[mlir][Linalg] Generalize identity generic op removal to handle trivial broadcasts.

[mravishankar]

Summary

The existing pattern to remove identity generic ops when input is just
copied to the output looks for strict identity of the indexing
maps. This could be generalized to handle another case where the
indexing maps for the operand and result match, except for existence
of unit dimensions in either. In such cases use reshapes to replace
the generic op.
The existing pattern for removing identity generic op is superceded by
this pattern for ops with tensor semantics. The existing pattern is
now only used for the case where the op has buffer semantics.

Depends on D119365

[mravishankar]

Rebase on dependent patch.

[mravishankar]

Rebase.

[gysit]

Looks good to me.

I have one question though. What happens if the operand and the result contain unit-dims. My understanding is the code generates either an expand shape or a collapse shape. But could there be operand result pairs that collapse and expand at the same time (meaning there are unit-dims in both operand and result shape)? Shouldn't this be checked some how, for example, by counting the number of collapsed and expanded dims (one of them should always be zero). Or does this already exist and I did not see it?

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

819

Agreed :)

923

is this needed due to the expand / collapse shape? copy / broadcast generic should not require a cast?

933

nit: on buffers?

938

maybe EraseBufferCopyOp or similar would by a better name here?

mlir/test/Dialect/Linalg/canonicalize.mlir

785

nit: newline

[mravishankar]

In D119904#3329638, @gysit wrote:

Looks good to me.

I have one question though. What happens if the operand and the result contain unit-dims. My understanding is the code generates either an expand shape or a collapse shape. But could there be operand result pairs that collapse and expand at the same time (meaning there are unit-dims in both operand and result shape)? Shouldn't this be checked some how, for example, by counting the number of collapsed and expanded dims (one of them should always be zero). Or does this already exist and I did not see it?

Yeah, you are probably right. Maybe I can split the reassociation analysis to first try to fold just the operand, or just the result cause the mixing in with preserved 1-extent dims in result and operand can be painful. Ill test this pattern a bit more before submitting.

[mravishankar]

Address comments.

[mravishankar]

Minor fixes to comments.

[mravishankar]

Upload missed changes.

[gysit]

LGTM!

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

886–888

The linalg op interface's getRank(OpOperand*) also returns rank 0 for scalar operands.

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

for (OpOperand *opOperand : llvm::reverse(inputOperands)) {
       payload.eraseArgument(operandNumber);
     }
 
     rewriter.replaceOp(genericOp, newOp->getResults());
     return success();
   }
 };
 
