Merge remote-tracking branch 'msft/main' into NEWBASE

Author: simoll

docs/DXIL.rst

@@ -2419,6 +2419,10 @@ ID  Name                                                  Description
 302 ReservedC9                                            reserved
 303 RawBufferVectorLoad                                   reads from a raw buffer and structured buffer
 304 RawBufferVectorStore                                  writes to a RWByteAddressBuffer or RWStructuredBuffer
+305 MatVecMul                                             Multiplies a MxK dimension matrix and a K sized input vector
+306 MatVecMulAdd                                          multiplies a MxK dimension matrix and a K sized input vector and adds an M-sized bias vector
+307 OuterProductAccumulate                                Computes the outer product between column vectors and an MxN matrix is accumulated component-wise atomically (with device scope) in memory
+308 VectorAccumulate                                      Accumulates the components of a vector component-wise atomically (with device scope) to the corresponding elements of an array in memory
 === ===================================================== =======================================================================================================================================================================================================================
 
 
@@ -3134,6 +3138,14 @@ INSTR.ILLEGALDXILOPCODE                               DXILOpCode must be [0..%0]
 INSTR.ILLEGALDXILOPFUNCTION                           '%0' is not a DXILOpFuncition for DXILOpcode '%1'.
 INSTR.IMMBIASFORSAMPLEB                               bias amount for sample_b must be in the range [%0,%1], but %2 was specified as an immediate.
 INSTR.INBOUNDSACCESS                                  Access to out-of-bounds memory is disallowed.
+INSTR.LINALGINTERPRETATIONPARAMARECONST               In Linalg operations, Interpretation value is a constant.
+INSTR.LINALGINVALIDMATRIXLAYOUTVALUEFORMATVECOPS      Matrix Layout for Linalg Mul/MulAdd operation must be valid.
+INSTR.LINALGINVALIDMEMORYINTERPVALUE                  In Memory Interpolation value must be valid.
+INSTR.LINALGINVALIDREGISTERINTERPVALUE                From Register Interpretation value must be valid.
+INSTR.LINALGMATRIXLAYOUTNOTTRANSPOSABLE               Row Major and Column Major matrix layouts are not transposable.
+INSTR.LINALGMATRIXSHAPEPARAMSARECONST                 Matrix Layout, Dimensions and isTranspose are constants
+INSTR.LINALGNOTANUNSIGNEDTYPE                         Unsigned flag set for a float signed type
+INSTR.MATVECOPISUNSIGNEDFLAGSARECONST                 In Linalg Mul/MulAdd functions, IsUnsigned flag is a constant.
 INSTR.MAYREORDERTHREADUNDEFCOHERENCEHINTPARAM         Use of undef coherence hint or num coherence hint bits in MaybeReorderThread.
 INSTR.MINPRECISIONNOTPRECISE                          Instructions marked precise may not refer to minprecision values.
 INSTR.MINPRECISONBITCAST                              Bitcast on minprecison types is not allowed.
@@ -3161,6 +3173,7 @@ INSTR.OPCODERESERVED                                  Instructions must not refe
 INSTR.OPCONST                                         DXIL intrinsic requires an immediate constant operand
 INSTR.OPCONSTRANGE                                    Constant values must be in-range for operation.
 INSTR.OPERANDRANGE                                    DXIL intrinsic operand must be within defined range
+INSTR.PARAMMULTIPLE                                   Parameter must be a valid multiple
 INSTR.PTRBITCAST                                      Pointer type bitcast must be have same size.
 INSTR.RESOURCECLASSFORLOAD                            load can only run on UAV/SRV resource.
 INSTR.RESOURCECLASSFORSAMPLERGATHER                   sample, lod and gather should be on srv resource.

docs/ReleaseNotes.md

@@ -21,7 +21,10 @@ The included licenses apply to the following files:
 
 Place release notes for the upcoming release below this line and remove this line upon naming this release.
 
+### Version 1.8.2505
+
 - Typed buffers (including ROV buffers) no longer accept types other than vectors and scalars. Any other types will produce descriptive errors. This removes support for appropriately sized matrices and structs. Though it worked in some contexts, code generated from such types was unreliable.
+- By default, the internal validator will be used instead of searching externally for an existing DXIL.dll.
 
