Implement lowering of aten.vstack

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

@@ -10300,6 +10300,29 @@ def Torch_AtenAtleast2dOp : Torch_Op<"aten.atleast_2d", [
   }];
 }
 
+def Torch_AtenVstackOp : Torch_Op<"aten.vstack", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::vstack : (Tensor[]) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchListOfTensorType:$tensors
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenVstackOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 1, 1);
+    }
+    void AtenVstackOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 1, 1);
+    }
+  }];
+}
+
 def Torch_AtenEinsumOp : Torch_Op<"aten.einsum", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -10568,6 +10568,21 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %2 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.vstack\"(%arg0: !torch.list<list<int>>) -> !torch.list<int> {\n"
+"    %true = torch.constant.bool true\n"
+"    %int0 = torch.constant.int 0\n"
+"    %0 = torch.prim.ListConstruct  : () -> !torch.list<list<int>>\n"
+"    %1 = torch.aten.len.t %arg0 : !torch.list<list<int>> -> !torch.int\n"
+"    torch.prim.Loop %1, %true, init() {\n"
+"    ^bb0(%arg1: !torch.int):\n"
+"      %3 = torch.aten.__getitem__.t %arg0, %arg1 : !torch.list<list<int>>, !torch.int -> !torch.list<int>\n"
+"      %4 = func.call @\"__torch_mlir_shape_fn.aten.atleast_2d\"(%3) : (!torch.list<int>) -> !torch.list<int>\n"
+"      %5 = torch.aten.append.t %0, %4 : !torch.list<list<int>>, !torch.list<int> -> !torch.list<list<int>>\n"
+"      torch.prim.Loop.condition %true, iter()\n"
+"    } : (!torch.int, !torch.bool) -> ()\n"
+"    %2 = call @__torch__.torch.jit._shape_functions.cat(%0, %int0) : (!torch.list<list<int>>, !torch.int) -> !torch.list<int>\n"
+"    return %2 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.stack\"(%arg0: !torch.list<list<int>>, %arg1: !torch.int) -> !torch.list<int> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.stack(%arg0, %arg1) : (!torch.list<list<int>>, !torch.int) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
@@ -15070,6 +15085,33 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.vstack\"(%arg0: !torch.list<tuple<int, int>>) -> !torch.int {\n"
+"    %true = torch.constant.bool true\n"
+"    %none = torch.constant.none\n"
+"    %str = torch.constant.str \"AssertionError: \"\n"
+"    %int0 = torch.constant.int 0\n"
+"    %0 = torch.prim.ListConstruct  : () -> !torch.list<optional<int>>\n"
+"    %1 = torch.prim.ListConstruct  : () -> !torch.list<int>\n"
+"    %2 = torch.aten.len.t %arg0 : !torch.list<tuple<int, int>> -> !torch.int\n"
+"    %3 = torch.aten.ne.int %2, %int0 : !torch.int, !torch.int -> !torch.bool\n"
+"    torch.prim.If %3 -> () {\n"
+"      torch.prim.If.yield\n"
+"    } else {\n"
+"      torch.prim.RaiseException %str, %none : !torch.str, !torch.none\n"
+"      torch.prim.If.yield\n"
+"    }\n"
+"    %4 = torch.aten.len.t %arg0 : !torch.list<tuple<int, int>> -> !torch.int\n"
+"    torch.prim.Loop %4, %true, init() {\n"
+"    ^bb0(%arg1: !torch.int):\n"
+"      %6 = torch.aten.__getitem__.t %arg0, %arg1 : !torch.list<tuple<int, int>>, !torch.int -> !torch.tuple<int, int>\n"
+"      %7:2 = torch.prim.TupleUnpack %6 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"      %8 = torch.aten.append.t %0, %7#0 : !torch.list<optional<int>>, !torch.int -> !torch.list<optional<int>>\n"
+"      %9 = torch.aten.append.t %1, %7#1 : !torch.list<int>, !torch.int -> !torch.list<int>\n"
+"      torch.prim.Loop.condition %true, iter()\n"
+"    } : (!torch.int, !torch.bool) -> ()\n"
+"    %5 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.promote_dtypes(%0, %1) : (!torch.list<optional<int>>, !torch.list<int>) -> !torch.int\n"
+"    return %5 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.einsum\"(%arg0: !torch.str, %arg1: !torch.list<tuple<int, int>>, %arg2: !torch.optional<list<int>>) -> !torch.int {\n"
 "    %true = torch.constant.bool true\n"
 "    %none = torch.constant.none\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -1846,6 +1846,97 @@ class DecomposeAtenAtleast2dOp : public OpRewritePattern<AtenAtleast2dOp> {
 };
 } // namespace
 
