[mlir][Linalg] Further improve codegen strategy and add a linalg.matmul_i8_i8_i32

[nicolasvasilache]

Summary

This revision adds a layer of SFINAE to the composable codegen strategy so it does
not have to require statically defined ops but instead can also be used with OpInterfaces, Operation* and an op name string.

A linalg.matmul_i8_i8_i32 is added to the .tc spec to demonstrate how all this works end to end.

[pifon2a]

mlir/include/mlir/Dialect/Linalg/Transforms/CodegenStrategy.h

24

do you need this constructor?

66

I thought only 1 arg constructors are usually marked "explicit"

161

nit: use make_unique here and below :

transformationSequence.push_back(std::make_unique<Tile<LinalgOpType>>(options,f));

[nicolasvasilache]

Address review.

[nicolasvasilache]

Address review.

mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOpsSpec.tc

 ods_def<MatmulOp>:
 def matmul(A: f32(M, K), B: f32(K, N)) -> (C: f32(M, N)) {
   C(m, n) = std_addf<k>(std_mulf(A(m, k), B(k, n)));
 }
 
 ods_def<MatmulColumnMajorOp>:
 def matmul_column_major(A: f32(K, M), B: f32(N, K)) -> (C: f32(N, M)) {
   C(n, m) = std_addf<k>(std_mulf(A(k, m), B(n, k)));
 }
 
+ods_def<MatmulI8I8I32Op>:
+def matmul_i8_i8_i32(A: i8(M, K), B: i8(K, N)) -> (C: i32(M, N)) {
+  // TODO: ideally something closer to
+  //   C(m, n) += cast<i32>(A(m, k)) * cast<i32>(B(k, n))
+  C(m, n) = std_addi<k>(std_sexti32(std_muli(A(m, k), B(k, n))));
+}
+
 ods_def<MatvecOp>:
 def matvec(A: f32(M, N), y: f32(N)) -> (x: f32(M)) {
   x(m) = std_addf<n>(std_mulf(A(m, n), y(n)));
 }
 
 ods_def<VecmatOp>:
 def vecmat(y: f32(M), A: f32(M, N)) -> (x: f32(N)) {
   x(n) = std_addf<m>(std_mulf(y(m), A(m, n)));

mlir/include/mlir/Dialect/Linalg/Transforms/CodegenStrategy.h

 
 class FuncOp;
 
 namespace linalg {
 
 /// Abstract Transformation class applied in a sequence that also handles state
 /// through markers.
 struct Transformation {
+  Transformation() {}
+  Transformation(linalg::LinalgTransformationFilter::FilterFunction f)
+      : filter(f) {}
   virtual ~Transformation() = default;
   virtual OwningRewritePatternList
-  buildRewritePatterns(MLIRContext *context, linalg::LinalgMarker m) = 0;
-  linalg::LinalgMarker marker;
+  buildRewritePatterns(MLIRContext *context,
+                       linalg::LinalgTransformationFilter m) = 0;
+  linalg::LinalgTransformationFilter::FilterFunction filter = nullptr;
 };
 
+/// SFINAE: Enqueue helper for ConcreteOpType that have a `getOperationName`.
+template <template <typename> typename PatternType, typename ConcreteOpType,
+          typename OptionsType,
+          typename = decltype(ConcreteOpType::getOperationName)>
+void sfinae_enqueue(OwningRewritePatternList &patterList, OptionsType options,
+                    MLIRContext *context, StringRef opName,
+                    linalg::LinalgTransformationFilter m) {
+  assert(opName.empty() ||
+         opName == ConcreteOpType::getOperationName() &&
+             "explicit name must match ConcreteOpType::getOperationName");
+  patterList.insert<PatternType<ConcreteOpType>>(context, options, m);
+}
+
+/// SFINAE: Enqueue helper for OpType that do not have a `getOperationName`
+/// (e.g. LinalgOp, other interfaces, Operation*).
+template <template <typename> typename PatternType, typename OpType,
+          typename OptionsType>
+void sfinae_enqueue(OwningRewritePatternList &patterList, OptionsType options,
+                    MLIRContext *context, StringRef opName,
+                    linalg::LinalgTransformationFilter m) {
+  assert(!opName.empty() && "opName must not be empty");
+  patterList.insert<PatternType<OpType>>(opName, context, options, m);
+}
+
 /// Promotion transformation enqueues a particular stage-1 pattern for
 /// `Tile<LinalgOpType>`with the appropriate `options`.
 template <typename LinalgOpType>
 struct Tile : public Transformation {
-  explicit Tile(linalg::LinalgTilingOptions options) : options(options) {}
+  explicit Tile(linalg::LinalgTilingOptions options,
+                linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(""), options(options) {}
+
+  explicit Tile(StringRef name, linalg::LinalgTilingOptions options,
+                linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(name), options(options) {}
 
   OwningRewritePatternList
-  buildRewritePatterns(MLIRContext *context, linalg::LinalgMarker m) override {
+  buildRewritePatterns(MLIRContext *context,
+                       linalg::LinalgTransformationFilter m) override {
     OwningRewritePatternList tilingPatterns;
-    tilingPatterns.insert<linalg::LinalgTilingPattern<LinalgOpType>>(
-        context, options, m);
+    sfinae_enqueue<linalg::LinalgTilingPattern, LinalgOpType>(
+        tilingPatterns, options, context, opName, m);
     return tilingPatterns;
   }
 
 private:
+  std::string opName;
   linalg::LinalgTilingOptions options;
 };
 
 /// Promotion transformation enqueues a particular stage-1 pattern for
 /// `Promote<LinalgOpType>`with the appropriate `options`.
 template <typename LinalgOpType>
 struct Promote : public Transformation {
-  explicit Promote(linalg::LinalgPromotionOptions options) : options(options) {}
+  explicit Promote(
+      linalg::LinalgPromotionOptions options,
+      linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(""), options(options) {}
+
+  explicit Promote(
+      StringRef name, linalg::LinalgPromotionOptions options,
+      linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(name), options(options) {}
 
   OwningRewritePatternList
-  buildRewritePatterns(MLIRContext *context, linalg::LinalgMarker m) override {
+  buildRewritePatterns(MLIRContext *context,
+                       linalg::LinalgTransformationFilter m) override {
     OwningRewritePatternList promotionPatterns;
-    promotionPatterns.insert<linalg::LinalgPromotionPattern<LinalgOpType>>(
-        context, options, m);
+    sfinae_enqueue<linalg::LinalgPromotionPattern, LinalgOpType>(
+        promotionPatterns, options, context, opName, m);
     return promotionPatterns;
   }
 
 private:
+  std::string opName;
   linalg::LinalgPromotionOptions options;
 };
 
 /// Vectorization transformation enqueues a particular stage-1 pattern for
 /// `LinalgVectorizationPattern<LinalgOpType>` as well as copy to vector
 /// transfer rewrite forwarding patterns.
 template <typename LinalgOpType>
 struct Vectorize : public Transformation {
+  explicit Vectorize(
+      linalg::LinalgVectorizationOptions options,
+      linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(""), options(options) {}
+
+  explicit Vectorize(
+      StringRef name, linalg::LinalgVectorizationOptions options,
+      linalg::LinalgTransformationFilter::FilterFunction f = nullptr)
+      : Transformation(f), opName(name), options(options) {}
+
   OwningRewritePatternList
-  buildRewritePatterns(MLIRContext *context, linalg::LinalgMarker m) override {
+  buildRewritePatterns(MLIRContext *context,
+                       linalg::LinalgTransformationFilter m) override {
     OwningRewritePatternList vectorizationPatterns;
-    // FillOp may interfere with forwarding patterns atm, so we bump up the
-    // priority of LinalgCopyVTRForwardingPattern /
-    // LinalgCopyVTWForwardingPattern.
-    vectorizationPatterns
-        .insert<linalg::LinalgVectorizationPattern<LinalgOpType>>(context, m);
+    sfinae_enqueue<linalg::LinalgVectorizationPattern, LinalgOpType>(
+        vectorizationPatterns, options, context, opName, m);
     vectorizationPatterns.insert<linalg::LinalgCopyVTRForwardingPattern,
                                  linalg::LinalgCopyVTWForwardingPattern>(
         context, /*benefit=*/2);
     return vectorizationPatterns;
   }
+
+private:
+  std::string opName;
+  linalg::LinalgVectorizationOptions options;
 };
 
 /// Codegen strategy controls how a Linalg op is progressively lowered.
 /// The application uses a 3-level staged patterns strategy which allows
-/// ordering transformations by using the Linalg `applyStagedPatterns` function,
-/// where:
+/// ordering transformations by using the Linalg `applyStagedPatterns`
+/// function, where:
 ///   1. The first stage consists of the successive `tile`, `promote` and
 ///   `vectorize` patterns, applied sequentially.
 ///   2. The second stage consists of common local canonicalization patterns
 ///   that are applied eagerly after each stage-1 pattern.
 ///   3. the third stage consists of more global transformation, also applied
 ///   eagerly, after all stage-2 patterns. Such more global transformations
 struct CodegenStrategy {
   /// Append a pattern to add a level of tiling for `LinalgOpType` with tiling
   /// `options`.
   template <typename LinalgOpType>
-  CodegenStrategy &tile(linalg::LinalgTilingOptions options) {
-    transformationSequence.emplace_back(new Tile<LinalgOpType>(options));
+  CodegenStrategy &
+  tile(linalg::LinalgTilingOptions options,
+       linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(new Tile<LinalgOpType>(options, f));
+    return *this;
+  }
+  /// Append a pattern to add a level of tiling for `LinalgOpType` with tiling
+  /// `options`.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  tile(StringRef opName, linalg::LinalgTilingOptions options,
+       linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(
+        new Tile<LinalgOpType>(opName, options, f));
     return *this;
   }
-  /// Conditionally append a pattern to add a level of tiling for `LinalgOpType`
-  /// with tiling `options`.
+  /// Conditionally append a pattern to add a level of tiling for
+  /// `LinalgOpType` with tiling `options`.
   template <typename LinalgOpType>
-  CodegenStrategy &tileIf(bool b, linalg::LinalgTilingOptions options) {
+  CodegenStrategy &
+  tileIf(bool b, linalg::LinalgTilingOptions options,
+         linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
     return b ? tile<LinalgOpType>(options) : *this;
   }
+  /// Conditionally append a pattern to add a level of tiling for
+  /// `LinalgOpType` with tiling `options`.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  tileIf(bool b, StringRef opName, linalg::LinalgTilingOptions options,
+         linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    return b ? tile<LinalgOpType>(opName, options) : *this;
+  }
   /// Append a pattern to add a level of promotion for `LinalgOpType` with
   /// promotion `options`.
   template <typename LinalgOpType>
-  CodegenStrategy &promote(linalg::LinalgPromotionOptions options) {
-    transformationSequence.emplace_back(new Promote<LinalgOpType>(options));
+  CodegenStrategy &
+  promote(linalg::LinalgPromotionOptions options,
+          linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(new Promote<LinalgOpType>(options, f));
+    return *this;
+  }
+  /// Append a pattern to add a level of promotion for `LinalgOpType` with
+  /// promotion `options`.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  promote(StringRef opName, linalg::LinalgPromotionOptions options,
+          linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(
+        new Promote<LinalgOpType>(opName, options, f));
     return *this;
   }
   /// Conditionally append a pattern to add a level of promotion for
   /// `LinalgOpType` with promotion `options`.
   template <typename LinalgOpType>
-  CodegenStrategy &promoteIf(bool b, linalg::LinalgPromotionOptions options) {
-    return b ? promote<LinalgOpType>(options) : *this;
+  CodegenStrategy &
+  promoteIf(bool b, StringRef opName, linalg::LinalgPromotionOptions options,
+            linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    return b ? promote<LinalgOpType>(opName, options, f) : *this;
+    return *this;
+  }
+  /// Conditionally append a pattern to add a level of promotion for
+  /// `LinalgOpType` with promotion `options`.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  promoteIf(bool b, linalg::LinalgPromotionOptions options,
+            linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    return b ? promote<LinalgOpType>(options, f) : *this;
     return *this;
   }
   /// Append a pattern to rewrite `LinalgOpType` as a vector operation.
   template <typename LinalgOpType>
-  CodegenStrategy &vectorize() {
-    transformationSequence.emplace_back(new Vectorize<LinalgOpType>());
+  CodegenStrategy &
+  vectorize(linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(
+        new Vectorize<LinalgOpType>(linalg::LinalgVectorizationOptions(), f));
+    return *this;
+  }
+  /// Append a pattern to rewrite `LinalgOpType` as a vector operation.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  vectorize(StringRef opName,
+            linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    transformationSequence.emplace_back(new Vectorize<LinalgOpType>(
+        opName, linalg::LinalgVectorizationOptions(), f));
+    return *this;
+  }
+  /// Conditionally append a pattern to rewrite `LinalgOpType` as a vector
+  /// operation.
+  template <typename LinalgOpType>
+  CodegenStrategy &
+  vectorizeIf(bool b,
+              linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    return b ? vectorize<LinalgOpType>(f) : *this;
     return *this;
   }
   /// Conditionally append a pattern to rewrite `LinalgOpType` as a vector
   /// operation.
   template <typename LinalgOpType>
-  CodegenStrategy &vectorizeIf(bool b) {
-    return b ? vectorize<LinalgOpType>() : *this;
+  CodegenStrategy &
+  vectorizeIf(bool b, StringRef opName,
+              linalg::LinalgTransformationFilter::FilterFunction f = nullptr) {
+    return b ? vectorize<LinalgOpType>(opName, f) : *this;
     return *this;
   }
   /// Configure the post staged-patterns late vector transformations.
   CodegenStrategy &
   setVectorTransformsOptions(vector::VectorTransformsOptions options) {
     vectorTransformsOptions = options;
     return *this;
   }
-  /// Configure the post staged-patterns late vector.transfer to scf conversion.
+  /// Configure the post staged-patterns late vector.transfer to scf
+  /// conversion.
   CodegenStrategy &
   setVectorTransferToSCFOptions(VectorTransferToSCFOptions options) {
     vectorToSCFOptions = options;
     return *this;
   }
 
