[torch] emit aten.Softshrink and aten.Hardshrink (llvm#3248)

as title

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

@@ -5019,6 +5019,54 @@ def Torch_AtenLogSigmoidOp : Torch_Op<"aten.log_sigmoid", [
   }];
 }
 
+def Torch_AtenHardshrinkOp : Torch_Op<"aten.hardshrink", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::hardshrink : (Tensor, Scalar) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchScalarType:$lambd
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenHardshrinkOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenHardshrinkOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
+def Torch_AtenSoftshrinkOp : Torch_Op<"aten.softshrink", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::softshrink : (Tensor, Scalar) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    AnyTorchScalarType:$lambd
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenSoftshrinkOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 2, 1);
+    }
+    void AtenSoftshrinkOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 2, 1);
+    }
+  }];
+}
+
 def Torch_AtenUnbindCopyIntOp : Torch_Op<"aten.unbind_copy.int", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -6514,6 +6514,14 @@ 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.hardshrink\"(%arg0: !torch.list<int>, %arg1: !torch.float) -> !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.softshrink\"(%arg0: !torch.list<int>, %arg1: !torch.float) -> !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.mish\"(%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"
@@ -9798,6 +9806,23 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %0#1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.hardshrink\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.number) -> !torch.int {\n"
+"    %int4 = torch.constant.int 4\n"
+"    %int11 = torch.constant.int 11\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    %1 = torch.aten.eq.int %0#1, %int11 : !torch.int, !torch.int -> !torch.bool\n"
+"    %2 = torch.prim.If %1 -> (!torch.int) {\n"
+"      torch.prim.If.yield %int4 : !torch.int\n"
+"    } else {\n"
+"      torch.prim.If.yield %0#1 : !torch.int\n"
+"    }\n"
+"    return %2 : !torch.int\n"
+"  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.softshrink\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.number) -> !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.logit\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.optional<float>) -> !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

@@ -1913,6 +1913,120 @@ class DecomposeAtenLogSigmoidOp : public OpRewritePattern<AtenLogSigmoidOp> {
 };
 } // namespace
 
