[MLIR][TORCH] Add aten.special.expm1 op lowering

This commit adds the support for torch.aten.special.expm1 op by
decomposing it into torch.aten.expm1 op.

Signed-off-by: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>

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

@@ -4610,6 +4610,29 @@ def Torch_AtenTrunc_Op : Torch_Op<"aten.trunc_", [
   }];
 }
 
+def Torch_AtenSpecialExpm1Op : Torch_Op<"aten.special_expm1", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::special_expm1 : (Tensor) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenSpecialExpm1Op::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 1, 1);
+    }
+    void AtenSpecialExpm1Op::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 1, 1);
+    }
+  }];
+}
+
 def Torch_AtenSignOp : Torch_Op<"aten.sign", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -6495,6 +6495,10 @@ 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.special_expm1\"(%arg0: !torch.list<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.isfinite\"(%arg0: !torch.list<int>) -> !torch.list<int> {\n"
 "    return %arg0 : !torch.list<int>\n"
 "  }\n"
@@ -11435,6 +11439,11 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %1 = call @__torch__._get_dtype_of_floating_point_op(%0#1) : (!torch.int) -> !torch.int\n"
 "    return %1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.special_expm1\"(%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"
+"    return %1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.isfinite\"(%arg0: !torch.tuple<int, int>) -> !torch.int {\n"
 "    %int11 = torch.constant.int 11\n"
 "    return %int11 : !torch.int\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -10501,6 +10501,19 @@ public:
 };
 } // namespace
 
+namespace {
+class DecomposeAtenSpecialExpm1Op
+    : public OpRewritePattern<AtenSpecialExpm1Op> {
+public:
+  using OpRewritePattern<AtenSpecialExpm1Op>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenSpecialExpm1Op op,
+                                PatternRewriter &rewriter) const override {
+    rewriter.replaceOpWithNewOp<AtenExpm1Op>(op, op.getType(), op.getSelf());
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 class DecomposeComplexOpsPass
     : public DecomposeComplexOpsBase<DecomposeComplexOpsPass> {
@@ -10776,6 +10789,7 @@ public:
     addPatternIfTargetOpIsIllegal<DecomposeAtenThresholdOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenFloatPowerTensorTensorOp>(
         patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenSpecialExpm1Op>(patterns);
 
     addPatternIfTargetOpIsIllegal<
         DecomposeAtenFMaxMinOp<AtenFmaxOp, AtenMaximumOp>>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -566,6 +566,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<AtenLinalgNormOp>();
   target.addIllegalOp<AtenFminOp>();
   target.addIllegalOp<AtenFmaxOp>();
+  target.addIllegalOp<AtenSpecialExpm1Op>();
 
