[mlir][vector] Use inferRankReducedResultType for subview type inference. (#84395)

Fixes iree-org/iree#16475

Author: hanhanW

mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp

@@ -1255,42 +1255,6 @@ getTransferFoldableInnerUnitDims(MemRefType srcType, VectorType vectorType) {
   return result;
 }
 
-/// Returns a MemRef type that drops inner `dimsToDrop` dimensions from
-/// `srcType`. E.g., if `srcType` is memref<512x16x1x1xf32> and `dimsToDrop` is
-/// two, it returns memref<512x16x16> type.
-static MemRefType getMemRefTypeWithDroppingInnerDims(OpBuilder &builder,
-                                                     MemRefType srcType,
-                                                     size_t dimsToDrop) {
-  MemRefLayoutAttrInterface layout = srcType.getLayout();
-  if (isa<AffineMapAttr>(layout) && layout.isIdentity()) {
-    return MemRefType::get(srcType.getShape().drop_back(dimsToDrop),
-                           srcType.getElementType(), nullptr,
-                           srcType.getMemorySpace());
-  }
-  MemRefLayoutAttrInterface updatedLayout;
-  if (auto strided = dyn_cast<StridedLayoutAttr>(layout)) {
-    auto strides = llvm::to_vector(strided.getStrides().drop_back(dimsToDrop));
-    updatedLayout = StridedLayoutAttr::get(strided.getContext(),
-                                           strided.getOffset(), strides);
-    return MemRefType::get(srcType.getShape().drop_back(dimsToDrop),
-                           srcType.getElementType(), updatedLayout,
-                           srcType.getMemorySpace());
-  }
-
-  // Non-strided layout case.
-  AffineMap map = srcType.getLayout().getAffineMap();
-  int numSymbols = map.getNumSymbols();
-  for (size_t i = 0; i < dimsToDrop; ++i) {
-    int dim = srcType.getRank() - i - 1;
-    map = map.replace(builder.getAffineDimExpr(dim),
-                      builder.getAffineConstantExpr(0), map.getNumDims() - 1,
-                      numSymbols);
-  }
-  return MemRefType::get(srcType.getShape().drop_back(dimsToDrop),
-                         srcType.getElementType(), updatedLayout,
-                         srcType.getMemorySpace());
-}
-
 /// Drop inner most contiguous unit dimensions from transfer_read operand.
 class DropInnerMostUnitDimsTransferRead
     : public OpRewritePattern<vector::TransferReadOp> {
@@ -1337,8 +1301,10 @@ class DropInnerMostUnitDimsTransferRead
                                       rewriter.getIndexAttr(0));
     SmallVector<OpFoldResult> strides(srcType.getRank(),
                                       rewriter.getIndexAttr(1));
-    MemRefType resultMemrefType =
-        getMemRefTypeWithDroppingInnerDims(rewriter, srcType, dimsToDrop);
+    auto resultMemrefType =
+        cast<MemRefType>(memref::SubViewOp::inferRankReducedResultType(
+            srcType.getShape().drop_back(dimsToDrop), srcType, offsets, sizes,
+            strides));
     ArrayAttr inBoundsAttr =
         readOp.getInBounds()
             ? rewriter.getArrayAttr(
@@ -1421,8 +1387,10 @@ class DropInnerMostUnitDimsTransferWrite
                                       rewriter.getIndexAttr(0));
     SmallVector<OpFoldResult> strides(srcType.getRank(),
                                       rewriter.getIndexAttr(1));
-    MemRefType resultMemrefType =
-        getMemRefTypeWithDroppingInnerDims(rewriter, srcType, dimsToDrop);
+    auto resultMemrefType =
+        cast<MemRefType>(memref::SubViewOp::inferRankReducedResultType(
+            srcType.getShape().drop_back(dimsToDrop), srcType, offsets, sizes,
+            strides));
     ArrayAttr inBoundsAttr =
         writeOp.getInBounds()
             ? rewriter.getArrayAttr(

mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir

@@ -16,22 +16,27 @@ func.func @contiguous_inner_most_view(%in: memref<1x1x8x1xf32, strided<[3072, 8,
 