+// SoftShrink(x, lambda) function:
+// Applies a shrinkage function where:
+// - If x > lambda, returns x - lambda
+// - If x < -lambda, returns x + lambda
+// - Otherwise, returns 0
+namespace {
+class DecomposeAtenSoftshrinkOp : public OpRewritePattern<AtenSoftshrinkOp> {
+public:
+  using OpRewritePattern<AtenSoftshrinkOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenSoftshrinkOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    Value lambdValue = op.getLambd();
+
+    auto resTy = dyn_cast<ValueTensorType>(op.getType());
+    if (!resTy || !resTy.hasDtype() || !resTy.hasSizes()) {
+      return rewriter.notifyMatchFailure(op,
+                                         "result should have dtype and size");
+    }
+
+    double lambd;
+    if (!matchPattern(lambdValue, m_TorchConstantFloat(&lambd))) {
+      return rewriter.notifyMatchFailure(
+          op, "expected lambd to be a constant float");
+    }
+
+    Value zero =
+        rewriter.create<ConstantFloatOp>(loc, rewriter.getF64FloatAttr(0.0));
+    Value neglambd = rewriter.create<Torch::ConstantFloatOp>(
+        loc, rewriter.getF64FloatAttr(-lambd));
+    Value poslambd = rewriter.create<Torch::ConstantFloatOp>(
+        loc, rewriter.getF64FloatAttr(lambd));
+
+    Value constOneFloat =
+        rewriter.create<ConstantFloatOp>(loc, rewriter.getF64FloatAttr(1.0));
+
+    auto boolResType =
+        resTy.getWithSizesAndDtype(resTy.getSizes(), rewriter.getI1Type());
+
+    Value posMask =
+        rewriter.create<AtenGtScalarOp>(loc, boolResType, self, poslambd);
+    Value negMask =
+        rewriter.create<AtenLtScalarOp>(loc, boolResType, self, neglambd);
+
+    Value posValue = rewriter.create<AtenSubScalarOp>(loc, resTy, self,
+                                                      poslambd, constOneFloat);
+    Value negValue = rewriter.create<AtenAddScalarOp>(loc, resTy, self,
+                                                      neglambd, constOneFloat);
+
+    Value result = rewriter.create<AtenWhereScalarOtherOp>(loc, resTy, posMask,
+                                                           posValue, zero);
+    result =
+        rewriter.create<AtenWhereSelfOp>(loc, resTy, negMask, negValue, result);
+
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+} // namespace
+
+// HardShrink(x, lambda) function:
+// Applies a shrinkage function where:
+// - If x > lambda, returns x
+// - If x < -lambda, returns x
+// - Otherwise, returns 0
+namespace {
+class DecomposeAtenHardshrinkOp : public OpRewritePattern<AtenHardshrinkOp> {
+public:
+  using OpRewritePattern<AtenHardshrinkOp>::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenHardshrinkOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value self = op.getSelf();
+    Value lambdValue = op.getLambd();
+
+    auto resTy = dyn_cast<ValueTensorType>(op.getType());
+    if (!resTy || !resTy.hasDtype() || !resTy.hasSizes()) {
+      return rewriter.notifyMatchFailure(op,
+                                         "result should have dtype and size");
+    }
+
+    double lambd;
+    if (!matchPattern(lambdValue, m_TorchConstantFloat(&lambd))) {
+      return rewriter.notifyMatchFailure(
+          op, "expected lambd to be a constant float");
+    }
+
+    Value zero =
+        rewriter.create<ConstantFloatOp>(loc, rewriter.getF64FloatAttr(0.0));
+    Value neglambd = rewriter.create<Torch::ConstantFloatOp>(
+        loc, rewriter.getF64FloatAttr(-lambd));
+    Value poslambd = rewriter.create<Torch::ConstantFloatOp>(
+        loc, rewriter.getF64FloatAttr(lambd));
+
+    auto boolResType =
+        resTy.getWithSizesAndDtype(resTy.getSizes(), rewriter.getI1Type());
+
+    Value posMask =
+        rewriter.create<AtenGtScalarOp>(loc, boolResType, self, poslambd);
+    Value negMask =
+        rewriter.create<AtenLtScalarOp>(loc, boolResType, self, neglambd);
+
+    Value result = rewriter.create<AtenWhereScalarOtherOp>(loc, resTy, posMask,
+                                                           self, zero);
+    result =
+        rewriter.create<AtenWhereSelfOp>(loc, resTy, negMask, self, result);
+
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+} // namespace
+
 // Decompose aten.matmul into: aten.mm and aten.bmm according to ranks.
 namespace {
 class DecomposeAtenMatmulOp : public OpRewritePattern<AtenMatmulOp> {
@@ -7664,6 +7778,8 @@ class DecomposeComplexOpsPass
     addPatternIfTargetOpIsIllegal<DecomposeAten_LogSoftmaxOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenLogSoftmaxIntOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenLogSigmoidOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenHardshrinkOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenSoftshrinkOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenEmptyLikeOp>(patterns);
     addPatternIfTargetOpIsIllegal<
         DecomposeConstantTensorAllocLikeOp<AtenOnesLikeOp, 1>>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -371,6 +371,8 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<Aten_LogSoftmaxOp>();
   target.addIllegalOp<AtenLogSoftmaxIntOp>();
   target.addIllegalOp<AtenLogSigmoidOp>();
+  target.addIllegalOp<AtenHardshrinkOp>();
+  target.addIllegalOp<AtenSoftshrinkOp>();
   target.addIllegalOp<AtenEmptyLikeOp>();
   target.addIllegalOp<AtenOnesLikeOp>();
   target.addIllegalOp<AtenZerosLikeOp>();

projects/pt1/e2e_testing/xfail_sets.py

@@ -1438,6 +1438,8 @@
     "ElementwiseTruncIntModule_basic",
     "ElementwiseTruncModule_basic",
     "ElementwiseLogSigmoidModule_basic",
+    "ElementwiseHardshrinkStaticModule_basic",
+    "ElementwiseSoftshrinkStaticModule_basic",
 }
 
