[Torch] Decompose torch.take op into AtenFlattenUsingInts and AtenSelectIndex

Decompose op inside DecomposeComplexOps.cpp

Add tests into slice_like.py

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

@@ -13148,6 +13148,30 @@ def Torch_AtenStftOp : Torch_Op<"aten.stft", [
   }];
 }
 
+def Torch_AtenTakeOp : Torch_Op<"aten.take", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::take : (Tensor, Tensor) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchTensorType:$index
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenTakeOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenTakeOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_AtenAliasCopyOp : Torch_Op<"aten.alias_copy", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -6591,6 +6591,38 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %3 = torch.prim.TupleConstruct %0, %1, %2 : !torch.int, !torch.int, !torch.int -> !torch.tuple<int, int, int>\n"
 "    return %3 : !torch.tuple<int, int, int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.take\"(%arg0: !torch.list<int>, %arg1: !torch.list<int>) -> !torch.list<int> {\n"
+"    return %arg1 : !torch.list<int>\n"
+"  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.take\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.tuple<int, int>) -> !torch.int {\n"
+"    %none = torch.constant.none\n"
+"    %str = torch.constant.str \"AssertionError: indexes must be integer types\"\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %1:2 = torch.prim.TupleUnpack %arg1 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %2 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.is_integer_dtype(%1#1) : (!torch.int) -> !torch.bool\n"
+"    torch.prim.If %2 -> () {\n"
+"      torch.prim.If.yield\n"
+"    } else {\n"
+"      torch.prim.RaiseException %str, %none : !torch.str, !torch.none\n"
+"      torch.prim.If.yield\n"
+"    }\n"
+"    return %0#1 : !torch.int\n"
+"  }\n"
+"  func.func @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.is_integer_dtype(%arg0: !torch.int) -> !torch.bool {\n"
+"    %0 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.all_integer_dtypes() : () -> !torch.list<int>\n"
+"    %1 = torch.aten.__contains__.int_list %0, %arg0 : !torch.list<int>, !torch.int -> !torch.bool\n"
+"    return %1 : !torch.bool\n"
+"  }\n"
+"  func.func @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.all_integer_dtypes() -> !torch.list<int> {\n"
+"    %int4 = torch.constant.int 4\n"
+"    %int3 = torch.constant.int 3\n"
+"    %int2 = torch.constant.int 2\n"
+"    %int1 = torch.constant.int 1\n"
+"    %int0 = torch.constant.int 0\n"
+"    %int11 = torch.constant.int 11\n"
+"    %0 = torch.prim.ListConstruct %int11, %int0, %int1, %int2, %int3, %int4 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>\n"
+"    return %0 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.linalg_slogdet\"(%arg0: !torch.list<int>) -> !torch.tuple<list<int>, list<int>> {\n"
 "    %int-2 = torch.constant.int -2\n"
 "    %int-1 = torch.constant.int -1\n"
@@ -11238,21 +11270,6 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %3 = call @__torch__._get_dtype_of_floating_point_op(%0#1) : (!torch.int) -> !torch.int\n"
 "    return %3 : !torch.int\n"
 "  }\n"
-"  func.func @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.is_integer_dtype(%arg0: !torch.int) -> !torch.bool {\n"
-"    %0 = call @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.all_integer_dtypes() : () -> !torch.list<int>\n"
-"    %1 = torch.aten.__contains__.int_list %0, %arg0 : !torch.list<int>, !torch.int -> !torch.bool\n"
-"    return %1 : !torch.bool\n"
-"  }\n"
-"  func.func @__torch__.torch_mlir.jit_ir_importer.build_tools.library_generator.all_integer_dtypes() -> !torch.list<int> {\n"
-"    %int4 = torch.constant.int 4\n"
-"    %int3 = torch.constant.int 3\n"
-"    %int2 = torch.constant.int 2\n"
-"    %int1 = torch.constant.int 1\n"
-"    %int0 = torch.constant.int 0\n"
-"    %int11 = torch.constant.int 11\n"
-"    %0 = torch.prim.ListConstruct %int11, %int0, %int1, %int2, %int3, %int4 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>\n"
-"    return %0 : !torch.list<int>\n"
-"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.sin\"(%arg0: !torch.tuple<int, int>) -> !torch.int {\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    %1 = call @__torch__._get_dtype_of_floating_point_op(%0#1) : (!torch.int) -> !torch.int\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -4113,6 +4113,38 @@ public:
 };
 } // namespace
 