 // -----
 
-func.func @contiguous_outer_dyn_inner_most_view(%in: memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x1xf32>{
+func.func @contiguous_outer_dyn_inner_most_view(%a: index, %b: index, %memref: memref<?x?x8x1xf32>) -> vector<8x1xf32> {
   %c0 = arith.constant 0 : index
-  %cst = arith.constant 0.0 : f32
-  %0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x8x1xf32>
-  return %0 : vector<1x8x1xf32>
+  %pad = arith.constant 0.0 : f32
+  %v = vector.transfer_read %memref[%a, %b, %c0, %c0], %pad {in_bounds = [true, true]} : memref<?x?x8x1xf32>, vector<8x1xf32>
+  return %v : vector<8x1xf32>
 }
-//      CHECK: func @contiguous_outer_dyn_inner_most_view(
+// CHECK: func.func @contiguous_outer_dyn_inner_most_view(
+// CHECK-SAME:   %[[IDX0:[a-zA-Z0-9]+]]
+// CHECK-SAME:   %[[IDX1:[a-zA-Z0-9]+]]
 // CHECK-SAME:   %[[SRC:[a-zA-Z0-9]+]]
-//  CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index
-//  CHECK-DAG:   %[[D0:.+]] = memref.dim %[[SRC]], %[[C0]]
-//      CHECK:   %[[SRC_0:.+]] = memref.subview %[[SRC]][0, 0, 0, 0] [%[[D0]], 1, 8, 1] [1, 1, 1, 1]
-// CHECK-SAME:    memref<?x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>> to memref<?x1x8xf32, strided<[3072, 8, 1], offset: ?>>
-//      CHECK:   %[[VEC:.+]] = vector.transfer_read %[[SRC_0]]
-// CHECK-SAME:    memref<?x1x8xf32, strided<[3072, 8, 1], offset: ?>>, vector<1x8xf32>
-//      CHECK:   %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
-//      CHECK:   return %[[RESULT]]
+// CHECK-DAG:    %[[C0:.+]] = arith.constant 0 : index
+// CHECK-DAG:    %[[C1:.+]] = arith.constant 1 : index
+// CHECK-DAG:    %[[PAD:.+]] = arith.constant 0.000000e+00 : f32
+// CHECK:        %[[D0:.+]] = memref.dim %[[SRC]], %[[C0]]
+// CHECK:        %[[D1:.+]] = memref.dim %[[SRC]], %[[C1]]
+// CHECK:        %[[VIEW:.+]] = memref.subview %[[SRC]][0, 0, 0, 0] [%[[D0]], %[[D1]], 8, 1] [1, 1, 1, 1]
+// CHECK-SAME:     memref<?x?x8x1xf32> to memref<?x?x8xf32, strided<[?, 8, 1], offset: ?>>
+// CHECK:        %[[VEC:.+]] = vector.transfer_read %[[VIEW]]
+// CHECK-SAME:     memref<?x?x8xf32, strided<[?, 8, 1], offset: ?>>, vector<8xf32>
+// CHECK:        %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
+// CHECK:        return %[[RESULT]]
 
 // -----
 
@@ -43,7 +48,7 @@ func.func @contiguous_inner_most_dim(%A: memref<16x1xf32>, %i:index, %j:index) -
 }
 //      CHECK: func @contiguous_inner_most_dim(%[[SRC:.+]]: memref<16x1xf32>, %[[I:.+]]: index, %[[J:.+]]: index) -> vector<8x1xf32>
 //      CHECK:   %[[SRC_0:.+]] = memref.subview %[[SRC]]
-// CHECK-SAME:     memref<16x1xf32> to memref<16xf32>
+// CHECK-SAME:     memref<16x1xf32> to memref<16xf32, strided<[1]>>
 //      CHECK:   %[[V:.+]] = vector.transfer_read %[[SRC_0]]
 //      CHECK:   %[[RESULT]] = vector.shape_cast %[[V]] : vector<8xf32> to vector<8x1xf32>
 //      CHECK:   return %[[RESULT]]
@@ -111,7 +116,7 @@ func.func @drop_two_inner_most_dim_for_transfer_write(%arg0: memref<1x512x16x1x1
 // CHECK-SAME:   %[[IDX:[a-zA-Z0-9]+]]
 // CHECK-DAG:    %[[C0:.+]] = arith.constant 0 : index
 // CHECK:        %[[SUBVIEW:.+]] = memref.subview %[[DEST]]
-// CHECK-SAME:     memref<1x512x16x1x1xf32> to memref<1x512x16xf32>
+// CHECK-SAME:     memref<1x512x16x1x1xf32> to memref<1x512x16xf32, strided<[8192, 16, 1]>>
 // CHECK:        %[[CAST:.+]] = vector.shape_cast %[[VEC]] : vector<1x16x16x1x1xf32> to vector<1x16x16xf32>
 // CHECK:        vector.transfer_write %[[CAST]], %[[SUBVIEW]]
 // CHECK-SAME:     [%[[C0]], %[[IDX]], %[[C0]]]