[Mlir-commits] [mlir] 1ede503 - [MLIR][Vector] Implement TransferReadOfExtractSliceOp as MaskableOpRewritePattern (#91960)

[llvmlistbot at llvm.org]

Author: Hugo Trachino
Date: 2024-05-16T21:40:56+01:00
New Revision: 1ede503910d3133ea1af8c084fee96ca98c66464

URL: https://github.com/llvm/llvm-project/commit/1ede503910d3133ea1af8c084fee96ca98c66464
DIFF: https://github.com/llvm/llvm-project/commit/1ede503910d3133ea1af8c084fee96ca98c66464.diff

LOG: [MLIR][Vector] Implement TransferReadOfExtractSliceOp as MaskableOpRewritePattern (#91960)

Split of https://github.com/llvm/llvm-project/pull/90835
Adds support for `TransferReadOfExtractSliceOpFolder` when the
`TransferReadOp` is inside a `MaskOp`.

Added: 
    

Modified: 
    mlir/lib/Dialect/Tensor/Transforms/CMakeLists.txt
    mlir/lib/Dialect/Tensor/Transforms/FoldTensorSubsetOps.cpp
    mlir/test/Dialect/Tensor/fold-tensor-subset-ops-into-vector-transfers.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/Tensor/Transforms/CMakeLists.txt b/mlir/lib/Dialect/Tensor/Transforms/CMakeLists.txt
index c6ef6ed86e0d9..0aabdaf667b9d 100644
--- a/mlir/lib/Dialect/Tensor/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/Tensor/Transforms/CMakeLists.txt
@@ -39,5 +39,6 @@ add_mlir_dialect_library(MLIRTensorTransforms
   MLIRTilingInterface
   MLIRTransforms
   MLIRVectorDialect
+  MLIRVectorUtils
   MLIRValueBoundsOpInterface
 )

diff  --git a/mlir/lib/Dialect/Tensor/Transforms/FoldTensorSubsetOps.cpp b/mlir/lib/Dialect/Tensor/Transforms/FoldTensorSubsetOps.cpp
index 3b8d3708bb731..5396531922aab 100644
--- a/mlir/lib/Dialect/Tensor/Transforms/FoldTensorSubsetOps.cpp
+++ b/mlir/lib/Dialect/Tensor/Transforms/FoldTensorSubsetOps.cpp
@@ -18,6 +18,7 @@
 #include "mlir/Dialect/Tensor/Transforms/Transforms.h"
 #include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
 #include "mlir/IR/AffineMap.h"
 #include "mlir/IR/BuiltinAttributes.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
@@ -48,12 +49,14 @@ static Value getTensorOperand(tensor::InsertSliceOp op) {
 namespace {
 /// Merge extract_slice operation with load/transferRead operation.
 class TransferReadOfExtractSliceOpFolder final
-    : public OpRewritePattern<vector::TransferReadOp> {
+    : public vector::MaskableOpRewritePattern<vector::TransferReadOp> {
 public:
-  using OpRewritePattern<vector::TransferReadOp>::OpRewritePattern;
+  using MaskableOpRewritePattern::MaskableOpRewritePattern;
 
-  LogicalResult matchAndRewrite(vector::TransferReadOp readOp,
-                                PatternRewriter &rewriter) const override;
+  FailureOr<mlir::Value>
+  matchAndRewriteMaskableOp(vector::TransferReadOp readOp,
+                            vector::MaskingOpInterface maskOp,
+                            PatternRewriter &rewriter) const override;
 };
 
 /// Merge insert_slice operation with store/transferWriteOp operation.
@@ -84,8 +87,10 @@ static LogicalResult preconditionsFoldExtractOrInsertWithTransferOp(
   return success();
 }
 
-LogicalResult TransferReadOfExtractSliceOpFolder::matchAndRewrite(
-    vector::TransferReadOp readOp, PatternRewriter &rewriter) const {
+FailureOr<mlir::Value>
+TransferReadOfExtractSliceOpFolder::matchAndRewriteMaskableOp(
+    vector::TransferReadOp readOp, vector::MaskingOpInterface maskOp,
+    PatternRewriter &rewriter) const {
   auto extractSliceOp =
       getTensorOperand(readOp).getDefiningOp<tensor::ExtractSliceOp>();
   if (!extractSliceOp)
@@ -95,7 +100,7 @@ LogicalResult TransferReadOfExtractSliceOpFolder::matchAndRewrite(
       preconditionsFoldExtractOrInsertWithTransferOp(rewriter, readOp,
                                                      extractSliceOp);
   if (failed(preconditionResult))
-    return preconditionResult;
+    return rewriter.notifyMatchFailure(readOp, "Failed preconditions");
 