+namespace {
+// Decompose aten.vstack into: aten.vstack and aten.cat See:
+// https://github.com/pytorch/pytorch/blob/9a8ab778d34bd24c5caceb340837483decc4c311/torch/_refs/__init__.py#L3887
+// @out_wrapper()
+// def vstack(tensors: TensorSequenceType) -> TensorLikeType:
+//     torch._check(len(tensors) > 0,
+//         lambda: "vstack expects a non-empty TensorList")
+//     aligned_tensors = atleast_2d(*tensors) return
+//     cat(aligned_tensors, 0)
+class DecomposeAtenVstackOp : public OpRewritePattern<AtenVstackOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenVstackOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+
+    SmallVector<Value> inputTensors;
+    if (!getListConstructElements(op.getTensors(), inputTensors)) {
+      return rewriter.notifyMatchFailure(
+          op, "input should come from a PrimListConstructOp");
+    }
+    if (inputTensors.empty()) {
+      return rewriter.notifyMatchFailure(
+          op, "vstack expects a non-empty TensorList");
+    }
+
+    for (Value tensor : inputTensors) {
+      BaseTensorType tensorType = cast<BaseTensorType>(tensor.getType());
+      if (!tensorType.hasSizes()) {
+        return rewriter.notifyMatchFailure(
+            op, "unimplemented: at least one tensor does not have known sizes");
+      }
+    }
+
+    // calculate dimensions of the tensors, after they've all had atleast_2d
+    // applied to them.
+    SmallVector<SmallVector<int64_t>> tensorShapes;
+    for (Value tensor : inputTensors) {
+      auto tensorType = cast<BaseTensorType>(tensor.getType());
+      SmallVector<int64_t> tensorShape{tensorType.getSizes()};
+      switch (tensorShape.size()) {
+      case 0:
+        tensorShape = SmallVector<int64_t>{1, 1};
+        break;
+      case 1:
+        int64_t x = tensorShape[0];
+        tensorShape = SmallVector<int64_t>{1, x};
+        break;
+      }
+      tensorShapes.push_back(tensorShape);
+    }
+
+    // check if all tensors match in all dimensions except the 0th dimension
+    SmallVector<int64_t> fstTensorShape = tensorShapes[0];
+    int64_t fstTensorRank = fstTensorShape.size();
+    for (auto tensorShape : tensorShapes) {
+      if (fstTensorShape.size() != tensorShape.size() ||
+          !std::equal(fstTensorShape.begin() + 1,
+                      fstTensorShape.begin() + fstTensorRank,
+                      tensorShape.begin() + 1)) {
+        return rewriter.notifyMatchFailure(
+            op, "tensors must have all matching dimensions except for 0");
+      }
+    }
+
+    // if so, proceed, and create all the necessary ops.
+    SmallVector<Value> atleast2dTensorOps;
+    for (size_t i = 0; i < inputTensors.size(); i++) {
+      auto tensorType = cast<BaseTensorType>(inputTensors[i].getType());
+      auto newTensorType = rewriter.getType<ValueTensorType>(
+          tensorShapes[i], tensorType.getOptionalDtype());
+      auto tensorOp =
+          rewriter.create<AtenAtleast2dOp>(loc, newTensorType, inputTensors[i]);
+      atleast2dTensorOps.push_back(tensorOp);
+    }
+
+    auto elemType = cast<BaseTensorType>(atleast2dTensorOps[0].getType())
+                        .getWithSizesAndDtype(std::nullopt, nullptr);
+    Value atleast2dTensorList = rewriter.create<PrimListConstructOp>(
+        loc, Torch::ListType::get(elemType), atleast2dTensorOps);
+
+    auto zero =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(0));
+    rewriter.replaceOpWithNewOp<AtenCatOp>(op, op.getType(),
+                                           atleast2dTensorList, zero);
+
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 // Decompose AtenEinsumOp to AtenMatmulOp, and supports possible reduce
 // operation and permute operation. Currently, this pass doesn't support
@@ -9477,6 +9568,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenCeluOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenAtleast1dOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenAtleast2dOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenVstackOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenEinsumOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTraceOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenHardswishOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -397,6 +397,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenTanhBackwardOp>();
   target.addIllegalOp<AtenAtleast1dOp>();
   target.addIllegalOp<AtenAtleast2dOp>();
+  target.addIllegalOp<AtenVstackOp>();
   target.addIllegalOp<AtenEinsumOp>();
   target.addIllegalOp<AtenTraceOp>();
   target.addIllegalOp<AtenAddmmOp>();

