Add a "kind" attribute to ContractionOp and OuterProductOp.

[pravnar]

Summary

Currently, vector.contract joins the intermediate result and the accumulator argument (of ranks K) using summation. We desire more joining operations --- such as max --- to help vector.contract express reductions. This change extends Vector_ContractionOp to take an optional attribute (called "kind", of enum type CombiningKind) specifying the joining operation to be add/mul/min/max for int/fp , and and/or/xor for int only. By default this attribute has value "add".

To implement this we also need to extend vector.outerproduct, since vector.contract gets transformed to vector.outerproduct (and that to vector.fma). The extension for vector.outerproduct is also an optional kind attribute that uses the same enum type and possible values. The default is "add". In case of max/min we transform vector.outerproduct to a combination of compare and select.

[mehdi_amini]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

Can we use an enum instead of string since the number of values is predefined?

[aartbik]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

If you are extending this, can we please extend this to all operations that currently are supported by reductions (add/mul/min/max for int/fp and and/or/xor for int only). Also, that way we can use one struct for both vector.contract and vector.reduction, which will address Mehdi's request.

[pravnar]

(Test) Changing the revision message.

[pravnar]

Add and use an enum to express combining functions

[pravnar]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

Thanks for the suggestions :). I've added a new enum for expressing the combining functions, and used this enum for defining vector.contract and vector.outerproduct. We can update vector.reduction to use this enum as a separate change.

[pravnar]

Update documentation for contract and outerproduct.

[mehdi_amini]

mlir/include/mlir/Dialect/Vector/VectorOps.td

234–235

Can you just write kind()? There should be a generated accessor here.

mlir/lib/Dialect/Vector/VectorOps.cpp

189

It may be best to change the build declaration to add a defaulted extra argument for kind to align with the usual generated TableGen builder?
@ftynse may have an opinion here.

mlir/lib/Dialect/Vector/VectorTransforms.cpp

1368

Nit: remove trivial braces (here and below / elsewhere)

[pravnar]

Use kind() and other edits.

[pravnar]

mlir/include/mlir/Dialect/Vector/VectorOps.td

234–235

That works, thanks.

[mehdi_amini]

@aartbik / @nicolasvasilache / @ftynse any more comment from the vector dialect point of view? Are you OK with this addition conceptually?

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

Actually it seems that you're still using a StringAttr, if we don't expect this to grow over 64 elements, could we use a BitEnumAttr instead?

[pravnar]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

Ah, I guess BitEnumAttr would provide better performance?

I see that BitEnumAttr is stored as an IntegerAttr. One reason I chose StrEnumAttr over something like I64EnumAttr was because the op attributes were printed like kind = "max" rather than kind = 4 : i64. In the latter case one has to look up what the "4th kind" is defined to be. How much do we care about that? IIUC, in the BitEnumAttr case it would look something like kind = 0x8 : i32.

[mehdi_amini]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

No it can be printed with textual support, take a look into FastmathFlags in LLVM dialect.

[pravnar]

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

I see! That looks great, thanks. I can change this to BitEnumAttr.

[aartbik]

thanks for leading this change!

mlir/include/mlir/Dialect/Vector/VectorOps.td

112

+1 on using bits (and symbolics) over strings after this change!

mlir/lib/Dialect/Vector/VectorOps.cpp

436

period at end of comment

1610

can we combine this with the similar code above in a helper?

mlir/lib/Dialect/Vector/VectorTransforms.cpp

1364

genMultX is probably a bit of a misnomer with the generalization

[nicolasvasilache]

mlir/include/mlir/Dialect/Vector/VectorOps.td

40

This is just expressing half of the semantics, we need 2 ops (see below).
Since some combinations are going to be invalid I'd recommend listing the supported cases explicitly rather than pair of ops.

AddMul,
MaxAdd,
etc..

110

the combining operator is only 1/2 of the op semantics.

902

This seems incorrect to me: it assumes multiply is always one of the correct operations but I don't think it is (e.g. {max, +}, {abs, +}, ...).

mlir/lib/Dialect/Vector/VectorTransforms.cpp

1402

If we could use llvm::TypeSwitch or an equivalent, things will look much nicer.
I am unclear it works on int value switching though, maybe @mehdi_amini has a suggestion?

[pravnar]

Implement CombiningKind using BitEnumAttr.
Add a wrapper CKAttr and define custom parsing/printing of the attribute.
Refactor duplicate verification functions into a helper.
Add a test for matvec transform with max.

[mehdi_amini]

mlir/include/mlir/Dialect/Vector/VectorOps.h

54

I'm not fond of the acronym usage, can we spell it out entirely?

[nicolasvasilache]

Ok, let me not hold you up here: this is already progress and we can separate the generalization of the mul to a followup CL, this is fine for now.

[pravnar]

s/CKAttr/CombiningKindAttr

[pravnar]

Add add_public_tablegen_target.

[pravnar]

Add new target MLIRVectorOpsEnumsIncGen to address build errors.

[pravnar]

Move kind attribute args to the end to make gen'ed build methods use default vals.

[pravnar]

Sync to latest repo version.

mlir/include/mlir/Dialect/Vector/CMakeLists.txt

 add_mlir_dialect(VectorOps vector)
+mlir_tablegen(VectorOpsEnums.h.inc -gen-enum-decls)
+mlir_tablegen(VectorOpsEnums.cpp.inc -gen-enum-defs)
 add_mlir_doc(VectorOps -gen-op-doc VectorOps Dialects/)

mlir/include/mlir/Dialect/Vector/VectorOps.h

 #include "mlir/IR/Attributes.h"
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/OpDefinition.h"
 #include "mlir/Interfaces/SideEffectInterfaces.h"
 #include "mlir/Interfaces/VectorInterfaces.h"
 #include "mlir/Interfaces/ViewLikeInterface.h"
 
+// Pull in all enum type definitions and utility function declarations.
+#include "mlir/Dialect/Vector/VectorOpsEnums.h.inc"
+
 namespace mlir {
 class MLIRContext;
 class OwningRewritePatternList;
 namespace vector {
 
 /// Collect a set of vector-to-vector canonicalization patterns.
 void populateVectorToVectorCanonicalizationPatterns(
     OwningRewritePatternList &patterns, MLIRContext *context);

 /// use for "slices" ops), this lowering removes all tuple related
 /// operations as well (through DCE and folding). If tuple values
 /// "leak" coming in, however, some tuple related ops will remain.
 void populateVectorSlicesLoweringPatterns(OwningRewritePatternList &patterns,
                                           MLIRContext *context);
 
 /// Enum to control the lowering of `vector.contract` operations.
 enum class VectorContractLowering {
   /// Progressively lower to finer grained `vector.contract` and dot-products.
   Dot = 0,
   /// Lower to `vector.matrix_multiply`, maps 1-1 to LLVM matrix intrinsics.
   Matmul = 1,
   /// Lower to `vector.outerproduct`.
   OuterProduct = 2,
 };
 /// Enum to control the lowering of `vector.transpose` operations.
 enum class VectorTransposeLowering {

mlir/include/mlir/Dialect/Vector/VectorOps.td

class Vector_Op<string mnemonic, list<OpTrait> traits = []> :
   //   * ParseResult parse${C++ class of Op}(OpAsmParser &parser,
   //                                         OperationState &result)
   // functions.
   let printer = [{ return ::print(p, *this); }];
   let verifier = [{ return ::verify(*this); }];
   let parser = [{ return ::parse$cppClass(parser, result); }];
 }
 
