[mlir][linalg] Avoid illegal elementwise fusion into reductions

[herhut]

Summary

Fusing into a reduction is only valid if doing so does not erase information on a reduction dimensions size.

[nicolasvasilache]

mlir/test/Dialect/Linalg/fusion-elementwise-ops.mlir

918

the test name is confusing, how does it relate to fusion?

926

what is the proposed outcome here?

[herhut]

Fix test

[herhut]

mlir/test/Dialect/Linalg/fusion-elementwise-ops.mlir

918

Yeah, I meant fusion. Got confused by the test above.

926

I added CHECK statements to ensure two linalg operations remain. A fusion would be illegal anyway, so the compiler aborts in that case.

[nicolasvasilache]

mlir/test/Dialect/Linalg/fusion-elementwise-ops.mlir

926

Ah yes, I missed it among the rest of the IR.

Do you mind putting those together at the top of the function block with a comment and a TODO: extend elementwise fusion to handle such cases ?

Thanks!

[herhut]

Add comment in test

[nicolasvasilache]

mlir/test/Dialect/Linalg/fusion-elementwise-ops.mlir

926

To close the loop on this as we discussed offlie:
A fusion would be illegal anyway, so the compiler aborts in that case.

We know how to express and fuse this, it is similar to the shape-carrying operand we use in maxpooling: it is just NYI atm.
This shape-carrying operand may be the proper motivation to have the proper shape ops piped through linalg (TL;DR the output tensor we discussed a while back was not a proper motivation).

mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp

 #include "mlir/IR/Matchers.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/Support/LLVM.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 
 using namespace mlir;
 using namespace mlir::linalg;
 
+/// Append to `fusedOpIndexingMapAttrs` the indexing maps for the operands of
+/// the `producer` to use in the fused operation given the indexing map of the
+/// result of the producer in the consumer.
+static AffineMap getIndexingMapOfProducerOperandsInCoordinatesOfFusedOp(
+    OpOperand *producerOpOperand, AffineMap producerResultIndexMap,
+    AffineMap fusedConsumerArgIndexMap) {
+  // The indexing map in the consumer op (fusedConsumerArgIndexMap) is a map
+  // from consumer loop -> consumer arg tensor index/producer result tensor
+  // index. The fused loop is same as the consumer loop. For each producer arg
+  // the indexing map to be computed is a map from consumer loop -> producer
+  // arg tensor index.
+  // producerResultIndexMap is a map from producer loop -> tensor index.
+  // Compute the inverse to get map from tensor index -> producer loop.
+  // The inverse is a map from producer result tensor index -> producer loop.
+  AffineMap invProducerResultIndexMap =
+      inversePermutation(producerResultIndexMap);
+  assert(invProducerResultIndexMap &&
+         "expected producer result indexig map to be invertible");
+
+  LinalgOp producer = cast<LinalgOp>(producerOpOperand->getOwner());
+  // argMap is a map from producer loop -> producer arg tensor index.
+  AffineMap argMap = producer.getTiedIndexingMap(producerOpOperand);
+
+  // Compose argMap with invProducerResultIndexMap to get a map from
+  // producer result tensor index -> producer arg tensor index.
+  AffineMap t1 = argMap.compose(invProducerResultIndexMap);
+
+  // Compose t1 with fusedConsumerArgIndexMap gives an indexing map from
+  // consumer loop/ fused loop -> producer arg tensor index.
+  return t1.compose(fusedConsumerArgIndexMap);
+}
+
 /// Conditions for elementwise fusion of generic operations.
 static bool areElementwiseOpsFusable(GenericOp producer, GenericOp consumer,
                                      OpOperand *consumerOpOperand) {
   // Producer and consumer must have tensor semantics.
   if (!producer.hasTensorSemantics() || !consumer.hasTensorSemantics())
     return false;
 
   // Verify that
static bool areElementwiseOpsFusable(GenericOp producer, GenericOp consumer,
   // Currently support only operations with single result.
   if (producer.getNumOutputs() != 1)
     return false;
 
   // Finally the index_map for the result must be invertible. For now just
   // verify it is a permutation.
   AffineMap producerResultIndexMap =
       producer.getTiedIndexingMap(producer.getOutputOperand(0));
-  return producerResultIndexMap.isPermutation();
-}
-
-/// Append to `fusedOpIndexingMapAttrs` the indexing maps for the operands of
-/// the `producer` to use in the fused operation given the indexing map of the
-/// result of the producer in the consumer.
-static AffineMap getIndexingMapOfProducerOperandsInCoordinatesOfFusedOp(
-    OpOperand *producerOpOperand, AffineMap producerResultIndexMap,
-    AffineMap fusedConsumerArgIndexMap) {
-  // The indexing map in the consumer op (fusedConsumerArgIndexMap) is a map
-  // from consumer loop -> consumer arg tensor index/producer result tensor
-  // index. The fused loop is same as the consumer loop. For each producer arg
-  // the indexing map to be computed is a map from consumer loop -> producer
-  // arg tensor index.
-  // producerResultIndexMap is a map from producer loop -> tensor index.
-  // Compute the inverse to get map from tensor index -> producer loop.
-  // The inverse is a map from producer result tensor index -> producer loop.
-  AffineMap invProducerResultIndexMap =
-      inversePermutation(producerResultIndexMap);
-  assert(invProducerResultIndexMap &&
-         "expected producer result indexig map to be invertible");
+  if (!producerResultIndexMap.isPermutation())
+    return false;
 
