[MLIR][TORCH] Add support for multiple indexing tensors for aten.index.Tensor

include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td

@@ -6619,6 +6619,7 @@ def Torch_AtenAddTOp : Torch_Op<"aten.add.t", [
       printDefaultTorchOp(printer, *this, 2, 1);
     }
   }];
+  let hasCanonicalizer = 1;
 }
 
 def Torch_AtenEqIntListOp : Torch_Op<"aten.eq.int_list", [

lib/Conversion/TorchToLinalg/IndirectDataMovement.cpp

@@ -244,13 +244,29 @@ class ConvertAtenIndexSelectOp : public OpConversionPattern<AtenIndexSelectOp> {
 };
 } // namespace
 
+// IndexTensor for multiple input tensors broadcasts their shapes to a common
+// shape and then replaces the indexed dims with the indices given by the
+// indexing tensors:
+// x[i_1, i_2, ..., i_M] = result
+// result[...] = x[i_1[...], i_2[...], ..., i_M[...]]
+//
+// where the result shape is computed as follows:
+// 1. broadcast i_1, i_2, ..., i_M to a common shape
+// 2. if i_1, i_2, ..., i_M is not contiguous, transpose the broadcasted
+//    shape to the beginning of the result shape, while removing the
+//    unchanged dims (marked by None)
+// 3. Otherwise replace the indexed dims with the broadcasted shape
+//
+// e.g. x: [2, 3]
+//      x[[4], [6, 1]] -> x[6, 4]
 namespace {
 class ConvertAtenIndexTensorOp : public OpConversionPattern<AtenIndexTensorOp> {
 public:
   using OpConversionPattern::OpConversionPattern;
   LogicalResult
   matchAndRewrite(AtenIndexTensorOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
+
     if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
       return failure();
 
@@ -266,78 +282,165 @@ class ConvertAtenIndexTensorOp : public OpConversionPattern<AtenIndexTensorOp> {
     SmallVector<Value> indicesVal =
         getTypeConvertedValues(rewriter, loc, getTypeConverter(), indicesTuple);
 
-    int indexTensorDim = -1;
+    // Identify the indices with non-None index tensors and determine if they
+    // are contiguous within the input list.
+    SmallVector<int> indexTensorDims;
+    SmallVector<Value> indexTensors;
+    bool contiguous = true;
     for (auto i : llvm::seq(0, (int)indicesVal.size())) {
       Value index = indicesVal[i];
       if (!index || failed(checkNotNone(rewriter, op, index)))
         continue;
-      if (indexTensorDim >= 0) {
-        return rewriter.notifyMatchFailure(
-            op, "unimplemented: only one index tensor allowed");
-      }
-      indexTensorDim = i;
+      if (!indexTensorDims.empty() && indexTensorDims.back() != i - 1)
+        contiguous = false;
+      indexTensorDims.push_back(i);
+      indexTensors.push_back(index);
     }
 
-    if (indexTensorDim == -1) {
+    if (indexTensors.empty()) {
       return rewriter.notifyMatchFailure(
-          op, "unimplemented: index tensor must not be None");
+          op, "aten.index.Tensor: index tensor must not be None");
     }
 
