[mlir][linalg] Support the empty anchor op string when padding.

[gysit]

Summary

Add support for an empty anchor op string in vectorization. An empty anchor op string is useful after fusion when there are multiple different operations to vectorize.

Depends On D114689

[gysit]

Rebase.

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

interchangeIf(bool b, ArrayRef<int64_t> iteratorInterchange,
     return b ? interchange(iteratorInterchange, f) : *this;
     return *this;
   }
   /// Append a pattern to rewrite `LinalgOpType` as a vector operation.
   CodegenStrategy &
   vectorize(StringRef opName,
             LinalgTransformationFilter::FilterFunction f = nullptr,
             bool vectorizePadding = false) {
-    assert(!opName.empty() && "expected an op name");
     transformationSequence.emplace_back(std::make_unique<Vectorize>(
         opName, linalg::LinalgVectorizationOptions(), f, vectorizePadding));
     return *this;
   }
   /// Conditionally append a pattern to rewrite `LinalgOpType` as a vector
   /// operation.
   CodegenStrategy &
   vectorizeIf(bool b, StringRef opName,

mlir/test/Dialect/Linalg/codegen-strategy.mlir

 // RUN: mlir-opt %s -test-linalg-codegen-strategy="anchor-func=matmul anchor-op=linalg.matmul tile-sizes=2,4,8 vectorize vectorize-contraction-to=matrixintrinsics unroll-vector-transfers=true" -split-input-file | FileCheck %s --check-prefix=CHECK-INTRINSIC
 // RUN: mlir-opt %s -test-linalg-codegen-strategy="anchor-func=matmul anchor-op=linalg.matmul tile-sizes=16,32,64 promote promote-full-tile-pad register-tile-sizes=2,4,8 vectorize vectorize-contraction-to=outerproduct split-transfers=true unroll-vector-transfers=false" -split-input-file | FileCheck %s --check-prefix=CHECK-OUTER
 // RUN: mlir-opt %s -test-linalg-codegen-strategy="anchor-func=matmul anchor-op=linalg.matmul tile-sizes=16,32,64 tile-interchange=1,2,0 generalize iterator-interchange=0,2,1" -split-input-file | FileCheck %s --check-prefix=CHECK-INTERCHANGE
 // RUN: mlir-opt %s -test-linalg-codegen-strategy="anchor-func=matmul anchor-op=linalg.matmul tile-sizes=16,32,64 pad pack-paddings=1,1,0 hoist-paddings=3,3,0" -split-input-file | FileCheck %s --check-prefix=CHECK-PAD
+// RUN: mlir-opt %s -test-linalg-codegen-strategy="anchor-func=matmul anchor-op=linalg.matmul tile-sizes=16,32,64 fuse pad vectorize" -split-input-file | FileCheck %s --check-prefix=CHECK-FUSE
 
 // CHECK-INTRINSIC: func @matmul(
 //     CHECK-OUTER: func @matmul(
 func @matmul(%arg0: memref<72x72xf32>, %arg1: memref<72x72xf32>, %arg2: memref<72x72xf32>) {
 
   // Check the matrix intrinsic lowering is triggered.
   //      CHECK-INTRINSIC: vector.matrix_multiply
   // CHECK-INTRINSIC-SAME: {lhs_columns = 8 : i32, lhs_rows = 2 : i32, rhs_columns = 4 : i32}
func @matmul(%arg0: tensor<72x72xf32>, %arg1: tensor<72x72xf32>, %arg2: tensor<72x72xf32>) -> tensor<72x72xf32> {
   // Check CSE eliminates the duplicate min operations introduced by tiling.
   //              CHECK-PAD: affine.min #[[MAP0]]
   //          CHECK-PAD-NOT: affine.min #[[MAP0]]
   //      CHECK-PAD-COUNT=2: scf.for
   //              CHECK-PAD: linalg.matmul
   %0 = linalg.matmul ins(%arg0, %arg1: tensor<72x72xf32>, tensor<72x72xf32>) outs(%arg2: tensor<72x72xf32>) -> tensor<72x72xf32>
   return %0 : tensor<72x72xf32>
 }
+
+// -----
+
+//         CHECK-FUSE: func @matmul(
+func @matmul(%arg0: tensor<72x72xf32>, %arg1: tensor<72x72xf32>, %arg2: tensor<72x72xf32>) -> tensor<72x72xf32> {
+
+  // Check the padding and vectorization applies to the fill operation due to the empty anchor op string.
+  //        CHECK-FUSE:  %[[CST:.*]] = arith.constant dense<0.000000e+00>
+  //        CHECK-FUSE:  vector.transfer_write %[[CST]]
+  %cst = arith.constant 0.0 : f32
+  %0 = linalg.fill(%cst, %arg0) : f32, tensor<72x72xf32> -> tensor<72x72xf32>
+
+  // Check the matmul is padded and vectorized despite the empty anchor op string.
+  //        CHECK-FUSE:  vector.outerproduct
+  %1 = linalg.matmul ins(%arg0, %arg1: tensor<72x72xf32>, tensor<72x72xf32>) outs(%0: tensor<72x72xf32>) -> tensor<72x72xf32>
+  return %1 : tensor<72x72xf32>
+}

mlir/test/lib/Dialect/Linalg/TestLinalgCodegenStrategy.cpp

void TestLinalgCodegenStrategy::runStrategy(
     LinalgTilingAndFusionOptions tilingAndFusionOptions,
     LinalgTilingOptions tilingOptions,
     LinalgTilingOptions registerTilingOptions,
     LinalgPaddingOptions paddingOptions,
     vector::VectorContractLowering vectorContractLowering,
     vector::VectorTransferSplit vectorTransferSplit) {
   assert(!anchorOpName.empty());
   CodegenStrategy strategy;
-  StringRef genericOpName = GenericOp::getOperationName();
   strategy
       .tileAndFuseIf(fuse && !tileSizes.empty(), anchorOpName,
                      tilingAndFusionOptions)
       .tileIf(!fuse && !tileSizes.empty(), anchorOpName, tilingOptions)
-      .promoteIf(promote, anchorOpName,
+      .promoteIf(!fuse && promote, anchorOpName,
                  LinalgPromotionOptions()
                      .setAlignment(16)
                      .setUseFullTileBuffersByDefault(promoteFullTile))
-      .tileIf(!registerTileSizes.empty(), anchorOpName, registerTilingOptions)
-      .promoteIf(registerPromote, anchorOpName,
+      .tileIf(!fuse && !registerTileSizes.empty(), anchorOpName,
+              registerTilingOptions)
+      .promoteIf(!fuse && registerPromote, anchorOpName,
                  LinalgPromotionOptions()
                      .setAlignment(16)
                      .setUseFullTileBuffersByDefault(registerPromoteFullTile))
-      .padIf(pad, anchorOpName, paddingOptions)
-      .generalizeIf(generalize, anchorOpName)
+      .padIf(pad, "", paddingOptions)
+      .generalizeIf(generalize, "")
       .interchangeIf(!iteratorInterchange.empty(), iteratorInterchange)
-      .vectorizeIf(vectorize, generalize ? genericOpName : anchorOpName)
+      .vectorizeIf(vectorize, "")
       .vectorLowering(
           LinalgVectorLoweringOptions()
               .setVectorTransformsOptions(
                   vector::VectorTransformsOptions()
                       .setVectorTransformsOptions(vectorContractLowering)
                       .setVectorTransferSplit(vectorTransferSplit))
               .setVectorTransferToSCFOptions(
                   VectorTransferToSCFOptions().enableFullUnroll(