+// The "kind" of combining function for contractions and reductions.
+def COMBINING_KIND_ADD : StrEnumAttrCase<"ADD", 1, "add">;
+def COMBINING_KIND_MUL : StrEnumAttrCase<"MUL", 2, "mul">;
+def COMBINING_KIND_MIN : StrEnumAttrCase<"MIN", 3, "min">;
+def COMBINING_KIND_MAX : StrEnumAttrCase<"MAX", 4, "max">;
+def COMBINING_KIND_AND : StrEnumAttrCase<"AND", 5, "and">;
+def COMBINING_KIND_OR  : StrEnumAttrCase<"OR",  6, "or">;
+def COMBINING_KIND_XOR : StrEnumAttrCase<"XOR", 7, "xor">;
+
+def CombiningKindAttr : StrEnumAttr<
+    "CombiningKind",
+    "Kind of combining function for contractions and reductions",
+    [COMBINING_KIND_ADD, COMBINING_KIND_MUL, COMBINING_KIND_MIN,
+     COMBINING_KIND_MAX, COMBINING_KIND_AND, COMBINING_KIND_OR,
+     COMBINING_KIND_XOR]> {
+  let cppNamespace = "::mlir";
+}
+
 // TODO: Add an attribute to specify a different algebra with operators other
 // than the current set: {*, +}.
 def Vector_ContractionOp :
   Vector_Op<"contract", [
       NoSideEffect,
       PredOpTrait<"lhs and rhs have same element type", TCopVTEtIsSameAs<0, 1>>,
       PredOpTrait<"third operand acc and result have same element type",
                   TCresVTEtIsSameAsOpBase<0, 2>>,
       DeclareOpInterfaceMethods<VectorUnrollOpInterface, ["getShapeForUnroll"]>
     ]>,
-    Arguments<(ins AnyVector:$lhs, AnyVector:$rhs, AnyType:$acc,
+    Arguments<(ins DefaultValuedAttr<CombiningKindAttr, "add">:$kind,
+               AnyVector:$lhs, AnyVector:$rhs, AnyType:$acc,
                Variadic<VectorOf<[I1]>>:$masks,
                AffineMapArrayAttr:$indexing_maps, ArrayAttr:$iterator_types)>,
     Results<(outs AnyType)> {
   let summary = "vector contraction operation";
   let description = [{
     Computes the sum of products of vector elements along contracting
     dimension pairs from 2 vectors of rank M and N respectively, adds this
     intermediate result to the accumulator argument of rank K, and returns a
*) "parallel": Batch dimensions are iterator type "parallel", and
                    lhs or rhs (but not both). The lhs and rhs free dimensions
                    are unrelated to each other and do not co-iterate, which
                    should be expressed in their indexing maps.
 
     An indexing map attribute list must be specified with an entry for lhs, rhs
     and acc arguments. An indexing map attribute specifies a mapping from each
     iterator in the iterator type list, to each dimension of an N-D vector.
 
+    An optional kind attribute may be used to specify the combining function
+    between the intermediate result and accumulator argument of rank K. This
+    attribute can take the values add/mul/min/max for int/fp, and/or/xor for
+    int only. The default is "add".
+
     Example:
 
     ```mlir
     // Simple DOT product (K = 0).
     #contraction_accesses = [
      affine_map<(i) -> (i)>,
      affine_map<(i) -> (i)>,
      affine_map<(i) -> ()>
let description = [{
 
     %5 = vector.contract #contraction_trait %0, %1, %2, %lhs_mask, %rhs_mask
        : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x15x8x5xf32>
 
     // Vector contraction with mixed typed. lhs/rhs have different element
     // types than accumulator/result.
     %6 = vector.contract #contraction_trait %0, %1, %2
       : vector<10xf16>, vector<10xf16> into f32
+
+    // Contract with max (K = 0).
+    #contraction_accesses = [
+     affine_map<(i) -> (i)>,
+     affine_map<(i) -> (i)>,
+     affine_map<(i) -> ()>
+    ]
+    #contraction_trait = {
+      indexing_maps = #contraction_accesses,
+      iterator_types = ["reduction"],
+      kind = "max"
+    }
+    %7 = vector.contract #contraction_trait %0, %1, %2
+      : vector<10xf32>, vector<10xf32> into f32
     ```
   }];
   let builders = [
     OpBuilderDAG<(ins "Value":$lhs, "Value":$rhs, "Value":$acc,
       "ArrayAttr":$indexingMaps, "ArrayAttr":$iteratorTypes)>,
     OpBuilderDAG<(ins "Value":$lhs, "Value":$rhs, "Value":$acc,
       "ArrayRef<ArrayRef<AffineExpr>>":$indexingExprs,
       "ArrayRef<StringRef>":$iteratorTypes)>
let extraClassDeclaration = [{
     // op indexing map attribute (i.e. one for each input and output, with
     // the output listed last). Each index map, maps from this operations
     // iteration space, to vector dimensions of the maps input/output.
     void getIterationIndexMap(
       std::vector<DenseMap<int64_t, int64_t>> &iterationIndexMap);
 
     std::vector<std::pair<int64_t, int64_t>> getContractingDimMap();
     std::vector<std::pair<int64_t, int64_t>> getBatchDimMap();
+
+    static constexpr StringRef getKindAttrName() { return "kind"; }
+
+    CombiningKind getCombiningKind() {
+      return symbolizeEnum<CombiningKind>(kind()).getValue();
+    }
+
+    static StringRef getDefaultKind() {
+      return stringifyCombiningKind(CombiningKind::ADD);
+    }
   }];
 }
 
 def Vector_ReductionOp :
   Vector_Op<"reduction", [NoSideEffect,
      PredOpTrait<"source operand and result have same element type",
                  TCresVTEtIsSameAsOpBase<0, 0>>]>,
     Arguments<(ins StrAttr:$kind, AnyVector:$vector, Variadic<AnyType>:$acc)>,

 }
 
 def Vector_OuterProductOp :
   Vector_Op<"outerproduct", [NoSideEffect,
     PredOpTrait<"lhs operand and result have same element type",
                 TCresVTEtIsSameAsOpBase<0, 0>>,
     PredOpTrait<"rhs operand and result have same element type",
                 TCresVTEtIsSameAsOpBase<0, 1>>]>,
-    Arguments<(ins AnyVector:$lhs, AnyType:$rhs, Variadic<AnyVector>:$acc)>,
+    Arguments<(ins AnyVector:$lhs, AnyType:$rhs,
+               DefaultValuedAttr<CombiningKindAttr, "add">:$kind,
+               Variadic<AnyVector>:$acc)>,
     Results<(outs AnyVector)> {
   let summary = "vector outerproduct with optional fused add";
   let description = [{
     Takes 2 1-D vectors and returns the 2-D vector containing the outer-product,
     as illustrated below:
     ```
      outer |   [c, d]
      ------+------------
let description = [{
     An optional extra vector argument with the same shape as the output
     vector may be specified in which case the operation returns the sum of
     the outer-product and the extra vector. In this multiply-accumulate
     scenario for floating-point arguments, the rounding mode is enforced
     by guaranteeing that a fused-multiply add operation is emitted. When
     lowered to the LLVMIR dialect, this form emits `llvm.intr.fma`, which
     is guaranteed to lower to actual `fma` instructions on x86.
 
+    An optional kind attribute may be specified to be add/mul/min/max
+    for int/fp, and and/or/xor for int only. The default is "add", in which
+    case the operation returns a fused multiply-add. In other cases it returns
+    a multiply followed by the appropriate operation (for example, a compare and
+    select for "max").
+
     Example:
 
     ```
     %2 = vector.outerproduct %0, %1: vector<4xf32>, vector<8xf32>
     return %2: vector<4x8xf32>
 
     %3 = vector.outerproduct %0, %1, %2:
       vector<4xf32>, vector<8xf32>, vector<4x8xf32>
     return %3: vector<4x8xf32>
 
+    %4 = vector.outerproduct %0, %1, %2 {kind = "max"}:
+      vector<4xf32>, vector<8xf32>, vector<4x8xf32>
+    return %3: vector<4x8xf32>
+
     %6 = vector.outerproduct %4, %5: vector<10xf32>, f32
     return %6: vector<10xf32>
 
     ```
   }];
   let builders = [
     // Build an op without mask, use the type of `acc` as the return type.
     OpBuilderDAG<(ins "Value":$lhs, "Value":$rhs, "Value":$acc)>
   ];
   let extraClassDeclaration = [{
     VectorType getOperandVectorTypeLHS() {
       return lhs().getType().cast<VectorType>();
     }
     Type getOperandTypeRHS() {
       return rhs().getType();
     }
     VectorType getOperandVectorTypeACC() {
       return (llvm::size(acc()) == 0)
         ? VectorType()
         : (*acc().begin()).getType().cast<VectorType>();
     }
     VectorType getVectorType() {
       return getResult().getType().cast<VectorType>();
     }
+    static constexpr StringRef getKindAttrName() {
+      return "kind";
+    }
+    CombiningKind getCombiningKind() {
+      return symbolizeEnum<CombiningKind>(kind()).getValue();
+    }
+    static StringRef getDefaultKind() {
+      return stringifyCombiningKind(CombiningKind::ADD);
+    }
   }];
 }
 
 // TODO: Add transformation which decomposes ReshapeOp into an optimized
 // sequence of vector rotate/shuffle/select operations.
 def Vector_ReshapeOp :
   Vector_Op<"reshape", [AttrSizedOperandSegments, NoSideEffect]>,
     Arguments<(ins AnyVector:$vector, Variadic<Index>:$input_shape,

mlir/include/mlir/IR/OpBase.td

class EnumAttrCaseInfo<string sym, int intVal, string strVal> {
   // to enumerators in the generated enum class.
   int value = intVal;
 