-    Value indexTensor = indicesVal[indexTensorDim];
     RankedTensorType inputType = input.getType().cast<RankedTensorType>();
-    RankedTensorType indexTensorType =
-        indexTensor.getType().cast<RankedTensorType>();
     RankedTensorType resultType = getTypeConverter()
                                       ->convertType(op->getResult(0).getType())
                                       .cast<RankedTensorType>();
     Type elementType = resultType.getElementType();
     int inputRank = inputType.getRank();
-    int indexTensorRank = indexTensorType.getRank();
+    int resultRank = resultType.getRank();
+    int firstIndexDim = indexTensorDims[0];
+    int replacedIndexCount = indexTensorDims.size();
+    int64_t startIndex = contiguous ? firstIndexDim : 0;
+
+    // Currently we only support statically sized index tensors
+    // when there is more than one index tensor.
+    // TODO: Add support for dynamic size index tensors. This will probably
+    // require broadcasting the index tensors to a common shape.
+    SmallVector<Value> broadcastedIndexShape;
+    if (indexTensors.size() > 1) {
+      int maxRank = -1;
+      for (auto indexTensor : indexTensors) {
+        RankedTensorType indexTensorType =
+            indexTensor.getType().cast<RankedTensorType>();
+        maxRank = std::max(maxRank, (int)indexTensorType.getRank());
+      }
+
+      // Because we are assuming static shapes, we can get the shape of the
+      // broadcasted index tensors from the shape refinement pass
+      auto refinedResultShape = resultType.getShape();
+      for (auto i : llvm::seq(startIndex, startIndex + maxRank)) {
+        auto resultDimSize = refinedResultShape[i];
+        if (ShapedType::isDynamic(resultDimSize)) {
+          return rewriter.notifyMatchFailure(
+              op, "unimplemented: index tensors must have static shape if "
+                  "there is more than one index tensor");
+        }
+        broadcastedIndexShape.push_back(
+            getConstant(rewriter, loc, resultDimSize, rewriter.getIndexType()));
+      }
+    } else {
+      // For a single indexing tensor we can simply use its (dynamic) sizes
+      broadcastedIndexShape =
+          getTensorSizes(rewriter, loc, indexTensors.front());
+    }
 
     // This result shape calculation assumes that there is only one
-    // index tensor of the input tensor. The calculation for arbitrary inputs is
-    // much more complex.
+    // index tensor, or all of the index tensors are statically shaped.
+    int broadcastRank = broadcastedIndexShape.size();
+
     SmallVector<Value> resultShape;
-    for (auto i : llvm::seq(0, indexTensorDim)) {
-      resultShape.push_back(getDimOp(rewriter, loc, input, i));
-    }
-    for (auto i : llvm::seq(0, indexTensorRank)) {
-      resultShape.push_back(getDimOp(rewriter, loc, indexTensor, i));
-    }
-    for (auto i : llvm::seq(indexTensorDim + 1, inputRank)) {
-      resultShape.push_back(getDimOp(rewriter, loc, input, i));
+    if (contiguous) {
+      for (auto i : llvm::seq(0, firstIndexDim)) {
+        resultShape.push_back(getDimOp(rewriter, loc, input, i));
+      }
+      resultShape.append(broadcastedIndexShape);
+      for (auto i : llvm::seq((int)resultShape.size(), resultRank)) {
+        resultShape.push_back(getDimOp(rewriter, loc, input,
+                                       i - broadcastRank + replacedIndexCount));
+      }
+    } else {
+      resultShape.append(broadcastedIndexShape);
+      int j = 0;
+      for (auto i : llvm::seq(0, inputRank)) {
+        if (j < replacedIndexCount && i == indexTensorDims[j]) {
+          j++;
+          continue;
+        }
+        resultShape.push_back(getDimOp(rewriter, loc, input, i));
+      }
     }
-    int resultRank = resultShape.size();
 
+    // Initialize the indexing maps for the generic op. Because we are assuming
+    // static shapes for the indexing tensors when there are more than 1, we can
+    // safely map all size 1 dims to 0 in the corresponding affine maps.
+    // TODO: For dynamic shapes, we have to either broadcast the index tensors
+    // to a common shape or introduce some form of control flow.
     Value initTensor =
         rewriter.create<linalg::InitTensorOp>(loc, resultShape, elementType);
-    SmallVector<AffineExpr> indicesExpr, resultExpr;
+    SmallVector<AffineMap> indexingMaps;
     SmallVector<StringRef> iteratorTypes;
 