-  /// Apply the transformation patterns in sequence with cleanup transformations
-  /// interleaved.
+  /// Apply the transformation patterns in sequence with cleanup
+  /// transformations interleaved.
   void transform(FuncOp func) const;
 
 private:
   LogicalResult postPatternTransforms(Operation *func) const;
 
   vector::VectorTransformsOptions vectorTransformsOptions;
   VectorTransferToSCFOptions vectorToSCFOptions;
   SmallVector<std::unique_ptr<Transformation>, 4> transformationSequence;
 };
 
 } // namespace linalg
 } // namespace mlir
 
 #endif // MLIR_DIALECT_LINALG_TRANSFORMS_CODEGENSTRATEGY_H_

mlir/include/mlir/Dialect/Linalg/Transforms/Transforms.h

 //===----------------------------------------------------------------------===//
 // Transformations exposed as rewrite patterns.
 //===----------------------------------------------------------------------===//
 // Marker used as attribute name in generated Linalg rewriting transformations.
 struct LinalgTransforms {
   static const StringLiteral kLinalgTransformMarker;
 };
 
-/// Helper class to control common attribute matching and setting behavior.
-struct LinalgMarker {
-  explicit LinalgMarker(ArrayRef<Identifier> matchDisjunction = {},
-                        Optional<Identifier> replacement = None);
-  LinalgMarker(LinalgMarker &&) = default;
-  LinalgMarker(const LinalgMarker &) = default;
+/// Helper class to control application of linalg transformation patterns.
+/// Control comes in 2 forms:
+///   1. attribute matching and setting behavior using the attribute named
+///      `kLinalgTransformMarker`. This can be used to build a state machine
+///      using attributes and incrementally applying patterns to advance states.
+///   2. filter function, which is a simple lambda on the Operation* that
+///      returns a LogicalResult.
+struct LinalgTransformationFilter {
+  using FilterFunction = std::function<LogicalResult(Operation *)>;
+
+  explicit LinalgTransformationFilter(
+      ArrayRef<Identifier> matchDisjunction = {},
+      Optional<Identifier> replacement = None);
+
+  explicit LinalgTransformationFilter(
+      FilterFunction f, ArrayRef<Identifier> matchDisjunction = {},
+
+  LinalgTransformationFilter(LinalgTransformationFilter &&) = default;
+  LinalgTransformationFilter(const LinalgTransformationFilter &) = default;
   LogicalResult checkAndNotify(PatternRewriter &rewriter, Operation *op) const;
-  void replaceLinalgMarker(PatternRewriter &rewriter, Operation *op) const;
+  void replaceLinalgTransformationFilter(PatternRewriter &rewriter,
+                                         Operation *op) const;
 
 private:
+  FilterFunction filter;
   SmallVector<Identifier, 4> matchDisjunction;
   Optional<Identifier> replacement;
 };
 
 ///
 /// Linalg tiling patterns.
 ///
 /// Apply the `tileLinalgOp` transformation as a pattern.

 /// when tiling is called programmatically.
 OwningRewritePatternList
 getLinalgTilingCanonicalizationPatterns(MLIRContext *ctx);
 void populateLinalgTilingCanonicalizationPatterns(
     OwningRewritePatternList &patterns, MLIRContext *ctx);
 
 struct LinalgBaseTilingPattern : public RewritePattern {
   // Entry point to match any LinalgOp OpInterface.
-  LinalgBaseTilingPattern(LinalgTilingOptions options,
-                          LinalgMarker marker = LinalgMarker(),
-                          PatternBenefit benefit = 1);
+  LinalgBaseTilingPattern(
+      LinalgTilingOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
   // Entry point to match a specific Linalg op.
-  LinalgBaseTilingPattern(StringRef opName, MLIRContext *context,
-                          LinalgTilingOptions options,
-                          LinalgMarker marker = LinalgMarker(),
-                          PatternBenefit benefit = 1);
+  LinalgBaseTilingPattern(
+      StringRef opName, MLIRContext *context, LinalgTilingOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
   LogicalResult matchAndRewriteBase(Operation *op, PatternRewriter &rewriter,
                                     TiledLinalgOp &result) const;
 
 private:
   /// LinalgTransformMarker handles special attribute manipulations.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
   /// Options to control tiling;
   LinalgTilingOptions options;
 };
 
 template <typename OpTy>
 struct LinalgTilingPattern : public LinalgBaseTilingPattern {
-  LinalgTilingPattern(MLIRContext *context, LinalgTilingOptions options,
-                      LinalgMarker marker = LinalgMarker(),
-                      PatternBenefit benefit = 1)
-      : LinalgBaseTilingPattern(OpTy::getOperationName(), context, options,
-                                marker, benefit) {}
+  /// SFINAE: This constructor can only trigger for concrete ops that have a
+  /// static `getOperationName` method.
+  template <typename ConcreateOpTy = OpTy>
+  LinalgTilingPattern(
+      MLIRContext *context, LinalgTilingOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      : LinalgBaseTilingPattern(ConcreateOpTy::getOperationName(), context,
+                                options, marker, benefit) {}
+
+  /// This constructor is available to anyone.
+  LinalgTilingPattern(
+      StringRef opName, MLIRContext *context, LinalgTilingOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      PatternBenefit benefit = 1)
+      : LinalgBaseTilingPattern(opName, context, options, marker, benefit) {}
+
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override {
     TiledLinalgOp tiledLinalgOp;
     if (failed(LinalgBaseTilingPattern::matchAndRewriteBase(op, rewriter,
                                                             tiledLinalgOp)))
       return failure();
     if (tiledLinalgOp.tensorResults.empty())
       rewriter.eraseOp(op);
     else
       rewriter.replaceOp(op, tiledLinalgOp.tensorResults);
     return success();
   }
 };
 
 struct LinalgFusionOptions {
   /// List of operands indices to use for fusion.
   llvm::SmallSet<unsigned, 1> indicesToFuse = {};
   LinalgFusionOptions &setIndicesToFuse(ArrayRef<int64_t> operands) {
     indicesToFuse.insert(operands.begin(), operands.end());
     return *this;
   }
 };
 
 struct LinalgBaseTileAndFusePattern : public RewritePattern {
-  LinalgBaseTileAndFusePattern(StringRef opName, MLIRContext *context,
-                               const LinalgDependenceGraph &dependenceGraph,
-                               LinalgTilingOptions tilingOptions,
-                               LinalgFusionOptions fusionOptions,
-                               LinalgMarker marker = LinalgMarker(),
-                               LinalgMarker fusedOpMarker = LinalgMarker(),
-                               LinalgMarker originalOpMarker = LinalgMarker(),
-                               PatternBenefit benefit = 1);
+  LinalgBaseTileAndFusePattern(
+      StringRef opName, MLIRContext *context,
+      LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      LinalgTransformationFilter fusedOpMarker = LinalgTransformationFilter(),
+      LinalgTransformationFilter originalOpMarker =
+          LinalgTransformationFilter(),
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override;
 
 private:
   /// Dependence graph needed for fusion.
   const LinalgDependenceGraph &dependenceGraph;
   /// Options to control tiling.
   LinalgTilingOptions tilingOptions;
   /// Options to control fusion.
   LinalgFusionOptions fusionOptions;
   /// Marker to control application of the pattern.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
   /// Marker set on the fused op after tile and fuse.
-  LinalgMarker fusedOpMarker;
+  LinalgTransformationFilter fusedOpMarker;
   /// The dependenceGraph is not modifiable, i.e. if the Linalg operations used
   /// to build the dependence graph changes then the dependenceGraph needs to be
   /// recomputed right now. To not invalidate the dependenceGraph as
   /// transformation happens, the original producer can be tagged with a marker
   /// that can be later used to delete the original operations.
-  LinalgMarker originalOpMarker;
+  LinalgTransformationFilter originalOpMarker;
 };
 
 template <typename OpTy>
 struct LinalgTileAndFusePattern : public LinalgBaseTileAndFusePattern {
-  LinalgTileAndFusePattern(MLIRContext *context,
-                           const LinalgDependenceGraph &dependenceGraph,
-                           LinalgTilingOptions tilingOptions,
-                           LinalgFusionOptions fusionOptions,
-                           LinalgMarker marker = LinalgMarker(),
-                           LinalgMarker fusedOpMarker = LinalgMarker(),
-                           LinalgMarker originalOpMarker = LinalgMarker(),
-                           PatternBenefit benefit = 1)
+  LinalgTileAndFusePattern(
+      MLIRContext *context, const LinalgDependenceGraph &dependenceGraph,
+      LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      LinalgTransformationFilter fusedOpMarker = LinalgTransformationFilter(),
+      LinalgTransformationFilter originalOpMarker =
+          LinalgTransformationFilter(),
       : LinalgBaseTileAndFusePattern(
             OpTy::getOperationName(), context, dependenceGraph, tilingOptions,
             fusionOptions, marker, fusedOpMarker, originalOpMarker, benefit) {}
 };
 