projects/pt1/e2e_testing/xfail_sets.py

@@ -882,6 +882,11 @@
     "Atleast2dModule0dInput_basic",
     "Atleast2dModule1dInput_basic",
     "Atleast2dModule2dInput_basic",
+    "AtenVstackModule_basic",
+    "AtenVstackStaticModule_basic",
+    "AtenVstackSingleElementListModule_basic",
+    "AtenVstackMatchingZerothDimModule_basic",
+    "AtenVstackPromoteDTypeModule_basic",
     "AtenLinear1D_basic",
     "AtenLinear2D_basic",
     "AtenLinear3DBias_basic",
@@ -1582,6 +1587,8 @@
     "Atleast2dModule0dInput_basic",
     "Atleast2dModule1dInput_basic",
     "Atleast2dModule2dInput_basic",
+    "AtenVstackStaticModule_basic",
+    "AtenVstackSingleElementListModule_basic",
     "AtenLinear2D_basic",
     "AtenLinear3DBias_basic",
     "ElementwiseAddScalar_NumToTensorFloat_Module_basic",
@@ -2419,6 +2426,11 @@
     "AtenSubFloatModule_basic",
     "AtenTopKModule_basic",
     "AtenTopKSmallestModule_basic",
+    "AtenVstackModule_basic",
+    "AtenVstackStaticModule_basic",
+    "AtenVstackSingleElementListModule_basic",
+    "AtenVstackMatchingZerothDimModule_basic",
+    "AtenVstackPromoteDTypeModule_basic",
     "Aten_EmbeddingBagExample_basic",
     "AvgPool2dWithoutPadModule_basic",
     "BatchMlpLayerModule_basic",
@@ -4538,6 +4550,11 @@
     "TensorsConcatModule_basic",
     "TensorsConcatNegativeDimModule_basic",
     "TensorsConcatPromoteDTypeModule_basic",
+    "AtenVstackModule_basic",
+    "AtenVstackStaticModule_basic",
+    "AtenVstackSingleElementListModule_basic",
+    "AtenVstackMatchingZerothDimModule_basic",
+    "AtenVstackPromoteDTypeModule_basic",
     "TensorsStackModule_basic",
     "TensorsStackNegativeDimModule_basic",
     "TensorsStackPromoteDTypeModule_basic",

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -2130,6 +2130,10 @@ def aten〇atleast_2d〡shape(self: List[int]) -> List[int]:
     else:
         return self
 
+def aten〇vstack〡shape(tensors: List[List[int]]) -> List[int]:
+    tensors_atleast2d = [ aten〇atleast_2d〡shape(tensor) for tensor in tensors ]
+    return upstream_shape_functions.cat(tensors_atleast2d, 0)
+
 def aten〇stack〡shape(tensors: List[List[int]], dim: int = 0) -> List[int]:
     return upstream_shape_functions.stack(tensors, dim)
 
@@ -5279,6 +5283,22 @@ def aten〇atleast_2d〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
+@check_dtype_function(
+    [Invocation([NonZeroDTensorWithDtype(torch.float32), NonZeroDTensorWithDtype(torch.int32)]),
+     Invocation([NonZeroDTensorWithDtype(torch.float16), NonZeroDTensorWithDtype(torch.float64)]),
+     Invocation([NonZeroDTensorWithDtype(torch.float32), NonZeroDTensorWithDtype(torch.int32),
+                 NonZeroDTensorWithDtype(torch.complex64)])])
+def aten〇vstack〡dtype(tensors_rank_dtype: List[Tuple[int, int]]) -> int:
+    ranks: List[Optional[int]] = []
+    dtypes: List[int] = []
+    assert len(tensors_rank_dtype) != 0
+    for tensor_rank_dtype in tensors_rank_dtype:
+        tensor_rank, tensor_dtype = tensor_rank_dtype
+        ranks.append(tensor_rank)
+        dtypes.append(tensor_dtype)
+
+    return promote_dtypes(ranks, dtypes)
+
 @check_dtype_function(
     [Invocation("i,j->ij", [TensorOfShape(1, dtype=torch.float32),
                             TensorOfShape(1, dtype=torch.int32)]),])

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py

@@ -796,6 +796,7 @@ def emit_with_mutating_variants(key, **kwargs):
     emit("aten::one_hot : (Tensor, int) -> (Tensor)")
     emit("aten::atleast_1d : (Tensor) -> (Tensor)")
     emit("aten::atleast_2d : (Tensor) -> (Tensor)")
+    emit("aten::vstack : (Tensor[]) -> (Tensor)")
     emit("aten::einsum : (str, Tensor[], int[]?) -> (Tensor)")
     emit("aten::trace : (Tensor) -> (Tensor)")
     emit("aten::bucketize.Tensor : (Tensor, Tensor, bool, bool) -> (Tensor)")