-    for (auto i : llvm::seq(indexTensorDim, indexTensorDim + indexTensorRank))
-      indicesExpr.push_back(rewriter.getAffineDimExpr(i));
+    for (auto indexTensor : indexTensors) {
+      RankedTensorType indexTensorType =
+          indexTensor.getType().cast<RankedTensorType>();
+      auto indexTensorShape = indexTensorType.getShape();
+      int rank = indexTensorShape.size();
+      SmallVector<AffineExpr> indicesExpr;
+      for (auto dim : llvm::seq(0, rank)) {
+        if (indexTensorShape[dim] == 1) {
+          indicesExpr.push_back(rewriter.getAffineConstantExpr(0));
+          continue;
+        }
+        indicesExpr.push_back(
+            rewriter.getAffineDimExpr(startIndex + broadcastRank - rank + dim));
+      }
+      indexingMaps.push_back(
+          AffineMap::get(resultRank, 0, indicesExpr, op->getContext()));
+    }
+
+    SmallVector<AffineExpr> resultExpr;
     for (auto i : llvm::seq(0, resultRank)) {
       resultExpr.push_back(rewriter.getAffineDimExpr(i));
       iteratorTypes.push_back(getParallelIteratorTypeName());
     }
-    auto indexingMaps = AffineMap::inferFromExprList({indicesExpr, resultExpr});
+
+    indexingMaps.push_back(
+        AffineMap::get(resultRank, 0, resultExpr, op->getContext()));
 
     Value finalRes =
         rewriter
             .create<linalg::GenericOp>(
-                loc, initTensor.getType(), indexTensor, initTensor,
+                loc, initTensor.getType(), indexTensors, initTensor,
                 indexingMaps, iteratorTypes,
                 [&](OpBuilder &b, Location loc, ValueRange args) {
-                  Value index = castIntToIndex(b, loc, args[0]);
                   SmallVector<Value> extractionIndices;
-                  int extra_dims = 0;
-                  for (auto i : llvm::seq(0, inputRank)) {
-                    if (i == indexTensorDim) {
-                      extractionIndices.push_back(index);
-                      extra_dims += indexTensorRank - 1;
-                    } else {
+                  if (contiguous) {
+                    for (auto i : llvm::seq(0, firstIndexDim)) {
+                      extractionIndices.push_back(
+                          b.create<linalg::IndexOp>(loc, i));
+                    }
+                    for (auto i : llvm::seq(0, (int)indexTensorDims.size())) {
                       extractionIndices.push_back(
-                          b.create<linalg::IndexOp>(loc, i + extra_dims));
+                          castIntToIndex(b, loc, args[i]));
+                    }
+                    for (auto i :
+                         llvm::seq((int)extractionIndices.size(), inputRank)) {
+                      extractionIndices.push_back(b.create<linalg::IndexOp>(
+                          loc, i + broadcastRank - replacedIndexCount));
+                    }
+                  } else {
+                    int indexCount = 0, unchanged = 0;
+                    for (auto i : llvm::seq(0, inputRank)) {
+                      if (indexCount < replacedIndexCount &&
+                          i == indexTensorDims[indexCount]) {
+                        extractionIndices.push_back(
+                            castIntToIndex(b, loc, args[indexCount++]));
+                        continue;
+                      }
+                      extractionIndices.push_back(b.create<linalg::IndexOp>(
+                          loc, broadcastRank + unchanged));
+                      unchanged++;
                     }
                   }
                   Value extractedElement = b.create<tensor::ExtractOp>(

lib/Dialect/Torch/IR/TorchOps.cpp

@@ -1479,6 +1479,35 @@ void Aten__Getitem__TOp::getCanonicalizationPatterns(
   });
 }
 
