[mlir][vector] Support distributing transfer op with permutation map

[nicolasvasilache]

mlir/lib/Dialect/Vector/VectorTransforms.cpp

2848

The growth in complexity really starts calling for docs in these patterns.

2864

nit: indexMap

2906

nit: indexMap

[ThomasRaoux]

Address comments

[ThomasRaoux]

mlir/lib/Dialect/Vector/VectorTransforms.cpp

2848

Added docs

mlir/lib/Dialect/Vector/VectorTransforms.cpp

Optional<mlir::vector::DistributeOps> mlir::vector::distributPointwiseVectorOp(
   DistributeOps ops;
   ops.extract =
       builder.create<vector::ExtractMapOp>(loc, result, ids, multiplicity, map);
   ops.insert =
       builder.create<vector::InsertMapOp>(loc, ops.extract, result, ids);
   return ops;
 }
 
 struct TransferReadExtractPattern
     : public OpRewritePattern<vector::TransferReadOp> {
   TransferReadExtractPattern(MLIRContext *context)
       : OpRewritePattern<vector::TransferReadOp>(context) {}
   LogicalResult matchAndRewrite(vector::TransferReadOp read,
                                 PatternRewriter &rewriter) const override {
     if (!read.getResult().hasOneUse())
       return failure();
     auto extract =
         dyn_cast<vector::ExtractMapOp>(*read.getResult().getUsers().begin());
     if (!extract)
       return failure();
     if (read.mask())
       return failure();
 
     SmallVector<Value, 4> indices(read.indices().begin(), read.indices().end());
-    AffineMap map = extract.map();
+    AffineMap Indexmap = extract.map().compose(read.permutation_map());
     unsigned idCount = 0;
     ImplicitLocOpBuilder lb(read.getLoc(), rewriter);
-    for (auto expr : map.getResults()) {
+    for (auto it :
+         llvm::zip(Indexmap.getResults(), extract.map().getResults())) {
       AffineExpr d0, d1;
       bindDims(read.getContext(), d0, d1);
-      unsigned pos = expr.cast<AffineDimExpr>().getPosition();
+      auto indexExpr = std::get<0>(it).dyn_cast<AffineDimExpr>();
+      if (!indexExpr)
+        continue;
+      unsigned indexPos = indexExpr.getPosition();
+      unsigned vectorPos = std::get<1>(it).cast<AffineDimExpr>().getPosition();
       auto scale = getAffineConstantExpr(
-          extract.getResultType().getDimSize(pos), read.getContext());
-      indices[pos] =
-          makeComposedAffineApply(rewriter, read.getLoc(), d0 + scale * d1,
-                                  {indices[pos], extract.ids()[idCount++]});
+          extract.getResultType().getDimSize(vectorPos), read.getContext());
+      indices[indexPos] = makeComposedAffineApply(
+          rewriter, read.getLoc(), d0 + scale * d1,
+          {indices[indexPos], extract.ids()[idCount++]});
     }
     Value newRead = lb.create<vector::TransferReadOp>(
         extract.getType(), read.source(), indices, read.permutation_map(),
         read.padding(), read.in_boundsAttr());
     Value dest = lb.create<ConstantOp>(read.getType(),
                                        rewriter.getZeroAttr(read.getType()));
     newRead = lb.create<vector::InsertMapOp>(newRead, dest, extract.ids());
     rewriter.replaceOp(read, newRead);
LogicalResult matchAndRewrite(vector::TransferWriteOp write,
                                 PatternRewriter &rewriter) const override {
     auto insert = write.vector().getDefiningOp<vector::InsertMapOp>();
     if (!insert)
       return failure();
     if (write.mask())
       return failure();
     SmallVector<Value, 4> indices(write.indices().begin(),
                                   write.indices().end());
-    AffineMap map = insert.map();
+    AffineMap Indexmap = insert.map().compose(write.permutation_map());
     unsigned idCount = 0;
     Location loc = write.getLoc();
-    for (auto expr : map.getResults()) {
+    for (auto it :
+         llvm::zip(Indexmap.getResults(), insert.map().getResults())) {
       AffineExpr d0, d1;
       bindDims(write.getContext(), d0, d1);
-      unsigned pos = expr.cast<AffineDimExpr>().getPosition();
+      auto indexExpr = std::get<0>(it).dyn_cast<AffineDimExpr>();
+      if (!indexExpr)
+        continue;
+      unsigned indexPos = indexExpr.getPosition();
+      unsigned vectorPos = std::get<1>(it).cast<AffineDimExpr>().getPosition();
       auto scale = getAffineConstantExpr(
-          insert.getSourceVectorType().getDimSize(pos), write.getContext());
-      indices[pos] =
+          insert.getSourceVectorType().getDimSize(vectorPos),
+          write.getContext());
+      indices[indexPos] =
           makeComposedAffineApply(rewriter, loc, d0 + scale * d1,
-                                  {indices[pos], insert.ids()[idCount++]});
+                                  {indices[indexPos], insert.ids()[idCount++]});
     }
     rewriter.create<vector::TransferWriteOp>(
         loc, insert.vector(), write.source(), indices, write.permutation_map(),
         write.in_boundsAttr());
     rewriter.eraseOp(write);
     return success();
   }
 };