 ///
 /// Linalg interchange patterns.
 ///
 /// Apply the `interchange` transformation as a pattern.
 /// `marker` controls LinalgTransformMarker matching and update when specified.
 /// See `interchange` for more details.
 struct LinalgBaseInterchangePattern : public RewritePattern {
-  LinalgBaseInterchangePattern(StringRef opName, MLIRContext *context,
-                               ArrayRef<unsigned> interchangeVector,
-                               LinalgMarker marker = LinalgMarker(),
-                               PatternBenefit benefit = 1);
+  LinalgBaseInterchangePattern(
+      StringRef opName, MLIRContext *context,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override;
 
 private:
   /// LinalgTransformMarker handles special attribute manipulations.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
   /// The interchange vector to reorder the iterators and indexing_maps dims.
   SmallVector<unsigned, 8> interchangeVector;
 };
 
 template <typename OpTy>
 struct LinalgInterchangePattern : public LinalgBaseInterchangePattern {
-  LinalgInterchangePattern(MLIRContext *context,
-                           ArrayRef<unsigned> interchangeVector,
-                           LinalgMarker marker = LinalgMarker(),
-                           PatternBenefit benefit = 1)
+  LinalgInterchangePattern(
+      MLIRContext *context, ArrayRef<unsigned> interchangeVector,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
       : LinalgBaseInterchangePattern(OpTy::getOperationName(), context,
                                      interchangeVector, marker, benefit) {}
 };
 
 ///
 /// Linalg promotion patterns.
 ///
 /// Apply the `promoteSubViews` transformation as a pattern.
 /// `marker` controls LinalgTransformMarker matching and update when specified.
 /// See `promoteSubViews` for more details.
 struct LinalgBasePromotionPattern : public RewritePattern {
-  LinalgBasePromotionPattern(StringRef opName, MLIRContext *context,
-                             LinalgPromotionOptions options,
-                             LinalgMarker marker = LinalgMarker(),
-                             PatternBenefit benefit = 1);
+  LinalgBasePromotionPattern(
+      StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override;
 
 private:
   /// LinalgTransformMarker handles special attribute manipulations.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
   /// Promotion options.
   LinalgPromotionOptions options;
 };
 
 template <typename OpTy>
 struct LinalgPromotionPattern : public LinalgBasePromotionPattern {
-  LinalgPromotionPattern(MLIRContext *context, LinalgPromotionOptions options,
-                         LinalgMarker marker = LinalgMarker(),
-                         PatternBenefit benefit = 1)
+  /// SFINAE: This constructor can only trigger for concrete ops that have a
+  /// static `getOperationName` method.
+  template <typename ConcreateOpTy = OpTy>
+  LinalgPromotionPattern(
+      MLIRContext *context, LinalgPromotionOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
       : LinalgBasePromotionPattern(OpTy::getOperationName(), context, options,
                                    marker, benefit) {}
+  /// This constructor is available to anyone.
+  LinalgPromotionPattern(
+      StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      PatternBenefit benefit = 1)
+      : LinalgBasePromotionPattern(opName, context, options, marker, benefit) {}
 };
 
 ///
 /// Linalg vectorization patterns.
 ///
 /// Apply the `vectorizeLinalgOp` transformation as a pattern.
 /// `marker` controls LinalgTransformMarker matching and update when specified.
 /// See `vectorizeLinalgOp` for more details.
+
+/// Empty for now, used for SFINAE purposes only.
+struct LinalgVectorizationOptions {};
+
 struct LinalgBaseVectorizationPattern : public RewritePattern {
-  LinalgBaseVectorizationPattern(StringRef opName, MLIRContext *context,
-                                 LinalgMarker marker = LinalgMarker(),
-                                 PatternBenefit benefit = 1);
+  LinalgBaseVectorizationPattern(
+      StringRef opName, MLIRContext *context,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override;
 
 private:
   /// LinalgTransformMarker handles special attribute manipulations.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
 };
 
 template <typename OpTy>
 struct LinalgVectorizationPattern : public LinalgBaseVectorizationPattern {
-  LinalgVectorizationPattern(MLIRContext *context,
-                             LinalgMarker marker = LinalgMarker(),
-                             PatternBenefit benefit = 1)
+  /// SFINAE: This constructor can only trigger for concrete ops that have a
+  /// static `getOperationName` method.
+  template <typename ConcreateOpTy = OpTy>
+  LinalgVectorizationPattern(
+      MLIRContext *context,
+      LinalgVectorizationOptions options = LinalgVectorizationOptions(),
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
       : LinalgBaseVectorizationPattern(OpTy::getOperationName(), context,
                                        marker, benefit) {}
+  /// This constructor is available to anyone.
+  LinalgVectorizationPattern(
+      StringRef opName, MLIRContext *context,
+      LinalgVectorizationOptions options = LinalgVectorizationOptions(),
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      PatternBenefit benefit = 1)
+      : LinalgBaseVectorizationPattern(opName, context, marker, benefit) {}
 };
 
 ///
 /// Linalg lowering patterns.
 ///
 /// Apply the `linalgLowerOpToLoops` transformation as a pattern.
 /// `marker` controls LinalgTransformMarker matching and update when specified.
 /// See `linalgLowerOpToLoops` for more details.
 enum class LinalgLoweringType {
   LibraryCall = 0,
   Loops = 1,
   AffineLoops = 2,
   ParallelLoops = 3
 };
 
 template <typename OpTy>
 struct LinalgLoweringPattern : public RewritePattern {
-  LinalgLoweringPattern(MLIRContext *context, LinalgLoweringType loweringType,
-                        LinalgMarker marker = LinalgMarker(),
-                        ArrayRef<unsigned> interchangeVector = {},
-                        PatternBenefit benefit = 1)
+  LinalgLoweringPattern(
+      MLIRContext *context, LinalgLoweringType loweringType,
+      LinalgTransformationFilter marker = LinalgTransformationFilter(),
+      ArrayRef<unsigned> interchangeVector = {}, PatternBenefit benefit = 1)
       : RewritePattern(OpTy::getOperationName(), {}, benefit, context),
         marker(marker), loweringType(loweringType),
         interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
 
   // TODO: Move implementation to .cpp once named ops are auto-generated.
   LogicalResult matchAndRewrite(Operation *op,
                                 PatternRewriter &rewriter) const override {
     LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
LogicalResult matchAndRewrite(Operation *op,
     }
 
     rewriter.eraseOp(op);
     return success();
   }
 
 private:
   /// LinalgTransformMarker handles special attribute manipulations.
-  LinalgMarker marker;
+  LinalgTransformationFilter marker;
   /// Controls whether the pattern lowers to library calls, scf.for, affine.for
   /// or scf.parallel.
   LinalgLoweringType loweringType;
   /// Permutated loop order in the generated loop nest.
   SmallVector<unsigned, 4> interchangeVector;
 };
 
 /// Linalg generalization patterns
 
 /// Populates `patterns` with patterns to convert spec-generated named ops to
 /// linalg.generic ops.
 void populateLinalgNamedOpsGeneralizationPatterns(
     MLIRContext *context, OwningRewritePatternList &patterns,
-    LinalgMarker marker = LinalgMarker());
+    LinalgTransformationFilter marker = LinalgTransformationFilter());
 
 /// Populates `patterns` with patterns to convert linalg.conv ops to
 /// linalg.generic ops.
 void populateLinalgConvGeneralizationPatterns(
     MLIRContext *context, OwningRewritePatternList &patterns,
-    LinalgMarker marker = LinalgMarker());
+    LinalgTransformationFilter marker = LinalgTransformationFilter());
 
 //===----------------------------------------------------------------------===//
 // Op-specific patterns.
 //===----------------------------------------------------------------------===//
 /// Match and rewrite for the pattern:
 /// ```
 ///    %alloc = ...
 ///    [optional] %view = std.view %alloc ...

mlir/include/mlir/Dialect/StandardOps/EDSC/Intrinsics.h

 using std_index_cast = ValueBuilder<IndexCastOp>;
 using std_muli = ValueBuilder<MulIOp>;
 using std_mulf = ValueBuilder<MulFOp>;
 using std_memref_cast = ValueBuilder<MemRefCastOp>;
 using std_ret = OperationBuilder<ReturnOp>;
 using std_rsqrt = ValueBuilder<RsqrtOp>;
 using std_select = ValueBuilder<SelectOp>;
 using std_load = ValueBuilder<LoadOp>;
+using std_sign_extendi = ValueBuilder<SignExtendIOp>;
 using std_splat = ValueBuilder<SplatOp>;
 using std_store = OperationBuilder<StoreOp>;
 using std_subf = ValueBuilder<SubFOp>;
 using std_subi = ValueBuilder<SubIOp>;
 using std_sub_view = ValueBuilder<SubViewOp>;
 using std_tanh = ValueBuilder<TanhOp>;
 using std_tensor_load = ValueBuilder<TensorLoadOp>;
 using std_tensor_store = OperationBuilder<TensorStoreOp>;
 using std_view = ValueBuilder<ViewOp>;
 using std_zero_extendi = ValueBuilder<ZeroExtendIOp>;
-using std_sign_extendi = ValueBuilder<SignExtendIOp>;
 using tensor_extract = ValueBuilder<tensor::ExtractOp>;
 
+template <int N>
+struct SExtiValueBuilder : public ValueBuilder<SignExtendIOp> {
+  using ValueBuilder<SignExtendIOp>::ValueBuilder;
+  template <typename... Args>
+  SExtiValueBuilder(Args... args)
+      : ValueBuilder<SignExtendIOp>(ScopedContext::getBuilderRef().getI32Type(),
+                                    args...) {}
+};
+
+using std_sexti32 = SExtiValueBuilder<32>;
+
 /// Branches into `block` with `operands`.
 BranchOp std_br(Block *block, ValueRange operands);
 
 /// Branches into `trueBranch` with `trueOperands` if `cond` evaluates to `true`
 /// or to `falseBranch` and `falseOperand` if `cond` evaluates to `false`.
 CondBranchOp std_cond_br(Value cond, Block *trueBranch, ValueRange trueOperands,
                          Block *falseBranch, ValueRange falseOperands);
 
 /// Provide an index notation around sdt_load and std_store.
 using StdIndexedValue =
     TemplatedIndexedValue<intrinsics::std_load, intrinsics::std_store>;
 } // namespace intrinsics
 } // namespace edsc
 } // namespace mlir
 
 #endif // MLIR_DIALECT_STANDARDOPS_EDSC_INTRINSICS_H_

mlir/integration_test/Dialect/Linalg/CPU/benchmark_matmul.mlir

 // RUN: export M=24 && export K=64 && export N=192 && export ITERS=10 && \
 // RUN: cat %s | sed 's@${M}@'"$M"'@g'| sed 's@${K}@'"$K"'@g' | sed 's@${N}@'"$N"'@g'| sed 's@${ITERS}@'"$ITERS"'@g'| \
 // RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.matmul register-tile-sizes=12,32,16 vectorize" | \
 // RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.fill register-tile-sizes=4,32 vectorize" | \
 // RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.copy register-tile-sizes=4,32 vectorize" | \
 
 // RUN: mlir-opt -canonicalize -convert-vector-to-scf -lower-affine -convert-linalg-to-loops | \
 // RUN: mlir-opt -canonicalize -convert-scf-to-std -convert-vector-to-llvm | \
 // RUN: mlir-cpu-runner -O3 -e main -entry-point-result=void \
 // Activate to dump assembly
 // R_UN:   -dump-object-file -object-filename=/tmp/a.o \
 // RUN:   -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext | \
 // Use tee to both print to stderr and FileCheck
 // RUN: tee -a /dev/stderr | FileCheck %s
 
 
-!row_major_A = type memref<${M}x${K}xf32>
-!row_major_B = type memref<${K}x${N}xf32>
-!row_major_C = type memref<${M}x${N}xf32>
+!elem_type_a = f32
+!elem_type_b = f32
+!elem_type_c = f32
+!row_major_A = type memref<${M}x${K}x!elem_type_a>
+!row_major_B = type memref<${K}x${N}x!elem_type_b>
+!row_major_C = type memref<${M}x${N}x!elem_type_c>
 
 func @matmul(%a: !row_major_A, %b: !row_major_B, %c: !row_major_C)
 // TODO: activate manually for now.
 // attributes { passthrough = [["target-cpu", "skylake-avx512"], ["prefer-vector-width", "512"]]}
 {
   linalg.matmul ins(%a, %b : !row_major_A, !row_major_B)
     outs(%c: !row_major_C)
   return
func @print_perf(%iters: index, %total_time: f64) {
   %flops_f = sitofp %flops_i64 : i64 to f64
   %flops_per_s = divf %flops_f, %total_time : f64
   vector.print %flops_per_s : f64
 
   return
 }
 
 func @main() {
-  %f0 = constant 0.0 : f32
-  %f1 = constant 1.0 : f32
+  %v0 = constant 0.0 : !elem_type_a
+  %v1 = constant 1.0 : !elem_type_a
 