-  LinalgOp producer = cast<LinalgOp>(producerOpOperand->getOwner());
-  // argMap is a map from producer loop -> producer arg tensor index.
-  AffineMap argMap = producer.getTiedIndexingMap(producerOpOperand);
+  // Ensure that the fusion does not remove size information required to
+  // get the loop bounds. For non-reduction generics, this is trivially the
+  // case due to the output operand. For reductions, we need to check that after
+  // the fusion, each loop dimension has at least one input that defines it.
+  if ((consumer.getNumReductionLoops())) {
+    llvm::BitVector coveredDims(consumer.getNumLoops(), false);
+
+    auto addToCoveredDims = [&](AffineMap map) {
+      for (auto result : map.getResults())
+        if (auto dimExpr = result.dyn_cast<AffineDimExpr>())
+          coveredDims[dimExpr.getPosition()] = true;
+    };
+
+    for (auto pair :
+         llvm::zip(consumer->getOperands(), consumer.getIndexingMaps())) {
+      Value operand = std::get<0>(pair);
+      if (operand == consumerOpOperand->get())
+        continue;
+      AffineMap operandMap = std::get<1>(pair);
+      addToCoveredDims(operandMap);
+    }
 
-  // Compose argMap with invProducerResultIndexMap to get a map from
-  // producer result tensor index -> producer arg tensor index.
-  AffineMap t1 = argMap.compose(invProducerResultIndexMap);
+    for (OpOperand *operand : producer.getInputOperands()) {
+      AffineMap newIndexingMap =
+          getIndexingMapOfProducerOperandsInCoordinatesOfFusedOp(
+              operand, producerResultIndexMap, consumerIndexMap);
+      addToCoveredDims(newIndexingMap);
+    }
+    if (!coveredDims.all())
+      return false;
+  }
 
-  // Compose t1 with fusedConsumerArgIndexMap gives an indexing map from
-  // consumer loop/ fused loop -> producer arg tensor index.
-  return t1.compose(fusedConsumerArgIndexMap);
+  return true;
 }
 
 /// Generate the region of the fused tensor operation. The region of the fused
 /// op must be empty.
 static void
 generateFusedElementwiseOpRegion(PatternRewriter &rewriter, GenericOp fusedOp,
                                  AffineMap consumerToProducerLoopsMap,
                                  OpOperand *consumerOpOperand,

mlir/test/Dialect/Linalg/fusion-elementwise-ops.mlir

%1 = linalg.generic {
     iterator_types = ["parallel", "parallel"]
   } ins(%input : tensor<2x3xf32>) outs(%init : tensor<3x2xf32>) {
   ^bb0(%arg1: f32, %arg2: f32):
     %add = arith.addf %arg1, %arg1 : f32
     linalg.yield %add : f32
   } -> tensor<3x2xf32>
   return %1 : tensor<3x2xf32>
 }
+
+// -----
+
+#map0 = affine_map<(d0, d1) -> ()>
+#map1 = affine_map<(d0, d1) -> (d0, d1)>
+#map2 = affine_map<(d0, d1) -> (d1, d0)>
+#map3 = affine_map<(d0, d1) -> (d0)>
+// CHECK-LABEL: @no_fusion_missing_reduction_shape
+func @no_fusion_missing_reduction_shape(%arg0: tensor<f32>, %arg1: index) -> tensor<?xf32> {
+  %cst = arith.constant 0xFF800000 : f32
+  %4 = linalg.init_tensor [%arg1, %arg1] : tensor<?x?xf32>
+  // CHECK: linalg.generic
+  %5 = linalg.generic {
+    indexing_maps = [#map0, #map1],
+    iterator_types = ["parallel", "parallel"]
+  } ins(%arg0 : tensor<f32>) outs(%4 : tensor<?x?xf32>) {
+  ^bb0(%arg2: f32, %arg3: f32):  // no predecessors
+    linalg.yield %arg2 : f32
+  } -> tensor<?x?xf32>
+  %6 = linalg.init_tensor [%arg1] : tensor<?xf32>
+  %7 = linalg.fill(%cst, %6) : f32, tensor<?xf32> -> tensor<?xf32>
+  // CHECK: linalg.generic
+  %8 = linalg.generic {
+    indexing_maps = [#map2, #map3],
+    iterator_types = ["parallel", "reduction"]
+  } ins(%5 : tensor<?x?xf32>) outs(%7 : tensor<?xf32>) {
+  ^bb0(%arg2: f32, %arg3: f32):  // no predecessors
+    %9 = maxf %arg2, %arg3 : f32
+    linalg.yield %9 : f32
+  } -> tensor<?xf32>
+  return %8 : tensor<?xf32>
+}