[MLIR][TORCH] Add E2E support for aten.index.Tensor_hacked_twin op

This commit adds lowering of `index.Tensor_hacked_twin` op.

Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>

e2e_testing/xfail_sets.py

@@ -447,6 +447,9 @@ LTC_XFAIL_SET = {
     "IndexTensorMultiInputNonContiguousOneDimDynamic_basic",
     "IndexTensorMultiInputNonContiguousDynamic_basic",
     "IndexTensorMultiInputNonContiguousMultipleStaticDims_basic",
+    "IndexTensorHackedTwinModule_basic",
+    "IndexTensorHackedTwinModule3dInput_basic",
+    "IndexTensorHackedTwinMultiInputNonContiguousMultipleStaticDims_basic",
     "Matmul_dot",
     "Matmul_matvec",
     "MulIntModule_basic",

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

@@ -5661,6 +5661,30 @@ def Torch_AtenIndexTensorOp : Torch_Op<"aten.index.Tensor", [
   }];
 }
 
+def Torch_AtenIndexTensorHackedTwinOp : Torch_Op<"aten.index.Tensor_hacked_twin", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::index.Tensor_hacked_twin : (Tensor, Tensor[]) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchListOfTensorType:$indices
+  );
+  let results = (outs
+    AnyTorchTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenIndexTensorHackedTwinOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenIndexTensorHackedTwinOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_AtenIndexSelectOp : Torch_Op<"aten.index_select", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -2762,6 +2762,21 @@ class DecomposeAtenLiftFreshCopyOp
 };
 } // namespace
 
+namespace {
+// Decompose `aten.index.Tensor_hacked_twin` op into `aten.index.Tensor` op.
+class DecomposeAtenIndexTensorHackedTwinOp
+    : public OpRewritePattern<AtenIndexTensorHackedTwinOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenIndexTensorHackedTwinOp op,
+                                PatternRewriter &rewriter) const override {
+    rewriter.replaceOpWithNewOp<AtenIndexTensorOp>(op, op.getType(), op.self(),
+                                                   op.indices());
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeComplexOpsPass
     : public DecomposeComplexOpsBase<DecomposeComplexOpsPass> {
@@ -2949,6 +2964,8 @@ public:
     target.addIllegalOp<Aten_EmbeddingBagOp>();
     patterns.add<DecomposeAtenLiftFreshCopyOp>(context);
     target.addIllegalOp<AtenLiftFreshCopyOp>();
+    patterns.add<DecomposeAtenIndexTensorHackedTwinOp>(context);
+    target.addIllegalOp<AtenIndexTensorHackedTwinOp>();
 
     for (std::string opName : legalOps) {
       target.addLegalOp(OperationName(opName, context));

lib/Dialect/Torch/Transforms/RefineTypes.cpp

@@ -665,7 +665,8 @@ void TypeAnalysis::visitOperation(Operation *op,
           AtenIndexPutOp, ValsemVariantAtenCopyOp, AtenZeroOp,
           AtenIndexPutHackedTwinOp, AtenMaskedFillScalarOp, AtenFlipOp,
           PrimAbsScalarOp, AtenNumpyTOp, AtenTriuOp, AtenMaskedFillTensorOp,
-          AtenRollOp, AtenPowTensorTensorOp, AtenLiftFreshCopyOp>(op)) {
+          AtenRollOp, AtenPowTensorTensorOp, AtenLiftFreshCopyOp,
+          AtenIndexTensorHackedTwinOp>(op)) {
     return incorporateKnowledge(op->getResult(0), operands[0]->getValue());
   }
 

lib/Dialect/Torch/Transforms/ShapeLibrary.cpp

@@ -6822,6 +6822,10 @@ module {
     return %0 : !torch.list<int>
   }
   func.func @"__torch_mlir_shape_fn.aten.index.Tensor"(%arg0: !torch.list<int>, %arg1: !torch.list<optional<list<int>>>) -> !torch.list<int> {
+    %0 = call @__torch__.index_tensor_like(%arg0, %arg1) : (!torch.list<int>, !torch.list<optional<list<int>>>) -> !torch.list<int>
+    return %0 : !torch.list<int>
+  }
+  func.func @__torch__.index_tensor_like(%arg0: !torch.list<int>, %arg1: !torch.list<optional<list<int>>>) -> !torch.list<int> {
     %false = torch.constant.bool false
     %int-1 = torch.constant.int -1
     %true = torch.constant.bool true
@@ -6932,6 +6936,19 @@ module {
     }
     return %9 : !torch.list<int>
   }
+  func.func @"__torch_mlir_shape_fn.aten.index.Tensor_hacked_twin"(%arg0: !torch.list<int>, %arg1: !torch.list<list<int>>) -> !torch.list<int> {
+    %true = torch.constant.bool true
+    %0 = torch.prim.ListConstruct  : () -> !torch.list<optional<list<int>>>
+    %1 = torch.aten.len.t %arg1 : !torch.list<list<int>> -> !torch.int
+    torch.prim.Loop %1, %true, init() {
+    ^bb0(%arg2: !torch.int):
+      %3 = torch.aten.__getitem__.t %arg1, %arg2 : !torch.list<list<int>>, !torch.int -> !torch.list<int>
+      %4 = torch.aten.append.t %0, %3 : !torch.list<optional<list<int>>>, !torch.list<int> -> !torch.list<optional<list<int>>>
+      torch.prim.Loop.condition %true, iter()
+    } : (!torch.int, !torch.bool) -> ()
+    %2 = call @__torch__.index_tensor_like(%arg0, %0) : (!torch.list<int>, !torch.list<optional<list<int>>>) -> !torch.list<int>
+    return %2 : !torch.list<int>
+  }
   func.func @"__torch_mlir_shape_fn.aten.cat"(%arg0: !torch.list<list<int>>, %arg1: !torch.int) -> !torch.list<int> {
     %0 = call @__torch__.torch.jit._shape_functions.cat(%arg0, %arg1) : (!torch.list<list<int>>, !torch.int) -> !torch.list<int>
     return %0 : !torch.list<int>