   // The string representation of the enumerant. May be the same as symbol.
   string str = strVal;
 }
 
 // An enum attribute case stored with StringAttr.
-class StrEnumAttrCase<string sym, int val = -1> :
-    EnumAttrCaseInfo<sym, val, sym>,
+class StrEnumAttrCase<string sym, int val = -1, string str = sym> :
+    EnumAttrCaseInfo<sym, val, str>,
     StringBasedAttr<
-      CPred<"$_self.cast<::mlir::StringAttr>().getValue() == \"" # sym # "\"">,
-      "case " # sym>;
+      CPred<"$_self.cast<::mlir::StringAttr>().getValue() == \"" # str # "\"">,
+      "case " # str>;
 
 // An enum attribute case stored with IntegerAttr, which has an integer value,
 // its representation as a string and a C++ symbol name which may be different.
 class IntEnumAttrCaseBase<I intType, string sym, string strVal, int intVal> :
     EnumAttrCaseInfo<sym, intVal, strVal>,
     SignlessIntegerAttrBase<intType, "case " # strVal> {
   let predicate =
     CPred<"$_self.cast<::mlir::IntegerAttr>().getInt() == " # intVal>;

mlir/lib/Dialect/Vector/VectorOps.cpp

 #include "mlir/IR/OpImplementation.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Support/LLVM.h"
 #include "mlir/Support/MathExtras.h"
 #include "llvm/ADT/StringSet.h"
 #include <numeric>
 
+// Pull in all enum type and utility function definitions.
+#include "mlir/Dialect/Vector/VectorOpsEnums.cpp.inc"
+
 using namespace mlir;
 using namespace mlir::vector;
 
 /// Helper enum to classify mask value.
 enum class MaskFormat {
   AllTrue = 0,
   AllFalse = 1,
   Unknown = 2,
void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
                                   ArrayRef<StringRef> iteratorTypes) {
   result.addOperands({lhs, rhs, acc});
   result.addTypes(acc.getType());
   result.addAttribute(getIndexingMapsAttrName(),
                       builder.getAffineMapArrayAttr(
                           AffineMap::inferFromExprList(indexingExprs)));
   result.addAttribute(getIteratorTypesAttrName(),
                       builder.getStrArrayAttr(iteratorTypes));
+  result.addAttribute(ContractionOp::getKindAttrName(),
+                      builder.getStringAttr(ContractionOp::getDefaultKind()));
 }
 
 void vector::ContractionOp::build(OpBuilder &builder, OperationState &result,
                                   Value lhs, Value rhs, Value acc,
                                   ArrayAttr indexingMaps,
                                   ArrayAttr iteratorTypes) {
   result.addOperands({lhs, rhs, acc});
   result.addTypes(acc.getType());
   result.addAttribute(getIndexingMapsAttrName(), indexingMaps);
   result.addAttribute(getIteratorTypesAttrName(), iteratorTypes);
+  result.addAttribute(ContractionOp::getKindAttrName(),
+                      builder.getStringAttr(ContractionOp::getDefaultKind()));
 }
 
 static ParseResult parseContractionOp(OpAsmParser &parser,
                                       OperationState &result) {
   OpAsmParser::OperandType lhsInfo;
   OpAsmParser::OperandType rhsInfo;
   OpAsmParser::OperandType accInfo;
   SmallVector<OpAsmParser::OperandType, 2> masksInfo;
if (parser.parseAttribute(dictAttr, "_", result.attributes) ||
       parser.parseKeywordType("into", resultType) ||
       parser.resolveOperand(lhsInfo, types[0], result.operands) ||
       parser.resolveOperand(rhsInfo, types[1], result.operands) ||
       parser.resolveOperand(accInfo, resultType, result.operands) ||
       parser.addTypeToList(resultType, result.types))
     return failure();
   result.attributes.assign(dictAttr.getValue().begin(),
                            dictAttr.getValue().end());
+  if (!result.attributes.get(ContractionOp::getKindAttrName())) {
+    result.addAttribute(
+        ContractionOp::getKindAttrName(),
+        StringAttr::get(ContractionOp::getDefaultKind(), result.getContext()));
+  }
   if (masksInfo.empty())
     return success();
   if (masksInfo.size() != 2)
     return parser.emitError(parser.getNameLoc(),
                             "expected zero or exactly 2 vector mask operands");
   auto lhsType = types[0].cast<VectorType>();
   auto rhsType = types[1].cast<VectorType>();
   auto maskElementType = parser.getBuilder().getI1Type();
static LogicalResult verify(ContractionOp op) {
   if ((lhsMaskType && !rhsMaskType) || (!lhsMaskType && rhsMaskType))
     return op.emitOpError("invalid number of vector masks specified");
   if (lhsMaskType && rhsMaskType) {
     // Verify mask rank == argument rank.
     if (lhsMaskType.getShape().size() != lhsType.getShape().size() ||
         rhsMaskType.getShape().size() != rhsType.getShape().size())
       return op.emitOpError("invalid vector mask rank");
   }
+
+  // Verify supported combining kind
+  CombiningKind combiningKind = op.getCombiningKind();
+  auto vectorType = resType.dyn_cast<VectorType>();
+  auto elementType = vectorType ? vectorType.getElementType() : resType;
+  switch (combiningKind) {
+  case CombiningKind::ADD:
+  case CombiningKind::MUL:
+  case CombiningKind::MIN:
+  case CombiningKind::MAX:
+    if (!elementType.isIntOrIndexOrFloat()) {
+      return op.emitOpError("unsupported contraction type");
+    }
+    break;
+  case CombiningKind::AND:
+  case CombiningKind::OR:
+  case CombiningKind::XOR:
+    if (!elementType.isIntOrIndex()) {
+      return op.emitOpError("unsupported contraction type");
+    }
+  }
+
   return success();
 }
 
 ArrayRef<StringRef> ContractionOp::getTraitAttrNames() {
-  static constexpr StringRef names[2] = {getIndexingMapsAttrName(),
-                                         getIteratorTypesAttrName()};
+  static constexpr StringRef names[3] = {getIndexingMapsAttrName(),
+                                         getIteratorTypesAttrName(),
+                                         ContractionOp::getKindAttrName()};
   return llvm::makeArrayRef(names);
 }
 
 static int64_t getResultIndex(AffineMap map, AffineExpr targetExpr) {
   for (int64_t i = 0, e = map.getNumResults(); i < e; ++i)
     if (targetExpr == map.getResult(i))
       return i;
   return -1;

 // OuterProductOp
 //===----------------------------------------------------------------------===//
 
 /// Build an op without mask, use the type of `acc` as the return type.
 void OuterProductOp::build(OpBuilder &builder, OperationState &result,
                            Value lhs, Value rhs, Value acc) {
   result.addOperands({lhs, rhs, acc});
   result.addTypes(acc.getType());
+  result.addAttribute(
+      OuterProductOp::getKindAttrName(),
+      StringAttr::get(OuterProductOp::getDefaultKind(), result.getContext()));
 }
 
 static void print(OpAsmPrinter &p, OuterProductOp op) {
   p << op.getOperationName() << " " << op.lhs() << ", " << op.rhs();
-  if (!op.acc().empty())
+  if (!op.acc().empty()) {
     p << ", " << op.acc();
+    p.printOptionalAttrDict(op.getAttrs());
+  }
   p << " : " << op.lhs().getType() << ", " << op.rhs().getType();
 }
 
 static ParseResult parseOuterProductOp(OpAsmParser &parser,
                                        OperationState &result) {
   SmallVector<OpAsmParser::OperandType, 3> operandsInfo;
   Type tLHS, tRHS;
-  if (parser.parseOperandList(operandsInfo) || parser.parseColonType(tLHS) ||
-      parser.parseComma() || parser.parseType(tRHS))
+  if (parser.parseOperandList(operandsInfo) ||
+      parser.parseOptionalAttrDict(result.attributes) ||
+      parser.parseColonType(tLHS) || parser.parseComma() ||
+      parser.parseType(tRHS))
     return failure();
   if (operandsInfo.size() < 2)
     return parser.emitError(parser.getNameLoc(),
                             "expected at least 2 operands");
   VectorType vLHS = tLHS.dyn_cast<VectorType>();
   VectorType vRHS = tRHS.dyn_cast<VectorType>();
   if (!vLHS)
     return parser.emitError(parser.getNameLoc(),
                             "expected vector type for operand #1");
   VectorType resType =
       vRHS ? VectorType::get({vLHS.getDimSize(0), vRHS.getDimSize(0)},
                              vLHS.getElementType())
            : VectorType::get({vLHS.getDimSize(0)}, vLHS.getElementType());
+
+  if (!result.attributes.get(OuterProductOp::getKindAttrName())) {
+    result.attributes.append(
+        OuterProductOp::getKindAttrName(),
+        StringAttr::get(OuterProductOp::getDefaultKind(), result.getContext()));
+  }
+
   return failure(
       parser.resolveOperand(operandsInfo[0], tLHS, result.operands) ||
       parser.resolveOperand(operandsInfo[1], tRHS, result.operands) ||
       (operandsInfo.size() > 2 &&
        parser.resolveOperand(operandsInfo[2], resType, result.operands)) ||
       parser.addTypeToList(resType, result.types));
 }
 