-/// Remove generic operations (on tensors) that are just copying
-/// the values from inputs to the results. Requirements are
-/// 1) All iterator types are parallel
-/// 2) The body contains just a yield operation with the yielded values being
-///    the arguments corresponding to the operands.
-struct EraseIdentityGenericOp : public OpRewritePattern<GenericOp> {
+/// Given the operand and result indexing maps and types check if all
+/// expressions used to access the operand and the result are the same, except
+/// for unit dimensions.
+static Optional<SmallVector<ReassociationIndices>>
+checkForTrivialBroadcast(AffineMap operandIndexingMap, Type operandType,
+                         AffineMap resultIndexingMap, Type resultType) {
+  if (!operandIndexingMap.isProjectedPermutation() ||
+      !resultIndexingMap.isProjectedPermutation())
+    return llvm::None;
+  unsigned operandPos = 0, resultPos = 0;
+  ArrayRef<int64_t> operandShape = {}, resultShape = {};
+  if (auto operandTensorType = operandType.dyn_cast<RankedTensorType>())
+    operandShape = operandTensorType.getShape();
+  if (auto resultTensorType = resultType.dyn_cast<RankedTensorType>())
+    resultShape = resultTensorType.getShape();
+  ArrayRef<AffineExpr> operandExprs = operandIndexingMap.getResults(),
+                       resultExprs = resultIndexingMap.getResults();
+  SmallVector<ReassociationIndices> reassociation;
+  ReassociationIndices currReassociation;
+  unsigned foldedDim = 0;
+  while (operandPos < operandShape.size() && resultPos < resultShape.size()) {
+    // If the access expressions are the same, it could still be a copy.
+    if (operandExprs[operandPos] == resultExprs[resultPos]) {
+      currReassociation.push_back(foldedDim++);
+      reassociation.emplace_back(ReassociationIndices{});
+      std::swap(currReassociation, reassociation.back());
+      operandPos++, resultPos++;
+      continue;
+    }
+    // Check if the operand is unit-dimension, if so, could still be a copy.
+    if (operandShape[operandPos] == 1) {
+      currReassociation.push_back(foldedDim++);
+      operandPos++;
+      continue;
+    }
+    // Check if the operand is unit-dimension, if so, could still be a copy.
+    if (resultShape[resultPos] == 1) {
+      currReassociation.push_back(foldedDim++);
+      resultPos++;
+      continue;
+    }
+    // If none of these hold, not a trivial broadcast or copy.
+    return llvm::None;
+  }
+  // currReassociation should always be empty at this stage. If not just push it
+  // into the list.
+  if (!currReassociation.empty()) {
+    reassociation.emplace_back(std::move(currReassociation));
+  }
+
+  // For the case where operand or result is a 0D tensor (or scalar) the
+  // reassociation will be empty. In other cases, check for trailing dimensions
+  // in operand/result and append folding to the last ReassociationIndices.
+  ReassociationIndices trailingDimFolding;
+  auto checkForTrailingUnitDims =
+      [&](unsigned pos, ArrayRef<int64_t> shape) -> LogicalResult {
+    while (pos < shape.size()) {
+      if (shape[pos] != 1)
+        return failure();
+      pos++;
+      trailingDimFolding.push_back(foldedDim++);
+    }
+    return success();
+  };
+  if (failed(checkForTrailingUnitDims(operandPos, operandShape)) ||
+      failed(checkForTrailingUnitDims(resultPos, resultShape)))
+    return llvm::None;
+
+  if (!reassociation.empty())
+    reassociation.back().append(trailingDimFolding);
+
+  return reassociation;
+}
+
+/// Erase generic ops that are either copies or trivial broadcasts (where the
+/// broadcast dimension is 1) of the operands.
+struct EraseTrivialCopyOrBroadcastOp : public OpRewritePattern<GenericOp> {
   using OpRewritePattern<GenericOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(GenericOp genericOp,
                                 PatternRewriter &rewriter) const override {
-    // Check all indexing maps are identity.
-    if (llvm::any_of(genericOp.getIndexingMaps(),
-                     [](AffineMap map) { return !map.isIdentity(); }))
+    LinalgOp linalgOp = cast<LinalgOp>(genericOp.getOperation());
+    if (!genericOp.hasTensorSemantics())
       return failure();
 
-    // Check that the body of the linalg operation is just a linalg.yield
-    // operation.
+    // The body must be just a yield operation.
     Block &body = genericOp.region().front();
     if (!llvm::hasSingleElement(body))
       return failure();
     auto yieldOp = dyn_cast<linalg::YieldOp>(body.getTerminator());
     if (!yieldOp)
       return failure();
 