 STABLEHLO_CRASHING_SET = {
@@ -1667,6 +1669,10 @@
     "ElementwiseSeluModule_basic",
     "ElementwiseSigmoidModule_basic",
     "ElementwiseSignModule_basic",
+    "ElementwiseHardshrinkModule_basic",
+    "ElementwiseHardshrinkStaticModule_basic",
+    "ElementwiseSoftshrinkModule_basic",
+    "ElementwiseSoftshrinkStaticModule_basic",
     "ElementwiseSqrtIntModule_basic",
     "ElementwiseSqrtModule_basic",
     "ElementwiseSubScalarFloatModule_basic",

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

@@ -254,6 +254,12 @@ def aten〇log〡shape(self: List[int]) -> List[int]:
 def aten〇log_sigmoid〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇hardshrink〡shape(self: List[int], lambd: float = 0.5) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
+def aten〇softshrink〡shape(self: List[int], lambd: float = 0.5) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇mish〡shape(self: List[int]) -> List[int]:
     return upstream_shape_functions.unary(self)
 
@@ -2098,6 +2104,18 @@ def aten〇log_sigmoid〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     assert not self_dtype == torch.bool
     return self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, lambd=0.5))
+def aten〇hardshrink〡dtype(self_rank_dtype: Tuple[int, int], lambd: Union[int, float, complex] = 0.5) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    if self_dtype == torch.bool:
+        return torch.int64
+    return self_dtype
+
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1, lambd=0.5))
+def aten〇softshrink〡dtype(self_rank_dtype: Tuple[int, int], lambd: Union[int, float, complex] = 0.5) -> 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logit〡dtype(self_rank_dtype: Tuple[int, int], eps: Optional[float] = None) -> int:
     self_rank, self_dtype = self_rank_dtype

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

@@ -480,6 +480,8 @@ def emit_with_mutating_variants(key, **kwargs):
     emit("aten::isclose : (Tensor, Tensor, float, float, bool) -> (Tensor)")
     emit("aten::glu : (Tensor, int) -> (Tensor)")
     emit("aten::log_sigmoid : (Tensor) -> (Tensor)")
+    emit("aten::hardshrink : (Tensor, Scalar) -> (Tensor)")
+    emit("aten::softshrink : (Tensor, Scalar) -> (Tensor)")
 
     # Ops with dynamic number of outputs
     emit("aten::unbind_copy.int : (Tensor, int) -> (Tensor[])")

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

@@ -2205,6 +2205,98 @@ def ElementwiseLogSigmoidModule_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class ElementwiseSoftshrinkModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.softshrink(a)
+
+
+@register_test_case(module_factory=lambda: ElementwiseSoftshrinkModule())
+def ElementwiseSoftshrinkModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4))
+
+
+# ==============================================================================
+
+
+class ElementwiseSoftshrinkStaticModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([4, 5, 6], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.softshrink(a, 2.0)
+
+
+@register_test_case(module_factory=lambda: ElementwiseSoftshrinkStaticModule())
+def ElementwiseSoftshrinkStaticModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4, 5, 6))
+
+
+# ==============================================================================
+
+
+class ElementwiseHardshrinkModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.hardshrink(a, 1.0)
+
+
+@register_test_case(module_factory=lambda: ElementwiseHardshrinkModule())
+def ElementwiseHardshrinkModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 4, 5))
+
+
+# ==============================================================================
+
+
+class ElementwiseHardshrinkStaticModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([4, 5, 6], torch.float32, True),
+        ]
+    )
+    def forward(self, a):
+        return torch.ops.aten.hardshrink(a, 2.0)
+
+
+@register_test_case(module_factory=lambda: ElementwiseHardshrinkStaticModule())
+def ElementwiseHardshrinkStaticModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4, 5, 6))
+
+
+# ==============================================================================
+
+
 class ElementwiseErfModule(torch.nn.Module):
     def __init__(self):
         super().__init__()