   SmallVector<Value> indices(readOp.getIndices().begin(),
                              readOp.getIndices().end());
@@ -105,15 +110,17 @@ LogicalResult TransferReadOfExtractSliceOpFolder::matchAndRewrite(
       extractSliceOp.getMixedStrides(), extractSliceOp.getDroppedDims(),
       indices, sourceIndices);
 
-  rewriter.replaceOpWithNewOp<vector::TransferReadOp>(
-      readOp, readOp.getVectorType(), extractSliceOp.getSource(), sourceIndices,
+  Operation *newOp = rewriter.create<vector::TransferReadOp>(
+      readOp.getLoc(), readOp.getVectorType(), extractSliceOp.getSource(),
+      sourceIndices,
       AffineMapAttr::get(expandDimsToRank(
           readOp.getPermutationMap(), extractSliceOp.getSourceType().getRank(),
           extractSliceOp.getDroppedDims())),
       readOp.getPadding(),
       /*mask=*/Value(), readOp.getInBoundsAttr());
-
-  return success();
+  if (maskOp)
+    newOp = mlir::vector::maskOperation(rewriter, newOp, maskOp.getMask());
+  return newOp->getResults()[0];
 }
 
 LogicalResult InsertSliceOfTransferWriteOpFolder::matchAndRewrite(

diff  --git a/mlir/test/Dialect/Tensor/fold-tensor-subset-ops-into-vector-transfers.mlir b/mlir/test/Dialect/Tensor/fold-tensor-subset-ops-into-vector-transfers.mlir
index 6213db3956f9a..e562cf8efc356 100644
--- a/mlir/test/Dialect/Tensor/fold-tensor-subset-ops-into-vector-transfers.mlir
+++ b/mlir/test/Dialect/Tensor/fold-tensor-subset-ops-into-vector-transfers.mlir
@@ -75,6 +75,21 @@ func.func @transfer_read_of_extract_slice_non_leading_rank_reduction(%t : tensor
   return %1 : vector<5x6xf32>
 }
 
+// CHECK-LABEL: func @masked_transfer_read_of_extract_slice
+//  CHECK-SAME:     %[[t:.*]]: tensor<?x?xf32>, %[[s1:.*]]: index, %[[s2:.*]]: index
+//   CHECK-DAG: %[[m:.*]] = vector.create_mask{{.*}} : vector<5x6xi1>
+//   CHECK-DAG: %[[a:.*]] = affine.apply {{.*}}[[s1]]
+//       CHECK: vector.mask %[[m]] { vector.transfer_read %[[t]]{{.*}}: tensor<?x?xf32>, vector<5x6xf32> } : vector<5x6xi1> -> vector<5x6xf32>
+func.func @masked_transfer_read_of_extract_slice(%t : tensor<?x?xf32>, %s1 : index, %s2 : index) -> vector<5x6xf32> {
+  %c3 = arith.constant 3 : index
+  %c4 = arith.constant 4 : index
+  %cst = arith.constant 0.0 : f32
+  %0 = tensor.extract_slice %t[5, %s1] [10, %s2] [1, 1] : tensor<?x?xf32> to tensor<10x?xf32>
+  %mask = vector.create_mask %c3, %c4 : vector<5x6xi1>
+  %1 = vector.mask %mask {vector.transfer_read %0[%c3, %c4], %cst {in_bounds = [true, true]} : tensor<10x?xf32>, vector<5x6xf32>} : vector<5x6xi1> -> vector<5x6xf32>
+  return %1 : vector<5x6xf32>
+}
+
 // CHECK-LABEL: func @insert_slice_of_transfer_write(
 //  CHECK-SAME:     %[[t1:.*]]: tensor<?x12xf32>, %[[v:.*]]: vector<5x6xf32>, %[[s:.*]]: index
 //       CHECK:   %[[c3:.*]] = arith.constant 3 : index