-    // In the buffer case, we need to check exact buffer equality.
-    if (genericOp.hasBufferSemantics()) {
-      if (genericOp.getNumInputs() == 1 && genericOp.getNumOutputs() == 1 &&
-          genericOp.getInputOperand(0)->get() ==
-              genericOp.getOutputOperand(0)->get()) {
-        rewriter.eraseOp(genericOp);
-        return success();
-      }
-      return failure();
-    }
-
-    // Get the argument number of the returned values. That is the operand
-    // number to use for replacing uses of this operation.
-    SmallVector<Value> returnedArgs;
+    struct ReplacementInfo {
+      unsigned operandNumber;
+      SmallVector<ReassociationIndices> reassociation;
+    };
+    SmallVector<ReplacementInfo> replacementInfo;
+    // Check if all operands are copy or trivial-broadcasts. Collect information
+    // needed, while not creating any ops (to avoid dead code generation on
+    // failure).
+    // TODO: It is possible to keep track of the non-copy/non-trivial-broadcast
+    // operands and create a new generic op. Avoiding it to reduce complexity
+    // here.
     for (const auto &yieldVal : llvm::enumerate(yieldOp.values())) {
       auto yieldArg = yieldVal.value().dyn_cast<BlockArgument>();
       if (!yieldArg || yieldArg.getOwner() != &body)
         return failure();
       unsigned argumentNumber = yieldArg.getArgNumber();
-      Value returnedArg = genericOp->getOperand(argumentNumber);
-      Type resultType = genericOp->getResult(yieldVal.index()).getType();
-      // The input can have a different type than the result, e.g. a dynamic
-      // input dimension can be turned into a static output dimension.
-      if (returnedArg.getType() != resultType)
-        returnedArg = rewriter.create<tensor::CastOp>(genericOp.getLoc(),
-                                                      resultType, returnedArg);
-      returnedArgs.push_back(returnedArg);
+      OpOperand &operand = genericOp->getOpOperand(argumentNumber);
+      OpResult result = genericOp->getResult(yieldVal.index());
+
+      Optional<SmallVector<ReassociationIndices>> reassociation =
+          checkForTrivialBroadcast(
+              linalgOp.getTiedIndexingMap(&operand), operand.get().getType(),
+              linalgOp.getTiedIndexingMapForResult(result), result.getType());
+
+      if (!reassociation)
+        return failure();
+      ReplacementInfo info{operand.getOperandNumber(),
+                           std::move(reassociation.getValue())};
+      replacementInfo.emplace_back(std::move(info));
+    }
+
+    // All operands are copy or trivial-broadcasts, find replacement values. They are either the operand
+    // value, or operand followed by a reshape to match the result shape.
+    SmallVector<Value> replacements;
+    Location loc = genericOp->getLoc();
+    replacements.reserve(genericOp->getNumResults());
+    for (const auto &info : llvm::enumerate(replacementInfo)) {
+      Value operand = genericOp->getOperand(info.value().operandNumber);
+      Value result = genericOp->getResult(info.index());
+      Value replacement = operand;
+      int64_t operandRank = 0, resultRank = 0;
+      if (auto operandTensorType =
+              operand.getType().dyn_cast<RankedTensorType>())
+        operandRank = operandTensorType.getRank();
+      if (auto resultTensorType = result.getType().dyn_cast<RankedTensorType>())
+        resultRank = resultTensorType.getRank();
+      if (operandRank > resultRank) {
+        replacement = rewriter.create<tensor::CollapseShapeOp>(
+            loc, replacement, info.value().reassociation);
+      } else if (operandRank < resultRank) {
+        replacement = rewriter.create<tensor::ExpandShapeOp>(
+            loc, result.getType(), replacement, info.value().reassociation);
+      }
+      if (replacement.getType() != result.getType()) {
+        replacement =
+            rewriter.create<tensor::CastOp>(loc, result.getType(), replacement);
       }
+      replacements.push_back(replacement);
+    }
+
+    rewriter.replaceOp(genericOp, replacements);
+    return success();
+  }
+};
+
+/// Remove generic operations (on tensors) that are just copying
+/// the values from inputs to the results. Requirements are
+/// 1) All iterator types are parallel
+/// 2) The body contains just a yield operation with the yielded values being
+///    the arguments corresponding to the operands.
+struct EraseIdentityGenericOp : public OpRewritePattern<GenericOp> {
+  using OpRewritePattern<GenericOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(GenericOp genericOp,
+                                PatternRewriter &rewriter) const override {
+    // In the buffer case, we need to check exact buffer equality.
+    if (!genericOp.hasBufferSemantics())
+      return failure();
+    if (genericOp.getNumInputs() != 1 || genericOp.getNumOutputs() != 1 ||
+        genericOp.getInputOperand(0)->get() !=
+            genericOp.getOutputOperand(0)->get())
+      return failure();
 
-    if (returnedArgs.size() != genericOp->getNumResults())
+    // The body must be just a yield operation.
+    Block &body = genericOp.region().front();
+    if (!llvm::hasSingleElement(body))
       return failure();
-    rewriter.replaceOp(genericOp, returnedArgs);
+    auto yieldOp = dyn_cast<linalg::YieldOp>(body.getTerminator());
+    if (!yieldOp)
+      return failure();
+
+    rewriter.eraseOp(genericOp);
     return success();
   }
 };
 } // namespace
 
 void GenericOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                             MLIRContext *context) {
-  results.add<DeduplicateGenericOpInputs, EraseIdentityGenericOp>(context);
+  results.add<DeduplicateGenericOpInputs, EraseIdentityGenericOp,
+              EraseTrivialCopyOrBroadcastOp>(context);
 }
 