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

@@ -1081,20 +1081,7 @@ def aten〇constant_pad_nd(self: List[int], pad: List[int], value: float = 0) ->
 def aten〇pad(self: List[int], pad: List[int], mode: str = "constant", value: Optional[float] = None) -> List[int]:
     return pad_shape_fn(self, pad)
 
-# See https://numpy.org/doc/stable/user/basics.indexing.html
+def index_tensor_like(self: List[int], indices: List[Optional[List[int]]]) -> List[int]:
-@check_shape_function([
-    Invocation(TensorOfShape(2), [LongTensorOfShape(4)]), # Basic case.
-    Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4), LongTensorOfShape(4)]), # More dimensions.
-    Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4), LongTensorOfShape(6, 4)]), # Multidimensional index tensor along a dimension.
-    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.
-])
-def aten〇index〇Tensor(self: List[int], indices: List[Optional[List[int]]]) -> List[int]:
     assert len(indices) <= len(self), "More indices than dimensions to index"
     broadcasted_shape: List[int] = []
     unused_dim_sizes: List[int] = []
@@ -1134,6 +1121,26 @@ def aten〇index〇Tensor(self: List[int], indices: List[Optional[List[int]]]) -
         result_shape.append(unused_dim_sizes[i])
     return result_shape
 
+# See https://numpy.org/doc/stable/user/basics.indexing.html
+@check_shape_function([
+    Invocation(TensorOfShape(2), [LongTensorOfShape(4)]), # Basic case.
+    Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4), LongTensorOfShape(4)]), # More dimensions.
+    Invocation(TensorOfShape(2, 3), [LongTensorOfShape(4), LongTensorOfShape(6, 4)]), # Multidimensional index tensor along a dimension.
+    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.
+])
+def aten〇index〇Tensor(self: List[int], indices: List[Optional[List[int]]]) -> List[int]:
+    return index_tensor_like(self, indices)
+
+def aten〇index〇Tensor_hacked_twin(self: List[int], indices: List[List[int]]) -> List[int]:
+    optional_indices: List[Optional[List[int]]] = [x for x in indices]
+    return index_tensor_like(self, optional_indices)
+
 def aten〇cat(tensors: List[List[int]], dim: int = 0) -> List[int]:
     return upstream_shape_functions.cat(tensors, dim)
 

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

@@ -446,6 +446,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit("aten::expand_as : (Tensor, Tensor) -> (Tensor)")
     emit("aten::broadcast_to : (Tensor, int[]) -> (Tensor)")
     emit("aten::index.Tensor : (Tensor, Tensor?[]) -> (Tensor)")
+    emit("aten::index.Tensor_hacked_twin : (Tensor, Tensor[]) -> (Tensor)")
     emit("aten::index_select : (Tensor, int, Tensor) -> (Tensor)")
     emit("aten::_index_put_impl_ : (Tensor, Tensor?[], Tensor, bool, bool) -> (Tensor)")
     emit("aten::item : (Tensor) -> (Scalar)")

python/torch_mlir_e2e_test/test_suite/basic.py

@@ -1963,6 +1963,83 @@ def IndexTensorMultiInputContiguousCenter_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class IndexTensorHackedTwinModule(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1], torch.float32, True),
+        ([-1, -1], torch.int64, True),
+    ])
+    def forward(self, x, index):
+        return torch.ops.aten.index(x, [index])
+
+
+@register_test_case(module_factory=lambda: IndexTensorHackedTwinModule())
+def IndexTensorHackedTwinModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5), tu.randint(2, 3, high=4))
+
+
+# ==============================================================================
+
+
+class IndexTensorHackedTwinModule3dInput(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1, -1], torch.float32, True),
+        ([-1, -1], torch.int64, True),
+    ])
+    def forward(self, x, index):
+        return torch.ops.aten.index(x, [index])
+
+
+@register_test_case(
+    module_factory=lambda: IndexTensorHackedTwinModule3dInput())
+def IndexTensorHackedTwinModule3dInput_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5, 4, 3), tu.randint(2, 3, high=3))
+
+
+# ==============================================================================
+
+
+class IndexTensorHackedTwinMultiInputNonContiguousMultipleStaticDims(
+        torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args([
+        None,
+        ([-1, -1, -1, -1], torch.float32, True),
+        ([4, 1], torch.int64, True),
+        ([1, 3], torch.int64, True),
+        ([-1, 3], torch.int64, True),
+    ])
+    def forward(self, x, index1, index2, index3):
+        return torch.ops.aten.index(x, [index1, index2, index3])
+
+
+@register_test_case(
+    module_factory=lambda:
+    IndexTensorHackedTwinMultiInputNonContiguousMultipleStaticDims())
+def IndexTensorHackedTwinMultiInputNonContiguousMultipleStaticDims_basic(
+        module, tu: TestUtils):
+    module.forward(tu.rand(5, 4, 3, 2), tu.randint(4, 1, high=3),
+                   tu.randint(1, 3, high=1), tu.randint(4, 3, high=1))
+
+
+# ==============================================================================
+
+
 class SquareModule(torch.nn.Module):
 
     def __init__(self):