   %A = alloc() : !row_major_A
   %B = alloc() : !row_major_B
   %C = alloc() : !row_major_C
 
-  linalg.fill(%A, %f1) : !row_major_A, f32
-  linalg.fill(%B, %f1) : !row_major_B, f32
-  linalg.fill(%C, %f0) : !row_major_C, f32
+  linalg.fill(%A, %v1) : !row_major_A, !elem_type_a
+  linalg.fill(%B, %v1) : !row_major_B, !elem_type_b
+  linalg.fill(%C, %v0) : !row_major_C, !elem_type_c
 
   %c0 = constant 0: index
   %c1 = constant 1: index
   %iters = constant ${ITERS}: index
 
   /// Run and dump performance for matmul.
   /// Preheating run:
   scf.for %arg0 = %c0 to %iters step %c1 {
-    linalg.fill(%C, %f0) : !row_major_C, f32
+    %z = constant 0.0 : !elem_type_c
+    linalg.fill(%C, %z) : !row_major_C, !elem_type_c
     call @matmul(%A, %B, %C) : (!row_major_A, !row_major_B, !row_major_C) -> ()
   }
   %t_start_matmul = call @rtclock() : () -> f64
   scf.for %arg0 = %c0 to %iters step %c1 {
     // linalg.matmul writes %C in place, need to reset it to zero every time.
     // This is accounts for about 10-15% perf hit on small sizes.
     // Once linalg on tensors is ready, fusing fill at teh register level will
     // be easy.
-    linalg.fill(%C, %f0) : !row_major_C, f32
+    %z = constant 0.0 : !elem_type_c
+    linalg.fill(%C, %z) : !row_major_C, !elem_type_c
     call @matmul(%A, %B, %C) : (!row_major_A, !row_major_B, !row_major_C) -> ()
   }
   %t_end_matmul = call @rtclock() : () -> f64
   %tmatmul = subf %t_end_matmul, %t_start_matmul: f64
   call @print_perf(%iters, %tmatmul) : (index, f64) -> ()
 
   %res = load %C[%c0, %c0]: !row_major_C
   // CHECK: 64

mlir/integration_test/Dialect/Linalg/CPU/benchmark_matmul_column_major.mlir

 // RUN: mlir-opt -canonicalize -convert-scf-to-std -convert-vector-to-llvm | \
 // RUN: mlir-cpu-runner -O3 -e main -entry-point-result=void \
 // Activate to dump assembly
 // R_UN:   -dump-object-file -object-filename=/tmp/a.o \
 // RUN:   -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext | \
 // Use tee to both print to stderr and FileCheck
 // RUN: tee -a /dev/stderr | FileCheck %s
 
-!row_major_A = type memref<${M}x${K}xf32>
-!row_major_B = type memref<${K}x${N}xf32>
-!row_major_C = type memref<${M}x${N}xf32>
-!column_major_A = type memref<${K}x${M}xf32>
-!column_major_B = type memref<${N}x${K}xf32>
-!column_major_C = type memref<${N}x${M}xf32>
+!elem_type_a = f32
+!elem_type_b = f32
+!elem_type_c = f32
+!row_major_A = type memref<${M}x${K}x!elem_type_a>
+!row_major_B = type memref<${K}x${N}x!elem_type_b>
+!row_major_C = type memref<${M}x${N}x!elem_type_c>
+!column_major_A = type memref<${K}x${M}x!elem_type_a>
+!column_major_B = type memref<${N}x${K}x!elem_type_b>
+!column_major_C = type memref<${N}x${M}x!elem_type_c>
 
 func @matmul_column_major(%a: !column_major_A, %b: !column_major_B, %c: !column_major_C)
 // TODO: activate manually for now.
 // attributes { passthrough = [["target-cpu", "skylake-avx512"], ["prefer-vector-width", "512"]]}
 {
   linalg.matmul_column_major ins(%a, %b : !column_major_A, !column_major_B)
     outs(%c: !column_major_C)
   return
func @print_perf(%iters: index, %total_time: f64) {
   %flops_f = sitofp %flops_i64 : i64 to f64
   %flops_per_s = divf %flops_f, %total_time : f64
   vector.print %flops_per_s : f64
 
   return
 }
 
 func @main() {
-  %f0 = constant 0.0 : f32
-  %f1 = constant 1.0 : f32
+  %f0 = constant 0.0 : !elem_type_c
+  %f1 = constant 1.0 : !elem_type_a
 
   %cA = alloc() : !column_major_A
   %cB = alloc() : !column_major_B
   %cC = alloc() : !column_major_C
 
-  linalg.fill(%cA, %f1) : !column_major_A, f32
-  linalg.fill(%cB, %f1) : !column_major_B, f32
-  linalg.fill(%cC, %f0) : !column_major_C, f32
+  linalg.fill(%cA, %f1) : !column_major_A, !elem_type_a
+  linalg.fill(%cB, %f1) : !column_major_B, !elem_type_b
+  linalg.fill(%cC, %f0) : !column_major_C, !elem_type_c
 
   %c0 = constant 0: index
   %c1 = constant 1: index
   %iters = constant ${ITERS}: index
 
   /// Run and dump performance for matmul_column_major.
   %t_start_matmul_column_major = call @rtclock() : () -> f64
   scf.for %arg0 = %c0 to %iters step %c1 {
     // linalg.matmul writes %C in place, need to reset it to zero every time.
     // This is accounts for about 10-15% perf hit on small sizes.
     // Once linalg on tensors is ready, fusing fill at teh register level will
     // be easy.
-    linalg.fill(%cC, %f0) : !column_major_C, f32
+    linalg.fill(%cC, %f0) : !column_major_C, !elem_type_c
     call @matmul_column_major(%cA, %cB, %cC) : (!column_major_A, !column_major_B, !column_major_C) -> ()
   }
   %t_end_matmul_column_major = call @rtclock() : () -> f64
   %tmatmul_column_major = subf %t_end_matmul_column_major, %t_start_matmul_column_major: f64
   call @print_perf(%iters, %tmatmul_column_major) : (index, f64) -> ()
 
   %res = load %cC[%c0, %c0]: !column_major_C
   // CHECK: 64
-  vector.print %res: f32
+  vector.print %res: !elem_type_c
 
   dealloc %cA : !column_major_A
   dealloc %cB : !column_major_B
   dealloc %cC : !column_major_C
 
   return
 }
 
 func private @rtclock() -> f64
 
 // TODO: init with random, run and check output.
 // func private @fill_random_f32(memref<*xf32>)

mlir/integration_test/Dialect/Linalg/CPU/benchmark_matmul_column_major_as_row_major.mlir

 // RUN: mlir-opt -canonicalize -convert-scf-to-std -convert-vector-to-llvm | \
 // RUN: mlir-cpu-runner -O3 -e main -entry-point-result=void \
 // Activate to dump assembly
 // R_UN:   -dump-object-file -object-filename=/tmp/a.o \
 // RUN:   -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext | \
 // Use tee to both print to stderr and FileCheck
 // RUN: tee -a /dev/stderr | FileCheck %s
 
-!row_major_A = type memref<${M}x${K}xf32>
-!row_major_B = type memref<${K}x${N}xf32>
-!row_major_C = type memref<${M}x${N}xf32>
-!column_major_A = type memref<${K}x${M}xf32>
-!column_major_B = type memref<${N}x${K}xf32>
-!column_major_C = type memref<${N}x${M}xf32>
+!elem_type_a = f32
+!elem_type_b = f32
+!elem_type_c = f32
+!row_major_A = type memref<${M}x${K}x!elem_type_a>
+!row_major_B = type memref<${K}x${N}x!elem_type_b>
+!row_major_C = type memref<${M}x${N}x!elem_type_c>
+!column_major_A = type memref<${K}x${M}x!elem_type_a>
+!column_major_B = type memref<${N}x${K}x!elem_type_b>
+!column_major_C = type memref<${N}x${M}x!elem_type_c>
 
 func @matmul_column_major_as_row_major(
   %ca: !column_major_A, %cb: !column_major_B, %cc: !column_major_C,
    %a: !row_major_A,     %b: !row_major_B,     %c: !row_major_C)
 // TODO: activate manually for now.
 // attributes { passthrough = [["target-cpu", "skylake-avx512"], ["prefer-vector-width", "512"]]}
 {
   linalg.copy(%ca, %a) {inputPermutation = affine_map<(i, j) -> (j, i)> } : !column_major_A, !row_major_A
func @print_perf(%iters: index, %total_time: f64) {
   %flops_f = sitofp %flops_i64 : i64 to f64
   %flops_per_s = divf %flops_f, %total_time : f64
   vector.print %flops_per_s : f64
 
   return
 }
 
 func @main() {
-  %f0 = constant 0.0 : f32
-  %f1 = constant 1.0 : f32
+  %f0 = constant 0.0 : !elem_type_c
+  %f1 = constant 1.0 : !elem_type_a
 
   %cA = alloc() : !column_major_A
   %cB = alloc() : !column_major_B
   %cC = alloc() : !column_major_C
 
-  linalg.fill(%cA, %f1) : !column_major_A, f32
-  linalg.fill(%cB, %f1) : !column_major_B, f32
-  linalg.fill(%cC, %f0) : !column_major_C, f32
+  linalg.fill(%cA, %f1) : !column_major_A, !elem_type_a
+  linalg.fill(%cB, %f1) : !column_major_B, !elem_type_b
+  linalg.fill(%cC, %f0) : !column_major_C, !elem_type_c
 
   %c0 = constant 0: index
   %c1 = constant 1: index
   %iters = constant ${ITERS}: index
 
   /// Run and dump performance for matmul_column_major as a row-major
   %A = alloc() : !row_major_A
   %B = alloc() : !row_major_B
   %C = alloc() : !row_major_C
   %t_start_matmul_column_major_as_row_major = call @rtclock() : () -> f64
   scf.for %arg0 = %c0 to %iters step %c1 {
     // linalg.matmul writes %C in place, need to reset it to zero every time.
     // This is accounts for about 10-15% perf hit on small sizes.
     // Once linalg on tensors is ready, fusing fill at teh register level will
     // be easy.
-    linalg.fill(%C, %f0) : !row_major_C, f32
+    linalg.fill(%C, %f0) : !row_major_C, !elem_type_c
     call @matmul_column_major_as_row_major(%cA, %cB, %cC, %A, %B, %C) :
       (!column_major_A, !column_major_B, !column_major_C,
        !row_major_A, !row_major_B, !row_major_C) -> ()
   }
   %t_end_matmul_column_major_as_row_major = call @rtclock() : () -> f64
   %tmatmul_column_major_as_row_major = subf %t_end_matmul_column_major_as_row_major, %t_start_matmul_column_major_as_row_major: f64
   call @print_perf(%iters, %tmatmul_column_major_as_row_major) : (index, f64) -> ()
 
