[mlir] Add lowering for IsBroadcastable to Std dialect.

[tpopp]

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

571

Just a note. I have deleted this locally.

[frgossen]

Looks good!

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

228

If you want to enforce this, you can compare the types below to shape::getExtentTensorType.

mlir/test/Conversion/ShapeToStandard/shape-to-standard.mlir

394

nit: indentation

403

Maybe test this with one shape of statically known rank?

409

Name, e.g. I

412

SMALLER_EXTENT_INDEX?

[tpopp]

I'm going to land this with these changes because the code was already in another commit and didn't receive any strong negative feedback, so I'm hoping there are only nits left if anything. I will follow up on any comments.

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

228

Right now this is enforced by checking that neither input is a Shape. I'm going to leave this to ease future find replace when this is changed because I copied this from Broadcast lowering in this file.

mlir/lib/Conversion/ShapeToStandard/ShapeToStandard.cpp

 LogicalResult ConstSizeOpConversion::matchAndRewrite(
     ConstSizeOp op, ArrayRef<Value> operands,
     ConversionPatternRewriter &rewriter) const {
   rewriter.replaceOpWithNewOp<ConstantIndexOp>(op, op.value().getSExtValue());
   return success();
 }
 
 namespace {
+struct IsBroadcastableOpConverter
+    : public OpConversionPattern<IsBroadcastableOp> {
+  using OpConversionPattern<IsBroadcastableOp>::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(IsBroadcastableOp op, ArrayRef<Value> operands,
+                  ConversionPatternRewriter &rewriter) const override;
+};
+} // namespace
+
+// A lowering to remove the implicit broadcasting logic of
+// shape.cstr_broadcastable without removing the benefits of constraints
+// compared yields. To actually remove the constraint, look to
+// ConvertShapeConstraints.cpp.
+LogicalResult IsBroadcastableOpConverter::matchAndRewrite(
+    IsBroadcastableOp op, ArrayRef<Value> operands,
+    ConversionPatternRewriter &rewriter) const {
+  // For now, this lowering is only defined on `tensor<?xindex>` operands, not
+  // on shapes.
+  IsBroadcastableOp::Adaptor transformed(operands);
+  if (transformed.lhs().getType().isa<ShapeType>() ||
+      transformed.rhs().getType().isa<ShapeType>())
+    return failure();
+
+  auto loc = op.getLoc();
+  Value zero = rewriter.create<ConstantIndexOp>(loc, 0);
+  Value one = rewriter.create<ConstantIndexOp>(loc, 1);
+
+  // Find smaller and greater rank and extent tensor.
+  Value lhsRank = rewriter.create<DimOp>(loc, transformed.lhs(), zero);
+  Value rhsRank = rewriter.create<DimOp>(loc, transformed.rhs(), zero);
+  Value lhsRankULE =
+      rewriter.create<CmpIOp>(loc, CmpIPredicate::ule, lhsRank, rhsRank);
+  Type indexTy = rewriter.getIndexType();
+  Value lesserRank =
+      rewriter.create<SelectOp>(loc, lhsRankULE, lhsRank, rhsRank);
+  Value greaterRank =
+      rewriter.create<SelectOp>(loc, lhsRankULE, rhsRank, lhsRank);
+  auto erasedRankType =
+      RankedTensorType::get({ShapedType::kDynamicSize}, indexTy);
+  Value rankErasedLhs =
+      rewriter.create<TensorCastOp>(loc, erasedRankType, transformed.lhs());
+  Value rankErasedRhs =
+      rewriter.create<TensorCastOp>(loc, erasedRankType, transformed.rhs());
+  Value lesserRankOperand =
+      rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedLhs, rankErasedRhs);
+  Value greaterRankOperand =
+      rewriter.create<SelectOp>(loc, lhsRankULE, rankErasedRhs, rankErasedLhs);
+  Value rankDiff =
+      rewriter.create<SubIOp>(loc, indexTy, greaterRank, lesserRank);
+  Type i1Ty = rewriter.getI1Type();
+  Value init =
+      rewriter.create<ConstantOp>(loc, i1Ty, rewriter.getBoolAttr(true));
+
+  // Determine if all overlapping extents are broadcastable.
+  auto reduceResult = rewriter.create<ForOp>(
+      loc, rankDiff, greaterRank, one, ValueRange{init},
+      [&](OpBuilder &b, Location loc, Value iv, ValueRange iterArgs) {
+        Value greaterRankOperandExtent =
+            b.create<ExtractElementOp>(loc, greaterRankOperand, ValueRange{iv});
+        Value greaterRankOperandExtentIsOne = b.create<CmpIOp>(
+            loc, CmpIPredicate::eq, greaterRankOperandExtent, one);
+        Value ivShifted = b.create<SubIOp>(loc, indexTy, iv, rankDiff);
+        Value lesserRankOperandExtent = b.create<ExtractElementOp>(
+            loc, lesserRankOperand, ValueRange{ivShifted});
+        Value lesserRankOperandExtentIsOne = b.create<CmpIOp>(
+            loc, CmpIPredicate::eq, lesserRankOperandExtent, one);
+        Value extentsAreEqual =
+            b.create<CmpIOp>(loc, CmpIPredicate::eq, greaterRankOperandExtent,
+                             lesserRankOperandExtent);
+        Value broadcastableExtents = b.create<AndOp>(
+            loc, iterArgs[0],
+            b.create<OrOp>(loc,
+                           b.create<OrOp>(loc, greaterRankOperandExtentIsOne,
+                                          lesserRankOperandExtentIsOne),
+                           extentsAreEqual));
+        b.create<scf::YieldOp>(loc, broadcastableExtents);
+      });
+
+  rewriter.replaceOp(op, reduceResult.results().front());
+  return success();
+}
+
+namespace {
 class GetExtentOpConverter : public OpConversionPattern<GetExtentOp> {
   using OpConversionPattern<GetExtentOp>::OpConversionPattern;
 
