[mlir][linalg] Simplifying the SplitReduction logic that uses the control to get the dimension where the extra parallel dimension is inserted

[vmurali]

Summary

Currently, the innerParallel and non innerParallel strategies use two different ways to fix for where the extra loop is inserted and where the extra dimension for the intermediate result is inserted - innerParallel adds the extra (parallel) loop right after the pre-existing reduction loop, whereas non innerParallel adds the reduction loop in the successor to the index supplied by control, and the parallel loop in the index supplied by the control. The semantics of the index supplied by the control is supposed to only control where the extra tensor dimension is inserted in the intermediate tensor. Conflating this index with where the reduction (and parallel) loops are inserted leads to more complex (and confusing) logic overall. This differential removes conflating the two uses of the index, and keeps the reduction and parallel loops in the same vicinity and uses the supplied index to only determine the position of the extra tensor dimension. It also simplifies the code by merging the two strategies in a lot more places.

[ThomasRaoux]

LGTM, thanks!

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

 FailureOr<SplitReductionResult> mlir::linalg::splitReduction(
     PatternRewriter &b, LinalgOp op,
     const ControlSplitReductionFn &controlSplitReductionFn, bool useAlloc) {
   OpBuilder::InsertionGuard guard(b);
   b.setInsertionPoint(op);
 
   SplitReductionOptions control = controlSplitReductionFn(op);
   int64_t ratio = control.ratio;
-  unsigned insertSplitDimension = control.index;
+  unsigned insertSplitIndex = control.index;
   if (ratio <= 1)
     return b.notifyMatchFailure(op, "split ratio needs to be greater than 1");
 
   SmallVector<unsigned> dims;
   op.getReductionDims(dims);
   assert(dims.size() == 1);
   unsigned reductionDim = dims[0];
   SmallVector<int64_t, 4> loopRanges = op.getStaticLoopRanges();
   int64_t reductionDimSize = loopRanges[reductionDim];
   if (reductionDimSize == ShapedType::kDynamicSize ||
-      reductionDimSize % ratio != 0 ||
-      insertSplitDimension >= loopRanges.size())
+      reductionDimSize % ratio != 0)
     return b.notifyMatchFailure(
         op, "Reduction dimension not divisible by split ratio");
+  if (op.getNumDpsInits() != 1)
+    return b.notifyMatchFailure(op, "More than one output in split reduction");
+  if (insertSplitIndex > op.getShape(op.getDpsInitOperand(0)).size())
+    return b.notifyMatchFailure(op, "Insert dimension position too large "
+                                    "compared to intermediate tensor size");
 
   SmallVector<Operation *, 4> combinerOps;
   if (!matchReduction(op.getRegionOutputArgs(), 0, combinerOps) ||
       combinerOps.size() != 1)
     return b.notifyMatchFailure(op, "Cannot match the reduction pattern");
 
   Operation *reductionOp = combinerOps[0];
   Optional<Attribute> identity = getNeutralElement(reductionOp);
for (unsigned idx : llvm::seq<unsigned>(0, map.getNumResults())) {
       if (reductionDim == dim) {
         if (control.innerParallel) {
           newShape.push_back(op.getShape(operand)[idx] / ratio);
           newShape.push_back(ratio);
         } else {
           newShape.push_back(ratio);
           newShape.push_back(op.getShape(operand)[idx] / ratio);
         }
-        reassociation.push_back({index++, index++});
-        if (control.innerParallel) {
-          exprs.push_back(b.getAffineDimExpr(reductionDim));
-          exprs.push_back(b.getAffineDimExpr(reductionDim + 1));
-        } else {
-          exprs.push_back(b.getAffineDimExpr(insertSplitDimension));
-          exprs.push_back(
-              b.getAffineDimExpr(dim < insertSplitDimension ? dim : dim + 1));
-        }
+        reassociation.push_back({index++, index++});
         continue;
       }
       newShape.push_back(op.getShape(operand)[idx]);
-      if (control.innerParallel) {
-        exprs.push_back(
-            b.getAffineDimExpr(dim <= reductionDim ? dim : dim + 1));
-      } else {
-        exprs.push_back(
-            b.getAffineDimExpr(dim < insertSplitDimension ? dim : dim + 1));
-      }
+      exprs.push_back(b.getAffineDimExpr(dim < reductionDim ? dim : dim + 1));
       reassociation.push_back({index++});
     }
     newMaps.push_back(
         AffineMap::get(map.getNumDims() + 1, 0, exprs, op.getContext()));
     // If the shape is unchanged the input doesn't change.
     if (newShape == op.getShape(operand)) {
       newInputs.push_back(operand->get());
       continue;
     }
     Type newType = RankedTensorType::get(
         newShape,
         operand->get().getType().cast<RankedTensorType>().getElementType());
     Value newInput = b.create<tensor::ExpandShapeOp>(
         loc, newType, operand->get(), reassociation);
     newInputs.push_back(newInput);
   }
 