   for (auto &opName : backendLegalOpsSet) {
     target.addLegalOp(

projects/pt1/e2e_testing/xfail_sets.py

@@ -501,8 +501,6 @@ FX_IMPORTER_XFAIL_SET = {
     "AdaptiveMaxPool1dStatic_basic",
     "CrossEntropyLossModule_basic",
     "CrossEntropyLossNoReductionModule_basic",
-    "ElementwiseExpm1IntModule_basic",
-    "ElementwiseExpm1Module_basic",
     "IndexPutImpl1DFloatAccumulateModule_basic",
     "IndexPutImpl1DFloatNonAccumulateModule_basic",
     "IndexPutImpl1DIntAccumulateModule_basic",
@@ -928,8 +926,6 @@ FX_IMPORTER_STABLEHLO_XFAIL_SET = {
     "AtenItemIntOpModule_basic",
     "CrossEntropyLossModule_basic",
     "CrossEntropyLossNoReductionModule_basic",
-    "ElementwiseExpm1IntModule_basic",
-    "ElementwiseExpm1Module_basic",
     "InterpolateDynamicModule_sizes_nearest",
     "IouOfModule_basic",
     "IscloseStaticModuleTrue_basic",
@@ -1226,6 +1222,8 @@ STABLEHLO_PASS_SET = {
     "ElementwiseRsqrtModule_basic",
     "ElementwiseSigmoidModule_basic",
     "ElementwiseSinModule_basic",
+    "ElementwiseSpecialExpm1IntModule_basic",
+    "ElementwiseSpecialExpm1Module_basic",
     "ElementwiseSqrtModule_basic",
     "ElementwiseTanIntModule_basic",
     "ElementwiseTanModule_basic",
@@ -2913,6 +2911,8 @@ ONNX_XFAIL_SET = {
     "ElementwiseEluNonDefaultModule_basic",
     "ElementwiseExpm1IntModule_basic",
     "ElementwiseExpm1Module_basic",
+    "ElementwiseSpecialExpm1IntModule_basic",
+    "ElementwiseSpecialExpm1Module_basic",
     "ElementwiseFmodTensor_Int_basic",
     "ElementwiseCreateComplexModule_basic",
     "ElementwiseMulTensorComplexModule_basic",
@@ -3630,6 +3630,8 @@ FX_IMPORTER_TOSA_XFAIL_SET = {
     "ElementwiseSigmoidIntModule_basic",
     "ElementwiseSinhIntModule_basic",
     "ElementwiseSinhModule_basic",
+    "ElementwiseSpecialExpm1IntModule_basic",
+    "ElementwiseSpecialExpm1Module_basic",
     "ElementwiseTanIntModule_basic",
     "ElementwiseTanModule_basic",
     "ElementwiseToDtypeF32ToI64Module_basic",
@@ -4342,6 +4344,8 @@ ONNX_TOSA_XFAIL_SET = {
     "ElementwiseSinIntModule_basic",
     "ElementwiseSinhIntModule_basic",
     "ElementwiseSinhModule_basic",
+    "ElementwiseSpecialExpm1IntModule_basic",
+    "ElementwiseSpecialExpm1Module_basic",
     "ElementwiseSqrtIntModule_basic",
     "ElementwiseSubScalarIntModule_basic",
     "ElementwiseTanIntModule_basic",

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

@@ -222,6 +222,9 @@ def aten〇exp2〡shape(self: List[int]) -> List[int]:
 def aten〇expm1〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇special_expm1〡shape(self: List[int]) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇isfinite〡shape(self: List[int]) -> List[int]:
     return self
 
@@ -2656,6 +2659,11 @@ def aten〇expm1〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype
     return _get_dtype_of_floating_point_op(self_dtype)
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
+def aten〇special_expm1〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return _get_dtype_of_floating_point_op(self_dtype)
+
 def aten〇isfinite〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     return torch.bool
 

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

@@ -452,6 +452,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     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::trunc : (Tensor) -> (Tensor)", has_folder=True)
+    emit("aten::special_expm1 : (Tensor) -> (Tensor)")
     emit_with_mutating_variants(
         "aten::sign : (Tensor) -> (Tensor)", has_canonicalizer=True
     )

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

@@ -5207,7 +5207,7 @@ class ElementwiseExpm1Module(torch.nn.Module):
         ]
     )
     def forward(self, a):
-        return torch.special.expm1(a)
+        return torch.expm1(a)
 
 
 @register_test_case(module_factory=lambda: ElementwiseExpm1Module())
@@ -5230,7 +5230,7 @@ class ElementwiseExpm1IntModule(torch.nn.Module):
         ]
     )
     def forward(self, a):
-        return torch.special.expm1(a)
+        return torch.expm1(a)
 
 
 @register_test_case(module_factory=lambda: ElementwiseExpm1IntModule())
@@ -5241,6 +5241,51 @@ def ElementwiseExpm1IntModule_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class ElementwiseSpecialExpm1Module(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.special.expm1(a)
+
+
+@register_test_case(module_factory=lambda: ElementwiseSpecialExpm1Module())
+def ElementwiseSpecialExpm1Module_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4))
+
+
+# ==============================================================================
+
+
+class ElementwiseSpecialExpm1IntModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.int32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.special.expm1(a)
+
+
+@register_test_case(module_factory=lambda: ElementwiseSpecialExpm1IntModule())
+def ElementwiseSpecialExpm1IntModule_basic(module, tu: TestUtils):
+    module.forward(tu.randint(3, 4, low=1, high=10).to(torch.int32))
+
+# ==============================================================================
+
+
 class ElementwiseRad2DegModule(torch.nn.Module):
     def __init__(self):
         super().__init__()