if (vRHS) {
     if (vRES.getRank() != 1)
       return op.emitOpError("expected 1-d vector result");
     if (vLHS.getDimSize(0) != vRES.getDimSize(0))
       return op.emitOpError("expected #1 operand dim to match result dim #1");
   }
 
   if (vACC && vACC != vRES)
     return op.emitOpError("expected operand #3 of same type as result type");
+
+  // Verify supported combining kind
+  CombiningKind combiningKind = op.getCombiningKind();
+  auto elementType = vRES.getElementType();
+  switch (combiningKind) {
+  case CombiningKind::ADD:
+  case CombiningKind::MUL:
+  case CombiningKind::MIN:
+  case CombiningKind::MAX:
+    if (!elementType.isIntOrIndexOrFloat()) {
+      return op.emitOpError("unsupported outerproduct type");
+    }
+    break;
+  case CombiningKind::AND:
+  case CombiningKind::OR:
+  case CombiningKind::XOR:
+    if (!elementType.isIntOrIndex()) {
+      return op.emitOpError("unsupported outerproduct type");
+    }
+  }
+
   return success();
 }
 
 //===----------------------------------------------------------------------===//
 // ReshapeOp
 //===----------------------------------------------------------------------===//
 
 static LogicalResult verify(ReshapeOp op) {

mlir/lib/Dialect/Vector/VectorTransforms.cpp

LogicalResult matchAndRewrite(vector::OuterProductOp op,
     auto loc = op.getLoc();
 
     VectorType lhsType = op.getOperandVectorTypeLHS();
     VectorType rhsType = op.getOperandTypeRHS().dyn_cast<VectorType>();
     VectorType resType = op.getVectorType();
     Type eltType = resType.getElementType();
     bool isInt = eltType.isa<IntegerType>();
     Value acc = (op.acc().empty()) ? nullptr : op.acc()[0];
+    CombiningKind kind = op.getCombiningKind();
 
     if (!rhsType) {
       // Special case: AXPY operation.
       Value b = rewriter.create<vector::BroadcastOp>(loc, lhsType, op.rhs());
-      rewriter.replaceOp(op, genMult(loc, op.lhs(), b, acc, isInt, rewriter));
+      Optional<Value> mult =
+          isInt ? genMultI(loc, op.lhs(), b, acc, kind, rewriter)
+                : genMultF(loc, op.lhs(), b, acc, kind, rewriter);
+      if (!mult.hasValue())
+        return failure();
+      rewriter.replaceOp(op, mult.getValue());
       return success();
     }
 
     Value result = rewriter.create<ConstantOp>(loc, resType,
                                                rewriter.getZeroAttr(resType));
     for (int64_t d = 0, e = resType.getDimSize(0); d < e; ++d) {
       auto pos = rewriter.getI64ArrayAttr(d);
       Value x = rewriter.create<vector::ExtractOp>(loc, eltType, op.lhs(), pos);
       Value a = rewriter.create<vector::BroadcastOp>(loc, rhsType, x);
       Value r = nullptr;
       if (acc)
         r = rewriter.create<vector::ExtractOp>(loc, rhsType, acc, pos);
-      Value m = genMult(loc, a, op.rhs(), r, isInt, rewriter);
-      result = rewriter.create<vector::InsertOp>(loc, resType, m, result, pos);
+      Optional<Value> m = isInt ? genMultI(loc, a, op.rhs(), r, kind, rewriter)
+                                : genMultF(loc, a, op.rhs(), r, kind, rewriter);
+      if (!m.hasValue())
+        return failure();
+      result = rewriter.create<vector::InsertOp>(loc, resType, m.getValue(),
+                                                 result, pos);
     }
     rewriter.replaceOp(op, result);
     return success();
   }
 
 private:
-  static Value genMult(Location loc, Value x, Value y, Value acc, bool isInt,
+  static Optional<Value> genMultI(Location loc, Value x, Value y, Value acc,
+                                  CombiningKind kind,
                                   PatternRewriter &rewriter) {
-    if (acc) {
-      if (isInt)
-        return rewriter.create<AddIOp>(loc, rewriter.create<MulIOp>(loc, x, y),
-                                       acc);
-      return rewriter.create<vector::FMAOp>(loc, x, y, acc);
+    MulIOp mul = rewriter.create<MulIOp>(loc, x, y);
+
+    if (!acc)
+      return Optional<Value>(mul);
+
+    Value combinedResult;
+    switch (kind) {
+    case CombiningKind::ADD:
+      combinedResult = rewriter.create<AddIOp>(loc, mul, acc);
+      break;
+    case CombiningKind::MUL:
+      combinedResult = rewriter.create<MulIOp>(loc, mul, acc);
+      break;
+    case CombiningKind::MIN:
+      combinedResult = rewriter.create<SelectOp>(
+          loc, rewriter.create<CmpIOp>(loc, CmpIPredicate::slt, mul, acc), mul,
+      break;
+    case CombiningKind::MAX:
+      combinedResult = rewriter.create<SelectOp>(
+          loc, rewriter.create<CmpIOp>(loc, CmpIPredicate::sge, mul, acc), mul,
+          acc);
+      break;
+    case CombiningKind::AND:
+      combinedResult = rewriter.create<AndOp>(loc, mul, acc);
+      break;
+    case CombiningKind::OR:
+      combinedResult = rewriter.create<OrOp>(loc, mul, acc);
+      break;
+    case CombiningKind::XOR:
+      combinedResult = rewriter.create<XOrOp>(loc, mul, acc);
+      break;
+    }
+    return Optional<Value>(combinedResult);
+  }
+
+  static Optional<Value> genMultF(Location loc, Value x, Value y, Value acc,
+                                  CombiningKind kind,
+                                  PatternRewriter &rewriter) {
+    // Special case for fused multiply-add.
+    if (acc && kind == CombiningKind::ADD) {
+      return Optional<Value>(rewriter.create<vector::FMAOp>(loc, x, y, acc));
+    }
+
+    MulFOp mul = rewriter.create<MulFOp>(loc, x, y);
+
+    if (!acc)
+      return Optional<Value>(mul);
+
+    Value combinedResult;
+    switch (kind) {
+    case CombiningKind::MUL:
+      combinedResult = rewriter.create<MulFOp>(loc, mul, acc);
+      break;
+    case CombiningKind::MIN:
+      combinedResult = rewriter.create<SelectOp>(
+          loc, rewriter.create<CmpFOp>(loc, CmpFPredicate::OLE, mul, acc), mul,
+          acc);
+      break;
+    case CombiningKind::MAX:
+      combinedResult = rewriter.create<SelectOp>(
+          loc, rewriter.create<CmpFOp>(loc, CmpFPredicate::OGT, mul, acc), mul,
+          acc);
+      break;
+    case CombiningKind::ADD: // Already handled this special case above.
+    case CombiningKind::AND: // Only valid for integer types.
+    case CombiningKind::OR:  // Only valid for integer types.
+    case CombiningKind::XOR: // Only valid for integer types.
+      return Optional<Value>();
     }
-    if (isInt)
-      return rewriter.create<MulIOp>(loc, x, y);
-    return rewriter.create<MulFOp>(loc, x, y);
+    return Optional<Value>(combinedResult);
   }
 };
 
 /// Progressive lowering of ConstantMaskOp.
 /// One:
 ///   %x = vector.constant_mask [a,b]
 /// is replaced by:
 ///   %z = zero-result
if (isParallelIterator(iteratorTypes[0]) &&
     return failure();
   }
   assert(reductionSize > 0);
 