   %res = load %cC[%c0, %c0]: !column_major_C
   // CHECK: 64
-  vector.print %res: f32
+  vector.print %res: !elem_type_c
   %res2 = load %C[%c0, %c0]: !row_major_C
   // CHECK: 64
-  vector.print %res2: f32
+  vector.print %res2: !elem_type_c
 
   dealloc %A : !row_major_A
   dealloc %B : !row_major_B
   dealloc %C : !row_major_C
 
   dealloc %cA : !column_major_A
   dealloc %cB : !column_major_B
   dealloc %cC : !column_major_C
 
   return
 }
 
 func private @rtclock() -> f64
 
 // TODO: init with random, run and check output.
 // func private @fill_random_f32(memref<*xf32>)

mlir/integration_test/Dialect/Linalg/CPU/benchmark_matmul_i8_i8_i32.mlir

 // RUN: export M=24 && export K=64 && export N=192 && export ITERS=10 && \
 // RUN: cat %s | sed 's@${M}@'"$M"'@g'| sed 's@${K}@'"$K"'@g' | sed 's@${N}@'"$N"'@g'| sed 's@${ITERS}@'"$ITERS"'@g'| \
-// RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.matmul register-tile-sizes=12,32,16 vectorize" | \
+// TODO: extend vectorization with interfaces so that it works with sexti
+// RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.matmul_i8_i8_i32 register-tile-sizes=12,32,16" | \
 // RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.fill register-tile-sizes=4,32 vectorize" | \
 // RUN: mlir-opt -test-linalg-codegen-strategy="anchor-op=linalg.copy register-tile-sizes=4,32 vectorize" | \
-
 // RUN: mlir-opt -canonicalize -convert-vector-to-scf -lower-affine -convert-linalg-to-loops | \
+
 // RUN: mlir-opt -canonicalize -convert-scf-to-std -convert-vector-to-llvm | \
 // RUN: mlir-cpu-runner -O3 -e main -entry-point-result=void \
 // Activate to dump assembly
 // R_UN:   -dump-object-file -object-filename=/tmp/a.o \
 // RUN:   -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext | \
 // Use tee to both print to stderr and FileCheck
 // RUN: tee -a /dev/stderr | FileCheck %s
 
 
-!row_major_A = type memref<${M}x${K}xf32>
-!row_major_B = type memref<${K}x${N}xf32>
-!row_major_C = type memref<${M}x${N}xf32>
+!elem_type_a = type i8
+!elem_type_b = type i8
+!elem_type_c = type i32
+!row_major_A = type memref<${M}x${K}x!elem_type_a>
+!row_major_B = type memref<${K}x${N}x!elem_type_b>
+!row_major_C = type memref<${M}x${N}x!elem_type_c>
 
 func @matmul(%a: !row_major_A, %b: !row_major_B, %c: !row_major_C)
 // TODO: activate manually for now.
 // attributes { passthrough = [["target-cpu", "skylake-avx512"], ["prefer-vector-width", "512"]]}
 {
-  linalg.matmul ins(%a, %b : !row_major_A, !row_major_B)
+  linalg.matmul_i8_i8_i32 ins(%a, %b : !row_major_A, !row_major_B)
     outs(%c: !row_major_C)
   return
 }
 
 func @print_perf(%iters: index, %total_time: f64) {
   %c2 = constant 2 : index
   %cM = constant ${M} : index
   %cN = constant ${N} : index
func @print_perf(%iters: index, %total_time: f64) {
   %flops_f = sitofp %flops_i64 : i64 to f64
   %flops_per_s = divf %flops_f, %total_time : f64
   vector.print %flops_per_s : f64
 
   return
 }
 
 func @main() {
-  %f0 = constant 0.0 : f32
-  %f1 = constant 1.0 : f32
+  %v0 = constant 0 : !elem_type_c
+  %v1 = constant 1 : !elem_type_a
 
   %A = alloc() : !row_major_A
   %B = alloc() : !row_major_B
   %C = alloc() : !row_major_C
 
-  linalg.fill(%A, %f1) : !row_major_A, f32
-  linalg.fill(%B, %f1) : !row_major_B, f32
-  linalg.fill(%C, %f0) : !row_major_C, f32
+  linalg.fill(%A, %v1) : !row_major_A, !elem_type_a
+  linalg.fill(%B, %v1) : !row_major_B, !elem_type_b
+  linalg.fill(%C, %v0) : !row_major_C, !elem_type_c
 
   %c0 = constant 0: index
   %c1 = constant 1: index
   %iters = constant ${ITERS}: index
 
   /// Run and dump performance for matmul.
   /// Preheating run:
   scf.for %arg0 = %c0 to %iters step %c1 {
-    linalg.fill(%C, %f0) : !row_major_C, f32
+    linalg.fill(%C, %v0) : !row_major_C, !elem_type_c
     call @matmul(%A, %B, %C) : (!row_major_A, !row_major_B, !row_major_C) -> ()
   }
   %t_start_matmul = call @rtclock() : () -> f64
   scf.for %arg0 = %c0 to %iters step %c1 {
     // linalg.matmul writes %C in place, need to reset it to zero every time.
     // This is accounts for about 10-15% perf hit on small sizes.
     // Once linalg on tensors is ready, fusing fill at teh register level will
     // be easy.
-    linalg.fill(%C, %f0) : !row_major_C, f32
+    linalg.fill(%C, %v0) : !row_major_C, !elem_type_c
     call @matmul(%A, %B, %C) : (!row_major_A, !row_major_B, !row_major_C) -> ()
   }
   %t_end_matmul = call @rtclock() : () -> f64
   %tmatmul = subf %t_end_matmul, %t_start_matmul: f64
   call @print_perf(%iters, %tmatmul) : (index, f64) -> ()
 
   %res = load %C[%c0, %c0]: !row_major_C
   // CHECK: 64

mlir/lib/Dialect/Linalg/Transforms/CodegenStrategy.cpp

void mlir::linalg::CodegenStrategy::transform(FuncOp func) const {
   // state transition.
   unsigned stepCount = 0;
   SmallVector<FrozenRewritePatternList, 4> stage1Patterns;
   auto zeroState = Identifier::get(std::to_string(stepCount), context);
   auto currentState = zeroState;
   for (const std::unique_ptr<Transformation> &t : transformationSequence) {
     auto nextState = Identifier::get(std::to_string(++stepCount), context);
     auto marker = (currentState == zeroState)
-                      ? linalg::LinalgMarker({}, nextState)
-                      : linalg::LinalgMarker(currentState, nextState);
+                      ? linalg::LinalgTransformationFilter(
+                            t->filter, ArrayRef<Identifier>{}, nextState)
+                      : linalg::LinalgTransformationFilter(
+                            t->filter, currentState, nextState);
     stage1Patterns.emplace_back(t->buildRewritePatterns(context, marker));
     currentState = nextState;
   }
 
   OwningRewritePatternList stage2Patterns =
       linalg::getLinalgTilingCanonicalizationPatterns(context);
   stage2Patterns.insert<AffineMinSCFCanonicalizationPattern>(context);
 

mlir/lib/Dialect/Linalg/Transforms/Generalization.cpp

 /// Base class for all linalg generalization patterns. A subclass must provide
 /// the following method:
 ///   linalg::GenericOp createGenericOp(RootOp, PatternRewriter &)
 /// for creating the generic op.
 // TODO: remove this pattern after migrating all manually-written named ops
 // into auto-generated ones.
 template <typename ConcretePattern, typename RootOp>
 struct LinalgGeneralizationPattern : OpRewritePattern<RootOp> {
-  LinalgGeneralizationPattern(MLIRContext *context, linalg::LinalgMarker marker,
+  LinalgGeneralizationPattern(MLIRContext *context,
+                              linalg::LinalgTransformationFilter marker,
                               PatternBenefit benefit = 1)
       : OpRewritePattern<RootOp>(context, benefit), marker(std::move(marker)) {}
 
   LogicalResult matchAndRewrite(RootOp rootOp,
                                 PatternRewriter &rewriter) const override {
     auto linalgOp = dyn_cast<linalg::LinalgOp>(rootOp.getOperation());
     if (!linalgOp)
       return failure();
     if (failed(marker.checkAndNotify(rewriter, linalgOp)))
       return failure();
 
     auto *pattern = static_cast<const ConcretePattern *>(this);
     linalg::GenericOp genericOp = pattern->createGenericOp(rootOp, rewriter);
     if (!genericOp)
       return failure();
 
     rewriter.replaceOp(rootOp, genericOp.getResults());
-    marker.replaceLinalgMarker(rewriter, genericOp.getOperation());
+    marker.replaceLinalgTransformationFilter(rewriter,
+                                             genericOp.getOperation());
     return success();
   }
 
 private:
-  linalg::LinalgMarker marker;
+  linalg::LinalgTransformationFilter marker;
 };
 
 struct GeneralizeConvOp
     : public LinalgGeneralizationPattern<GeneralizeConvOp, linalg::ConvOp> {
   using LinalgGeneralizationPattern::LinalgGeneralizationPattern;
 
   linalg::GenericOp createGenericOp(linalg::ConvOp, OpBuilder &rewriter) const;
 };
 
 /// Catch-all pattern for converting all named ops with a region builder into
 /// linalg.generic.
 struct LinalgNamedOpGeneralizationPattern : RewritePattern {
   LinalgNamedOpGeneralizationPattern(MLIRContext *context,
-                                     linalg::LinalgMarker marker,
+                                     linalg::LinalgTransformationFilter marker,
                                      PatternBenefit benefit = 1)
       : RewritePattern(benefit, MatchAnyOpTypeTag()),
         marker(std::move(marker)) {}
 
   LogicalResult matchAndRewrite(Operation *rootOp,
                                 PatternRewriter &rewriter) const override {
     auto linalgOp = dyn_cast<linalg::LinalgOp>(rootOp);
     if (!linalgOp)
       return failure();
     if (failed(marker.checkAndNotify(rewriter, linalgOp)))
       return failure();
 
     // No nothing to do for linalg.generic and linalg.indexed_generic.
     if (isa<linalg::GenericOp, linalg::IndexedGenericOp>(rootOp))
       return failure();
 
     linalg::GenericOp genericOp =
         createGenericOpFromNamedOp(linalgOp, rewriter);
     if (!genericOp)
       return failure();
 
     rewriter.replaceOp(rootOp, genericOp.getResults());
-    marker.replaceLinalgMarker(rewriter, genericOp.getOperation());
+    marker.replaceLinalgTransformationFilter(rewriter,
+                                             genericOp.getOperation());
     return success();
   }
 
 private:
-  linalg::LinalgMarker marker;
+  linalg::LinalgTransformationFilter marker;
 };
 
 struct LinalgGeneralizationPass
     : public LinalgGeneralizationBase<LinalgGeneralizationPass> {
   void runOnFunction() override;
 };
 
 } // namespace
return builder.create<linalg::GenericOp>(
             bodyBuilder.create<MulFOp>(bodyLoc, bodyArgs[0], bodyArgs[1]);
         Value add = bodyBuilder.create<AddFOp>(bodyLoc, mul, bodyArgs[2]);
         bodyBuilder.create<linalg::YieldOp>(bodyLoc, add);
       });
 }
 
 void mlir::linalg::populateLinalgConvGeneralizationPatterns(
     MLIRContext *context, OwningRewritePatternList &patterns,
-    linalg::LinalgMarker marker) {
+    linalg::LinalgTransformationFilter marker) {
   patterns.insert<GeneralizeConvOp>(context, marker);
 }
 
 void mlir::linalg::populateLinalgNamedOpsGeneralizationPatterns(
     MLIRContext *context, OwningRewritePatternList &patterns,
-    linalg::LinalgMarker marker) {
+    linalg::LinalgTransformationFilter marker) {
   patterns.insert<LinalgNamedOpGeneralizationPattern>(context, marker);
 }
 
 std::unique_ptr<OperationPass<FuncOp>> mlir::createLinalgGeneralizationPass() {
   return std::make_unique<LinalgGeneralizationPass>();
 }

mlir/lib/Dialect/Linalg/Transforms/Tiling.cpp

 };
 
 template <typename OpTy, typename... OpTypes>
 class RewritePatternList<OpTy, OpTypes...> {
 public:
   static void insert(OwningRewritePatternList &patterns,
                      const LinalgTilingOptions &options, MLIRContext *ctx) {
     patterns.insert<LinalgTilingPattern<OpTy>>(
-        ctx, options, LinalgMarker({}, Identifier::get("tiled", ctx)));
+        ctx, options,
+        LinalgTransformationFilter(ArrayRef<Identifier>{},
+                                   Identifier::get("tiled", ctx)));
     RewritePatternList<OpTypes...>::insert(patterns, options, ctx);
   }
 };
 } // namespace
 