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

func @vector_add_transfer_3d(%id0 : index, %id1 : index, %A: memref<64x64x64xf32>,
   %cf0 = constant 0.0 : f32
   %a = vector.transfer_read %A[%c0, %c0, %c0], %cf0: memref<64x64x64xf32>, vector<64x4x32xf32>
   %b = vector.transfer_read %B[%c0, %c0, %c0], %cf0: memref<64x64x64xf32>, vector<64x4x32xf32>
   %acc = addf %a, %b: vector<64x4x32xf32>
   vector.transfer_write %acc, %C[%c0, %c0, %c0]: vector<64x4x32xf32>, memref<64x64x64xf32>
   return
 }
 
+// -----
+
+#map0 = affine_map<(d0, d1, d2, d3) -> (d3, 0, 0)>
+#map1 = affine_map<(d0, d1, d2, d3) -> (0, d3, d0)>
+#map2 = affine_map<(d0, d1, d2, d3) -> (d3, d2, d1)>
+
+// CHECK-DAG: #[[MAP0:.*]] = affine_map<()[s0] -> (s0 * 2)>
+// CHECK-DAG: #[[MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, 0, 0)>
+// CHECK-DAG: #[[MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (0, d3, d0)>
+// CHECK-DAG: #[[MAP3:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, d2, d1)>
 
+//       CHECK: func @vector_add_transfer_permutation
+//  CHECK-SAME: (%[[ID_0:.*]]: index, %[[ID_1:.*]]: index
+//       CHECK:    %[[C0:.*]] = constant 0 : index
+//       CHECK:    %[[ID2:.*]] = affine.apply #[[MAP0]]()[%[[ID_0]]]
+//  CHECK-NEXT:    %[[EXA:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]], %[[C0]], %[[ID2]]], %{{.*}} {permutation_map = #[[MAP1]]} : memref<?x?x?x?xf32>, vector<2x4x1xf32>
+//  CHECK-NEXT:    %[[EXB:.*]] = vector.transfer_read %{{.*}}[%[[ID_0]], %[[C0]], %[[C0]], %[[C0]]], %{{.*}} {permutation_map = #[[MAP2]]} : memref<?x?x?x?xf32>, vector<2x4x1xf32>
+//  CHECK-NEXT:    %[[ADD:.*]] = addf %[[EXA]], %[[EXB]] : vector<2x4x1xf32>
+//  CHECK-NEXT:    %[[ID3:.*]] = affine.apply #[[MAP0]]()[%[[ID_0]]]
+//  CHECK-NEXT:    vector.transfer_write %[[ADD]], %{{.*}}[%[[C0]], %[[ID_1]], %[[C0]], %[[ID3]]] {permutation_map = #[[MAP3]]} : vector<2x4x1xf32>, memref<?x?x?x?xf32>
+//  CHECK-NEXT:    return
+func @vector_add_transfer_permutation(%id0 : index, %id1 : index, %A: memref<?x?x?x?xf32>,
+  %B: memref<?x?x?x?xf32>, %C: memref<?x?x?x?xf32>) {
+  %c0 = constant 0 : index
+  %cf0 = constant 0.0 : f32
+  %a = vector.transfer_read %A[%c0, %c0, %c0, %c0], %cf0 {permutation_map = #map0} : memref<?x?x?x?xf32>, vector<64x4x32xf32>
+  %b = vector.transfer_read %B[%c0, %c0, %c0, %c0], %cf0 {permutation_map = #map1}: memref<?x?x?x?xf32>, vector<64x4x32xf32>
+  %acc = addf %a, %b: vector<64x4x32xf32>
+  vector.transfer_write %acc, %C[%c0, %c0, %c0, %c0] {permutation_map = #map2}: vector<64x4x32xf32>, memref<?x?x?x?xf32>
+  return
+}