+//===----------------------------------------------------------------------===//
+// AtenAddTOp
+//===----------------------------------------------------------------------===//
+
+void AtenAddTOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
+                                             MLIRContext *context) {
+  patterns.add(+[](AtenAddTOp op, PatternRewriter &rewriter) {
+    auto lhsListConstruct = op.a().getDefiningOp<Torch::PrimListConstructOp>();
+    if (!lhsListConstruct || isListPotentiallyMutated(lhsListConstruct))
+      return failure();
+
+    auto rhsListConstruct = op.b().getDefiningOp<Torch::PrimListConstructOp>();
+    if (!rhsListConstruct || isListPotentiallyMutated(rhsListConstruct))
+      return failure();
+
+    SmallVector<Value> concatenatedList;
+    for (auto a : lhsListConstruct.getOperands()) {
+      concatenatedList.push_back(a);
+    }
+    for (auto b : rhsListConstruct.getOperands()) {
+      concatenatedList.push_back(b);
+    }
+
+    rewriter.replaceOpWithNewOp<Torch::PrimListConstructOp>(op, op.getType(),
+                                                            concatenatedList);
+    return success();
+  });
+}
+
 //===----------------------------------------------------------------------===//
 // AtenEqIntListOp
 //===----------------------------------------------------------------------===//

lib/Dialect/Torch/Transforms/ShapeLibrary.cpp

@@ -6590,30 +6590,30 @@ module {
       %10 = torch.aten.len.t %arg1 : !torch.list<optional<list<int>>> -> !torch.int
       %11 = torch.prim.ListConstruct %int9223372036854775807, %10 : (!torch.int, !torch.int) -> !torch.list<int>
       %12 = torch.prim.min.self_int %11 : !torch.list<int> -> !torch.int
-      %13:3 = torch.prim.Loop %12, %true, init(%true, %int-1, %int-1) {
-      ^bb0(%arg2: !torch.int, %arg3: !torch.bool, %arg4: !torch.int, %arg5: !torch.int):
+      %13:2 = torch.prim.Loop %12, %true, init(%true, %int-1) {
+      ^bb0(%arg2: !torch.int, %arg3: !torch.bool, %arg4: !torch.int):
         %16 = torch.aten.__getitem__.t %arg1, %arg2 : !torch.list<optional<list<int>>>, !torch.int -> !torch.optional<list<int>>
         %17 = torch.aten.__isnot__ %16, %none : !torch.optional<list<int>>, !torch.none -> !torch.bool
-        %18:3 = torch.prim.If %17 -> (!torch.bool, !torch.int, !torch.int) {
+        %18:2 = torch.prim.If %17 -> (!torch.bool, !torch.int) {
           %19 = torch.aten.eq.int %arg4, %int-1 : !torch.int, !torch.int -> !torch.bool
-          %20:3 = torch.prim.If %19 -> (!torch.bool, !torch.int, !torch.int) {
-            torch.prim.If.yield %arg3, %arg2, %arg2 : !torch.bool, !torch.int, !torch.int
+          %20:2 = torch.prim.If %19 -> (!torch.bool, !torch.int) {
+            torch.prim.If.yield %arg3, %arg2 : !torch.bool, !torch.int
           } else {
-            %21 = torch.aten.sub.int %arg2, %arg5 : !torch.int, !torch.int -> !torch.int
+            %21 = torch.aten.sub.int %arg2, %arg4 : !torch.int, !torch.int -> !torch.int
             %22 = torch.aten.ne.int %21, %int1 : !torch.int, !torch.int -> !torch.bool
             %23 = torch.prim.If %22 -> (!torch.bool) {
               torch.prim.If.yield %false : !torch.bool
             } else {
               torch.prim.If.yield %arg3 : !torch.bool
             }
-            torch.prim.If.yield %23, %arg4, %arg5 : !torch.bool, !torch.int, !torch.int
+            torch.prim.If.yield %23, %arg4 : !torch.bool, !torch.int
           }
-          torch.prim.If.yield %20#0, %20#1, %20#2 : !torch.bool, !torch.int, !torch.int
+          torch.prim.If.yield %20#0, %20#1 : !torch.bool, !torch.int
         } else {
-          torch.prim.If.yield %arg3, %arg4, %arg5 : !torch.bool, !torch.int, !torch.int
+          torch.prim.If.yield %arg3, %arg4 : !torch.bool, !torch.int
         }
-        torch.prim.Loop.condition %true, iter(%18#0, %18#1, %18#2 : !torch.bool, !torch.int, !torch.int)
-      } : (!torch.int, !torch.bool, !torch.bool, !torch.int, !torch.int) -> (!torch.bool, !torch.int, !torch.int)
+        torch.prim.Loop.condition %true, iter(%18#0, %18#1 : !torch.bool, !torch.int)
+      } : (!torch.int, !torch.bool, !torch.bool, !torch.int) -> (!torch.bool, !torch.int)
       %14 = torch.aten.__not__ %13#0 : !torch.bool -> !torch.bool
       %15 = torch.prim.If %14 -> (!torch.list<int>) {
         %16 = torch.aten.add.t %6, %4 : !torch.list<int>, !torch.list<int> -> !torch.list<int>