 OwningRewritePatternList
 mlir::linalg::getLinalgTilingCanonicalizationPatterns(MLIRContext *ctx) {
   OwningRewritePatternList patterns;

mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp

 
 //===----------------------------------------------------------------------===//
 // Transformations exposed as rewrite patterns.
 //===----------------------------------------------------------------------===//
 // Marker used as attribute name in generated Linalg rewriting transformations.
 const StringLiteral mlir::linalg::LinalgTransforms::kLinalgTransformMarker =
     "__internal_linalg_transform__";
 
-mlir::linalg::LinalgMarker::LinalgMarker(ArrayRef<Identifier> matchDisjunction,
-                                         Optional<Identifier> replacement)
-    : matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()),
+mlir::linalg::LinalgTransformationFilter::LinalgTransformationFilter(
+    ArrayRef<Identifier> matchDisjunction, Optional<Identifier> replacement)
+    : LinalgTransformationFilter([](Operation *) { return success(); },
+                                 matchDisjunction, replacement) {}
+
+mlir::linalg::LinalgTransformationFilter::LinalgTransformationFilter(
+    FilterFunction f, ArrayRef<Identifier> matchDisjunction,
+    : filter(f),
+      matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()),
       replacement(replacement) {}
 
-LogicalResult
-mlir::linalg::LinalgMarker::checkAndNotify(PatternRewriter &rewriter,
-                                           Operation *op) const {
+LogicalResult mlir::linalg::LinalgTransformationFilter::checkAndNotify(
+    PatternRewriter &rewriter, Operation *op) const {
+  if (filter && failed(filter(op)))
+    return failure();
+
   auto attr = op->template getAttrOfType<StringAttr>(
       LinalgTransforms::kLinalgTransformMarker);
 
   if (!attr) {
     // 1. Has no marker case and matchDisjunction is empty.
     if (matchDisjunction.empty())
       return success();
 
LogicalResult mlir::linalg::LinalgTransformationFilter::checkAndNotify(
 
   // 5. Fail to match.
   return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
     diag << " does not have any marker from list: ";
     interleaveComma(matchDisjunction, diag);
   });
 }
 
-void mlir::linalg::LinalgMarker::replaceLinalgMarker(PatternRewriter &rewriter,
-                                                     Operation *op) const {
+void mlir::linalg::LinalgTransformationFilter::
+    replaceLinalgTransformationFilter(PatternRewriter &rewriter,
   if (replacement.hasValue())
     op->setAttr(LinalgTransforms::kLinalgTransformMarker,
                 rewriter.getStringAttr(replacement.getValue()));
   else
     op->removeAttr(Identifier::get(LinalgTransforms::kLinalgTransformMarker,
                                    rewriter.getContext()));
 }
 
tileSizeComputationFunction = [tileSizes](OpBuilder &b, Operation *op) {
     }));
   };
   return *this;
 }
 
 /// Linalg base tiling pattern.
 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
     StringRef opName, MLIRContext *context, LinalgTilingOptions options,
-    LinalgMarker marker, PatternBenefit benefit)
+    LinalgTransformationFilter marker, PatternBenefit benefit)
     : RewritePattern(opName, {}, benefit, context), marker(marker),
       options(options) {}
 
 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
-    LinalgTilingOptions options, LinalgMarker marker, PatternBenefit benefit)
+    LinalgTilingOptions options, LinalgTransformationFilter marker,
+    PatternBenefit benefit)
     : RewritePattern(benefit, MatchAnyOpTypeTag()), marker(marker),
       options(options) {}
 
 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewriteBase(
     Operation *op, PatternRewriter &rewriter, TiledLinalgOp &result) const {
   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
   if (!linalgOp)
     return failure();
   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
     return failure();
 
   Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options);
 
   if (!res)
     return failure();
 
   // Return relevant information to derived pattern.
   result = *res;
 
   // New marker if specified.
-  marker.replaceLinalgMarker(rewriter, res->op.getOperation());
+  marker.replaceLinalgTransformationFilter(rewriter, res->op.getOperation());
   return success();
 }
 
 mlir::linalg::LinalgBaseTileAndFusePattern::LinalgBaseTileAndFusePattern(
     StringRef opName, MLIRContext *context,
     const LinalgDependenceGraph &dependenceGraph,
     LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
-    LinalgMarker marker, LinalgMarker fusedOpMarker,
-    LinalgMarker originalOpMarker, PatternBenefit benefit)
+    LinalgTransformationFilter marker, LinalgTransformationFilter fusedOpMarker,
+    LinalgTransformationFilter originalOpMarker, PatternBenefit benefit)
     : RewritePattern(opName, {}, benefit, context),
       dependenceGraph(dependenceGraph), tilingOptions(tilingOptions),
       fusionOptions(fusionOptions), marker(marker),
       fusedOpMarker(fusedOpMarker), originalOpMarker(originalOpMarker) {}
 
 LogicalResult mlir::linalg::LinalgBaseTileAndFusePattern::matchAndRewrite(
     Operation *op, PatternRewriter &rewriter) const {
   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
if (llvm::any_of(unfusedLoopTileSizes, [](Value val) {
     Optional<TiledLinalgOp> unfusedTiledOp =
         tileLinalgOp(rewriter, tiledAndFusedOps->op, unfusedTilingOptions);
     if (!unfusedTiledOp)
       return failure();
     rewriter.eraseOp(tiledAndFusedOps->op);
     tiledAndFusedOps->op = unfusedTiledOp->op;
   }
 
-  marker.replaceLinalgMarker(rewriter, tiledAndFusedOps->op.getOperation());
+  marker.replaceLinalgTransformationFilter(rewriter,
+                                           tiledAndFusedOps->op.getOperation());
   for (auto fusedOp : tiledAndFusedOps->fusedProducers) {
-    fusedOpMarker.replaceLinalgMarker(rewriter, fusedOp.getOperation());
+    fusedOpMarker.replaceLinalgTransformationFilter(rewriter,
+                                                    fusedOp.getOperation());
   }
   for (auto origProducerOp : ArrayRef<LinalgOp>(fusionOps).drop_back()) {
-    originalOpMarker.replaceLinalgMarker(rewriter,
-                                         origProducerOp.getOperation());
+    originalOpMarker.replaceLinalgTransformationFilter(
+        rewriter, origProducerOp.getOperation());
   }
-  rewriter.updateRootInPlace(
-      op, [&]() { originalOpMarker.replaceLinalgMarker(rewriter, op); });
+  rewriter.updateRootInPlace(op, [&]() {
+    originalOpMarker.replaceLinalgTransformationFilter(rewriter, op);
+  });
   return success();
 }
 
 /// Linalg base interchange pattern.
 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern(
     StringRef opName, MLIRContext *context,
-    ArrayRef<unsigned> interchangeVector, LinalgMarker marker,
+    ArrayRef<unsigned> interchangeVector, LinalgTransformationFilter marker,
     PatternBenefit benefit)
     : RewritePattern(opName, {}, benefit, context), marker(marker),
       interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
 
 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite(
     Operation *op, PatternRewriter &rewriter) const {
   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
   if (!linalgOp)
     return failure();
   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
     return failure();
   if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector)))
     return failure();
 
   // TODO: figure out how this interplays with named ops. In particular this
   // should break the named op property.
   rewriter.updateRootInPlace(op, [&]() {
     interchange(linalgOp, interchangeVector);
     // New marker if specified.
-    marker.replaceLinalgMarker(rewriter, op);
+    marker.replaceLinalgTransformationFilter(rewriter, op);
   });
   return success();
 }
 
 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern(
     StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
-    LinalgMarker marker, PatternBenefit benefit)
+    LinalgTransformationFilter marker, PatternBenefit benefit)
     : RewritePattern(opName, {}, benefit, context), marker(marker),
       options(options) {}
 
 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite(
     Operation *op, PatternRewriter &rewriter) const {
   if (failed(marker.checkAndNotify(rewriter, op)))
     return failure();
   if (failed(promoteSubviewsPrecondition(op, options)))
     return failure();
 
   // TODO: We cannot use root update here. This pattern is creating other ops,
   // so if the promotion fails, those need to be cleaned up, which doesnt seem
   // to be happening here. So to fail properly, we should be cloning the op and
   // deleting the previous op. This needs more investigation.
   rewriter.startRootUpdate(op);
   Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options);
   if (!promotedOp) {
     rewriter.cancelRootUpdate(op);
     return op->emitError("subview promotion failed");
   }
   rewriter.finalizeRootUpdate(op);
-  marker.replaceLinalgMarker(rewriter, op);
+  marker.replaceLinalgTransformationFilter(rewriter, op);
   return success();
 }
 
 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern(
-    StringRef opName, MLIRContext *context, LinalgMarker marker,
+    StringRef opName, MLIRContext *context, LinalgTransformationFilter marker,
     PatternBenefit benefit)
     : RewritePattern(opName, {}, benefit, context), marker(marker) {}
 
 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite(
     Operation *op, PatternRewriter &rewriter) const {
   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
   if (!linalgOp)
     return failure();

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

populateVectorizationPatterns(OwningRewritePatternList &tilingPatterns,
                               MLIRContext *context) {
   if (tileSizes.size() < N)
     return;
 
   constexpr static StringRef kTiledMarker = "TILED";
   constexpr static StringRef kPromotedMarker = "PROMOTED";
   tilingPatterns.insert<LinalgTilingPattern<ConvOp>>(
       context, LinalgTilingOptions().setTileSizes(tileSizes),
-      LinalgMarker({}, Identifier::get(kTiledMarker, context)));
+      LinalgTransformationFilter(ArrayRef<Identifier>{},
+                                 Identifier::get(kTiledMarker, context)));
 
   promotionPatterns.insert<LinalgPromotionPattern<ConvOp>>(
       context, LinalgPromotionOptions().setUseFullTileBuffersByDefault(true),
-      LinalgMarker(Identifier::get(kTiledMarker, context),
+      LinalgTransformationFilter(Identifier::get(kTiledMarker, context),
                                  Identifier::get(kPromotedMarker, context)));
 
   SmallVector<bool, 4> mask(N);
   int offset = tileSizes.size() - N;
   std::transform(tileSizes.begin() + offset, tileSizes.end(), mask.begin(),
                  [](int64_t i) -> bool { return i > 1; });
 
   vectorizationPatterns.insert<ConvOpVectorization<ConvOp, N>>(context, mask);
 }

mlir/test/lib/Transforms/TestLinalgCodegenStrategy.cpp

registry.insert<AffineDialect,
                     gpu::GPUDialect,
                     linalg::LinalgDialect,
                     scf::SCFDialect,
                     StandardOpsDialect,
                     vector::VectorDialect>();
     // clang-format on
   }
 
-  template <typename LinalgNamedOp> void applyStrategyToNamedLinalgOp();
-
   void runOnFunction() override;
 
   ListOption<int64_t> tileSizes{*this, "tile-sizes",
                                 llvm::cl::MiscFlags::CommaSeparated,
                                 llvm::cl::desc("Specifies the tile sizes.")};
   Option<bool> promote{
       *this, "promote",
       llvm::cl::desc("Promote the tile into a small aligned memory buffer."),
Option<std::string> anchorOpName{
           "\tlinalg.matmul: anchor on linalg.matmul\n"
           "\tlinalg.matmul_column_major: anchor on linalg.matmul_column_major\n"
           "\tlinalg.copy: anchor on linalg.copy\n"
           "\tlinalg.fill: anchor on linalg.fill\n"),
       llvm::cl::init("")};
 };
 } // end anonymous namespace
 