 //===----------------------------------------------------------------------===//
 // InitTensorOp
 //===----------------------------------------------------------------------===//
 
 void InitTensorOp::build(OpBuilder &b, OperationState &result,
                          ArrayRef<OpFoldResult> sizes, Type elementType,

mlir/test/Dialect/Linalg/canonicalize.mlir

func @no_fold_pad_fill_value_mismatch() -> tensor<412x276xf32> {
   %fill = linalg.fill(%f0, %init) : f32, tensor<400x273xf32> -> tensor<400x273xf32>
   // CHECK: tensor.pad
   %pad = tensor.pad %fill low[4, 1] high[8, 2] {
   ^bb0(%arg1: index, %arg2: index):
     tensor.yield %f1 : f32
   } : tensor<400x273xf32> to tensor<412x276xf32>
   return %pad : tensor<412x276xf32>
 }
+
+// -----
+
+#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d0, d2, d3, d4, d1, d6, d5, d7, d8, d9, d10)>
+#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d8, d1, d6)>
+#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d3, d1, d6, d8)>
+func @trivial_broadcast(%arg0 : tensor<1x1x3x1x4x5x1x1x6x1x1xf32>, %arg1 : tensor<6x4x5xf32>) -> 
+  (tensor<6x4x5xf32>, tensor<3x4x5x6xf32>) {
+  %init = linalg.init_tensor [3, 4, 5, 6] : tensor<3x4x5x6xf32>
+  %0:2 = linalg.generic {
+    indexing_maps = [#map0, #map1, #map1, #map2],
+    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel",
+                      "parallel", "parallel", "parallel", "parallel", "parallel"]}
+    ins(%arg0, %arg1 : tensor<1x1x3x1x4x5x1x1x6x1x1xf32>, tensor<6x4x5xf32>)
+    outs(%arg1, %init : tensor<6x4x5xf32>, tensor<3x4x5x6xf32>) {
+      ^bb0(%b0: f32, %b1 : f32, %b2 : f32, %b3 : f32):
+        linalg.yield %b1, %b0 : f32, f32
+    } -> (tensor<6x4x5xf32>, tensor<3x4x5x6xf32>)
+  return %0#0, %0#1 : tensor<6x4x5xf32>, tensor<3x4x5x6xf32>
+}
+//      CHECK: func @trivial_broadcast(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<1x1x3x1x4x5x1x1x6x1x1xf32>
+// CHECK-SAME:     %[[ARG1:.+]]: tensor<6x4x5xf32>
+//      CHECK:   %[[RESHAPE:.+]] = tensor.collapse_shape %[[ARG0]] {{\[}}[0, 1, 2], [3, 4], [5], [6, 7, 8, 9, 10]{{\]}}
+//      CHECK:   return %[[ARG1]], %[[RESHAPE]]
+
+// -----
+
+#map0 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d0, d2, d3, d4, d1, d6, d5, d7, d8, d9, d10)>
+#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d8, d1, d6)>
+#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6, d7, d8, d9, d10) -> (d3, d1, d6, d8)>
+func @trivial_broadcast_result(%arg0 : tensor<3x4x5x6xf32>, %arg1 : tensor<6x4x5xf32>) -> 
+  (tensor<6x4x5xf32>, tensor<1x1x3x1x4x5x1x1x6x1x1xf32>) {
+  %init = linalg.init_tensor [1, 1, 3, 1, 4, 5, 1, 1, 6, 1, 1] : tensor<1x1x3x1x4x5x1x1x6x1x1xf32>
+  %0:2 = linalg.generic {
+    indexing_maps = [#map2, #map1, #map1, #map0],
+    iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel", "parallel",
+                      "parallel", "parallel", "parallel", "parallel", "parallel"]}
+    ins(%arg0, %arg1 : tensor<3x4x5x6xf32>, tensor<6x4x5xf32>)
+    outs(%arg1, %init : tensor<6x4x5xf32>, tensor<1x1x3x1x4x5x1x1x6x1x1xf32>) {
+      ^bb0(%b0: f32, %b1 : f32, %b2 : f32, %b3 : f32):
+        linalg.yield %b1, %b0 : f32, f32
+    } -> (tensor<6x4x5xf32>, tensor<1x1x3x1x4x5x1x1x6x1x1xf32>)
+  return %0#0, %0#1 : tensor<6x4x5xf32>, tensor<1x1x3x1x4x5x1x1x6x1x1xf32>
+}
+//      CHECK: func @trivial_broadcast_result(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<3x4x5x6xf32>