   // Unroll outer-products along reduction.
   for (int64_t k = 0; k < reductionSize; ++k) {
     Value a = rewriter.create<vector::ExtractOp>(op.getLoc(), lhs, k);
     Value b = rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, k);
-    res = rewriter.create<vector::OuterProductOp>(op.getLoc(), a, b, res);
+    res = rewriter.create<vector::OuterProductOp>(op.getLoc(), res.getType(), a,
+                                                  b, op.kind(), res);
   }
   rewriter.replaceOp(op, res);
   return success();
 }
 
 LogicalResult
 ContractionOpToDotLowering::matchAndRewrite(vector::ContractionOp op,
                                             PatternRewriter &rewriter) const {

mlir/test/Dialect/Linalg/vectorization.mlir

^bb(%a: f32, %b: f32, %c: f32) :
       linalg.yield %e : f32
   }
   return
 }
 // CHECK-LABEL: func @vectorization_test
 //       CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x16xf32>
 //       CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<16x32xf32>
 //       CHECK: vector.transfer_read %{{.*}} : memref<8x32xf32>, vector<8x32xf32>
-//       CHECK: vector.contract {indexing_maps = [#[[$mk]], #[[$kn]], #[[$mn]]], iterator_types = ["parallel", "parallel", "reduction"]} %{{.*}}, %{{.*}}, %{{.*}} : vector<8x16xf32>, vector<16x32xf32> into vector<8x32xf32>
+//       CHECK: vector.contract {indexing_maps = [#[[$mk]], #[[$kn]], #[[$mn]]], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %{{.*}}, %{{.*}}, %{{.*}} : vector<8x16xf32>, vector<16x32xf32> into vector<8x32xf32>
 //       CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<8x32xf32>
 
 func @vectorization_test_integer(%A: memref<8x16xi32>, %B: memref<16x32xi32>,
                                  %C: memref<8x32xi32>) {
   linalg.generic #matmul_trait
     ins(%A, %B : memref<8x16xi32>, memref<16x32xi32>)
    outs(%C : memref<8x32xi32>) {
     ^bb(%a: i32, %b: i32, %c: i32) :
       %d = muli %a, %b: i32
       %e = addi %c, %d: i32
       linalg.yield %e : i32
   }
   return
 }
 // CHECK-LABEL: func @vectorization_test_integer
 //       CHECK: vector.transfer_read %{{.*}} : memref<8x16xi32>, vector<8x16xi32>
 //       CHECK: vector.transfer_read %{{.*}} : memref<16x32xi32>, vector<16x32xi32>
 //       CHECK: vector.transfer_read %{{.*}} : memref<8x32xi32>, vector<8x32xi32>
-//       CHECK: vector.contract {indexing_maps = [#[[$mk]], #[[$kn]], #[[$mn]]], iterator_types = ["parallel", "parallel", "reduction"]} %{{.*}}, %{{.*}}, %{{.*}} : vector<8x16xi32>, vector<16x32xi32> into vector<8x32xi32>
+//       CHECK: vector.contract {indexing_maps = [#[[$mk]], #[[$kn]], #[[$mn]]], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %{{.*}}, %{{.*}}, %{{.*}} : vector<8x16xi32>, vector<16x32xi32> into vector<8x32xi32>
 //       CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xi32>, memref<8x32xi32>
 
 func @vectorization_test_2(%A: memref<8x16xf32>, %B: memref<16x32xf32>,
                          %C: memref<8x32xf32>) {
   linalg.matmul
     ins(%A, %B: memref<8x16xf32>, memref<16x32xf32>)
    outs(%C: memref<8x32xf32>)
   return

 
 // CHECK-LABEL: func @matmul_tensors
 //  CHECK-SAME: (%[[ARG0:.*]]: tensor<8x4xf32>, %[[ARG1:.*]]: tensor<4x12xf32>,
 //  CHECK-SAME:  %[[ARG2:.*]]: tensor<8x12xf32>) -> tensor<8x12xf32>
 //       CHECK:   %[[C0:.*]] = constant 0 : index
 //       CHECK:   %[[V0:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x4xf32>, vector<8x4xf32>
 //       CHECK:   %[[V1:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x12xf32>, vector<4x12xf32>
 //       CHECK:   %[[V2:.*]] = vector.transfer_read %[[ARG2]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x12xf32>, vector<8x12xf32>
-//       CHECK:   %[[C:.*]] = vector.contract {{.*}} iterator_types = ["parallel", "parallel", "reduction"]} %[[V0]], %[[V1]], %[[V2]] : vector<8x4xf32>, vector<4x12xf32> into vector<8x12xf32>
+//       CHECK:   %[[C:.*]] = vector.contract {{.*}} iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[V0]], %[[V1]], %[[V2]] : vector<8x4xf32>, vector<4x12xf32> into vector<8x12xf32>
 //       CHECK:   %[[W:.*]] = vector.transfer_write %[[C]], %[[ARG2]][%[[C0]], %[[C0]]] {masked = [false, false]} : vector<8x12xf32>, tensor<8x12xf32>
 //       CHECK:   return %[[W]] : tensor<8x12xf32>

mlir/test/Dialect/Vector/ops.mlir

 #contraction_to_scalar_trait = {
   indexing_maps = #contraction_to_scalar_accesses,
   iterator_types = ["reduction"]
 }
 // CHECK-LABEL: @contraction_to_scalar
 func @contraction_to_scalar(%arg0: vector<10xf32>, %arg1: vector<10xf32>) -> f32 {
   // CHECK:      %[[C0:.*]] = constant 0.000000e+00 : f32
   %f0 = constant 0.0: f32
-  // CHECK:      %[[X:.*]] = vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["reduction"]} %{{.*}}, %{{.*}}, %[[C0]] : vector<10xf32>, vector<10xf32> into f32
+  // CHECK:      %[[X:.*]] = vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["reduction"], kind = "add"} %{{.*}}, %{{.*}}, %[[C0]] : vector<10xf32>, vector<10xf32> into f32
   %0 = vector.contract #contraction_to_scalar_trait %arg0, %arg1, %f0
     : vector<10xf32>, vector<10xf32> into f32
   // CHECK:      return %[[X]] : f32
   return %0 : f32
 }
 
+#contraction_to_scalar_max_accesses = [
+  affine_map<(i) -> (i)>,
+  affine_map<(i) -> (i)>,
+  affine_map<(i) -> ()>
+]
+#contraction_to_scalar_max_trait = {
+  indexing_maps = #contraction_to_scalar_max_accesses,
+  iterator_types = ["reduction"],
+  kind = "max"
+}
+// CHECK-LABEL: @contraction_to_scalar_with_max
+func @contraction_to_scalar_with_max(%arg0: vector<10xf32>, %arg1: vector<10xf32>) -> f32 {
+  // CHECK:      %[[C0:.*]] = constant 0.000000e+00 : f32
+  %f0 = constant 0.0: f32
+  // CHECK:      %[[X:.*]] = vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["reduction"], kind = "max"} %{{.*}}, %{{.*}}, %[[C0]] : vector<10xf32>, vector<10xf32> into f32
+  %0 = vector.contract #contraction_to_scalar_max_trait %arg0, %arg1, %f0
+    : vector<10xf32>, vector<10xf32> into f32
+  // CHECK:      return %[[X]] : f32
+  return %0 : f32
+}
+
 #contraction_accesses0 = [
   affine_map<(b0, f0, f1, c0, c1) -> (c0, b0, c1, f0)>,
   affine_map<(b0, f0, f1, c0, c1) -> (b0, c1, c0, f1)>,
   affine_map<(b0, f0, f1, c0, c1) -> (b0, f0, f1)>
 ]
 #contraction_trait0 = {
   indexing_maps = #contraction_accesses0,
   iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]
 }
 #contraction_accesses1 = [              // 7,  8, 16, 15
   affine_map<(f0, f1, f2, f3, c0, c1) -> (c0, f0, c1, f2)>,
                                         // 8, 16,  7,  5
   affine_map<(f0, f1, f2, f3, c0, c1) -> (f1, c1, c0, f3)>,
                                         // 8,  8, 15,  5
   affine_map<(f0, f1, f2, f3, c0, c1) -> (f0, f1, f2, f3)>
 ]
+#iterator_types1 = ["parallel", "parallel", "parallel", "parallel", "reduction",
+                    "reduction"]
 #contraction_trait1 = {
   indexing_maps = #contraction_accesses1,
-  iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction",
-                    "reduction"]
+  iterator_types = #iterator_types1
+}
+#contraction_trait2 = {
+  indexing_maps = #contraction_accesses1,
+  iterator_types = #iterator_types1,
+  kind = "max"
 }
 // CHECK-LABEL: @contraction
 func @contraction(%arg0 : vector<7x8x16x15xf32>, %arg1 : vector<8x16x7x5xf32>,
                   %arg2 : vector<8x15x5xf32>, %arg3 : vector<8x8x15x5xf32>,
                   %arg4 : vector<7x8x16x15xf16>, %arg5 : vector<8x16x7x5xf16>) {
   // Test contraction with batch and contracting dims.
-  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x15x5xf32>
+  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"], kind = "add"} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x15x5xf32>
   %0 = vector.contract #contraction_trait0 %arg0, %arg1, %arg2
       : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x15x5xf32>
   // Test contraction with only contracting dims. In this case the lhs/rhs
   // dimension of size 8 will be considered a parallel dim for lhs/rhs and will
   // appear twice in the output.
-  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"]} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
+  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"], kind = "add"} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
   %1 = vector.contract #contraction_trait1 %arg0, %arg1, %arg3
       : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
   // Test contraction with optional vector mask arguments.
   %lhs_mask = vector.constant_mask [7, 8, 16, 15] : vector<7x8x16x15xi1>
   %rhs_mask = vector.constant_mask [8, 16, 7, 5] : vector<8x16x7x5xi1>
-  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"]} {{.*}}, {{.*}}, {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
+  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"], kind = "add"} {{.*}}, {{.*}}, {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
   %2 = vector.contract #contraction_trait1 %arg0, %arg1, %arg3, %lhs_mask,
                                            %rhs_mask
       : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
   // Test contraction with mixed type.
-  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"]} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf16>, vector<8x16x7x5xf16> into vector<8x8x15x5xf32>
+  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"], kind = "add"} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf16>, vector<8x16x7x5xf16> into vector<8x8x15x5xf32>
   %3 = vector.contract #contraction_trait1 %arg4, %arg5, %arg3
       : vector<7x8x16x15xf16>, vector<8x16x7x5xf16> into vector<8x8x15x5xf32>
+  // Test contraction with "max" instead of "add".
+  // CHECK: vector.contract {indexing_maps = [#{{.*}}, #{{.*}}, #{{.*}}], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction"], kind = "max"} {{.*}}, {{.*}}, {{.*}} : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
+  %4 = vector.contract #contraction_trait2 %arg0, %arg1, %arg3
+      : vector<7x8x16x15xf32>, vector<8x16x7x5xf32> into vector<8x8x15x5xf32>
   return
 }
 