-template <typename LinalgNamedOp>
-void TestLinalgCodegenStrategy::applyStrategyToNamedLinalgOp() {
+/// Apply transformations specified as patterns.
+void TestLinalgCodegenStrategy::runOnFunction() {
+  linalg::LinalgTransformationFilter::FilterFunction filterOpName =
+      [&](Operation *op) -> LogicalResult {
+    return success(op->getName().getStringRef() == anchorOpName);
+  };
   LinalgTilingOptions tilingOptions;
   if (!tileSizes.empty())
     tilingOptions = tilingOptions.setTileSizes(tileSizes);
 
   LinalgTilingOptions registerTilingOptions;
   if (!registerTileSizes.empty())
     registerTilingOptions =
         registerTilingOptions.setTileSizes(registerTileSizes);
vector::VectorTransferSplit vectorTransferSplit =
       llvm::StringSwitch<vector::VectorTransferSplit>(
           splitVectorTransfersTo.getValue())
           .Case("none", vector::VectorTransferSplit::None)
           .Case("linalg-copy", vector::VectorTransferSplit::LinalgCopy)
           .Case("vector-transfers", vector::VectorTransferSplit::VectorTransfer)
           .Default(vector::VectorTransferSplit::None);
 
   CodegenStrategy strategy;
-  strategy.template tileIf<LinalgNamedOp>(!tileSizes.empty(), tilingOptions)
-      .template promoteIf<LinalgNamedOp>(
-          promote, LinalgPromotionOptions()
+  strategy.tileIf<LinalgOp>(!tileSizes.empty(), anchorOpName, tilingOptions)
+      .promoteIf<LinalgOp>(promote, anchorOpName,
+                           LinalgPromotionOptions()
                                .setAlignment(16)
-                       .setUseFullTileBuffersByDefault(promoteFullTile))
-      .template tileIf<LinalgNamedOp>(!registerTileSizes.empty(),
-                                      registerTilingOptions)
-      .template promoteIf<LinalgNamedOp>(
-          registerPromote,
+                               .setUseFullTileBuffersByDefault(promoteFullTile),
+                           filterOpName)
+      .tileIf<LinalgOp>(!registerTileSizes.empty(), anchorOpName,
+      .promoteIf<LinalgOp>(
+          registerPromote, anchorOpName,
           LinalgPromotionOptions()
               .setAlignment(16)
               .setUseFullTileBuffersByDefault(registerPromoteFullTile))
-      .template vectorizeIf<LinalgNamedOp>(vectorize)
+      .vectorizeIf<LinalgOp>(vectorize, anchorOpName)
       .setVectorTransformsOptions(
           vector::VectorTransformsOptions()
               .setVectorTransformsOptions(vectorContractLowering)
               .setVectorTransferSplit(vectorTransferSplit))
       .setVectorTransferToSCFOptions(
           VectorTransferToSCFOptions().setUnroll(unrollVectorTransfers));
   strategy.transform(getFunction());
 }
 
-/// Apply transformations specified as patterns.
-void TestLinalgCodegenStrategy::runOnFunction() {
-  if (anchorOpName == MatmulOp::getOperationName())
-    applyStrategyToNamedLinalgOp<MatmulOp>();
-  else if (anchorOpName == MatmulColumnMajorOp::getOperationName())
-    applyStrategyToNamedLinalgOp<MatmulColumnMajorOp>();
-  else if (anchorOpName == CopyOp::getOperationName())
-    applyStrategyToNamedLinalgOp<CopyOp>();
-  else if (anchorOpName == FillOp::getOperationName())
-    applyStrategyToNamedLinalgOp<FillOp>();
-  else
-    llvm_unreachable("Unsupported anchor op");
-}
-
 namespace mlir {
 namespace test {
 void registerTestLinalgCodegenStrategy() {
   PassRegistration<TestLinalgCodegenStrategy> testLinalgCodegenStrategyPass(
       "test-linalg-codegen-strategy", "Test Linalg Codegen Strategy.");
 }
 } // namespace test
 } // namespace mlir

mlir/test/lib/Transforms/TestLinalgFusionTransforms.cpp

static void fillFusionPatterns(MLIRContext *context,
                                OwningRewritePatternList &patterns) {
   patterns.insert<LinalgTileAndFusePattern<MatmulOp>,
                   LinalgTileAndFusePattern<ConvOp>>(
       context, dependenceGraph,
       LinalgTilingOptions()
           .setTileSizes({32, 64, 16})
           .setLoopType(LinalgTilingLoopType::ParallelLoops),
       LinalgFusionOptions().setIndicesToFuse({2}),
-      LinalgMarker(Identifier::get("basic_fusion", context),
-                   Identifier::get("after_basic_fusion", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_basic_fusion_producer", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_basic_fusion_original", context)));
+      LinalgTransformationFilter(
+          Identifier::get("basic_fusion", context),
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
 
   patterns.insert<LinalgTileAndFusePattern<MatmulOp>>(
       context, dependenceGraph,
       LinalgTilingOptions()
           .setTileSizes({32, 64, 16})
           .setLoopType(LinalgTilingLoopType::ParallelLoops),
       LinalgFusionOptions().setIndicesToFuse({0}),
-      LinalgMarker(Identifier::get("lhs_fusion", context),
+      LinalgTransformationFilter(Identifier::get("lhs_fusion", context),
                                  Identifier::get("after_lhs_fusion", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_lhs_fusion_producer", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_lhs_fusion_original", context)));
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
 
   patterns.insert<LinalgTileAndFusePattern<MatmulOp>>(
       context, dependenceGraph,
       LinalgTilingOptions()
           .setTileSizes({32, 64, 16})
           .setLoopType(LinalgTilingLoopType::ParallelLoops),
       LinalgFusionOptions().setIndicesToFuse({1}),
-      LinalgMarker(Identifier::get("rhs_fusion", context),
+      LinalgTransformationFilter(Identifier::get("rhs_fusion", context),
                                  Identifier::get("after_rhs_fusion", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_rhs_fusion_producer", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_rhs_fusion_original", context)));
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
 
   patterns.insert<LinalgTileAndFusePattern<MatmulOp>>(
       context, dependenceGraph,
       LinalgTilingOptions()
           .setTileSizes({32, 64, 16})
           .setLoopType(LinalgTilingLoopType::ParallelLoops),
       LinalgFusionOptions().setIndicesToFuse({0, 2}),
-      LinalgMarker(Identifier::get("two_operand_fusion", context),
-                   Identifier::get("after_two_operand_fusion", context)),
-      LinalgMarker(
+      LinalgTransformationFilter(
+          Identifier::get("two_operand_fusion", context),
+      LinalgTransformationFilter(
           ArrayRef<Identifier>(),
           Identifier::get("after_two_operand_fusion_producer", context)),
-      LinalgMarker(
+      LinalgTransformationFilter(
           ArrayRef<Identifier>(),
           Identifier::get("after_two_operand_fusion_original", context)));
 
   patterns.insert<LinalgTileAndFusePattern<GenericOp>>(
       context, dependenceGraph,
       LinalgTilingOptions().setTileSizes({32, 64}).setLoopType(
           LinalgTilingLoopType::ParallelLoops),
       LinalgFusionOptions().setIndicesToFuse({0, 1}),
-      LinalgMarker(Identifier::get("transpose_fusion", context),
-                   Identifier::get("after_transpose_fusion", context)),
-      LinalgMarker(ArrayRef<Identifier>(),
-                   Identifier::get("after_transpose_fusion_producer", context)),
-      LinalgMarker(
+      LinalgTransformationFilter(
+          Identifier::get("transpose_fusion", context),
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>(),
+      LinalgTransformationFilter(
           ArrayRef<Identifier>(),
           Identifier::get("after_transpose_fusion_original", context)));
 }
 
 static void applyFusionPatterns(MLIRContext *context, FuncOp funcOp) {
   OwningRewritePatternList fusionPatterns;
   Aliases alias;
   LinalgDependenceGraph dependenceGraph =

mlir/test/lib/Transforms/TestLinalgTransforms.cpp

static void applyPatterns(FuncOp funcOp) {
   MLIRContext *ctx = funcOp.getContext();
   OwningRewritePatternList patterns;
 
   //===--------------------------------------------------------------------===//
   // Linalg tiling patterns.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({2000, 3000, 4000}),
-      LinalgMarker(Identifier::get("MEM", ctx), Identifier::get("L3", ctx)));
+      LinalgTransformationFilter(Identifier::get("MEM", ctx),
+                                 Identifier::get("L3", ctx)));
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({200, 300, 400}),
-      LinalgMarker(Identifier::get("L3", ctx), Identifier::get("L2", ctx)));
+      LinalgTransformationFilter(Identifier::get("L3", ctx),
+                                 Identifier::get("L2", ctx)));
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({20, 30, 40}),
-      LinalgMarker(Identifier::get("L2", ctx), Identifier::get("L1", ctx)));
+      LinalgTransformationFilter(Identifier::get("L2", ctx),
+                                 Identifier::get("L1", ctx)));
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({2, 3, 4}),
-      LinalgMarker(Identifier::get("L1", ctx), Identifier::get("REG", ctx)));
+      LinalgTransformationFilter(Identifier::get("L1", ctx),
+                                 Identifier::get("REG", ctx)));
 
   patterns.insert<LinalgTilingPattern<MatvecOp>>(
       ctx,
       LinalgTilingOptions().setTileSizes({5, 6}).setLoopType(
           LinalgTilingLoopType::ParallelLoops),
-      LinalgMarker({}, Identifier::get("L1", ctx)));
+      LinalgTransformationFilter(ArrayRef<Identifier>{},
+                                 Identifier::get("L1", ctx)));
 
   patterns.insert<LinalgTilingPattern<DotOp>>(
       ctx, LinalgTilingOptions().setTileSizes(8000),
-      LinalgMarker(ArrayRef<Identifier>{Identifier::get("MEM", ctx),
-                                        Identifier::get("L3", ctx),
-                                        Identifier::get("L2", ctx)},
-                   Identifier::get("REG", ctx)));
+      LinalgTransformationFilter(
+          ArrayRef<Identifier>{Identifier::get("MEM", ctx),
 
   //===--------------------------------------------------------------------===//
   // Linalg tiling and permutation patterns.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx,
       LinalgTilingOptions()
           .setTileSizes({2000, 3000, 4000})
           .setInterchange({1, 2, 0}),
-      LinalgMarker(Identifier::get("__with_perm__", ctx),
+      LinalgTransformationFilter(Identifier::get("__with_perm__", ctx),
                                  Identifier::get("L2__with_perm__", ctx)));
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx,
       LinalgTilingOptions()
           .setTileSizes({200, 300, 400})
           .setInterchange({1, 0, 2}),
-      LinalgMarker(Identifier::get("L2__with_perm__", ctx),
+      LinalgTransformationFilter(Identifier::get("L2__with_perm__", ctx),
                                  Identifier::get("L1__with_perm__", ctx)));
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({20, 30, 40}),
-      LinalgMarker(Identifier::get("L1__with_perm__", ctx),
+      LinalgTransformationFilter(Identifier::get("L1__with_perm__", ctx),
                                  Identifier::get("REG__with_perm__", ctx)));
 
   patterns.insert<LinalgTilingPattern<MatvecOp>>(
       ctx, LinalgTilingOptions().setTileSizes({5, 6}).setInterchange({1, 0}),
-      LinalgMarker(Identifier::get("__with_perm__", ctx),
+      LinalgTransformationFilter(Identifier::get("__with_perm__", ctx),
                                  Identifier::get("L1__with_perm__", ctx)));
 
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx,
       LinalgTilingOptions()
           .setTileSizes({16, 8, 4})
           .setInterchange({1, 2, 0})
           .setLoopType(LinalgTilingLoopType::ParallelLoops),
-      LinalgMarker(Identifier::get("par__with_perm__", ctx),
-                   Identifier::get("after_par__with_perm__", ctx)));
+      LinalgTransformationFilter(
+          Identifier::get("par__with_perm__", ctx),
 