+// CHECK-SAME:     %[[ARG1:.+]]: tensor<6x4x5xf32>
+//      CHECK:   %[[RESHAPE:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0, 1, 2], [3, 4], [5], [6, 7, 8, 9, 10]{{\]}}
+//      CHECK:   return %[[ARG1]], %[[RESHAPE]]
+
+// -----
+
+func @scalar_copy(%arg0 : tensor<f32>) -> tensor<1x1xf32> {
+  %init = linalg.init_tensor [1, 1] : tensor<1x1xf32>
+  %0 = linalg.generic {
+    indexing_maps = [affine_map<(d0, d1) -> ()>, affine_map<(d0, d1) -> (d1, d0)>],
+    iterator_types = ["parallel", "parallel"]}
+    ins(%arg0 : tensor<f32>) outs(%init : tensor<1x1xf32>) {
+      ^bb0(%b0 : f32, %b1 : f32):
+        linalg.yield %b0 : f32
+    } -> tensor<1x1xf32>
+  return %0 : tensor<1x1xf32>
+}
+//      CHECK: func @scalar_copy(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<f32>
+//      CHECK:   %[[RESHAPE:.+]] = tensor.expand_shape %[[ARG0]] []
+//      CHECK:   return %[[RESHAPE]]
+
+// -----
+
+// Op cannot be removed if it is not just a copy.
+func @no_remove_op(%arg0 : tensor<?xf32>) -> tensor<?xf32> {
+  %0 = linalg.generic {
+    indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>],
+    iterator_types = ["parallel"]}
+    ins(%arg0 : tensor<?xf32>) outs(%arg0 : tensor<?xf32>) {
+      ^bb0(%b0 : f32, %b1 : f32):
+        %0 = arith.addf %b0, %b0 : f32
+        linalg.yield %0 : f32
+    } -> tensor<?xf32>
+  return %0 : tensor<?xf32>
+}
+//      CHECK: func @no_remove_op(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: tensor<?xf32>)
+// CHECK-NEXT:   %[[RETURN:.+]] = linalg.generic
+// CHECK-SAME:       ins(%[[ARG0]] :
+// CHECK-SAME:       outs(%[[ARG0]] :
+//      CHECK:   return %[[RETURN]]
+
+// -----
+
+func @no_remove_op_memref(%arg0 : memref<?xf32>, %arg1 : memref<?xf32>) {
+  linalg.generic {
+    indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> (d0)>],
+    iterator_types = ["parallel"]}
+    ins(%arg0 : memref<?xf32>) outs(%arg1 : memref<?xf32>) {
+      ^bb0(%b0 : f32, %b1 : f32):
+        %0 = arith.addf %b0, %b0 : f32
+        linalg.yield %0 : f32
+    }
+  return
+}
+//      CHECK: func @no_remove_op_memref(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]: memref<?xf32>
+// CHECK-SAME:     %[[ARG1:[a-zA-Z0-9]+]]: memref<?xf32>
+// CHECK-NEXT:   linalg.generic
+// CHECK-SAME:       ins(%[[ARG0]] :
+// CHECK-SAME:       outs(%[[ARG1]] :
+//      CHECK:   return
+
+// -----
+
+// Check that op is not canonicalized as a copy when one of the operands is not 
+// a trivial copy or broadcast.
+#map0 = affine_map<(d0, d1) -> (d0, d1)>
+#map1 = affine_map<(d0, d1) -> (d1)>
+func @no_remove_op_operand_not_trivial_broadcast(%arg0 : tensor<?x?xf32>,
+    %arg1 : tensor<?xf32>) -> (tensor<?x?xf32>, tensor<?x?xf32>) {
+  %0:2 = linalg.generic {
+    indexing_maps = [#map0, #map1, #map0, #map0],
+    iterator_types = ["parallel", "parallel"]}
+    ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?xf32>)
+    outs(%arg0, %arg0 : tensor<?x?xf32>, tensor<?x?xf32>) {
+      ^bb0(%b0 : f32, %b1 : f32, %b2 : f32, %b3 : f32):
+        linalg.yield %b0, %b1 : f32, f32
+    } -> (tensor<?x?xf32>, tensor<?x?xf32>)
+  return %0#0, %0#1 : tensor<?x?xf32>, tensor<?x?xf32>
+}
+// CHECK-LABEL: func @no_remove_op_operand_not_trivial_broadcast(
+//       CHECK:   %[[RETURN:.+]]:2 = linalg.generic
+//       CHECK:   return %[[RETURN]]#0, %[[RETURN]]#1
 No newline at end of file