 // CHECK-LABEL: @create_vector_mask
 func @create_vector_mask() {
   // CHECK:      %[[C2:.*]] = constant 2 : index
   %c2 = constant 2 : index
   // CHECK-NEXT: %[[C3:.*]] = constant 3 : index

mlir/test/Dialect/Vector/vector-contract-matvec-transforms.mlir

 // CHECK-SAME: %[[B:.*1]]: memref<vector<2xf32>>
 // CHECK-SAME: %[[C:.*2]]: memref<vector<2xf32>>
 // CHECK: %[[T0:.*]] = load %[[A]][] : memref<vector<2x2xf32>>
 // CHECK: %[[T1:.*]] = load %[[B]][] : memref<vector<2xf32>>
 // CHECK: %[[T2:.*]] = load %[[C]][] : memref<vector<2xf32>>
 // CHECK: %[[T3:.*]] = vector.transpose %[[T0]], [1, 0] : vector<2x2xf32> to vector<2x2xf32>
 // CHECK: %[[T4:.*]] = vector.extract %[[T3]][0] : vector<2x2xf32>
 // CHECK: %[[T5:.*]] = vector.extract %[[T1]][0] : vector<2xf32>
-// CHECK: %[[T6:.*]] = vector.outerproduct %[[T4]], %[[T5]], %[[T2]] : vector<2xf32>, f32
+// CHECK: %[[T6:.*]] = vector.outerproduct %[[T4]], %[[T5]], %[[T2]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: %[[T7:.*]] = vector.extract %[[T3]][1] : vector<2x2xf32>
 // CHECK: %[[T8:.*]] = vector.extract %[[T1]][1] : vector<2xf32>
-// CHECK: %[[T9:.*]] = vector.outerproduct %[[T7]], %[[T8]], %[[T6]] : vector<2xf32>, f32
+// CHECK: %[[T9:.*]] = vector.outerproduct %[[T7]], %[[T8]], %[[T6]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: store %[[T9]], %[[C]][] : memref<vector<2xf32>>
 // CHECK: return
 func @matvec2x2(%arg0: memref<vector<2x2xf32>>, %arg1: memref<vector<2xf32>>,
                                                 %arg2: memref<vector<2xf32>>) {
   %A = load %arg0[] : memref<vector<2x2xf32>>
   %x = load %arg1[] : memref<vector<2xf32>>
   %b = load %arg2[] : memref<vector<2xf32>>
   %0 = vector.contract #matvec_trait %A, %x, %b : vector<2x2xf32>, vector<2xf32> into vector<2xf32>
   store %0, %arg2[] : memref<vector<2xf32>>
   return
 }
 
 // CHECK-LABEL: func @mattransvec2x2
 // CHECK-SAME: %[[A:.*0]]: memref<vector<2x2xf32>>
 // CHECK-SAME: %[[B:.*1]]: memref<vector<2xf32>>
 // CHECK-SAME: %[[C:.*2]]: memref<vector<2xf32>>
 // CHECK: %[[T0:.*]] = load %[[A]][] : memref<vector<2x2xf32>>
 // CHECK: %[[T1:.*]] = load %[[B]][] : memref<vector<2xf32>>
 // CHECK: %[[T2:.*]] = load %[[C]][] : memref<vector<2xf32>>
 // CHECK: %[[T3:.*]] = vector.extract %[[T0]][0] : vector<2x2xf32>
 // CHECK: %[[T4:.*]] = vector.extract %[[T1]][0] : vector<2xf32>
-// CHECK: %[[T5:.*]] = vector.outerproduct %[[T3]], %[[T4]], %[[T2]] : vector<2xf32>, f32
+// CHECK: %[[T5:.*]] = vector.outerproduct %[[T3]], %[[T4]], %[[T2]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: %[[T6:.*]] = vector.extract %[[T0]][1] : vector<2x2xf32>
 // CHECK: %[[T7:.*]] = vector.extract %[[T1]][1] : vector<2xf32>
-// CHECK: %[[T8:.*]] = vector.outerproduct %[[T6]], %[[T7]], %[[T5]] : vector<2xf32>, f32
+// CHECK: %[[T8:.*]] = vector.outerproduct %[[T6]], %[[T7]], %[[T5]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: store %[[T8]], %[[C]][] : memref<vector<2xf32>>
 // CHECK: return
 func @mattransvec2x2(%arg0: memref<vector<2x2xf32>>, %arg1: memref<vector<2xf32>>,
                                                      %arg2: memref<vector<2xf32>>) {
   %A = load %arg0[] : memref<vector<2x2xf32>>
   %x = load %arg1[] : memref<vector<2xf32>>
   %b = load %arg2[] : memref<vector<2xf32>>
   %0 = vector.contract #mattransvec_trait %A, %x, %b : vector<2x2xf32>, vector<2xf32> into vector<2xf32>
   store %0, %arg2[] : memref<vector<2xf32>>
   return
 }
 
 // CHECK-LABEL: func @vecmat2x2
 // CHECK-SAME: %[[A:.*0]]: memref<vector<2x2xf32>>
 // CHECK-SAME: %[[B:.*1]]: memref<vector<2xf32>>
 // CHECK-SAME: %[[C:.*2]]: memref<vector<2xf32>>
 // CHECK: %[[T0:.*]] = load %[[A]][] : memref<vector<2x2xf32>>
 // CHECK: %[[T1:.*]] = load %[[B]][] : memref<vector<2xf32>>
 // CHECK: %[[T2:.*]] = load %[[C]][] : memref<vector<2xf32>>
 // CHECK: %[[T3:.*]] = vector.transpose %[[T0]], [1, 0] : vector<2x2xf32> to vector<2x2xf32>
 // CHECK: %[[T4:.*]] = vector.extract %[[T3]][0] : vector<2x2xf32>
 // CHECK: %[[T5:.*]] = vector.extract %[[T1]][0] : vector<2xf32>
-// CHECK: %[[T6:.*]] = vector.outerproduct %[[T4]], %[[T5]], %[[T2]] : vector<2xf32>, f32
+// CHECK: %[[T6:.*]] = vector.outerproduct %[[T4]], %[[T5]], %[[T2]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: %[[T7:.*]] = vector.extract %[[T3]][1] : vector<2x2xf32>
 // CHECK: %[[T8:.*]] = vector.extract %[[T1]][1] : vector<2xf32>
-// CHECK: %[[T9:.*]] = vector.outerproduct %[[T7]], %[[T8]], %[[T6]] : vector<2xf32>, f32
+// CHECK: %[[T9:.*]] = vector.outerproduct %[[T7]], %[[T8]], %[[T6]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: store %[[T9]], %[[C]][] : memref<vector<2xf32>>
 // CHECK: return
 func @vecmat2x2(%arg0: memref<vector<2x2xf32>>, %arg1: memref<vector<2xf32>>,
                                                 %arg2: memref<vector<2xf32>>) {
   %A = load %arg0[] : memref<vector<2x2xf32>>
   %x = load %arg1[] : memref<vector<2xf32>>
   %b = load %arg2[] : memref<vector<2xf32>>
   %0 = vector.contract #vecmat_trait %x, %A, %b : vector<2xf32>, vector<2x2xf32> into vector<2xf32>
   store %0, %arg2[] : memref<vector<2xf32>>
   return
 }
 