   //===--------------------------------------------------------------------===//
   // Linalg to loops patterns.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgLoweringPattern<DotOp>>(
       ctx,
       /*loweringType=*/LinalgLoweringType::Loops,
-      LinalgMarker(Identifier::get("REG", ctx)));
+      LinalgTransformationFilter(Identifier::get("REG", ctx)));
 
   //===--------------------------------------------------------------------===//
   // Linalg distribution patterns.
   //===--------------------------------------------------------------------===//
   LinalgLoopDistributionOptions distributionOptions;
 
   //===--------------------------------------------------------------------===//
   // Linalg to vector contraction patterns.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgVectorizationPattern<MatmulOp>,
                   LinalgVectorizationPattern<FillOp>,
                   LinalgVectorizationPattern<CopyOp>,
                   LinalgVectorizationPattern<GenericOp>>(
-      ctx, LinalgMarker(Identifier::get("VECTORIZE", ctx)));
+      ctx, LinalgVectorizationOptions(),
+      LinalgTransformationFilter(Identifier::get("VECTORIZE", ctx)));
 
   //===--------------------------------------------------------------------===//
   // Linalg generic permutation patterns.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgInterchangePattern<GenericOp>>(
       ctx,
       /*interchangeVector=*/ArrayRef<unsigned>{1, 2, 0},
-      LinalgMarker({}, Identifier::get("PERMUTED", ctx)));
+      LinalgTransformationFilter(ArrayRef<Identifier>{},
+                                 Identifier::get("PERMUTED", ctx)));
   patterns.insert<LinalgInterchangePattern<IndexedGenericOp>>(
       ctx,
       /*interchangeVector=*/ArrayRef<unsigned>{1, 2, 0},
-      LinalgMarker({}, Identifier::get("PERMUTED", ctx)));
+      LinalgTransformationFilter(ArrayRef<Identifier>{},
+                                 Identifier::get("PERMUTED", ctx)));
 
   //===--------------------------------------------------------------------===//
   // Linalg subview operands promotion.
   //===--------------------------------------------------------------------===//
   patterns.insert<LinalgPromotionPattern<MatmulOp>>(
       ctx, LinalgPromotionOptions().setUseFullTileBuffersByDefault(true),
-      LinalgMarker(Identifier::get("_promote_views_", ctx),
+      LinalgTransformationFilter(Identifier::get("_promote_views_", ctx),
                                  Identifier::get("_views_promoted_", ctx)));
   patterns.insert<LinalgPromotionPattern<MatmulOp>>(
       ctx,
       LinalgPromotionOptions()
           .setOperandsToPromote({0})
           .setUseFullTileBuffersByDefault(true),
-      LinalgMarker(Identifier::get("_promote_first_view_", ctx),
-                   Identifier::get("_first_view_promoted_", ctx)));
+      LinalgTransformationFilter(
+          Identifier::get("_promote_first_view_", ctx),
   patterns.insert<LinalgPromotionPattern<FillOp>>(
       ctx,
       LinalgPromotionOptions()
           .setOperandsToPromote({0})
           .setUseFullTileBuffers({true})
           .setAlignment(32),
-      LinalgMarker(Identifier::get("_promote_views_aligned_", ctx),
-                   Identifier::get("_views_aligned_promoted_", ctx)));
+      LinalgTransformationFilter(
+          Identifier::get("_promote_views_aligned_", ctx),
 
   applyPatternsAndFoldGreedily(funcOp, std::move(patterns));
 
   // Drop the marker.
   funcOp.walk([](LinalgOp op) {
     op.removeAttr(LinalgTransforms::kLinalgTransformMarker);
   });
 }
 
 static void fillL1TilingAndMatmulToVectorPatterns(
     FuncOp funcOp, StringRef startMarker,
     SmallVectorImpl<OwningRewritePatternList> &patternsVector) {
   MLIRContext *ctx = funcOp.getContext();
   patternsVector.emplace_back(std::make_unique<LinalgTilingPattern<MatmulOp>>(
       ctx,
       LinalgTilingOptions().setTileSizes({8, 12, 16}).setInterchange({1, 0, 2}),
-      LinalgMarker(Identifier::get(startMarker, ctx),
+      LinalgTransformationFilter(Identifier::get(startMarker, ctx),
                                  Identifier::get("L1", ctx))));
 
   patternsVector.emplace_back(
       std::make_unique<LinalgPromotionPattern<MatmulOp>>(
           ctx, LinalgPromotionOptions().setUseFullTileBuffersByDefault(true),
-          LinalgMarker(Identifier::get("L1", ctx),
+          LinalgTransformationFilter(Identifier::get("L1", ctx),
                                      Identifier::get("VEC", ctx))));
 
   patternsVector.emplace_back(
       std::make_unique<LinalgVectorizationPattern<MatmulOp>>(
-          ctx, LinalgMarker(Identifier::get("VEC", ctx))));
+          ctx, LinalgVectorizationOptions(),
+          LinalgTransformationFilter(Identifier::get("VEC", ctx))));
   patternsVector.back()
       .insert<LinalgVectorizationPattern<FillOp>,
               LinalgVectorizationPattern<CopyOp>>(ctx);
 }
 
 //===----------------------------------------------------------------------===//
 // Test promotion callbacks
 //===----------------------------------------------------------------------===//
static LogicalResult copyCallBackFn(OpBuilder &b, Value src, Value dst,
   b.create<CopyOp>(src.getLoc(), src, dst);
   return success();
 }
 
 static void fillPromotionCallBackPatterns(MLIRContext *ctx,
                                           OwningRewritePatternList &patterns) {
   patterns.insert<LinalgTilingPattern<MatmulOp>>(
       ctx, LinalgTilingOptions().setTileSizes({16, 16, 16}),
-      LinalgMarker(Identifier::get("START", ctx),
+      LinalgTransformationFilter(Identifier::get("START", ctx),
                                  Identifier::get("PROMOTE", ctx)));
   patterns.insert<LinalgPromotionPattern<MatmulOp>>(
       ctx,
       LinalgPromotionOptions()
           .setOperandsToPromote({0, 2})
           .setUseFullTileBuffers({false, false})
           .setAllocationDeallocationFns(allocCallBackFn, deallocCallBackFn)
           .setCopyInOutFns(
               [](OpBuilder &b, Value src, Value dst) -> LogicalResult {
                 copyCallBackFn(b, src, dst, false);
                 return success();
               },
               [](OpBuilder &b, Value src, Value dst) -> LogicalResult {
                 copyCallBackFn(b, src, dst, true);
                 return success();
               }),
-      LinalgMarker(Identifier::get("PROMOTE", ctx)));
+      LinalgTransformationFilter(Identifier::get("PROMOTE", ctx)));
 }
 
 template <typename IdOp, typename NProcsOp>
 static SmallVector<ProcInfo, 2>
 getGpuProcIds(OpBuilder &b, Location loc, ArrayRef<Range> parallelLoopRanges) {
   Type indexType = b.getIndexType();
   SmallVector<ProcInfo, 2> procInfo(2);
   procInfo[0] = {b.create<IdOp>(loc, indexType, b.getStringAttr("y")),
static void fillTileAndDistributePatterns(MLIRContext *context,
     cyclicNprocsEqNiters.procInfo =
         getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsEqNiters),
-        LinalgMarker(Identifier::get("distribute1", context),
-                     Identifier::get("after_distribute1", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute1", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsGeNiters;
     cyclicNprocsGeNiters.distributionMethod.resize(
         2, DistributionMethod::CyclicNumProcsGeNumIters);
     cyclicNprocsGeNiters.procInfo =
         getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsGeNiters),
-        LinalgMarker(Identifier::get("distribute2", context),
-                     Identifier::get("after_distribute2", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute2", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsDefault;
     cyclicNprocsDefault.distributionMethod.resize(2,
                                                   DistributionMethod::Cyclic);
     cyclicNprocsDefault.procInfo =
         getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsDefault),
-        LinalgMarker(Identifier::get("distribute3", context),
-                     Identifier::get("after_distribute3", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute3", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsMixed1;
     cyclicNprocsMixed1.distributionMethod = {
         DistributionMethod::CyclicNumProcsEqNumIters,
         DistributionMethod::CyclicNumProcsGeNumIters};
     cyclicNprocsMixed1.procInfo = getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsMixed1),
-        LinalgMarker(Identifier::get("distribute4", context),
-                     Identifier::get("after_distribute4", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute4", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsMixed2;
     cyclicNprocsMixed2.distributionMethod = {
         DistributionMethod::CyclicNumProcsGeNumIters,
         DistributionMethod::Cyclic};
     cyclicNprocsMixed2.procInfo = getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsMixed2),
-        LinalgMarker(Identifier::get("distribute5", context),
-                     Identifier::get("after_distribute5", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute5", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsMixed3;
     cyclicNprocsMixed3.distributionMethod = {
         DistributionMethod::Cyclic,
         DistributionMethod::CyclicNumProcsEqNumIters};
     cyclicNprocsMixed3.procInfo = getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
 
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::ParallelLoops)
             .setDistributionOptions(cyclicNprocsMixed3),
-        LinalgMarker(Identifier::get("distribute6", context),
-                     Identifier::get("after_distribute6", context)));
+        LinalgTransformationFilter(
+            Identifier::get("distribute6", context),
   }
 
   {
     LinalgLoopDistributionOptions cyclicNprocsEqNiters;
     cyclicNprocsEqNiters.distributionMethod.resize(2,
                                                    DistributionMethod::Cyclic);
     cyclicNprocsEqNiters.procInfo =
         getGpuProcIds<gpu::BlockIdOp, gpu::GridDimOp>;
     patterns.insert<LinalgTilingPattern<MatmulOp>>(
         context,
         LinalgTilingOptions()
             .setTileSizes({8, 8, 4})
             .setLoopType(LinalgTilingLoopType::Loops)
             .setDistributionOptions(cyclicNprocsEqNiters),
-        LinalgMarker(Identifier::get("tensors_distribute1", context),
-                     Identifier::get("tensors_after_distribute1", context)));
+        LinalgTransformationFilter(
+            Identifier::get("tensors_distribute1", context),
   }
 }
 
 static void
 applyMatmulToVectorPatterns(FuncOp funcOp,
                             bool testMatmulToVectorPatterns1dTiling,
                             bool testMatmulToVectorPatterns2dTiling) {
   MLIRContext *ctx = funcOp.getContext();
   SmallVector<OwningRewritePatternList, 4> stage1Patterns;
   if (testMatmulToVectorPatterns1dTiling) {
     fillL1TilingAndMatmulToVectorPatterns(funcOp, Identifier::get("START", ctx),
                                           stage1Patterns);
   } else if (testMatmulToVectorPatterns2dTiling) {
     stage1Patterns.emplace_back(std::make_unique<LinalgTilingPattern<MatmulOp>>(
         ctx,
         LinalgTilingOptions()
             .setTileSizes({768, 264, 768})
             .setInterchange({1, 2, 0}),
-        LinalgMarker(Identifier::get("START", ctx),
+        LinalgTransformationFilter(Identifier::get("START", ctx),
                                    Identifier::get("L2", ctx))));
     fillL1TilingAndMatmulToVectorPatterns(funcOp, Identifier::get("L2", ctx),
                                           stage1Patterns);
   }
   SmallVector<FrozenRewritePatternList, 4> frozenStage1Patterns;
   llvm::move(stage1Patterns, std::back_inserter(frozenStage1Patterns));
   FrozenRewritePatternList stage2Patterns =
       getLinalgTilingCanonicalizationPatterns(ctx);
   applyStagedPatterns(funcOp, frozenStage1Patterns, std::move(stage2Patterns));