   // Calculate the new output map and shape, we insert the new dimension based
   // on the index returned by `controlSplitReductionFn`.
   SmallVector<int64_t> newOutputShape;
   AffineMap oldOutputMap = op.getMatchingIndexingMap(op.getDpsInitOperand(0));
   ArrayRef<int64_t> oldShape = op.getShape(op.getDpsInitOperand(0));
   SmallVector<AffineExpr> outputExpr;
-  for (unsigned idx :
-       llvm::seq<unsigned>(0, oldOutputMap.getNumResults() + 1)) {
-    if (idx == insertSplitDimension) {
+  for (unsigned idx : llvm::seq<unsigned>(0, oldShape.size() + 1)) {
+    if (insertSplitIndex == idx) {
       newOutputShape.push_back(ratio);
       if (control.innerParallel) {
         outputExpr.push_back(b.getAffineDimExpr(reductionDim + 1));
       } else {
-        outputExpr.push_back(b.getAffineDimExpr(insertSplitDimension));
+        outputExpr.push_back(b.getAffineDimExpr(reductionDim));
       }
-      continue;
     }
-    unsigned oldIdx = idx < insertSplitDimension ? idx : idx - 1;
-    newOutputShape.push_back(oldShape[oldIdx]);
-    unsigned dim = oldOutputMap.getDimPosition(oldIdx);
-    if (control.innerParallel) {
-      outputExpr.push_back(
-          b.getAffineDimExpr(dim <= reductionDim ? dim : dim + 1));
-    } else {
+    if (idx < oldShape.size()) {
+      newOutputShape.push_back(oldShape[idx]);
+      unsigned dim = oldOutputMap.getDimPosition(idx);
       outputExpr.push_back(
-          b.getAffineDimExpr(dim < insertSplitDimension ? dim : dim + 1));
+          b.getAffineDimExpr(dim < reductionDim ? dim : dim + 1));
     }
   }
   Value emptyOrAllocTensor;
   if (useAlloc) {
     emptyOrAllocTensor = b.create<bufferization::AllocTensorOp>(
         loc,
         RankedTensorType::get(newOutputShape,
                               op.getRegionOutputArgs()[0].getType()),
         ValueRange{});
   } else {
     emptyOrAllocTensor = b.create<tensor::EmptyOp>(
         loc, newOutputShape, op.getRegionOutputArgs()[0].getType());
   }
   Value constantOp = b.create<arith::ConstantOp>(loc, *identity);
   Value identityTensor =
       b.create<linalg::FillOp>(op->getLoc(), constantOp, emptyOrAllocTensor)
           .getResult(0);
 