 // CHECK-LABEL: func @vecmattrans2x2
 // CHECK-SAME: %[[A:.*0]]: memref<vector<2x2xf32>>
 // CHECK-SAME: %[[B:.*1]]: memref<vector<2xf32>>
 // CHECK-SAME: %[[C:.*2]]: memref<vector<2xf32>>
 // CHECK: %[[T0:.*]] = load %[[A]][] : memref<vector<2x2xf32>>
 // CHECK: %[[T1:.*]] = load %[[B]][] : memref<vector<2xf32>>
 // CHECK: %[[T2:.*]] = load %[[C]][] : memref<vector<2xf32>>
 // CHECK: %[[T3:.*]] = vector.extract %[[T0]][0] : vector<2x2xf32>
 // CHECK: %[[T4:.*]] = vector.extract %[[T1]][0] : vector<2xf32>
-// CHECK: %[[T5:.*]] = vector.outerproduct %[[T3]], %[[T4]], %[[T2]] : vector<2xf32>, f32
+// CHECK: %[[T5:.*]] = vector.outerproduct %[[T3]], %[[T4]], %[[T2]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: %[[T6:.*]] = vector.extract %[[T0]][1] : vector<2x2xf32>
 // CHECK: %[[T7:.*]] = vector.extract %[[T1]][1] : vector<2xf32>
-// CHECK: %[[T8:.*]] = vector.outerproduct %[[T6]], %[[T7]], %[[T5]] : vector<2xf32>, f32
+// CHECK: %[[T8:.*]] = vector.outerproduct %[[T6]], %[[T7]], %[[T5]] {kind = "add"} : vector<2xf32>, f32
 // CHECK: store %[[T8]], %[[C]][] : memref<vector<2xf32>>
 // CHECK: return
 func @vecmattrans2x2(%arg0: memref<vector<2x2xf32>>, %arg1: memref<vector<2xf32>>,
                                                      %arg2: memref<vector<2xf32>>) {
   %A = load %arg0[] : memref<vector<2x2xf32>>
   %x = load %arg1[] : memref<vector<2xf32>>
   %b = load %arg2[] : memref<vector<2xf32>>
   %0 = vector.contract #vecmattrans_trait %x, %A, %b : vector<2xf32>, vector<2x2xf32> into vector<2xf32>
   store %0, %arg2[] : memref<vector<2xf32>>
   return
 }

mlir/test/Dialect/Vector/vector-transforms.mlir

 // CHECK-NEXT: %[[ES4:.*]] = vector.extract_slices %[[LMASK]], [2, 2], [1, 1] : vector<4x6xi1> into tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[ES5:.*]] = vector.extract_slices %[[RMASK]], [2, 2], [1, 1] : vector<6x4xi1> into tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
 
 // CHECK-NEXT: %[[TG1:.*]] = vector.tuple_get %[[ES1]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG2:.*]] = vector.tuple_get %[[ES2]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG3:.*]] = vector.tuple_get %[[ES3]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG4:.*]] = vector.tuple_get %[[ES4]], 0 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[TG5:.*]] = vector.tuple_get %[[ES5]], 0 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R1S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG1]], %[[TG2]], %[[TG3]], %[[TG4]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG1]], %[[TG2]], %[[TG3]], %[[TG4]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG6:.*]] = vector.tuple_get %[[ES1]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG7:.*]] = vector.tuple_get %[[ES2]], 2 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG8:.*]] = vector.tuple_get %[[ES4]], 1 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[TG9:.*]] = vector.tuple_get %[[ES5]], 2 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R2S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG6]], %[[TG7]], %[[R1S00]], %[[TG8]], %[[TG9]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R2S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG6]], %[[TG7]], %[[R1S00]], %[[TG8]], %[[TG9]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG10:.*]] = vector.tuple_get %[[ES1]], 2 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG11:.*]] = vector.tuple_get %[[ES2]], 4 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG12:.*]] = vector.tuple_get %[[ES4]], 2 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[TG13:.*]] = vector.tuple_get %[[ES5]], 4 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R3S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG10]], %[[TG11]], %[[R2S00]], %[[TG12]], %[[TG13]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R3S00:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG10]], %[[TG11]], %[[R2S00]], %[[TG12]], %[[TG13]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [0, 2]
 
 // CHECK-NEXT: %[[TG14:.*]] = vector.tuple_get %[[ES2]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG15:.*]] = vector.tuple_get %[[ES3]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG16:.*]] = vector.tuple_get %[[ES5]], 1 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R1S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG1]], %[[TG14]], %[[TG15]], %[[TG4]], %[[TG16]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG1]], %[[TG14]], %[[TG15]], %[[TG4]], %[[TG16]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG17:.*]] = vector.tuple_get %[[ES2]], 3 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG18:.*]] = vector.tuple_get %[[ES5]], 3 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R2S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG6]], %[[TG17]], %[[R1S02]], %[[TG8]], %[[TG18]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R2S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG6]], %[[TG17]], %[[R1S02]], %[[TG8]], %[[TG18]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG19:.*]] = vector.tuple_get %[[ES2]], 5 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG20:.*]] = vector.tuple_get %[[ES5]], 5 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R3S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG10]], %[[TG19]], %[[R2S02]], %[[TG12]], %[[TG20]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R3S02:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG10]], %[[TG19]], %[[R2S02]], %[[TG12]], %[[TG20]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [2, 0]
 
 // CHECK-NEXT: %[[TG21:.*]] = vector.tuple_get %[[ES1]], 3 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG22:.*]] = vector.tuple_get %[[ES3]], 2 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG23:.*]] = vector.tuple_get %[[ES4]], 3 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT: %[[R1S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG21]], %[[TG2]], %[[TG22]], %[[TG23]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG21]], %[[TG2]], %[[TG22]], %[[TG23]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG24:.*]] = vector.tuple_get %[[ES1]], 4 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG25:.*]] = vector.tuple_get %[[ES4]], 4 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT:  %[[R2S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG24]], %[[TG7]], %[[R1S20]], %[[TG25]], %[[TG9]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R2S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG24]], %[[TG7]], %[[R1S20]], %[[TG25]], %[[TG9]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[TG26:.*]] = vector.tuple_get %[[ES1]], 5 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG27:.*]] = vector.tuple_get %[[ES4]], 5 : tuple<vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT:  %[[R3S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG26]], %[[TG11]], %[[R2S20]], %[[TG27]], %[[TG13]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R3S20:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG26]], %[[TG11]], %[[R2S20]], %[[TG27]], %[[TG13]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [2, 2]
 
 // CHECK-NEXT: %[[TG28:.*]] = vector.tuple_get %[[ES3]], 3 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
-// CHECK-NEXT: %[[R1S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG21]], %[[TG14]], %[[TG28]], %[[TG23]], %[[TG16]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R2S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG24]], %[[TG17]], %[[R1S22]], %[[TG25]], %[[TG18]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R3S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"]} %[[TG26]], %[[TG19]], %[[R2S22]], %[[TG27]], %[[TG20]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG21]], %[[TG14]], %[[TG28]], %[[TG23]], %[[TG16]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R2S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG24]], %[[TG17]], %[[R1S22]], %[[TG25]], %[[TG18]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R3S22:.*]] = vector.contract {indexing_maps = [#map0, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = "add"} %[[TG26]], %[[TG19]], %[[R2S22]], %[[TG27]], %[[TG20]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[RES0:.*]] = vector.tuple %[[R3S00]], %[[R3S02]], %[[R3S20]], %[[R3S22]] : vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[RES1:.*]] = vector.insert_slices %[[RES0]], [2, 2], [1, 1] : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>> into vector<4x4xf32>
 // CHECK-NEXT:  return %[[RES1]] : vector<4x4xf32>
 
 func @contraction4x4_ijk(%arg0 : vector<4x6xf32>, %arg1 : vector<6x4xf32>,
                          %arg2 : vector<4x4xf32>, %arg3 : index)
                          -> (vector<4x4xf32>) {

