const unsigned kMatVecMulMatrixOffsetIdx = 7;
+const unsigned kMatVecMulMatrixInterpretationIdx = 8;
+const unsigned kMatVecMulMatrixMIdx = 9;
+const unsigned kMatVecMulMatrixKIdx = 10;
+const unsigned kMatVecMulMatrixLayoutIdx = 11;
+const unsigned kMatVecMulMatrixTransposeIdx = 12;
+const unsigned kMatVecMulMatrixStrideIdx = 13;
+
+
+// MatVecMulAdd
+const unsigned kMatVecMulAddOutputVectorIdx = 1;
+const unsigned kMatVecMulAddIsOutputUnsignedIdx = 2;
+const unsigned kMatVecMulAddInputVectorIdx = 3;
+const unsigned kMatVecMulAddIsInputUnsignedIdx = 4;
+const unsigned kMatVecMulAddInputInterpretationIdx = 5;
+const unsigned kMatVecMulAddMatrixBufferIdx = 6;
+const unsigned kMatVecMulAddMatrixOffsetIdx = 7;
+const unsigned kMatVecMulAddMatrixInterpretationIdx = 8;
+const unsigned kMatVecMulAddMatrixMIdx = 9;
+const unsigned kMatVecMulAddMatrixKIdx = 10;
+const unsigned kMatVecMulAddMatrixLayoutIdx = 11;
+const unsigned kMatVecMulAddMatrixTransposeIdx = 12;
+const unsigned kMatVecMulAddMatrixStrideIdx = 13;
+const unsigned kMatVecMulAddBiasBufferIdx = 14;
+const unsigned kMatVecMulAddBiasOffsetIdx = 15;
+const unsigned kMatVecMulAddBiasInterpretationIdx = 16;
+
+
+// OuterProductAccumulate
+const unsigned kOuterProdAccInputVec1Idx = 1;
+const unsigned kOuterProdAccInputVec2Idx = 2;
+const unsigned kOuterProdAccMatrixIdx = 3;
+const unsigned kOuterProdAccMatrixOffsetIdx = 4;
+const unsigned kOuterProdAccMatrixInterpretationIdx = 5;
+const unsigned kOuterProdAccMatrixLayoutIdx = 6;
+const unsigned kOuterProdAccMatrixStrideIdx = 7;
+
+// Vector Accumulate
+const unsigned kVectorAccInputVecIdx = 1;
+const unsigned kVectorAccMatrixIdx = 2;
+const unsigned kVectorAccMatrixOffsetIdx = 3;
 } // namespace HLOperandIndex
 
 llvm::Function *GetOrCreateHLFunction(llvm::Module &M,

include/dxc/HlslIntrinsicOp.h

@@ -333,8 +333,6 @@ enum class IntrinsicOp {
   MOP_TraceRayInline = 325,
   MOP_WorldRayDirection = 326,
   MOP_WorldRayOrigin = 327,
-  MOP_DxHitObject_MakeNop = 358,
-  IOP_DxMaybeReorderThread = 359,
   MOP_Count = 328,
   MOP_FinishedCrossGroupSharing = 329,
   MOP_GetGroupNodeOutputRecords = 330,
@@ -366,7 +364,13 @@ enum class IntrinsicOp {
   IOP_usign = 355,
   MOP_InterlockedUMax = 356,
   MOP_InterlockedUMin = 357,
-  Num_Intrinsics = 360,
+  MOP_DxHitObject_MakeNop = 358,
+  IOP_DxMaybeReorderThread = 359,
+  IOP___builtin_MatVecMul = 360,
+  IOP___builtin_MatVecMulAdd = 361,
+  IOP___builtin_OuterProductAccumulate = 362,
+  IOP___builtin_VectorAccumulate = 363,
+  Num_Intrinsics = 364,
 };
 inline bool HasUnsignedIntrinsicOpcode(IntrinsicOp opcode) {
   switch (opcode) {