Fix CombineContractBroadcast folding reduction iterators.

Fix CombineContractBroadcast folding reduction iterators.

Differential Revision: https://reviews.llvm.org/D128739

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

@@ -32,6 +32,7 @@
 #include "llvm/ADT/DenseSet.h"
 #include "llvm/ADT/MapVector.h"
 #include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallBitVector.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
@@ -1126,6 +1127,21 @@ struct CombineContractBroadcast
       // relaxed we can remove this case.
       if (innerDimBroadcast)
         continue;
+
+      // It would be incorrect to fold a broadcast onto a reduction dimension
+      // of non-unit size.
+      bool nonUnitDimReductionBroadcast = false;
+      for (int64_t i = 0; i < rankDiff; ++i) {
+        if (broadcast.getVectorType().getDimSize(i) != 1 &&
+            isReductionIterator(contractOp.getIteratorTypes()
+                                    .getValue()[map.getDimPosition(i)])) {
+          nonUnitDimReductionBroadcast = true;
+          break;
+        }
+      }
+      if (nonUnitDimReductionBroadcast)
+        continue;
+
       AffineMap broadcastMap =
           AffineMap::get(broadcast.getVectorType().getRank(), 0, originalDims,
                          contractOp.getContext());
@@ -1133,11 +1149,40 @@ struct CombineContractBroadcast
       *operand = broadcast.getSource();
       changed = true;
     }
+
     if (!changed)
       return failure();
+
+    // Determine which dims are usused, now that the maps have been composed
+    // with the broadcast maps.
+    unsigned numDims = maps[0].getNumDims();
+    llvm::SmallBitVector unusedDims(numDims, true);
+    for (const auto &m : maps) {
+      for (unsigned i = 0; i < numDims; ++i) {
+        if (m.isFunctionOfDim(i))
+          unusedDims.reset(i);
+      }
+    }
+    // Compress unused dims.
+    for (auto &m : maps)
+      m = compressDims(m, unusedDims);
+    // Compute the combined iterators.
+    SmallVector<Attribute, 4> iterators;
+    for (unsigned i = 0; i < numDims; ++i) {
+      if (!unusedDims.test(i))
+        iterators.push_back(contractOp.getIteratorTypes().getValue()[i]);
+    }
+    // Check that compressing unused dims isn't removing all reduction
+    // iterators. For example, if the vector.contract had only one reduction
+    // iterator and that was a unit-dimension created by a broadcast,
+    // then we should bail here, otherwise we would create a contract without
+    // a reduction iterator.
+    if (!llvm::any_of(iterators, isReductionIterator))
+      return failure();
+
     rewriter.replaceOpWithNewOp<vector::ContractionOp>(
         contractOp, lhs, rhs, contractOp.getAcc(),
-        rewriter.getAffineMapArrayAttr(maps), contractOp.getIteratorTypes());
+        rewriter.getAffineMapArrayAttr(maps), rewriter.getArrayAttr(iterators));
     return success();
   }
 };

mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir

@@ -86,6 +86,98 @@ func.func @contract_broadcast(
   return %1 : vector<8x32xf32>
 }
 