+namespace {
+
+class DecomposeAtenTakeOp : public OpRewritePattern<AtenTakeOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenTakeOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    Value index = op.getIndex();
+    auto selfTy = cast<BaseTensorType>(self.getType());
+    auto resType = cast<BaseTensorType>(op.getType());
+    int64_t selfNumel = getTensorNumel(self).value(); // as selfTy has sizes
+
+    auto flattenType = selfTy.getWithSizesAndDtype(
+        /*optionalSizes=*/{selfNumel}, resType.getDtype());
+    Value constMinusOne =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(-1));
+    Value constZero =
+        rewriter.create<ConstantIntOp>(loc, rewriter.getI64IntegerAttr(0));
+    Value flattenSelf = rewriter.create<AtenFlattenUsingIntsOp>(
+        loc, flattenType, self, constZero, constMinusOne);
+
+    Value result = rewriter.create<Torch::AtenIndexSelectOp>(
+        loc, resType, flattenSelf, constZero, index);
+
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+} // namespace
+
 // decompose aten.repeat_interleave.self_int into following ops:
 // aten.flatten.using_ints, aten.unsqueeze, aten.tile, aten.reshape
 namespace {
@@ -9660,6 +9692,7 @@ public:
     legalOpsSet.clear();
     legalOpsSet.insert(legalOps.begin(), legalOps.end());
 
+    addPatternIfTargetOpIsIllegal<DecomposeAtenTakeOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAten_WeightNormInterfaceOp>(
         patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenSoftmaxIntOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -414,6 +414,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenLinalgSlogdetOp>();
   target.addIllegalOp<AtenPixelShuffleOp>();
   target.addIllegalOp<AtenTOp>();
+  target.addIllegalOp<AtenTakeOp>();
   target.addIllegalOp<Aten_LogSoftmaxBackwardDataOp>();
   target.addDynamicallyLegalOp<AtenMatmulOp>([](AtenMatmulOp op) {
     std::optional<unsigned> lhsRank = getTensorRank(op.getSelf());

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

@@ -257,6 +257,16 @@ def aten〇_linalg_det〡shape(A: List[int]) -> Tuple[List[int], List[int], List
 def aten〇_linalg_det〡dtype(A_rank_dtype: Tuple[int, int]) -> Tuple[int, int, int]:
     return (A_rank_dtype[1], A_rank_dtype[1], A_rank_dtype[1])
 
+def aten〇take〡shape(self: List[int], index: List[int]) -> List[int]:
+    return index
+
+def aten〇take〡dtype(self_rank_dtype: Tuple[int, int], index_rank_dtype: Tuple[int, int]) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    index_rank, index_dtype = index_rank_dtype
+    assert is_integer_dtype(index_dtype), "indexes must be integer types"
+    return self_dtype
+
+
 def aten〇linalg_slogdet〡shape(A: List[int]) -> Tuple[List[int], List[int]]:
     assert len(A) == 2 or len(A) == 3
     assert A[-1] == A[-2]

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

@@ -964,6 +964,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit(
         "aten::stft : (Tensor, int, int?, int?, Tensor?, bool, bool?, bool?) -> (Tensor)"
     )
+    emit("aten::take : (Tensor, Tensor) -> (Tensor)")
 
     # Functionalization ops
     emit("aten::alias_copy : (Tensor) -> (Tensor)")

projects/pt1/python/torch_mlir_e2e_test/test_suite/slice_like.py

@@ -1121,3 +1121,50 @@ class TensorSplitSections_ListUnpackModule(torch.nn.Module):
 @register_test_case(module_factory=lambda: TensorSplitSections_ListUnpackModule())
 def TensorSplitSections_ListUnpackModule_basic(module, tu: TestUtils):
     module.forward(tu.rand(2, 5))
+
+
+# ==============================================================================
+
+
+class TakeModule(torch.nn.Module):
+    @export
+    @annotate_args([None, [(4, 4), torch.float32, True], [(4,), torch.int64, True]])
+    def forward(self, input, index):
+        return torch.take(input, index)
+
+
+@register_test_case(module_factory=lambda: TakeModule())
+def TakeModule_F32(module, tu: TestUtils):
+    A = tu.rand(4, 4).to(dtype=torch.float32)
+    index = tu.rand(4, low=0, high=torch.numel(A)).to(dtype=torch.int64)
+    module.forward(A, index)
+
+
+class TakeBatchModule(torch.nn.Module):
+    @export
+    @annotate_args([None, [(4, 4, 4), torch.float32, True], [(4,), torch.int64, True]])
+    def forward(self, input, index):
+        return torch.take(input, index)
+
+
+@register_test_case(module_factory=lambda: TakeBatchModule())
+def TakeModuleBatched_F32(module, tu: TestUtils):
+    A = tu.rand(4, 4, 4).to(dtype=torch.float32)
+    index = tu.rand(4, low=0, high=torch.numel(A)).to(dtype=torch.int64)
+    module.forward(A, index)
+
+
+class TakeDynamicModule(torch.nn.Module):
+    @export
+    @annotate_args(
+        [None, [(-1, -1, -1), torch.float32, True], [(4,), torch.int64, True]]
+    )
+    def forward(self, input, index):
+        return torch.take(input, index)
+
+
+@register_test_case(module_factory=lambda: TakeDynamicModule())
+def TakeModuleDynamic_F32(module, tu: TestUtils):
+    A = tu.rand(4, 4, 8).to(dtype=torch.float32)
+    index = tu.rand(4, low=0, high=torch.numel(A)).to(dtype=torch.int64)
+    module.forward(A, index)