   LogicalResult
   matchAndRewrite(GetExtentOp op, ArrayRef<Value> operands,
                   ConversionPatternRewriter &rewriter) const override;
 };
 } // namespace
matchAndRewrite(ToExtentTensorOp op, ArrayRef<Value> operands,
     if (!adaptor.input().getType().isa<RankedTensorType>())
       return rewriter.notifyMatchFailure(op, "input needs to be a tensor");
 
     rewriter.replaceOpWithNewOp<TensorCastOp>(op, adaptor.input(),
                                               op.getType());
     return success();
   }
 };
+
 } // namespace
 
 namespace {
 /// Conversion pass.
 class ConvertShapeToStandardPass
     : public ConvertShapeToStandardBase<ConvertShapeToStandardPass> {
 
   void runOnOperation() override;
void mlir::populateShapeToStandardConversionPatterns(
   // clang-format off
   patterns.insert<
       AnyOpConversion,
       BinaryOpConversion<AddOp, AddIOp>,
       BinaryOpConversion<MulOp, MulIOp>,
       BroadcastOpConverter,
       ConstShapeOpConverter,
       ConstSizeOpConversion,
+      IsBroadcastableOpConverter,
       GetExtentOpConverter,
       RankOpConverter,
       ReduceOpConverter,
       ShapeEqOpConverter,
       ShapeOfOpConversion,
       ToExtentTensorOpConversion>(ctx);
   // clang-format on
 }
 
 std::unique_ptr<OperationPass<ModuleOp>>
 mlir::createConvertShapeToStandardPass() {
   return std::make_unique<ConvertShapeToStandardPass>();
 }

