[Torch] Add support for aten.round.decimals op (#4166)

* Added decomposition to aten.round
* Added test to
projects/pt1/python/torch_mlir_e2e_test/test_suite/elementwise.py

Author: mmanzoorTT

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

@@ -4562,6 +4562,54 @@ def Torch_AtenRound_Op : Torch_Op<"aten.round_", [
   }];
 }
 
+def Torch_AtenRoundDecimalsOp : Torch_Op<"aten.round.decimals", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::round.decimals : (Tensor, int) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    Torch_IntType:$decimals
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenRoundDecimalsOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenRoundDecimalsOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+  let hasFolder = 1;
+}
+
+def Torch_AtenRound_DecimalsOp : Torch_Op<"aten.round_.decimals", [
+    IsTrailingUnderscoreInplaceVariant,
+    AllowsTypeRefinement
+  ]> {
+  let summary = "Generated op for `aten::round_.decimals : (Tensor, int) -> (Tensor)`";
+  let arguments = (ins
+    Torch_NonValueTensorType:$self,
+    Torch_IntType:$decimals
+  );
+  let results = (outs
+    AnyTorchOptionalNonValueTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenRound_DecimalsOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenRound_DecimalsOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_AtenTruncOp : Torch_Op<"aten.trunc", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/IR/TorchOps.cpp

@@ -1992,6 +1992,19 @@ OpFoldResult AtenCeilOp::fold(FoldAdaptor adaptor) {
   return {};
 }
 
+//===----------------------------------------------------------------------===//
+// AtenRoundDecimalsOp
+//===----------------------------------------------------------------------===//
+
+OpFoldResult AtenRoundDecimalsOp::fold(FoldAdaptor adaptor) {
+  auto resultType = dyn_cast<ValueTensorType>(getType());
+  if (resultType && resultType.hasDtype() &&
+      isa<mlir::IntegerType>(resultType.getDtype())) {
+    return getSelf();
+  }
+  return {};
+}
+
 //===----------------------------------------------------------------------===//
 // AtenRoundOp
 //===----------------------------------------------------------------------===//

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -6754,6 +6754,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.round.decimals\"(%arg0: !torch.list<int>, %arg1: !torch.int) -> !torch.list<int> {\n"
+"    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
+"    return %0 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.glu\"(%arg0: !torch.list<int>, %arg1: !torch.int) -> !torch.list<int> {\n"
 "    %none = torch.constant.none\n"
 "    %str = torch.constant.str \"AssertionError: glu's dim size must be multiply of 2\"\n"
@@ -13007,6 +13011,10 @@ 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.round.decimals\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.int) -> !torch.int {\n"
+"    %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.glu\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.int) -> !torch.int {\n"
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -11976,6 +11976,54 @@ class DecomposeAten_AssertScalarOp
 };
 } // namespace
 
+namespace {
+class DecomposeAtenRoundDecimalsOp
+    : public OpRewritePattern<AtenRoundDecimalsOp> {
+public:
+  using OpRewritePattern<AtenRoundDecimalsOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenRoundDecimalsOp op,
+                                PatternRewriter &rewriter) const override {
+    // AtenRoundDecimalsOp is decomposed as follows if the decimals value is
+    // non-zero: scale = 10 ** decimals return round(x * scale) / scale
+    // otherwise:
+    // return round(x)
+
+    auto loc = op.getLoc();
+    auto input = op.getSelf();
+    auto inputType = cast<BaseTensorType>(input.getType());
+
+    if (!inputType.hasDtype() || !isa<mlir::FloatType>(inputType.getDtype())) {
+      return rewriter.notifyMatchFailure(
+          op, "unimplemented: non-floating point dtype");
+    }
+
+    int64_t decimals;
+    if (!matchPattern(op.getDecimals(), m_TorchConstantInt(&decimals))) {
+      return rewriter.notifyMatchFailure(
+          op, "non-constant decimal point is not supported.");
+    }
+
+    Value newOp = op->getOperand(0);
+    Value scale;
+    if (decimals) {
+      auto scaleVal = pow(10, decimals);
+      scale = rewriter.create<ConstantFloatOp>(
+          loc, rewriter.getF64FloatAttr(scaleVal));
+      newOp = rewriter.create<AtenMulScalarOp>(loc, op.getType(), input, scale);
+    }
+
+    newOp = rewriter.create<AtenRoundOp>(loc, op.getType(), newOp);
+
+    if (decimals) {
+      newOp = rewriter.create<AtenDivScalarOp>(loc, op.getType(), newOp, scale);
+    }
+
+    rewriter.replaceOp(op, newOp);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeComplexOpsPass
     : public DecomposeComplexOpsBase<DecomposeComplexOpsPass> {
@@ -12291,6 +12339,7 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAtenConstrainRangeForSizeOp>(
         patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAten_AssertScalarOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenRoundDecimalsOp>(patterns);
 
     GreedyRewriteConfig config;
     config.setUseTopDownTraversal(true);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -384,6 +384,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenHstackOp>();
   target.addIllegalOp<AtenColumnStackOp>();
   target.addIllegalOp<AtenRollOp>();
+  target.addIllegalOp<AtenRoundDecimalsOp>();
   target.addIllegalOp<AtenRepeatOp>();
   target.addIllegalOp<AtenRepeatInterleaveSelfIntOp>();
   target.addIllegalOp<AtenExpandOp>();

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

@@ -352,6 +352,9 @@ def aten〇relu6〡shape(self: List[int]) -> List[int]:
 def aten〇round〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇round〇decimals〡shape(self: List[int], decimals: int) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇glu〡shape(self: List[int], dim: int = -1) -> List[int]:
     if dim < 0:
         dim += len(self)
@@ -3674,6 +3677,11 @@ def aten〇round〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, decimals=0))
+def aten〇round〇decimals〡dtype(self_rank_dtype: Tuple[int, int], decimals: int) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(_check_tensors_with_the_same_dtype(tensor_shapes=[(100,)], dim=0))
 def aten〇glu〡dtype(self_rank_dtype: Tuple[int, int], dim: int = -1) -> int:
     self_rank, self_dtype = self_rank_dtype

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

@@ -451,6 +451,9 @@ def emit_with_mutating_variants(key, **kwargs):
     emit_with_mutating_variants("aten::floor : (Tensor) -> (Tensor)", has_folder=True)
     emit_with_mutating_variants("aten::ceil : (Tensor) -> (Tensor)", has_folder=True)
     emit_with_mutating_variants("aten::round : (Tensor) -> (Tensor)", has_folder=True)
+    emit_with_mutating_variants(
+        "aten::round.decimals : (Tensor, int) -> (Tensor)", has_folder=True
+    )
     emit_with_mutating_variants("aten::trunc : (Tensor) -> (Tensor)", has_folder=True)
     emit_with_mutating_variants("aten::fix : (Tensor) -> (Tensor)")
     emit("aten::special_expm1 : (Tensor) -> (Tensor)")

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

@@ -6570,6 +6570,26 @@ def AtenRoundIntModule_basic(module, tu: TestUtils):
     module.forward(tu.randint(5, 5, low=-10))
 
 
+class AtenRoundFloatDecimalsModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return torch.ops.aten.round(x, decimals=2)
+
+
+@register_test_case(module_factory=lambda: AtenRoundFloatDecimalsModule())
+def AtenRoundFloatDecimalsModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(5, 5, low=-3.0, high=3.0))
+
+
 # ==============================================================================