+// -----
+// Test that CombineContractBroadcast is able to combine a broadcast that
+// creates a unit dim that is consumed by a reduction iterator, dropping that
+// reduction iterator, as long as there is another reduction iterator left.
+
+#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3)>
+#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d3)>
+#map2 = affine_map<(d0, d1, d2, d3) -> (d1, d2)>
+
+// CHECK-DAG: #[[$map0:.*]] = affine_map<(d0, d1, d2) -> (d0, d2)>
+// CHECK-DAG: #[[$map1:.*]] = affine_map<(d0, d1, d2) -> (d1, d2)>
+// CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)>
+
+// CHECK-LABEL: contract_broadcast_unit_dim_reduction
+//  CHECK-SAME: (%[[ARG0:.+]]: vector<8x4xi32>, %[[ARG1:.+]]: vector<8x4xi32>, %[[ARG2:.+]]: vector<8x8xi32>)
+//  CHECK: vector.contract
+//  CHECK-SAME: indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]]
+//  CHECK-SAME: iterator_types = ["parallel", "parallel", "reduction"]
+//  CHECK-SAME: %[[ARG0]], %[[ARG1]], %[[ARG2]] : vector<8x4xi32>, vector<8x4xi32> into vector<8x8xi32>
+func.func @contract_broadcast_unit_dim_reduction(%arg0 : vector<8x4xi32>, %arg1 : vector<8x4xi32>, %arg2 : vector<8x8xi32>) -> vector<8x8xi32> {
+    %0 = vector.broadcast %arg0 : vector<8x4xi32> to vector<1x8x4xi32>
+    %1 = vector.broadcast %arg1 : vector<8x4xi32> to vector<1x8x4xi32>
+    %result = vector.contract {
+        indexing_maps = [#map0, #map1, #map2],
+        iterator_types = ["reduction", "parallel", "parallel", "reduction"],
+        kind = #vector.kind<add>
+      } %0, %1, %arg2 : vector<1x8x4xi32>, vector<1x8x4xi32> into vector<8x8xi32>
+    return %result : vector<8x8xi32>
+}
+
+// -----
+// Test that CombineContractBroadcast will not combine a broadcast that creates
+// a non-unit dim that is consumed by a reduction iterator.
+// Moreover, the affine_map's are permuting the position of that reduction
+// iterator with the position of a parallel iterator to ensure that
+// the logic guarding that case does not mix up dimensions here.
+
+#map0 = affine_map<(d0, d1, d2, d3) -> (d1, d0, d3)>
+#map1 = affine_map<(d0, d1, d2, d3) -> (d1, d2, d3)>
+#map2 = affine_map<(d0, d1, d2, d3) -> (d0, d2)>
+
+// CHECK-DAG: #[[$map0:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, d0, d3)>
+// CHECK-DAG: #[[$map1:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, d2, d3)>
+// CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d2)>
+
+// CHECK-LABEL: contract_broadcast_non_unit_dim_reduction_with_permutation
+//  CHECK-SAME: (%[[ARG0:.+]]: vector<8x4xi32>, %[[ARG1:.+]]: vector<8x4xi32>, %[[ARG2:.+]]: vector<8x8xi32>)
+//  CHECK: %[[BROADCAST0:.+]] = vector.broadcast %[[ARG0]] : vector<8x4xi32> to vector<2x8x4xi32>
+//  CHECK: %[[BROADCAST1:.+]] = vector.broadcast %[[ARG1]] : vector<8x4xi32> to vector<2x8x4xi32>
+//  CHECK: vector.contract
+//  CHECK-SAME: indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]]
+//  CHECK-SAME: iterator_types = ["parallel", "reduction", "parallel", "reduction"]
+//  CHECK-SAME: %[[BROADCAST0]], %[[BROADCAST1]], %[[ARG2]] : vector<2x8x4xi32>, vector<2x8x4xi32> into vector<8x8xi32>
+func.func @contract_broadcast_non_unit_dim_reduction_with_permutation(%arg0 : vector<8x4xi32>, %arg1 : vector<8x4xi32>, %arg2 : vector<8x8xi32>) -> vector<8x8xi32> {
+    %0 = vector.broadcast %arg0 : vector<8x4xi32> to vector<2x8x4xi32>
+    %1 = vector.broadcast %arg1 : vector<8x4xi32> to vector<2x8x4xi32>
+    %result = vector.contract {
+        indexing_maps = [#map0, #map1, #map2],
+        iterator_types = ["parallel", "reduction", "parallel", "reduction"],
+        kind = #vector.kind<add>
+      } %0, %1, %arg2 : vector<2x8x4xi32>, vector<2x8x4xi32> into vector<8x8xi32>
+    return %result : vector<8x8xi32>
+}
+
+// -----
+
+// Test that CombineContractBroadcast is not combining this case, as that would
+// result in dropping this contract's only reduction iterator.
+
+#map0 = affine_map<(d0, d1, d2) -> (d0, d1)>
+#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
+#map2 = affine_map<(d0, d1, d2) -> (d1, d2)>
+
+// CHECK-LABEL: contract_broadcast_unit_dim_reduction_as_only_reduction
+//  CHECK-SAME: (%[[ARG0:.+]]: vector<8xi32>, %[[ARG1:.+]]: vector<8xi32>, %[[ARG2:.+]]: vector<8x8xi32>)
+//  CHECK: %[[BROADCAST0:.+]] = vector.broadcast %[[ARG0]] : vector<8xi32> to vector<1x8xi32>
+//  CHECK: %[[BROADCAST1:.+]] = vector.broadcast %[[ARG1]] : vector<8xi32> to vector<1x8xi32>
+//  CHECK: vector.contract
+//  CHECK-SAME: indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]]
+//  CHECK-SAME: iterator_types = ["reduction", "parallel", "parallel"]
+//  CHECK-SAME: %[[BROADCAST0]], %[[BROADCAST1]], %[[ARG2]] : vector<1x8xi32>, vector<1x8xi32> into vector<8x8xi32>
+func.func @contract_broadcast_unit_dim_reduction_as_only_reduction(%arg0 : vector<8xi32>, %arg1 : vector<8xi32>, %arg2 : vector<8x8xi32>) -> vector<8x8xi32> {
+    %0 = vector.broadcast %arg0 : vector<8xi32> to vector<1x8xi32>
+    %1 = vector.broadcast %arg1 : vector<8xi32> to vector<1x8xi32>
+    %result = vector.contract {
+        indexing_maps = [#map0, #map1, #map2],
+        iterator_types = ["reduction", "parallel", "parallel"],
+        kind = #vector.kind<add>
+      } %0, %1, %arg2 : vector<1x8xi32>, vector<1x8xi32> into vector<8x8xi32>
+    return %result : vector<8x8xi32>
+}
+
 //===----------------------------------------------------------------------===//
 // Reorder casting ops and vector ops. The casting ops have almost identical
 // pattern, so only arith.extsi op is tested.