mlir/test/Conversion/ShapeToStandard/shape-to-standard.mlir

func @broadcast_known_different_extents(%a : tensor<2xindex>, %b : tensor<3xindex>) {
   // CHECK:             yield %[[OUTPUT_EXTENT:.*]] : index
   // CHECK:           } : tensor<?xindex>
   // CHECK:           return
   // CHECK:         }
   %0 = shape.broadcast %a, %b
       : tensor<2xindex>, tensor<3xindex> -> tensor<?xindex>
   return
 }
+
+// -----
+
+func @try_is_broadcastable(%a : tensor<?xindex>, %b : tensor<?xindex>) -> i1 {
+  %0 = shape.is_broadcastable %a, %b : tensor<?xindex>, tensor<?xindex>
+  return %0 : i1
+}
+
+// CHECK-LABEL:   func @try_is_broadcastable(
+// CHECK-SAME:                    %[[LHS:.*]]: tensor<?xindex>,
+// CHECK-SAME:                    %[[RHS:.*]]: tensor<?xindex>) -> i1 {
+// CHECK:           %[[C0:.*]] = constant 0 : index
+// CHECK:           %[[C1:.*]] = constant 1 : index
+// CHECK:           %[[LHS_RANK:.*]] = dim %[[LHS]], %[[C0]] : tensor<?xindex>
+// CHECK:           %[[RHS_RANK:.*]] = dim %[[RHS]], %[[C0]] : tensor<?xindex>
+// CHECK:           %[[LHS_SMALLER:.*]] = cmpi "ule", %[[LHS_RANK]], %[[RHS_RANK]] : index
+// CHECK:           %[[SMALLER_RANK:.*]] = select %[[LHS_SMALLER]], %[[LHS_RANK]], %[[RHS_RANK]] : index
+// CHECK:           %[[LARGER_RANK:.*]] = select %[[LHS_SMALLER]], %[[RHS_RANK]], %[[LHS_RANK]] : index
+// CHECK:           %[[RANK_ERASED_LHS:.*]] = tensor_cast %[[LHS]] : tensor<?xindex> to tensor<?xindex>
+// CHECK:           %[[RANK_ERASED_RHS:.*]] = tensor_cast %[[RHS]] : tensor<?xindex> to tensor<?xindex>
+// CHECK:           %[[SMALLER_SHAPE:.*]] = select %[[LHS_SMALLER]], %[[RANK_ERASED_LHS]], %[[RANK_ERASED_RHS]] : tensor<?xindex>
+// CHECK:           %[[LARGER_SHAPE:.*]] = select %[[LHS_SMALLER]], %[[RANK_ERASED_RHS]], %[[RANK_ERASED_LHS]] : tensor<?xindex>
+// CHECK:           %[[RANK_DIFF:.*]] = subi %[[LARGER_RANK]], %[[SMALLER_RANK]] : index
+// CHECK:           %[[TRUE:.*]] = constant true
+// CHECK:           %[[ALL_RESULT:.*]] = scf.for %[[VAL_16:.*]] = %[[RANK_DIFF]] to %[[LARGER_RANK]] step %[[C1]] iter_args(%[[ALL_SO_FAR:.*]] = %[[TRUE]]) -> (i1) {
+// CHECK:             %[[LARGER_EXTENT:.*]] = extract_element %[[LARGER_SHAPE]]{{\[}}%[[VAL_16]]] : tensor<?xindex>
+// CHECK:             %[[LARGER_EXTENT_IS_ONE:.*]] = cmpi "eq", %[[LARGER_EXTENT]], %[[C1]] : index
+// CHECK:             %[[LHS_EXTENT_INDEX:.*]] = subi %[[VAL_16]], %[[RANK_DIFF]] : index
+// CHECK:             %[[SMALLER_EXTENT:.*]] = extract_element %[[SMALLER_SHAPE]]{{\[}}%[[LHS_EXTENT_INDEX]]] : tensor<?xindex>
+// CHECK:             %[[SMALLER_EXTENT_IS_ONE:.*]] = cmpi "eq", %[[SMALLER_EXTENT]], %[[C1]] : index
+// CHECK:             %[[EXTENTS_ARE_EQUAL:.*]] = cmpi "eq", %[[LARGER_EXTENT]], %[[SMALLER_EXTENT]] : index
+// CHECK:             %[[EITHER_EXTENT_IS_ONE:.*]] = or %[[LARGER_EXTENT_IS_ONE]], %[[SMALLER_EXTENT_IS_ONE]] : i1
+// CHECK:             %[[OR_EXTENTS_ARE_EQUAL:.*]] = or %[[EITHER_EXTENT_IS_ONE]], %[[EXTENTS_ARE_EQUAL]] : i1
+// CHECK:             %[[NEW_ALL_SO_FAR:.*]] = and %[[ALL_SO_FAR]], %[[OR_EXTENTS_ARE_EQUAL]] : i1
+// CHECK:             scf.yield %[[NEW_ALL_SO_FAR]] : i1
+// CHECK:           }
+// CHECK:           return %[[ALL_RESULT]] : i1
+// CHECK:         }