 // CHECK-NEXT: %[[ES4:.*]] = vector.extract_slices %{{.*}}, [2, 2], [1, 1] : vector<4x2xi1> into tuple<vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[ES5:.*]] = vector.extract_slices %{{.*}}, [2, 2], [1, 1] : vector<2x4xi1> into tuple<vector<2x2xi1>, vector<2x2xi1>>
 
 // CHECK-NEXT: %[[TG1:.*]] = vector.tuple_get %[[ES1]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG2:.*]] = vector.tuple_get %[[ES2]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG3:.*]] = vector.tuple_get %[[ES3]], 0 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG4:.*]] = vector.tuple_get %[[ES4]], 0 : tuple<vector<2x2xi1>, vector<2x2xi1>>
 // CHECK-NEXT: %[[TG5:.*]] = vector.tuple_get %[[ES5]], 0 : tuple<vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT:  %[[R1S00:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[TG1]], %[[TG2]], %[[TG3]], %[[TG4]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R1S00:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[TG1]], %[[TG2]], %[[TG3]], %[[TG4]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [0, 2]
 
 // CHECK-NEXT: %[[TG6:.*]] = vector.tuple_get %[[ES2]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG7:.*]] = vector.tuple_get %[[ES3]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG8:.*]] = vector.tuple_get %[[ES5]], 1 : tuple<vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT:  %[[R1S02:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[TG1]], %[[TG6]], %[[TG7]], %[[TG4]], %[[TG8]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R1S02:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[TG1]], %[[TG6]], %[[TG7]], %[[TG4]], %[[TG8]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [2, 0]
 
 // CHECK-NEXT: %[[TG9:.*]] = vector.tuple_get %[[ES1]], 1 : tuple<vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG10:.*]] = vector.tuple_get %[[ES3]], 2 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
 // CHECK-NEXT: %[[TG11:.*]] = vector.tuple_get %[[ES4]], 1 : tuple<vector<2x2xi1>, vector<2x2xi1>>
-// CHECK-NEXT:  %[[R1S20:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[TG9]], %[[TG2]], %[[TG10]], %[[TG11]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R1S20:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[TG9]], %[[TG2]], %[[TG10]], %[[TG11]], %[[TG5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // Reducing output vector [2, 2]
 
 // CHECK-NEXT: %[[TG12:.*]] = vector.tuple_get %[[ES3]], 3 : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>>
-// CHECK-NEXT:  %[[R1S22:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[TG9]], %[[TG6]], %[[TG12]], %[[TG11]], %[[TG8]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT:  %[[R1S22:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[TG9]], %[[TG6]], %[[TG12]], %[[TG11]], %[[TG8]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[RES0:.*]] = vector.tuple %[[R1S00]], %[[R1S02]], %[[R1S20]], %[[R1S22]] : vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[RES1:.*]] = vector.insert_slices %[[RES0]], [2, 2], [1, 1] : tuple<vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>, vector<2x2xf32>> into vector<4x4xf32>
 // CHECK-NEXT:  return %[[RES1]] : vector<4x4xf32>
 
 func @contraction4x4_ikj(%arg0 : vector<4x2xf32>, %arg1 : vector<2x4xf32>,
                          %arg2 : vector<4x4xf32>, %arg3 : index)
                          -> (vector<4x4xf32>) {

 // CHECK-NEXT: %[[VTR2:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %{{.*}} : memref<2x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR3:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C2]]], %{{.*}} : memref<2x4xf32>, vector<2x2xf32>
 
 // CHECK-NEXT: %[[VTR4:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %{{.*}} : memref<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR5:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C2]]], %{{.*}} : memref<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR6:.*]] = vector.transfer_read %{{.*}}[%[[C2]], %[[C0]]], %{{.*}} : memref<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR7:.*]] = vector.transfer_read %{{.*}}[%[[C2]], %[[C2]]], %{{.*}} : memref<4x4xf32>, vector<2x2xf32>
 
-// CHECK-NEXT: %[[R0:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR0]], %[[VTR2]], %[[VTR4]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R1:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR0]], %[[VTR3]], %[[VTR5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R2:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR1]], %[[VTR2]], %[[VTR6]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R3:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR1]], %[[VTR3]], %[[VTR7]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R0:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR0]], %[[VTR2]], %[[VTR4]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR0]], %[[VTR3]], %[[VTR5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R2:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR1]], %[[VTR2]], %[[VTR6]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R3:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR1]], %[[VTR3]], %[[VTR7]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: vector.transfer_write %[[R0]], %{{.*}}[%[[C0]], %[[C0]]] {masked = [false, false]} : vector<2x2xf32>, memref<4x4xf32>
 // CHECK-NEXT: vector.transfer_write %[[R1]], %{{.*}}[%[[C0]], %[[C2]]] {masked = [false, false]} : vector<2x2xf32>, memref<4x4xf32>
 // CHECK-NEXT: vector.transfer_write %[[R2]], %{{.*}}[%[[C2]], %[[C0]]] {masked = [false, false]} : vector<2x2xf32>, memref<4x4xf32>
 // CHECK-NEXT: vector.transfer_write %[[R3]], %{{.*}}[%[[C2]], %[[C2]]] {masked = [false, false]} : vector<2x2xf32>, memref<4x4xf32>
 // CHECK-NEXT: return
 
 func @contraction4x4_ikj_xfer_read(%arg0 : memref<4x2xf32>,

 // CHECK-NEXT: %[[VTR2:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %{{.*}} : tensor<2x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR3:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C2]]], %{{.*}} : tensor<2x4xf32>, vector<2x2xf32>
 
 // CHECK-NEXT: %[[VTR4:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %{{.*}} : tensor<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR5:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C2]]], %{{.*}} : tensor<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR6:.*]] = vector.transfer_read %{{.*}}[%[[C2]], %[[C0]]], %{{.*}} : tensor<4x4xf32>, vector<2x2xf32>
 // CHECK-NEXT: %[[VTR7:.*]] = vector.transfer_read %{{.*}}[%[[C2]], %[[C2]]], %{{.*}} : tensor<4x4xf32>, vector<2x2xf32>
 
-// CHECK-NEXT: %[[R0:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR0]], %[[VTR2]], %[[VTR4]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R1:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR0]], %[[VTR3]], %[[VTR5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R2:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR1]], %[[VTR2]], %[[VTR6]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-// CHECK-NEXT: %[[R3:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"]} %[[VTR1]], %[[VTR3]], %[[VTR7]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R0:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR0]], %[[VTR2]], %[[VTR4]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R1:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR0]], %[[VTR3]], %[[VTR5]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R2:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR1]], %[[VTR2]], %[[VTR6]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+// CHECK-NEXT: %[[R3:.*]] = vector.contract {indexing_maps = [#map2, #map3, #map0], iterator_types = ["parallel", "reduction", "parallel"], kind = "add"} %[[VTR1]], %[[VTR3]], %[[VTR7]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
 // CHECK-NEXT: %[[VTW0:.*]] = vector.transfer_write %[[R0]], %{{.*}}[%[[C0]], %[[C0]]] {masked = [false, false]} : vector<2x2xf32>, tensor<4x4xf32>
 // CHECK-NEXT: %[[VTW1:.*]] = vector.transfer_write %[[R1]], %[[VTW0]][%[[C0]], %[[C2]]] {masked = [false, false]} : vector<2x2xf32>, tensor<4x4xf32>
 // CHECK-NEXT: %[[VTW2:.*]] = vector.transfer_write %[[R2]], %[[VTW1]][%[[C2]], %[[C0]]] {masked = [false, false]} : vector<2x2xf32>, tensor<4x4xf32>
 // CHECK-NEXT: %[[VTW3:.*]] = vector.transfer_write %[[R3]], %[[VTW2]][%[[C2]], %[[C2]]] {masked = [false, false]} : vector<2x2xf32>, tensor<4x4xf32>
 // CHECK-NEXT: return %[[VTW3]] : tensor<4x4xf32>
 
 func @contraction4x4_ikj_xfer_read_tensor(%arg0 : tensor<4x2xf32>,