lib/Dialect/Torch/Transforms/SimplifyShapeCalculations.cpp

@@ -418,6 +418,7 @@ class SimplifyShapeCalculationsPass
     Aten__Getitem__TOp::getCanonicalizationPatterns(patterns, context);
     AtenSizeOp::getCanonicalizationPatterns(patterns, context);
     AtenLenTOp::getCanonicalizationPatterns(patterns, context);
+    AtenAddTOp::getCanonicalizationPatterns(patterns, context);
 
     // TODO: Debug visitation order to make this more efficient.
     // A single linear scan should suffice.

python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/shape_lib_gen.py

@@ -1016,6 +1016,7 @@ def aten〇pad(self: List[int], pad: List[int], mode: str = "constant", value: O
     Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4), None]), # Explicit None value.
     Invocation(TensorOfShape(2, 3, 4, 5), [None, LongTensorOfShape(4), LongTensorOfShape(4)]), # Indexing tensors on consecutive dimensions.
     Invocation(TensorOfShape(2, 3, 4, 5), [None, LongTensorOfShape(4), None, LongTensorOfShape(4)]), # Indexing tensors on non-consecutive dimensions.
+    Invocation(TensorOfShape(2, 3, 4, 5), [LongTensorOfShape(4, 2), None, LongTensorOfShape(2)]), # Indexing tensors on non-consecutive dimensions.
     Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4, 5, 6), LongTensorOfShape(1, 5, 1)]), # Broadcasting of index tensors.
     Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4)]), # Fewer index tensors than dimensions.
     ErrorInvocation(TensorOfShape(2, 3), [LongTensorOfShape(4), LongTensorOfShape(4), LongTensorOfShape(4)]), # More index tensors than dimensions.
@@ -1037,15 +1038,13 @@ def aten〇index〇Tensor(self: List[int], indices: List[Optional[List[int]]]) -
     if len(unused_dim_sizes) == 0:
         return broadcasted_shape
 
-    prev_index_tensor_location = -1
     first_index_tensor_location = -1
     index_tensors_are_together = True
     for e, index_tensor_shape in enumerate(indices):
         if index_tensor_shape is not None:
             if first_index_tensor_location == -1:
                 first_index_tensor_location = e
-                prev_index_tensor_location = e
-            elif e - prev_index_tensor_location != 1:
+            elif e - first_index_tensor_location != 1:
                 index_tensors_are_together = False
 
     if not index_tensors_are_together:

python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/torch_ods_gen.py

@@ -489,7 +489,7 @@ def emit_with_mutating_variants(key, **kwargs):
     # List ops.
     emit("aten::cat : (Tensor[], int) -> (Tensor)")
     emit("aten::append.t : (t[], t) -> (t[])")
-    emit("aten::add.t : (t[], t[]) -> (t[])")
+    emit("aten::add.t : (t[], t[]) -> (t[])", has_canonicalizer=True)
     emit("aten::eq.int_list : (int[], int[]) -> (bool)", has_folder=True)
     emit("aten::list.t : (t[]) -> (t[])")
     emit("aten::slice.t : (t[], int?, int?, int) -> (t[])")