 ### Version 1.8.2502
 

include/dxc/DXIL/DxilConstants.h

@@ -162,24 +162,32 @@ const unsigned kDxilMaxOloadDims = 2;
 
 enum class ComponentType : uint32_t {
   Invalid = 0,
-  I1,
-  I16,
-  U16,
-  I32,
-  U32,
-  I64,
-  U64,
-  F16,
-  F32,
-  F64,
-  SNormF16,
-  UNormF16,
-  SNormF32,
-  UNormF32,
-  SNormF64,
-  UNormF64,
-  PackedS8x32,
-  PackedU8x32,
+  I1 = 1,
+  I16 = 2,
+  U16 = 3,
+  I32 = 4,
+  U32 = 5,
+  I64 = 6,
+  U64 = 7,
+  F16 = 8,
+  F32 = 9,
+  F64 = 10,
+  SNormF16 = 11,
+  UNormF16 = 12,
+  SNormF32 = 13,
+  UNormF32 = 14,
+  SNormF64 = 15,
+  UNormF64 = 16,
+  PackedS8x32 = 17,
+  PackedU8x32 = 18,
+
+  // BEGIN NEW FOR SM 6.9
+  U8 = 19,
+  I8 = 20,
+  F8_E4M3 = 21,
+  F8_E5M2 = 22,
+  // END
+
   LastEntry
 };
 
@@ -743,6 +751,19 @@ enum class OpCode : unsigned {
   CreateHandleForLib =
       160, // create resource handle from resource struct for library
 
+  // Linear Algebra Operations
+  MatVecMul =
+      305, // Multiplies a MxK dimension matrix and a K sized input vector
+  MatVecMulAdd = 306, // multiplies a MxK dimension matrix and a K sized input
+                      // vector and adds an M-sized bias vector
+  OuterProductAccumulate =
+      307, // Computes the outer product between column vectors and an MxN
+           // matrix is accumulated component-wise atomically (with device
+           // scope) in memory
+  VectorAccumulate = 308, // Accumulates the components of a vector
+                          // component-wise atomically (with device scope) to
+                          // the corresponding elements of an array in memory
+
   // Mesh shader instructions
   EmitIndices = 169, // emit a primitive's vertex indices in a mesh shader
   GetMeshPayload =
@@ -1060,7 +1081,7 @@ enum class OpCode : unsigned {
   NumOpCodes_Dxil_1_7 = 226,
   NumOpCodes_Dxil_1_8 = 258,
 
-  NumOpCodes = 305 // exclusive last value of enumeration
+  NumOpCodes = 309 // exclusive last value of enumeration
 };
 // OPCODE-ENUM:END
 
@@ -1201,6 +1222,12 @@ enum class OpCodeClass : unsigned {
   // Library create handle from resource struct (like HL intrinsic)
   CreateHandleForLib,
 
+  // Linear Algebra Operations
+  MatVecMul,
+  MatVecMulAdd,
+  OuterProductAccumulate,
+  VectorAccumulate,
+
   // Mesh shader instructions
   EmitIndices,
   GetMeshPayload,
@@ -1385,7 +1412,7 @@ enum class OpCodeClass : unsigned {
   NumOpClasses_Dxil_1_7 = 153,
   NumOpClasses_Dxil_1_8 = 174,
 
-  NumOpClasses = 190 // exclusive last value of enumeration
+  NumOpClasses = 194 // exclusive last value of enumeration
 };
 // OPCODECLASS-ENUM:END
 
@@ -1561,6 +1588,28 @@ const unsigned kHitObjectTraceRay_RayDescOpIdx = 7;
 const unsigned kHitObjectTraceRay_PayloadOpIdx = 15;
 const unsigned kHitObjectTraceRay_NumOp = 16;
 
+// MatVec Ops
+const unsigned kMatVecMulInputVectorIdx = 1;
+const unsigned kMatVecMulIsInputUnsignedIdx = 2;
+const unsigned kMatVecMulInputInterpretationIdx = 3;
+const unsigned kMatVecMulMatrixBufferIdx = 4;
+const unsigned kMatVecMulMatrixOffsetIdx = 5;
+const unsigned kMatVecMulMatrixInterpretationIdx = 6;
+const unsigned kMatVecMulMatrixMIdx = 7;
+const unsigned kMatVecMulMatrixKIdx = 8;
+const unsigned kMatVecMulMatrixLayoutIdx = 9;
+const unsigned kMatVecMulMatrixTransposeIdx = 10;
+const unsigned kMatVecMulMatrixStrideIdx = 11;
+const unsigned kMatVecMulIsOutputUnsignedIdx = 12;
+
+// MatVecAdd
+const unsigned kMatVecMulAddBiasInterpretation = 14;
+const unsigned kMatVecMulAddIsOutputUnsignedIdx = 15;
+
+// Outer Product Accumulate
+const unsigned kOuterProdAccMatrixInterpretation = 5;
+const unsigned kOuterProdAccMatrixLayout = 6;
+
 // TODO: add operand index for all the OpCodeClass.
 } // namespace OperandIndex
 
@@ -2132,6 +2181,13 @@ extern const char *kHostLayoutTypePrefix;
 
 extern const char *kWaveOpsIncludeHelperLanesString;
 
+enum class LinalgMatrixLayout : uint32_t {
+  RowMajor = 0,
+  ColumnMajor = 1,
+  MulOptimal = 2,
+  OuterProductOptimal = 3,
+};
+
 } // namespace DXIL
 
 } // namespace hlsl