   newMaps.push_back(AffineMap::get(oldOutputMap.getNumDims() + 1, 0, outputExpr,
                                    op.getContext()));
   SmallVector<StringRef> newIteratorTypes;
   for (auto &it : llvm::enumerate(op.getIteratorTypesArray())) {
-    if (insertSplitDimension == it.index() && !control.innerParallel)
+    if (reductionDim == it.index() && !control.innerParallel)
       newIteratorTypes.push_back(getParallelIteratorTypeName());
     newIteratorTypes.push_back(it.value());
-    if (insertSplitDimension == it.index() && control.innerParallel)
+    if (reductionDim == it.index() && control.innerParallel)
       newIteratorTypes.push_back(getParallelIteratorTypeName());
   }
   // Create the new op matching the original op with an extra parallel
   // dimension.
   GenericOp genericOp = b.create<GenericOp>(
       loc, TypeRange({emptyOrAllocTensor.getType()}), newInputs,
       ValueRange({identityTensor}), newMaps, newIteratorTypes);
   b.inlineRegionBefore(op->getRegion(0), genericOp.getRegion(),
                        genericOp.getRegion().begin());
 
   // Then create a new reduction that only reduce the newly added dimension
   // from the previous op.
   unsigned intermRank = newOutputShape.size();
   AffineMap inputMap = b.getMultiDimIdentityMap(intermRank);
   SmallVector<StringRef> reductionIteratorTypes;
   SmallVector<AffineExpr> exprs;
   for (unsigned i : llvm::seq<unsigned>(0, intermRank)) {
-    if (insertSplitDimension == i) {
+    if (insertSplitIndex == i) {
       reductionIteratorTypes.push_back(getReductionIteratorTypeName());
     } else {
       exprs.push_back(b.getAffineDimExpr(i));
       reductionIteratorTypes.push_back(getParallelIteratorTypeName());
     }
   }
   AffineMap outputMap = AffineMap::get(intermRank, 0, exprs, op.getContext());
   SmallVector<AffineMap> reductionMaps = {inputMap, outputMap};

mlir/test/Dialect/Linalg/transform-op-split-reduction.mlir

%0 = linalg.generic {
     ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):
       %3 = arith.addf %arg0, %arg1 : f32
       %4 = arith.maxf %3, %arg2 : f32
       linalg.yield %4 : f32
     } -> tensor<5x2xf32>
   return %0 : tensor<5x2xf32>
 }
 
-//  CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d2, d1, d0)>
-//  CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, d2, d1)>
-//  CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, d0, d2)>
+//  CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, d2, d0)>
+//  CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, d1, d2)>
+//  CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d3, d0, d1)>
 //  CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
 //  CHECK-DAG: #[[$MAP4:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)>
 // CHECK-LABEL:  func @generic_split_3d
 //  CHECK-DAG: %[[ID:.*]] = arith.constant 0xFF800000 : f32
 //  CHECK-DAG: %[[I1:.*]] = tensor.expand_shape %{{.*}}[0, 1], [2]] : tensor<32x2xf32> into tensor<4x8x2xf32>
 //  CHECK-DAG: %[[I2:.*]] = tensor.expand_shape %{{.*}}[0], [1, 2]] : tensor<5x32xf32> into tensor<5x4x8xf32>
 //  CHECK-DAG: %[[INI:.*]] = tensor.empty() : tensor<5x2x4xf32>
 //      CHECK: %[[F:.*]] = linalg.fill ins(%[[ID]] : f32) outs(%[[INI]] : tensor<5x2x4xf32>) -> tensor<5x2x4xf32>
-//      CHECK: %[[G:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP2]]], iterator_types = ["parallel", "reduction", "parallel", "parallel"]}
+//      CHECK: %[[G:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "reduction", "parallel"]}
 // CHECK-SAME:   ins(%[[I1]], %[[I2]] : tensor<4x8x2xf32>, tensor<5x4x8xf32>) outs(%[[F]] : tensor<5x2x4xf32>) {
 //      CHECK:   arith.addf
 //      CHECK:   arith.maxf
 //      CHECK:   linalg.yield
 //      CHECK: } -> tensor<5x2x4xf32>
 //      CHECK: %[[R:.*]] = linalg.generic {indexing_maps = [#[[$MAP3]], #[[$MAP4]]], iterator_types = ["parallel", "parallel", "reduction"]}
 // CHECK-SAME:   ins(%[[G]] : tensor<5x2x4xf32>) outs(%{{.*}} : tensor<5x2xf32>) {
 //      CHECK:   arith.maxf